initial commit

pjproject-2.11.1/third_party/yuv/source/compare.cc

@@ -0,0 +1,423 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/compare.h"
+
+#include <float.h>
+#include <math.h>
+#ifdef _OPENMP
+#include <omp.h>
+#endif
+
+#include "libyuv/basic_types.h"
+#include "libyuv/compare_row.h"
+#include "libyuv/cpu_id.h"
+#include "libyuv/row.h"
+#include "libyuv/video_common.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// hash seed of 5381 recommended.
+LIBYUV_API
+uint32 HashDjb2(const uint8* src, uint64 count, uint32 seed) {
+  const int kBlockSize = 1 << 15;  // 32768;
+  int remainder;
+  uint32 (*HashDjb2_SSE)(const uint8* src, int count, uint32 seed) = HashDjb2_C;
+#if defined(HAS_HASHDJB2_SSE41)
+  if (TestCpuFlag(kCpuHasSSE41)) {
+    HashDjb2_SSE = HashDjb2_SSE41;
+  }
+#endif
+#if defined(HAS_HASHDJB2_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    HashDjb2_SSE = HashDjb2_AVX2;
+  }
+#endif
+
+  while (count >= (uint64)(kBlockSize)) {
+    seed = HashDjb2_SSE(src, kBlockSize, seed);
+    src += kBlockSize;
+    count -= kBlockSize;
+  }
+  remainder = (int)count & ~15;
+  if (remainder) {
+    seed = HashDjb2_SSE(src, remainder, seed);
+    src += remainder;
+    count -= remainder;
+  }
+  remainder = (int)count & 15;
+  if (remainder) {
+    seed = HashDjb2_C(src, remainder, seed);
+  }
+  return seed;
+}
+
+static uint32 ARGBDetectRow_C(const uint8* argb, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    if (argb[0] != 255) {  // First byte is not Alpha of 255, so not ARGB.
+      return FOURCC_BGRA;
+    }
+    if (argb[3] != 255) {  // 4th byte is not Alpha of 255, so not BGRA.
+      return FOURCC_ARGB;
+    }
+    if (argb[4] != 255) {  // Second pixel first byte is not Alpha of 255.
+      return FOURCC_BGRA;
+    }
+    if (argb[7] != 255) {  // Second pixel 4th byte is not Alpha of 255.
+      return FOURCC_ARGB;
+    }
+    argb += 8;
+  }
+  if (width & 1) {
+    if (argb[0] != 255) {  // First byte is not Alpha of 255, so not ARGB.
+      return FOURCC_BGRA;
+    }
+    if (argb[3] != 255) {  // 4th byte is not Alpha of 255, so not BGRA.
+      return FOURCC_ARGB;
+    }
+  }
+  return 0;
+}
+
+// Scan an opaque argb image and return fourcc based on alpha offset.
+// Returns FOURCC_ARGB, FOURCC_BGRA, or 0 if unknown.
+LIBYUV_API
+uint32 ARGBDetect(const uint8* argb, int stride_argb, int width, int height) {
+  uint32 fourcc = 0;
+  int h;
+
+  // Coalesce rows.
+  if (stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    stride_argb = 0;
+  }
+  for (h = 0; h < height && fourcc == 0; ++h) {
+    fourcc = ARGBDetectRow_C(argb, width);
+    argb += stride_argb;
+  }
+  return fourcc;
+}
+
+// NEON version accumulates in 16 bit shorts which overflow at 65536 bytes.
+// So actual maximum is 1 less loop, which is 64436 - 32 bytes.
+
+LIBYUV_API
+uint64 ComputeHammingDistance(const uint8* src_a,
+                              const uint8* src_b,
+                              int count) {
+  const int kBlockSize = 1 << 15;  // 32768;
+  const int kSimdSize = 64;
+  // SIMD for multiple of 64, and C for remainder
+  int remainder = count & (kBlockSize - 1) & ~(kSimdSize - 1);
+  uint64 diff = 0;
+  int i;
+  uint32 (*HammingDistance)(const uint8* src_a, const uint8* src_b, int count) =
+      HammingDistance_C;
+#if defined(HAS_HAMMINGDISTANCE_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    HammingDistance = HammingDistance_NEON;
+  }
+#endif
+#if defined(HAS_HAMMINGDISTANCE_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    HammingDistance = HammingDistance_SSSE3;
+  }
+#endif
+#if defined(HAS_HAMMINGDISTANCE_SSE42)
+  if (TestCpuFlag(kCpuHasSSE42)) {
+    HammingDistance = HammingDistance_SSE42;
+  }
+#endif
+#if defined(HAS_HAMMINGDISTANCE_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    HammingDistance = HammingDistance_AVX2;
+  }
+#endif
+#if defined(HAS_HAMMINGDISTANCE_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    HammingDistance = HammingDistance_MSA;
+  }
+#endif
+#ifdef _OPENMP
+#pragma omp parallel for reduction(+ : diff)
+#endif
+  for (i = 0; i < (count - (kBlockSize - 1)); i += kBlockSize) {
+    diff += HammingDistance(src_a + i, src_b + i, kBlockSize);
+  }
+  src_a += count & ~(kBlockSize - 1);
+  src_b += count & ~(kBlockSize - 1);
+  if (remainder) {
+    diff += HammingDistance(src_a, src_b, remainder);
+    src_a += remainder;
+    src_b += remainder;
+  }
+  remainder = count & (kSimdSize - 1);
+  if (remainder) {
+    diff += HammingDistance_C(src_a, src_b, remainder);
+  }
+  return diff;
+}
+
+// TODO(fbarchard): Refactor into row function.
+LIBYUV_API
+uint64 ComputeSumSquareError(const uint8* src_a,
+                             const uint8* src_b,
+                             int count) {
+  // SumSquareError returns values 0 to 65535 for each squared difference.
+  // Up to 65536 of those can be summed and remain within a uint32.
+  // After each block of 65536 pixels, accumulate into a uint64.
+  const int kBlockSize = 65536;
+  int remainder = count & (kBlockSize - 1) & ~31;
+  uint64 sse = 0;
+  int i;
+  uint32 (*SumSquareError)(const uint8* src_a, const uint8* src_b, int count) =
+      SumSquareError_C;
+#if defined(HAS_SUMSQUAREERROR_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SumSquareError = SumSquareError_NEON;
+  }
+#endif
+#if defined(HAS_SUMSQUAREERROR_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    // Note only used for multiples of 16 so count is not checked.
+    SumSquareError = SumSquareError_SSE2;
+  }
+#endif
+#if defined(HAS_SUMSQUAREERROR_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    // Note only used for multiples of 32 so count is not checked.
+    SumSquareError = SumSquareError_AVX2;
+  }
+#endif
+#if defined(HAS_SUMSQUAREERROR_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    SumSquareError = SumSquareError_MSA;
+  }
+#endif
+#ifdef _OPENMP
+#pragma omp parallel for reduction(+ : sse)
+#endif
+  for (i = 0; i < (count - (kBlockSize - 1)); i += kBlockSize) {
+    sse += SumSquareError(src_a + i, src_b + i, kBlockSize);
+  }
+  src_a += count & ~(kBlockSize - 1);
+  src_b += count & ~(kBlockSize - 1);
+  if (remainder) {
+    sse += SumSquareError(src_a, src_b, remainder);
+    src_a += remainder;
+    src_b += remainder;
+  }
+  remainder = count & 31;
+  if (remainder) {
+    sse += SumSquareError_C(src_a, src_b, remainder);
+  }
+  return sse;
+}
+
+LIBYUV_API
+uint64 ComputeSumSquareErrorPlane(const uint8* src_a,
+                                  int stride_a,
+                                  const uint8* src_b,
+                                  int stride_b,
+                                  int width,
+                                  int height) {
+  uint64 sse = 0;
+  int h;
+  // Coalesce rows.
+  if (stride_a == width && stride_b == width) {
+    width *= height;
+    height = 1;
+    stride_a = stride_b = 0;
+  }
+  for (h = 0; h < height; ++h) {
+    sse += ComputeSumSquareError(src_a, src_b, width);
+    src_a += stride_a;
+    src_b += stride_b;
+  }
+  return sse;
+}
+
+LIBYUV_API
+double SumSquareErrorToPsnr(uint64 sse, uint64 count) {
+  double psnr;
+  if (sse > 0) {
+    double mse = (double)count / (double)sse;
+    psnr = 10.0 * log10(255.0 * 255.0 * mse);
+  } else {
+    psnr = kMaxPsnr;  // Limit to prevent divide by 0
+  }
+
+  if (psnr > kMaxPsnr)
+    psnr = kMaxPsnr;
+
+  return psnr;
+}
+
+LIBYUV_API
+double CalcFramePsnr(const uint8* src_a,
+                     int stride_a,
+                     const uint8* src_b,
+                     int stride_b,
+                     int width,
+                     int height) {
+  const uint64 samples = width * height;
+  const uint64 sse = ComputeSumSquareErrorPlane(src_a, stride_a, src_b,
+                                                stride_b, width, height);
+  return SumSquareErrorToPsnr(sse, samples);
+}
+
+LIBYUV_API
+double I420Psnr(const uint8* src_y_a,
+                int stride_y_a,
+                const uint8* src_u_a,
+                int stride_u_a,
+                const uint8* src_v_a,
+                int stride_v_a,
+                const uint8* src_y_b,
+                int stride_y_b,
+                const uint8* src_u_b,
+                int stride_u_b,
+                const uint8* src_v_b,
+                int stride_v_b,
+                int width,
+                int height) {
+  const uint64 sse_y = ComputeSumSquareErrorPlane(src_y_a, stride_y_a, src_y_b,
+                                                  stride_y_b, width, height);
+  const int width_uv = (width + 1) >> 1;
+  const int height_uv = (height + 1) >> 1;
+  const uint64 sse_u = ComputeSumSquareErrorPlane(
+      src_u_a, stride_u_a, src_u_b, stride_u_b, width_uv, height_uv);
+  const uint64 sse_v = ComputeSumSquareErrorPlane(
+      src_v_a, stride_v_a, src_v_b, stride_v_b, width_uv, height_uv);
+  const uint64 samples = width * height + 2 * (width_uv * height_uv);
+  const uint64 sse = sse_y + sse_u + sse_v;
+  return SumSquareErrorToPsnr(sse, samples);
+}
+
+static const int64 cc1 = 26634;   // (64^2*(.01*255)^2
+static const int64 cc2 = 239708;  // (64^2*(.03*255)^2
+
+static double Ssim8x8_C(const uint8* src_a,
+                        int stride_a,
+                        const uint8* src_b,
+                        int stride_b) {
+  int64 sum_a = 0;
+  int64 sum_b = 0;
+  int64 sum_sq_a = 0;
+  int64 sum_sq_b = 0;
+  int64 sum_axb = 0;
+
+  int i;
+  for (i = 0; i < 8; ++i) {
+    int j;
+    for (j = 0; j < 8; ++j) {
+      sum_a += src_a[j];
+      sum_b += src_b[j];
+      sum_sq_a += src_a[j] * src_a[j];
+      sum_sq_b += src_b[j] * src_b[j];
+      sum_axb += src_a[j] * src_b[j];
+    }
+
+    src_a += stride_a;
+    src_b += stride_b;
+  }
+
+  {
+    const int64 count = 64;
+    // scale the constants by number of pixels
+    const int64 c1 = (cc1 * count * count) >> 12;
+    const int64 c2 = (cc2 * count * count) >> 12;
+
+    const int64 sum_a_x_sum_b = sum_a * sum_b;
+
+    const int64 ssim_n = (2 * sum_a_x_sum_b + c1) *
+                         (2 * count * sum_axb - 2 * sum_a_x_sum_b + c2);
+
+    const int64 sum_a_sq = sum_a * sum_a;
+    const int64 sum_b_sq = sum_b * sum_b;
+
+    const int64 ssim_d =
+        (sum_a_sq + sum_b_sq + c1) *
+        (count * sum_sq_a - sum_a_sq + count * sum_sq_b - sum_b_sq + c2);
+
+    if (ssim_d == 0.0) {
+      return DBL_MAX;
+    }
+    return ssim_n * 1.0 / ssim_d;
+  }
+}
+
+// We are using a 8x8 moving window with starting location of each 8x8 window
+// on the 4x4 pixel grid. Such arrangement allows the windows to overlap
+// block boundaries to penalize blocking artifacts.
+LIBYUV_API
+double CalcFrameSsim(const uint8* src_a,
+                     int stride_a,
+                     const uint8* src_b,
+                     int stride_b,
+                     int width,
+                     int height) {
+  int samples = 0;
+  double ssim_total = 0;
+  double (*Ssim8x8)(const uint8* src_a, int stride_a, const uint8* src_b,
+                    int stride_b) = Ssim8x8_C;
+
+  // sample point start with each 4x4 location
+  int i;
+  for (i = 0; i < height - 8; i += 4) {
+    int j;
+    for (j = 0; j < width - 8; j += 4) {
+      ssim_total += Ssim8x8(src_a + j, stride_a, src_b + j, stride_b);
+      samples++;
+    }
+
+    src_a += stride_a * 4;
+    src_b += stride_b * 4;
+  }
+
+  ssim_total /= samples;
+  return ssim_total;
+}
+
+LIBYUV_API
+double I420Ssim(const uint8* src_y_a,
+                int stride_y_a,
+                const uint8* src_u_a,
+                int stride_u_a,
+                const uint8* src_v_a,
+                int stride_v_a,
+                const uint8* src_y_b,
+                int stride_y_b,
+                const uint8* src_u_b,
+                int stride_u_b,
+                const uint8* src_v_b,
+                int stride_v_b,
+                int width,
+                int height) {
+  const double ssim_y =
+      CalcFrameSsim(src_y_a, stride_y_a, src_y_b, stride_y_b, width, height);
+  const int width_uv = (width + 1) >> 1;
+  const int height_uv = (height + 1) >> 1;
+  const double ssim_u = CalcFrameSsim(src_u_a, stride_u_a, src_u_b, stride_u_b,
+                                      width_uv, height_uv);
+  const double ssim_v = CalcFrameSsim(src_v_a, stride_v_a, src_v_b, stride_v_b,
+                                      width_uv, height_uv);
+  return ssim_y * 0.8 + 0.1 * (ssim_u + ssim_v);
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif

pjproject-2.11.1/third_party/yuv/source/compare_common.cc

@@ -0,0 +1,98 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/basic_types.h"
+
+#include "libyuv/compare_row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if ORIGINAL_OPT
+uint32 HammingDistance_C1(const uint8* src_a, const uint8* src_b, int count) {
+  uint32 diff = 0u;
+
+  int i;
+  for (i = 0; i < count; ++i) {
+    int x = src_a[i] ^ src_b[i];
+    if (x & 1)
+      ++diff;
+    if (x & 2)
+      ++diff;
+    if (x & 4)
+      ++diff;
+    if (x & 8)
+      ++diff;
+    if (x & 16)
+      ++diff;
+    if (x & 32)
+      ++diff;
+    if (x & 64)
+      ++diff;
+    if (x & 128)
+      ++diff;
+  }
+  return diff;
+}
+#endif
+
+// Hakmem method for hamming distance.
+uint32 HammingDistance_C(const uint8* src_a, const uint8* src_b, int count) {
+  uint32 diff = 0u;
+
+  int i;
+  for (i = 0; i < count - 3; i += 4) {
+    uint32 x = *((uint32*)src_a) ^ *((uint32*)src_b);
+    uint32 u = x - ((x >> 1) & 0x55555555);
+    u = ((u >> 2) & 0x33333333) + (u & 0x33333333);
+    diff += ((((u + (u >> 4)) & 0x0f0f0f0f) * 0x01010101) >> 24);
+    src_a += 4;
+    src_b += 4;
+  }
+
+  for (; i < count; ++i) {
+    uint32 x = *src_a ^ *src_b;
+    uint32 u = x - ((x >> 1) & 0x55);
+    u = ((u >> 2) & 0x33) + (u & 0x33);
+    diff += (u + (u >> 4)) & 0x0f;
+    src_a += 1;
+    src_b += 1;
+  }
+
+  return diff;
+}
+
+uint32 SumSquareError_C(const uint8* src_a, const uint8* src_b, int count) {
+  uint32 sse = 0u;
+  int i;
+  for (i = 0; i < count; ++i) {
+    int diff = src_a[i] - src_b[i];
+    sse += (uint32)(diff * diff);
+  }
+  return sse;
+}
+
+// hash seed of 5381 recommended.
+// Internal C version of HashDjb2 with int sized count for efficiency.
+uint32 HashDjb2_C(const uint8* src, int count, uint32 seed) {
+  uint32 hash = seed;
+  int i;
+  for (i = 0; i < count; ++i) {
+    hash += (hash << 5) + src[i];
+  }
+  return hash;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif

pjproject-2.11.1/third_party/yuv/source/compare_gcc.cc

@@ -0,0 +1,355 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/basic_types.h"
+
+#include "libyuv/compare_row.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC x86 and x64.
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(__x86_64__) || (defined(__i386__) && !defined(_MSC_VER)))
+
+#if defined(__x86_64__)
+uint32 HammingDistance_SSE42(const uint8* src_a,
+                             const uint8* src_b,
+                             int count) {
+  uint64 diff = 0u;
+
+  asm volatile(
+      "xor        %3,%3                          \n"
+      "xor        %%r8,%%r8                      \n"
+      "xor        %%r9,%%r9                      \n"
+      "xor        %%r10,%%r10                    \n"
+
+      // Process 32 bytes per loop.
+      LABELALIGN
+      "1:                                        \n"
+      "mov        (%0),%%rcx                     \n"
+      "mov        0x8(%0),%%rdx                  \n"
+      "xor        (%1),%%rcx                     \n"
+      "xor        0x8(%1),%%rdx                  \n"
+      "popcnt     %%rcx,%%rcx                    \n"
+      "popcnt     %%rdx,%%rdx                    \n"
+      "mov        0x10(%0),%%rsi                 \n"
+      "mov        0x18(%0),%%rdi                 \n"
+      "xor        0x10(%1),%%rsi                 \n"
+      "xor        0x18(%1),%%rdi                 \n"
+      "popcnt     %%rsi,%%rsi                    \n"
+      "popcnt     %%rdi,%%rdi                    \n"
+      "add        $0x20,%0                       \n"
+      "add        $0x20,%1                       \n"
+      "add        %%rcx,%3                       \n"
+      "add        %%rdx,%%r8                     \n"
+      "add        %%rsi,%%r9                     \n"
+      "add        %%rdi,%%r10                    \n"
+      "sub        $0x20,%2                       \n"
+      "jg         1b                             \n"
+
+      "add        %%r8, %3                       \n"
+      "add        %%r9, %3                       \n"
+      "add        %%r10, %3                      \n"
+      : "+r"(src_a),  // %0
+        "+r"(src_b),  // %1
+        "+r"(count),  // %2
+        "=r"(diff)    // %3
+      :
+      : "memory", "cc", "rcx", "rdx", "rsi", "rdi", "r8", "r9", "r10");
+
+  return static_cast<uint32>(diff);
+}
+#else
+uint32 HammingDistance_SSE42(const uint8* src_a,
+                             const uint8* src_b,
+                             int count) {
+  uint32 diff = 0u;
+
+  asm volatile(
+      // Process 16 bytes per loop.
+      LABELALIGN
+      "1:                                        \n"
+      "mov        (%0),%%ecx                     \n"
+      "mov        0x4(%0),%%edx                  \n"
+      "xor        (%1),%%ecx                     \n"
+      "xor        0x4(%1),%%edx                  \n"
+      "popcnt     %%ecx,%%ecx                    \n"
+      "add        %%ecx,%3                       \n"
+      "popcnt     %%edx,%%edx                    \n"
+      "add        %%edx,%3                       \n"
+      "mov        0x8(%0),%%ecx                  \n"
+      "mov        0xc(%0),%%edx                  \n"
+      "xor        0x8(%1),%%ecx                  \n"
+      "xor        0xc(%1),%%edx                  \n"
+      "popcnt     %%ecx,%%ecx                    \n"
+      "add        %%ecx,%3                       \n"
+      "popcnt     %%edx,%%edx                    \n"
+      "add        %%edx,%3                       \n"
+      "add        $0x10,%0                       \n"
+      "add        $0x10,%1                       \n"
+      "sub        $0x10,%2                       \n"
+      "jg         1b                             \n"
+      : "+r"(src_a),  // %0
+        "+r"(src_b),  // %1
+        "+r"(count),  // %2
+        "+r"(diff)    // %3
+      :
+      : "memory", "cc", "ecx", "edx");
+
+  return diff;
+}
+#endif
+
+static vec8 kNibbleMask = {15, 15, 15, 15, 15, 15, 15, 15,
+                           15, 15, 15, 15, 15, 15, 15, 15};
+static vec8 kBitCount = {0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4};
+
+uint32 HammingDistance_SSSE3(const uint8* src_a,
+                             const uint8* src_b,
+                             int count) {
+  uint32 diff = 0u;
+
+  asm volatile(
+      "movdqa     %4,%%xmm2                      \n"
+      "movdqa     %5,%%xmm3                      \n"
+      "pxor       %%xmm0,%%xmm0                  \n"
+      "pxor       %%xmm1,%%xmm1                  \n"
+      "sub        %0,%1                          \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqa     (%0),%%xmm4                    \n"
+      "movdqa     0x10(%0), %%xmm5               \n"
+      "pxor       (%0,%1), %%xmm4                \n"
+      "movdqa     %%xmm4,%%xmm6                  \n"
+      "pand       %%xmm2,%%xmm6                  \n"
+      "psrlw      $0x4,%%xmm4                    \n"
+      "movdqa     %%xmm3,%%xmm7                  \n"
+      "pshufb     %%xmm6,%%xmm7                  \n"
+      "pand       %%xmm2,%%xmm4                  \n"
+      "movdqa     %%xmm3,%%xmm6                  \n"
+      "pshufb     %%xmm4,%%xmm6                  \n"
+      "paddb      %%xmm7,%%xmm6                  \n"
+      "pxor       0x10(%0,%1),%%xmm5             \n"
+      "add        $0x20,%0                       \n"
+      "movdqa     %%xmm5,%%xmm4                  \n"
+      "pand       %%xmm2,%%xmm5                  \n"
+      "psrlw      $0x4,%%xmm4                    \n"
+      "movdqa     %%xmm3,%%xmm7                  \n"
+      "pshufb     %%xmm5,%%xmm7                  \n"
+      "pand       %%xmm2,%%xmm4                  \n"
+      "movdqa     %%xmm3,%%xmm5                  \n"
+      "pshufb     %%xmm4,%%xmm5                  \n"
+      "paddb      %%xmm7,%%xmm5                  \n"
+      "paddb      %%xmm5,%%xmm6                  \n"
+      "psadbw     %%xmm1,%%xmm6                  \n"
+      "paddd      %%xmm6,%%xmm0                  \n"
+      "sub        $0x20,%2                       \n"
+      "jg         1b                             \n"
+
+      "pshufd     $0xaa,%%xmm0,%%xmm1            \n"
+      "paddd      %%xmm1,%%xmm0                  \n"
+      "movd       %%xmm0, %3                     \n"
+      : "+r"(src_a),       // %0
+        "+r"(src_b),       // %1
+        "+r"(count),       // %2
+        "=r"(diff)         // %3
+      : "m"(kNibbleMask),  // %4
+        "m"(kBitCount)     // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
+
+  return diff;
+}
+
+#ifdef HAS_HAMMINGDISTANCE_AVX2
+uint32 HammingDistance_AVX2(const uint8* src_a, const uint8* src_b, int count) {
+  uint32 diff = 0u;
+
+  asm volatile(
+      "vbroadcastf128 %4,%%ymm2                  \n"
+      "vbroadcastf128 %5,%%ymm3                  \n"
+      "vpxor      %%ymm0,%%ymm0,%%ymm0           \n"
+      "vpxor      %%ymm1,%%ymm1,%%ymm1           \n"
+      "sub        %0,%1                          \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqa    (%0),%%ymm4                    \n"
+      "vmovdqa    0x20(%0), %%ymm5               \n"
+      "vpxor      (%0,%1), %%ymm4, %%ymm4        \n"
+      "vpand      %%ymm2,%%ymm4,%%ymm6           \n"
+      "vpsrlw     $0x4,%%ymm4,%%ymm4             \n"
+      "vpshufb    %%ymm6,%%ymm3,%%ymm6           \n"
+      "vpand      %%ymm2,%%ymm4,%%ymm4           \n"
+      "vpshufb    %%ymm4,%%ymm3,%%ymm4           \n"
+      "vpaddb     %%ymm4,%%ymm6,%%ymm6           \n"
+      "vpxor      0x20(%0,%1),%%ymm5,%%ymm4      \n"
+      "add        $0x40,%0                       \n"
+      "vpand      %%ymm2,%%ymm4,%%ymm5           \n"
+      "vpsrlw     $0x4,%%ymm4,%%ymm4             \n"
+      "vpshufb    %%ymm5,%%ymm3,%%ymm5           \n"
+      "vpand      %%ymm2,%%ymm4,%%ymm4           \n"
+      "vpshufb    %%ymm4,%%ymm3,%%ymm4           \n"
+      "vpaddb     %%ymm5,%%ymm4,%%ymm4           \n"
+      "vpaddb     %%ymm6,%%ymm4,%%ymm4           \n"
+      "vpsadbw    %%ymm1,%%ymm4,%%ymm4           \n"
+      "vpaddd     %%ymm0,%%ymm4,%%ymm0           \n"
+      "sub        $0x40,%2                       \n"
+      "jg         1b                             \n"
+
+      "vpermq     $0xb1,%%ymm0,%%ymm1            \n"
+      "vpaddd     %%ymm1,%%ymm0,%%ymm0           \n"
+      "vpermq     $0xaa,%%ymm0,%%ymm1            \n"
+      "vpaddd     %%ymm1,%%ymm0,%%ymm0           \n"
+      "vmovd      %%xmm0, %3                     \n"
+      "vzeroupper                                \n"
+      : "+r"(src_a),       // %0
+        "+r"(src_b),       // %1
+        "+r"(count),       // %2
+        "=r"(diff)         // %3
+      : "m"(kNibbleMask),  // %4
+        "m"(kBitCount)     // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
+
+  return diff;
+}
+#endif  // HAS_HAMMINGDISTANCE_AVX2
+
+uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count) {
+  uint32 sse;
+  asm volatile (
+    "pxor      %%xmm0,%%xmm0                   \n"
+    "pxor      %%xmm5,%%xmm5                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x10, 0) ",%0          \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x10, 1) ",%1          \n"
+    "movdqa    %%xmm1,%%xmm3                   \n"
+    "psubusb   %%xmm2,%%xmm1                   \n"
+    "psubusb   %%xmm3,%%xmm2                   \n"
+    "por       %%xmm2,%%xmm1                   \n"
+    "movdqa    %%xmm1,%%xmm2                   \n"
+    "punpcklbw %%xmm5,%%xmm1                   \n"
+    "punpckhbw %%xmm5,%%xmm2                   \n"
+    "pmaddwd   %%xmm1,%%xmm1                   \n"
+    "pmaddwd   %%xmm2,%%xmm2                   \n"
+    "paddd     %%xmm1,%%xmm0                   \n"
+    "paddd     %%xmm2,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+
+    "pshufd    $0xee,%%xmm0,%%xmm1             \n"
+    "paddd     %%xmm1,%%xmm0                   \n"
+    "pshufd    $0x1,%%xmm0,%%xmm1              \n"
+    "paddd     %%xmm1,%%xmm0                   \n"
+    "movd      %%xmm0,%3                       \n"
+
+  : "+r"(src_a),      // %0
+    "+r"(src_b),      // %1
+    "+r"(count),      // %2
+    "=g"(sse)         // %3
+  :: "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+  return sse;
+}
+
+static uvec32 kHash16x33 = {0x92d9e201, 0, 0, 0};  // 33 ^ 16
+static uvec32 kHashMul0 = {
+    0x0c3525e1,  // 33 ^ 15
+    0xa3476dc1,  // 33 ^ 14
+    0x3b4039a1,  // 33 ^ 13
+    0x4f5f0981,  // 33 ^ 12
+};
+static uvec32 kHashMul1 = {
+    0x30f35d61,  // 33 ^ 11
+    0x855cb541,  // 33 ^ 10
+    0x040a9121,  // 33 ^ 9
+    0x747c7101,  // 33 ^ 8
+};
+static uvec32 kHashMul2 = {
+    0xec41d4e1,  // 33 ^ 7
+    0x4cfa3cc1,  // 33 ^ 6
+    0x025528a1,  // 33 ^ 5
+    0x00121881,  // 33 ^ 4
+};
+static uvec32 kHashMul3 = {
+    0x00008c61,  // 33 ^ 3
+    0x00000441,  // 33 ^ 2
+    0x00000021,  // 33 ^ 1
+    0x00000001,  // 33 ^ 0
+};
+
+uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed) {
+  uint32 hash;
+  asm volatile (
+    "movd      %2,%%xmm0                       \n"
+    "pxor      %%xmm7,%%xmm7                   \n"
+    "movdqa    %4,%%xmm6                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x10, 0) ",%0          \n"
+    "pmulld    %%xmm6,%%xmm0                   \n"
+    "movdqa    %5,%%xmm5                       \n"
+    "movdqa    %%xmm1,%%xmm2                   \n"
+    "punpcklbw %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm2,%%xmm3                   \n"
+    "punpcklwd %%xmm7,%%xmm3                   \n"
+    "pmulld    %%xmm5,%%xmm3                   \n"
+    "movdqa    %6,%%xmm5                       \n"
+    "movdqa    %%xmm2,%%xmm4                   \n"
+    "punpckhwd %%xmm7,%%xmm4                   \n"
+    "pmulld    %%xmm5,%%xmm4                   \n"
+    "movdqa    %7,%%xmm5                       \n"
+    "punpckhbw %%xmm7,%%xmm1                   \n"
+    "movdqa    %%xmm1,%%xmm2                   \n"
+    "punpcklwd %%xmm7,%%xmm2                   \n"
+    "pmulld    %%xmm5,%%xmm2                   \n"
+    "movdqa    %8,%%xmm5                       \n"
+    "punpckhwd %%xmm7,%%xmm1                   \n"
+    "pmulld    %%xmm5,%%xmm1                   \n"
+    "paddd     %%xmm4,%%xmm3                   \n"
+    "paddd     %%xmm2,%%xmm1                   \n"
+    "paddd     %%xmm3,%%xmm1                   \n"
+    "pshufd    $0xe,%%xmm1,%%xmm2              \n"
+    "paddd     %%xmm2,%%xmm1                   \n"
+    "pshufd    $0x1,%%xmm1,%%xmm2              \n"
+    "paddd     %%xmm2,%%xmm1                   \n"
+    "paddd     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%1                        \n"
+    "jg        1b                              \n"
+    "movd      %%xmm0,%3                       \n"
+  : "+r"(src),        // %0
+    "+r"(count),      // %1
+    "+rm"(seed),      // %2
+    "=g"(hash)        // %3
+  : "m"(kHash16x33),  // %4
+    "m"(kHashMul0),   // %5
+    "m"(kHashMul1),   // %6
+    "m"(kHashMul2),   // %7
+    "m"(kHashMul3)    // %8
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+  return hash;
+}
+#endif  // defined(__x86_64__) || (defined(__i386__) && !defined(__pic__)))
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif

pjproject-2.11.1/third_party/yuv/source/compare_neon.cc

@@ -0,0 +1,92 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/basic_types.h"
+
+#include "libyuv/compare_row.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__) && \
+    !defined(__aarch64__)
+
+// 256 bits at a time
+// uses short accumulator which restricts count to 131 KB
+uint32 HammingDistance_NEON(const uint8* src_a, const uint8* src_b, int count) {
+  uint32 diff;
+
+  asm volatile(
+      "vmov.u16   q4, #0                         \n"  // accumulator
+
+      "1:                                        \n"
+      "vld1.8     {q0, q1}, [%0]!                \n"
+      "vld1.8     {q2, q3}, [%1]!                \n"
+      "veor.32    q0, q0, q2                     \n"
+      "veor.32    q1, q1, q3                     \n"
+      "vcnt.i8    q0, q0                         \n"
+      "vcnt.i8    q1, q1                         \n"
+      "subs       %2, %2, #32                    \n"
+      "vadd.u8    q0, q0, q1                     \n"  // 16 byte counts
+      "vpadal.u8  q4, q0                         \n"  // 8 shorts
+      "bgt        1b                             \n"
+
+      "vpaddl.u16 q0, q4                         \n"  // 4 ints
+      "vpadd.u32  d0, d0, d1                     \n"
+      "vpadd.u32  d0, d0, d0                     \n"
+      "vmov.32    %3, d0[0]                      \n"
+
+      : "+r"(src_a), "+r"(src_b), "+r"(count), "=r"(diff)
+      :
+      : "cc", "q0", "q1", "q2", "q3", "q4");
+  return diff;
+}
+
+uint32 SumSquareError_NEON(const uint8* src_a, const uint8* src_b, int count) {
+  uint32 sse;
+  asm volatile(
+      "vmov.u8    q8, #0                         \n"
+      "vmov.u8    q10, #0                        \n"
+      "vmov.u8    q9, #0                         \n"
+      "vmov.u8    q11, #0                        \n"
+
+      "1:                                        \n"
+      "vld1.8     {q0}, [%0]!                    \n"
+      "vld1.8     {q1}, [%1]!                    \n"
+      "subs       %2, %2, #16                    \n"
+      "vsubl.u8   q2, d0, d2                     \n"
+      "vsubl.u8   q3, d1, d3                     \n"
+      "vmlal.s16  q8, d4, d4                     \n"
+      "vmlal.s16  q9, d6, d6                     \n"
+      "vmlal.s16  q10, d5, d5                    \n"
+      "vmlal.s16  q11, d7, d7                    \n"
+      "bgt        1b                             \n"
+
+      "vadd.u32   q8, q8, q9                     \n"
+      "vadd.u32   q10, q10, q11                  \n"
+      "vadd.u32   q11, q8, q10                   \n"
+      "vpaddl.u32 q1, q11                        \n"
+      "vadd.u64   d0, d2, d3                     \n"
+      "vmov.32    %3, d0[0]                      \n"
+      : "+r"(src_a), "+r"(src_b), "+r"(count), "=r"(sse)
+      :
+      : "memory", "cc", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11");
+  return sse;
+}
+
+#endif  // defined(__ARM_NEON__) && !defined(__aarch64__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif

pjproject-2.11.1/third_party/yuv/source/compare_neon64.cc

@@ -0,0 +1,86 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/basic_types.h"
+
+#include "libyuv/compare_row.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
+
+// 256 bits at a time
+// uses short accumulator which restricts count to 131 KB
+uint32 HammingDistance_NEON(const uint8* src_a, const uint8* src_b, int count) {
+  uint32 diff;
+  asm volatile(
+      "movi       v4.8h, #0                      \n"
+
+      "1:                                        \n"
+      "ld1        {v0.16b, v1.16b}, [%0], #32    \n"
+      "ld1        {v2.16b, v3.16b}, [%1], #32    \n"
+      "eor        v0.16b, v0.16b, v2.16b         \n"
+      "eor        v1.16b, v1.16b, v3.16b         \n"
+      "cnt        v0.16b, v0.16b                 \n"
+      "cnt        v1.16b, v1.16b                 \n"
+      "subs       %w2, %w2, #32                  \n"
+      "add        v0.16b, v0.16b, v1.16b         \n"
+      "uadalp     v4.8h, v0.16b                  \n"
+      "b.gt       1b                             \n"
+
+      "uaddlv     s4, v4.8h                      \n"
+      "fmov       %w3, s4                        \n"
+      : "+r"(src_a), "+r"(src_b), "+r"(count), "=r"(diff)
+      :
+      : "cc", "v0", "v1", "v2", "v3", "v4");
+  return diff;
+}
+
+uint32 SumSquareError_NEON(const uint8* src_a, const uint8* src_b, int count) {
+  uint32 sse;
+  asm volatile(
+      "eor        v16.16b, v16.16b, v16.16b      \n"
+      "eor        v18.16b, v18.16b, v18.16b      \n"
+      "eor        v17.16b, v17.16b, v17.16b      \n"
+      "eor        v19.16b, v19.16b, v19.16b      \n"
+
+      "1:                                        \n"
+      "ld1        {v0.16b}, [%0], #16            \n"
+      "ld1        {v1.16b}, [%1], #16            \n"
+      "subs       %w2, %w2, #16                  \n"
+      "usubl      v2.8h, v0.8b, v1.8b            \n"
+      "usubl2     v3.8h, v0.16b, v1.16b          \n"
+      "smlal      v16.4s, v2.4h, v2.4h           \n"
+      "smlal      v17.4s, v3.4h, v3.4h           \n"
+      "smlal2     v18.4s, v2.8h, v2.8h           \n"
+      "smlal2     v19.4s, v3.8h, v3.8h           \n"
+      "b.gt       1b                             \n"
+
+      "add        v16.4s, v16.4s, v17.4s         \n"
+      "add        v18.4s, v18.4s, v19.4s         \n"
+      "add        v19.4s, v16.4s, v18.4s         \n"
+      "addv       s0, v19.4s                     \n"
+      "fmov       %w3, s0                        \n"
+      : "+r"(src_a), "+r"(src_b), "+r"(count), "=r"(sse)
+      :
+      : "cc", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19");
+  return sse;
+}
+
+#endif  // !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif

pjproject-2.11.1/third_party/yuv/source/compare_win.cc

@@ -0,0 +1,270 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/basic_types.h"
+#include "libyuv/row.h"
+
+#if defined(_MSC_VER)
+#include <intrin.h>  // For __popcnt
+#endif
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER)
+
+/* Visual Studio 2005 doesn't support __popcnt(). */
+#if (_MSC_VER > 1400)
+uint32 HammingDistance_SSE42(const uint8* src_a,
+                             const uint8* src_b,
+                             int count) {
+  uint32 diff = 0u;
+
+  int i;
+  for (i = 0; i < count - 3; i += 4) {
+    uint32 x = *((uint32*)src_a) ^ *((uint32*)src_b);
+    src_a += 4;
+    src_b += 4;
+    diff += __popcnt(x);
+  }
+  return diff;
+}
+#endif
+
+#if (_MSC_VER >= 1900)
+__declspec(naked)
+#else
+__declspec(naked) __declspec(align(16))
+#endif
+uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count) {
+  __asm {
+    mov        eax, [esp + 4]    // src_a
+    mov        edx, [esp + 8]    // src_b
+    mov        ecx, [esp + 12]   // count
+    pxor       xmm0, xmm0
+    pxor       xmm5, xmm5
+
+    align      4
+  wloop:
+    movdqa     xmm1, [eax]
+    lea        eax,  [eax + 16]
+    movdqa     xmm2, [edx]
+    lea        edx,  [edx + 16]
+    sub        ecx, 16
+    movdqa     xmm3, xmm1  // abs trick
+    psubusb    xmm1, xmm2
+    psubusb    xmm2, xmm3
+    por        xmm1, xmm2
+    movdqa     xmm2, xmm1
+    punpcklbw  xmm1, xmm5
+    punpckhbw  xmm2, xmm5
+    pmaddwd    xmm1, xmm1
+    pmaddwd    xmm2, xmm2
+    paddd      xmm0, xmm1
+    paddd      xmm0, xmm2
+    jg         wloop
+
+    pshufd     xmm1, xmm0, 0xee
+    paddd      xmm0, xmm1
+    pshufd     xmm1, xmm0, 0x01
+    paddd      xmm0, xmm1
+    movd       eax, xmm0
+    ret
+  }
+}
+
+// Visual C 2012 required for AVX2.
+#if _MSC_VER >= 1700
+// C4752: found Intel(R) Advanced Vector Extensions; consider using /arch:AVX.
+#pragma warning(disable: 4752)
+#if (_MSC_VER >= 1900)
+__declspec(naked)
+#else
+__declspec(naked) __declspec(align(16))
+#endif
+uint32 SumSquareError_AVX2(const uint8* src_a, const uint8* src_b, int count) {
+  __asm {
+    mov        eax, [esp + 4]    // src_a
+    mov        edx, [esp + 8]    // src_b
+    mov        ecx, [esp + 12]   // count
+    vpxor      ymm0, ymm0, ymm0  // sum
+    vpxor      ymm5, ymm5, ymm5  // constant 0 for unpck
+    sub        edx, eax
+
+    align      4
+  wloop:
+    vmovdqu    ymm1, [eax]
+    vmovdqu    ymm2, [eax + edx]
+    lea        eax,  [eax + 32]
+    sub        ecx, 32
+    vpsubusb   ymm3, ymm1, ymm2  // abs difference trick
+    vpsubusb   ymm2, ymm2, ymm1
+    vpor       ymm1, ymm2, ymm3
+    vpunpcklbw ymm2, ymm1, ymm5  // u16.  mutates order.
+    vpunpckhbw ymm1, ymm1, ymm5
+    vpmaddwd   ymm2, ymm2, ymm2  // square + hadd to u32.
+    vpmaddwd   ymm1, ymm1, ymm1
+    vpaddd     ymm0, ymm0, ymm1
+    vpaddd     ymm0, ymm0, ymm2
+    jg         wloop
+
+    vpshufd    ymm1, ymm0, 0xee  // 3, 2 + 1, 0 both lanes.
+    vpaddd     ymm0, ymm0, ymm1
+    vpshufd    ymm1, ymm0, 0x01  // 1 + 0 both lanes.
+    vpaddd     ymm0, ymm0, ymm1
+    vpermq     ymm1, ymm0, 0x02  // high + low lane.
+    vpaddd     ymm0, ymm0, ymm1
+    vmovd      eax, xmm0
+    vzeroupper
+    ret
+  }
+}
+#endif  // _MSC_VER >= 1700
+
+#define HAS_HASHDJB2_SSE41
+static uvec32 kHash16x33 = { 0x92d9e201, 0, 0, 0 };  // 33 ^ 16
+static uvec32 kHashMul0 = {
+  0x0c3525e1,  // 33 ^ 15
+  0xa3476dc1,  // 33 ^ 14
+  0x3b4039a1,  // 33 ^ 13
+  0x4f5f0981,  // 33 ^ 12
+};
+static uvec32 kHashMul1 = {
+  0x30f35d61,  // 33 ^ 11
+  0x855cb541,  // 33 ^ 10
+  0x040a9121,  // 33 ^ 9
+  0x747c7101,  // 33 ^ 8
+};
+static uvec32 kHashMul2 = {
+  0xec41d4e1,  // 33 ^ 7
+  0x4cfa3cc1,  // 33 ^ 6
+  0x025528a1,  // 33 ^ 5
+  0x00121881,  // 33 ^ 4
+};
+static uvec32 kHashMul3 = {
+  0x00008c61,  // 33 ^ 3
+  0x00000441,  // 33 ^ 2
+  0x00000021,  // 33 ^ 1
+  0x00000001,  // 33 ^ 0
+};
+
+// 27: 66 0F 38 40 C6     pmulld      xmm0,xmm6
+// 44: 66 0F 38 40 DD     pmulld      xmm3,xmm5
+// 59: 66 0F 38 40 E5     pmulld      xmm4,xmm5
+// 72: 66 0F 38 40 D5     pmulld      xmm2,xmm5
+// 83: 66 0F 38 40 CD     pmulld      xmm1,xmm5
+#define pmulld(reg) _asm _emit 0x66 _asm _emit 0x0F _asm _emit 0x38 \
+    _asm _emit 0x40 _asm _emit reg
+
+#if (_MSC_VER >= 1900)
+__declspec(naked)
+#else
+__declspec(naked) __declspec(align(16))
+#endif
+uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed) {
+  __asm {
+    mov        eax, [esp + 4]    // src
+    mov        ecx, [esp + 8]    // count
+    movd       xmm0, [esp + 12]  // seed
+
+    pxor       xmm7, xmm7        // constant 0 for unpck
+    movdqa     xmm6, kHash16x33
+
+    align      4
+  wloop:
+    movdqu     xmm1, [eax]       // src[0-15]
+    lea        eax, [eax + 16]
+    pmulld(0xc6)                 // pmulld      xmm0,xmm6  hash *= 33 ^ 16
+    movdqa     xmm5, kHashMul0
+    movdqa     xmm2, xmm1
+    punpcklbw  xmm2, xmm7        // src[0-7]
+    movdqa     xmm3, xmm2
+    punpcklwd  xmm3, xmm7        // src[0-3]
+    pmulld(0xdd)                 // pmulld     xmm3, xmm5
+    movdqa     xmm5, kHashMul1
+    movdqa     xmm4, xmm2
+    punpckhwd  xmm4, xmm7        // src[4-7]
+    pmulld(0xe5)                 // pmulld     xmm4, xmm5
+    movdqa     xmm5, kHashMul2
+    punpckhbw  xmm1, xmm7        // src[8-15]
+    movdqa     xmm2, xmm1
+    punpcklwd  xmm2, xmm7        // src[8-11]
+    pmulld(0xd5)                 // pmulld     xmm2, xmm5
+    movdqa     xmm5, kHashMul3
+    punpckhwd  xmm1, xmm7        // src[12-15]
+    pmulld(0xcd)                 // pmulld     xmm1, xmm5
+    paddd      xmm3, xmm4        // add 16 results
+    paddd      xmm1, xmm2
+    sub        ecx, 16
+    paddd      xmm1, xmm3
+
+    pshufd     xmm2, xmm1, 0x0e  // upper 2 dwords
+    paddd      xmm1, xmm2
+    pshufd     xmm2, xmm1, 0x01
+    paddd      xmm1, xmm2
+    paddd      xmm0, xmm1
+    jg         wloop
+
+    movd       eax, xmm0         // return hash
+    ret
+  }
+}
+
+// Visual C 2012 required for AVX2.
+#if _MSC_VER >= 1700
+#if (_MSC_VER >= 1900)
+__declspec(naked)
+#else
+__declspec(naked) __declspec(align(16))
+#endif
+uint32 HashDjb2_AVX2(const uint8* src, int count, uint32 seed) {
+  __asm {
+    mov        eax, [esp + 4]    // src
+    mov        ecx, [esp + 8]    // count
+    movd       xmm0, [esp + 12]  // seed
+    movdqa     xmm6, kHash16x33
+
+    align      4
+  wloop:
+    vpmovzxbd  xmm3, dword ptr [eax]  // src[0-3]
+    pmulld     xmm0, xmm6  // hash *= 33 ^ 16
+    vpmovzxbd  xmm4, dword ptr [eax + 4]  // src[4-7]
+    pmulld     xmm3, kHashMul0
+    vpmovzxbd  xmm2, dword ptr [eax + 8]  // src[8-11]
+    pmulld     xmm4, kHashMul1
+    vpmovzxbd  xmm1, dword ptr [eax + 12]  // src[12-15]
+    pmulld     xmm2, kHashMul2
+    lea        eax, [eax + 16]
+    pmulld     xmm1, kHashMul3
+    paddd      xmm3, xmm4        // add 16 results
+    paddd      xmm1, xmm2
+    sub        ecx, 16
+    paddd      xmm1, xmm3
+    pshufd     xmm2, xmm1, 0x0e  // upper 2 dwords
+    paddd      xmm1, xmm2
+    pshufd     xmm2, xmm1, 0x01
+    paddd      xmm1, xmm2
+    paddd      xmm0, xmm1
+    jg         wloop
+
+    movd       eax, xmm0         // return hash
+    ret
+  }
+}
+#endif  // _MSC_VER >= 1700
+
+#endif  // !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif

pjproject-2.11.1/third_party/yuv/source/convert_jpeg.cc

@@ -0,0 +1,318 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/convert.h"
+#include "libyuv/convert_argb.h"
+
+#ifdef HAVE_JPEG
+#include "libyuv/mjpeg_decoder.h"
+#endif
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#ifdef HAVE_JPEG
+struct I420Buffers {
+  uint8* y;
+  int y_stride;
+  uint8* u;
+  int u_stride;
+  uint8* v;
+  int v_stride;
+  int w;
+  int h;
+};
+
+static void JpegCopyI420(void* opaque,
+                         const uint8* const* data,
+                         const int* strides,
+                         int rows) {
+  I420Buffers* dest = (I420Buffers*)(opaque);
+  I420Copy(data[0], strides[0], data[1], strides[1], data[2], strides[2],
+           dest->y, dest->y_stride, dest->u, dest->u_stride, dest->v,
+           dest->v_stride, dest->w, rows);
+  dest->y += rows * dest->y_stride;
+  dest->u += ((rows + 1) >> 1) * dest->u_stride;
+  dest->v += ((rows + 1) >> 1) * dest->v_stride;
+  dest->h -= rows;
+}
+
+static void JpegI422ToI420(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  I420Buffers* dest = (I420Buffers*)(opaque);
+  I422ToI420(data[0], strides[0], data[1], strides[1], data[2], strides[2],
+             dest->y, dest->y_stride, dest->u, dest->u_stride, dest->v,
+             dest->v_stride, dest->w, rows);
+  dest->y += rows * dest->y_stride;
+  dest->u += ((rows + 1) >> 1) * dest->u_stride;
+  dest->v += ((rows + 1) >> 1) * dest->v_stride;
+  dest->h -= rows;
+}
+
+static void JpegI444ToI420(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  I420Buffers* dest = (I420Buffers*)(opaque);
+  I444ToI420(data[0], strides[0], data[1], strides[1], data[2], strides[2],
+             dest->y, dest->y_stride, dest->u, dest->u_stride, dest->v,
+             dest->v_stride, dest->w, rows);
+  dest->y += rows * dest->y_stride;
+  dest->u += ((rows + 1) >> 1) * dest->u_stride;
+  dest->v += ((rows + 1) >> 1) * dest->v_stride;
+  dest->h -= rows;
+}
+
+static void JpegI400ToI420(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  I420Buffers* dest = (I420Buffers*)(opaque);
+  I400ToI420(data[0], strides[0], dest->y, dest->y_stride, dest->u,
+             dest->u_stride, dest->v, dest->v_stride, dest->w, rows);
+  dest->y += rows * dest->y_stride;
+  dest->u += ((rows + 1) >> 1) * dest->u_stride;
+  dest->v += ((rows + 1) >> 1) * dest->v_stride;
+  dest->h -= rows;
+}
+
+// Query size of MJPG in pixels.
+LIBYUV_API
+int MJPGSize(const uint8* sample, size_t sample_size, int* width, int* height) {
+  MJpegDecoder mjpeg_decoder;
+  LIBYUV_BOOL ret = mjpeg_decoder.LoadFrame(sample, sample_size);
+  if (ret) {
+    *width = mjpeg_decoder.GetWidth();
+    *height = mjpeg_decoder.GetHeight();
+  }
+  mjpeg_decoder.UnloadFrame();
+  return ret ? 0 : -1;  // -1 for runtime failure.
+}
+
+// MJPG (Motion JPeg) to I420
+// TODO(fbarchard): review w and h requirement. dw and dh may be enough.
+LIBYUV_API
+int MJPGToI420(const uint8* sample,
+               size_t sample_size,
+               uint8* y,
+               int y_stride,
+               uint8* u,
+               int u_stride,
+               uint8* v,
+               int v_stride,
+               int w,
+               int h,
+               int dw,
+               int dh) {
+  if (sample_size == kUnknownDataSize) {
+    // ERROR: MJPEG frame size unknown
+    return -1;
+  }
+
+  // TODO(fbarchard): Port MJpeg to C.
+  MJpegDecoder mjpeg_decoder;
+  LIBYUV_BOOL ret = mjpeg_decoder.LoadFrame(sample, sample_size);
+  if (ret &&
+      (mjpeg_decoder.GetWidth() != w || mjpeg_decoder.GetHeight() != h)) {
+    // ERROR: MJPEG frame has unexpected dimensions
+    mjpeg_decoder.UnloadFrame();
+    return 1;  // runtime failure
+  }
+  if (ret) {
+    I420Buffers bufs = {y, y_stride, u, u_stride, v, v_stride, dw, dh};
+    // YUV420
+    if (mjpeg_decoder.GetColorSpace() == MJpegDecoder::kColorSpaceYCbCr &&
+        mjpeg_decoder.GetNumComponents() == 3 &&
+        mjpeg_decoder.GetVertSampFactor(0) == 2 &&
+        mjpeg_decoder.GetHorizSampFactor(0) == 2 &&
+        mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+        mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+        mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+        mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegCopyI420, &bufs, dw, dh);
+      // YUV422
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceYCbCr &&
+               mjpeg_decoder.GetNumComponents() == 3 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 2 &&
+               mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI422ToI420, &bufs, dw, dh);
+      // YUV444
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceYCbCr &&
+               mjpeg_decoder.GetNumComponents() == 3 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI444ToI420, &bufs, dw, dh);
+      // YUV400
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceGrayscale &&
+               mjpeg_decoder.GetNumComponents() == 1 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI400ToI420, &bufs, dw, dh);
+    } else {
+      // TODO(fbarchard): Implement conversion for any other colorspace/sample
+      // factors that occur in practice.
+      // ERROR: Unable to convert MJPEG frame because format is not supported
+      mjpeg_decoder.UnloadFrame();
+      return 1;
+    }
+  }
+  return ret ? 0 : 1;
+}
+
+#ifdef HAVE_JPEG
+struct ARGBBuffers {
+  uint8* argb;
+  int argb_stride;
+  int w;
+  int h;
+};
+
+static void JpegI420ToARGB(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  ARGBBuffers* dest = (ARGBBuffers*)(opaque);
+  I420ToARGB(data[0], strides[0], data[1], strides[1], data[2], strides[2],
+             dest->argb, dest->argb_stride, dest->w, rows);
+  dest->argb += rows * dest->argb_stride;
+  dest->h -= rows;
+}
+
+static void JpegI422ToARGB(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  ARGBBuffers* dest = (ARGBBuffers*)(opaque);
+  I422ToARGB(data[0], strides[0], data[1], strides[1], data[2], strides[2],
+             dest->argb, dest->argb_stride, dest->w, rows);
+  dest->argb += rows * dest->argb_stride;
+  dest->h -= rows;
+}
+
+static void JpegI444ToARGB(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  ARGBBuffers* dest = (ARGBBuffers*)(opaque);
+  I444ToARGB(data[0], strides[0], data[1], strides[1], data[2], strides[2],
+             dest->argb, dest->argb_stride, dest->w, rows);
+  dest->argb += rows * dest->argb_stride;
+  dest->h -= rows;
+}
+
+static void JpegI400ToARGB(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  ARGBBuffers* dest = (ARGBBuffers*)(opaque);
+  I400ToARGB(data[0], strides[0], dest->argb, dest->argb_stride, dest->w, rows);
+  dest->argb += rows * dest->argb_stride;
+  dest->h -= rows;
+}
+
+// MJPG (Motion JPeg) to ARGB
+// TODO(fbarchard): review w and h requirement. dw and dh may be enough.
+LIBYUV_API
+int MJPGToARGB(const uint8* sample,
+               size_t sample_size,
+               uint8* argb,
+               int argb_stride,
+               int w,
+               int h,
+               int dw,
+               int dh) {
+  if (sample_size == kUnknownDataSize) {
+    // ERROR: MJPEG frame size unknown
+    return -1;
+  }
+
+  // TODO(fbarchard): Port MJpeg to C.
+  MJpegDecoder mjpeg_decoder;
+  LIBYUV_BOOL ret = mjpeg_decoder.LoadFrame(sample, sample_size);
+  if (ret &&
+      (mjpeg_decoder.GetWidth() != w || mjpeg_decoder.GetHeight() != h)) {
+    // ERROR: MJPEG frame has unexpected dimensions
+    mjpeg_decoder.UnloadFrame();
+    return 1;  // runtime failure
+  }
+  if (ret) {
+    ARGBBuffers bufs = {argb, argb_stride, dw, dh};
+    // YUV420
+    if (mjpeg_decoder.GetColorSpace() == MJpegDecoder::kColorSpaceYCbCr &&
+        mjpeg_decoder.GetNumComponents() == 3 &&
+        mjpeg_decoder.GetVertSampFactor(0) == 2 &&
+        mjpeg_decoder.GetHorizSampFactor(0) == 2 &&
+        mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+        mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+        mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+        mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI420ToARGB, &bufs, dw, dh);
+      // YUV422
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceYCbCr &&
+               mjpeg_decoder.GetNumComponents() == 3 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 2 &&
+               mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI422ToARGB, &bufs, dw, dh);
+      // YUV444
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceYCbCr &&
+               mjpeg_decoder.GetNumComponents() == 3 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI444ToARGB, &bufs, dw, dh);
+      // YUV400
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceGrayscale &&
+               mjpeg_decoder.GetNumComponents() == 1 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI400ToARGB, &bufs, dw, dh);
+    } else {
+      // TODO(fbarchard): Implement conversion for any other colorspace/sample
+      // factors that occur in practice.
+      // ERROR: Unable to convert MJPEG frame because format is not supported
+      mjpeg_decoder.UnloadFrame();
+      return 1;
+    }
+  }
+  return ret ? 0 : 1;
+}
+#endif
+
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif

pjproject-2.11.1/third_party/yuv/source/convert_to_argb.cc

@@ -0,0 +1,270 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/convert_argb.h"
+
+#include "libyuv/cpu_id.h"
+#ifdef HAVE_JPEG
+#include "libyuv/mjpeg_decoder.h"
+#endif
+#include "libyuv/rotate_argb.h"
+#include "libyuv/row.h"
+#include "libyuv/video_common.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Convert camera sample to ARGB with cropping, rotation and vertical flip.
+// src_width is used for source stride computation
+// src_height is used to compute location of planes, and indicate inversion
+// sample_size is measured in bytes and is the size of the frame.
+//   With MJPEG it is the compressed size of the frame.
+LIBYUV_API
+int ConvertToARGB(const uint8* sample,
+                  size_t sample_size,
+                  uint8* crop_argb,
+                  int argb_stride,
+                  int crop_x,
+                  int crop_y,
+                  int src_width,
+                  int src_height,
+                  int crop_width,
+                  int crop_height,
+                  enum RotationMode rotation,
+                  uint32 fourcc) {
+  uint32 format = CanonicalFourCC(fourcc);
+  int aligned_src_width = (src_width + 1) & ~1;
+  const uint8* src;
+  const uint8* src_uv;
+  int abs_src_height = (src_height < 0) ? -src_height : src_height;
+  int inv_crop_height = (crop_height < 0) ? -crop_height : crop_height;
+  int r = 0;
+
+  // One pass rotation is available for some formats. For the rest, convert
+  // to ARGB (with optional vertical flipping) into a temporary ARGB buffer,
+  // and then rotate the ARGB to the final destination buffer.
+  // For in-place conversion, if destination crop_argb is same as source sample,
+  // also enable temporary buffer.
+  LIBYUV_BOOL need_buf =
+      (rotation && format != FOURCC_ARGB) || crop_argb == sample;
+  uint8* dest_argb = crop_argb;
+  int dest_argb_stride = argb_stride;
+  uint8* rotate_buffer = NULL;
+  int abs_crop_height = (crop_height < 0) ? -crop_height : crop_height;
+
+  if (crop_argb == NULL || sample == NULL || src_width <= 0 ||
+      crop_width <= 0 || src_height == 0 || crop_height == 0) {
+    return -1;
+  }
+  if (src_height < 0) {
+    inv_crop_height = -inv_crop_height;
+  }
+
+  if (need_buf) {
+    int argb_size = crop_width * 4 * abs_crop_height;
+    rotate_buffer = (uint8*)malloc(argb_size); /* NOLINT */
+    if (!rotate_buffer) {
+      return 1;  // Out of memory runtime error.
+    }
+    crop_argb = rotate_buffer;
+    argb_stride = crop_width * 4;
+  }
+
+  switch (format) {
+    // Single plane formats
+    case FOURCC_YUY2:
+      src = sample + (aligned_src_width * crop_y + crop_x) * 2;
+      r = YUY2ToARGB(src, aligned_src_width * 2, crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_UYVY:
+      src = sample + (aligned_src_width * crop_y + crop_x) * 2;
+      r = UYVYToARGB(src, aligned_src_width * 2, crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_24BG:
+      src = sample + (src_width * crop_y + crop_x) * 3;
+      r = RGB24ToARGB(src, src_width * 3, crop_argb, argb_stride, crop_width,
+                      inv_crop_height);
+      break;
+    case FOURCC_RAW:
+      src = sample + (src_width * crop_y + crop_x) * 3;
+      r = RAWToARGB(src, src_width * 3, crop_argb, argb_stride, crop_width,
+                    inv_crop_height);
+      break;
+    case FOURCC_ARGB:
+      if (!need_buf && !rotation) {
+        src = sample + (src_width * crop_y + crop_x) * 4;
+        r = ARGBToARGB(src, src_width * 4, crop_argb, argb_stride, crop_width,
+                       inv_crop_height);
+      }
+      break;
+    case FOURCC_BGRA:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = BGRAToARGB(src, src_width * 4, crop_argb, argb_stride, crop_width,
+                     inv_crop_height);
+      break;
+    case FOURCC_ABGR:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = ABGRToARGB(src, src_width * 4, crop_argb, argb_stride, crop_width,
+                     inv_crop_height);
+      break;
+    case FOURCC_RGBA:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = RGBAToARGB(src, src_width * 4, crop_argb, argb_stride, crop_width,
+                     inv_crop_height);
+      break;
+    case FOURCC_RGBP:
+      src = sample + (src_width * crop_y + crop_x) * 2;
+      r = RGB565ToARGB(src, src_width * 2, crop_argb, argb_stride, crop_width,
+                       inv_crop_height);
+      break;
+    case FOURCC_RGBO:
+      src = sample + (src_width * crop_y + crop_x) * 2;
+      r = ARGB1555ToARGB(src, src_width * 2, crop_argb, argb_stride, crop_width,
+                         inv_crop_height);
+      break;
+    case FOURCC_R444:
+      src = sample + (src_width * crop_y + crop_x) * 2;
+      r = ARGB4444ToARGB(src, src_width * 2, crop_argb, argb_stride, crop_width,
+                         inv_crop_height);
+      break;
+    case FOURCC_I400:
+      src = sample + src_width * crop_y + crop_x;
+      r = I400ToARGB(src, src_width, crop_argb, argb_stride, crop_width,
+                     inv_crop_height);
+      break;
+
+    // Biplanar formats
+    case FOURCC_NV12:
+      src = sample + (src_width * crop_y + crop_x);
+      src_uv = sample + aligned_src_width * (src_height + crop_y / 2) + crop_x;
+      r = NV12ToARGB(src, src_width, src_uv, aligned_src_width, crop_argb,
+                     argb_stride, crop_width, inv_crop_height);
+      break;
+    case FOURCC_NV21:
+      src = sample + (src_width * crop_y + crop_x);
+      src_uv = sample + aligned_src_width * (src_height + crop_y / 2) + crop_x;
+      // Call NV12 but with u and v parameters swapped.
+      r = NV21ToARGB(src, src_width, src_uv, aligned_src_width, crop_argb,
+                     argb_stride, crop_width, inv_crop_height);
+      break;
+    case FOURCC_M420:
+      src = sample + (src_width * crop_y) * 12 / 8 + crop_x;
+      r = M420ToARGB(src, src_width, crop_argb, argb_stride, crop_width,
+                     inv_crop_height);
+      break;
+    // Triplanar formats
+    case FOURCC_I420:
+    case FOURCC_YV12: {
+      const uint8* src_y = sample + (src_width * crop_y + crop_x);
+      const uint8* src_u;
+      const uint8* src_v;
+      int halfwidth = (src_width + 1) / 2;
+      int halfheight = (abs_src_height + 1) / 2;
+      if (format == FOURCC_YV12) {
+        src_v = sample + src_width * abs_src_height +
+                (halfwidth * crop_y + crop_x) / 2;
+        src_u = sample + src_width * abs_src_height +
+                halfwidth * (halfheight + crop_y / 2) + crop_x / 2;
+      } else {
+        src_u = sample + src_width * abs_src_height +
+                (halfwidth * crop_y + crop_x) / 2;
+        src_v = sample + src_width * abs_src_height +
+                halfwidth * (halfheight + crop_y / 2) + crop_x / 2;
+      }
+      r = I420ToARGB(src_y, src_width, src_u, halfwidth, src_v, halfwidth,
+                     crop_argb, argb_stride, crop_width, inv_crop_height);
+      break;
+    }
+
+    case FOURCC_J420: {
+      const uint8* src_y = sample + (src_width * crop_y + crop_x);
+      const uint8* src_u;
+      const uint8* src_v;
+      int halfwidth = (src_width + 1) / 2;
+      int halfheight = (abs_src_height + 1) / 2;
+      src_u = sample + src_width * abs_src_height +
+              (halfwidth * crop_y + crop_x) / 2;
+      src_v = sample + src_width * abs_src_height +
+              halfwidth * (halfheight + crop_y / 2) + crop_x / 2;
+      r = J420ToARGB(src_y, src_width, src_u, halfwidth, src_v, halfwidth,
+                     crop_argb, argb_stride, crop_width, inv_crop_height);
+      break;
+    }
+
+    case FOURCC_I422:
+    case FOURCC_YV16: {
+      const uint8* src_y = sample + src_width * crop_y + crop_x;
+      const uint8* src_u;
+      const uint8* src_v;
+      int halfwidth = (src_width + 1) / 2;
+      if (format == FOURCC_YV16) {
+        src_v = sample + src_width * abs_src_height + halfwidth * crop_y +
+                crop_x / 2;
+        src_u = sample + src_width * abs_src_height +
+                halfwidth * (abs_src_height + crop_y) + crop_x / 2;
+      } else {
+        src_u = sample + src_width * abs_src_height + halfwidth * crop_y +
+                crop_x / 2;
+        src_v = sample + src_width * abs_src_height +
+                halfwidth * (abs_src_height + crop_y) + crop_x / 2;
+      }
+      r = I422ToARGB(src_y, src_width, src_u, halfwidth, src_v, halfwidth,
+                     crop_argb, argb_stride, crop_width, inv_crop_height);
+      break;
+    }
+    case FOURCC_I444:
+    case FOURCC_YV24: {
+      const uint8* src_y = sample + src_width * crop_y + crop_x;
+      const uint8* src_u;
+      const uint8* src_v;
+      if (format == FOURCC_YV24) {
+        src_v = sample + src_width * (abs_src_height + crop_y) + crop_x;
+        src_u = sample + src_width * (abs_src_height * 2 + crop_y) + crop_x;
+      } else {
+        src_u = sample + src_width * (abs_src_height + crop_y) + crop_x;
+        src_v = sample + src_width * (abs_src_height * 2 + crop_y) + crop_x;
+      }
+      r = I444ToARGB(src_y, src_width, src_u, src_width, src_v, src_width,
+                     crop_argb, argb_stride, crop_width, inv_crop_height);
+      break;
+    }
+#ifdef HAVE_JPEG
+    case FOURCC_MJPG:
+      r = MJPGToARGB(sample, sample_size, crop_argb, argb_stride, src_width,
+                     abs_src_height, crop_width, inv_crop_height);
+      break;
+#endif
+    default:
+      r = -1;  // unknown fourcc - return failure code.
+  }
+
+  if (need_buf) {
+    if (!r) {
+      r = ARGBRotate(crop_argb, argb_stride, dest_argb, dest_argb_stride,
+                     crop_width, abs_crop_height, rotation);
+    }
+    free(rotate_buffer);
+  } else if (rotation) {
+    src = sample + (src_width * crop_y + crop_x) * 4;
+    r = ARGBRotate(src, src_width * 4, crop_argb, argb_stride, crop_width,
+                   inv_crop_height, rotation);
+  }
+
+  return r;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif

pjproject-2.11.1/third_party/yuv/source/convert_to_i420.cc

@@ -0,0 +1,263 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stdlib.h>
+
+#include "libyuv/convert.h"
+
+#include "libyuv/video_common.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Convert camera sample to I420 with cropping, rotation and vertical flip.
+// src_width is used for source stride computation
+// src_height is used to compute location of planes, and indicate inversion
+// sample_size is measured in bytes and is the size of the frame.
+//   With MJPEG it is the compressed size of the frame.
+LIBYUV_API
+int ConvertToI420(const uint8* sample,
+                  size_t sample_size,
+                  uint8* y,
+                  int y_stride,
+                  uint8* u,
+                  int u_stride,
+                  uint8* v,
+                  int v_stride,
+                  int crop_x,
+                  int crop_y,
+                  int src_width,
+                  int src_height,
+                  int crop_width,
+                  int crop_height,
+                  enum RotationMode rotation,
+                  uint32 fourcc) {
+  uint32 format = CanonicalFourCC(fourcc);
+  int aligned_src_width = (src_width + 1) & ~1;
+  const uint8* src;
+  const uint8* src_uv;
+  const int abs_src_height = (src_height < 0) ? -src_height : src_height;
+  // TODO(nisse): Why allow crop_height < 0?
+  const int abs_crop_height = (crop_height < 0) ? -crop_height : crop_height;
+  int r = 0;
+  LIBYUV_BOOL need_buf =
+      (rotation && format != FOURCC_I420 && format != FOURCC_NV12 &&
+       format != FOURCC_NV21 && format != FOURCC_YV12) ||
+      y == sample;
+  uint8* tmp_y = y;
+  uint8* tmp_u = u;
+  uint8* tmp_v = v;
+  int tmp_y_stride = y_stride;
+  int tmp_u_stride = u_stride;
+  int tmp_v_stride = v_stride;
+  uint8* rotate_buffer = NULL;
+  const int inv_crop_height =
+      (src_height < 0) ? -abs_crop_height : abs_crop_height;
+
+  if (!y || !u || !v || !sample || src_width <= 0 || crop_width <= 0 ||
+      src_height == 0 || crop_height == 0) {
+    return -1;
+  }
+
+  // One pass rotation is available for some formats. For the rest, convert
+  // to I420 (with optional vertical flipping) into a temporary I420 buffer,
+  // and then rotate the I420 to the final destination buffer.
+  // For in-place conversion, if destination y is same as source sample,
+  // also enable temporary buffer.
+  if (need_buf) {
+    int y_size = crop_width * abs_crop_height;
+    int uv_size = ((crop_width + 1) / 2) * ((abs_crop_height + 1) / 2);
+    rotate_buffer = (uint8*)malloc(y_size + uv_size * 2); /* NOLINT */
+    if (!rotate_buffer) {
+      return 1;  // Out of memory runtime error.
+    }
+    y = rotate_buffer;
+    u = y + y_size;
+    v = u + uv_size;
+    y_stride = crop_width;
+    u_stride = v_stride = ((crop_width + 1) / 2);
+  }
+
+  switch (format) {
+    // Single plane formats
+    case FOURCC_YUY2:
+      src = sample + (aligned_src_width * crop_y + crop_x) * 2;
+      r = YUY2ToI420(src, aligned_src_width * 2, y, y_stride, u, u_stride, v,
+                     v_stride, crop_width, inv_crop_height);
+      break;
+    case FOURCC_UYVY:
+      src = sample + (aligned_src_width * crop_y + crop_x) * 2;
+      r = UYVYToI420(src, aligned_src_width * 2, y, y_stride, u, u_stride, v,
+                     v_stride, crop_width, inv_crop_height);
+      break;
+    case FOURCC_RGBP:
+      src = sample + (src_width * crop_y + crop_x) * 2;
+      r = RGB565ToI420(src, src_width * 2, y, y_stride, u, u_stride, v,
+                       v_stride, crop_width, inv_crop_height);
+      break;
+    case FOURCC_RGBO:
+      src = sample + (src_width * crop_y + crop_x) * 2;
+      r = ARGB1555ToI420(src, src_width * 2, y, y_stride, u, u_stride, v,
+                         v_stride, crop_width, inv_crop_height);
+      break;
+    case FOURCC_R444:
+      src = sample + (src_width * crop_y + crop_x) * 2;
+      r = ARGB4444ToI420(src, src_width * 2, y, y_stride, u, u_stride, v,
+                         v_stride, crop_width, inv_crop_height);
+      break;
+    case FOURCC_24BG:
+      src = sample + (src_width * crop_y + crop_x) * 3;
+      r = RGB24ToI420(src, src_width * 3, y, y_stride, u, u_stride, v, v_stride,
+                      crop_width, inv_crop_height);
+      break;
+    case FOURCC_RAW:
+      src = sample + (src_width * crop_y + crop_x) * 3;
+      r = RAWToI420(src, src_width * 3, y, y_stride, u, u_stride, v, v_stride,
+                    crop_width, inv_crop_height);
+      break;
+    case FOURCC_ARGB:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = ARGBToI420(src, src_width * 4, y, y_stride, u, u_stride, v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_BGRA:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = BGRAToI420(src, src_width * 4, y, y_stride, u, u_stride, v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_ABGR:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = ABGRToI420(src, src_width * 4, y, y_stride, u, u_stride, v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_RGBA:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = RGBAToI420(src, src_width * 4, y, y_stride, u, u_stride, v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_I400:
+      src = sample + src_width * crop_y + crop_x;
+      r = I400ToI420(src, src_width, y, y_stride, u, u_stride, v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    // Biplanar formats
+    case FOURCC_NV12:
+      src = sample + (src_width * crop_y + crop_x);
+      src_uv = sample + (src_width * src_height) +
+               ((crop_y / 2) * aligned_src_width) + ((crop_x / 2) * 2);
+      r = NV12ToI420Rotate(src, src_width, src_uv, aligned_src_width, y,
+                           y_stride, u, u_stride, v, v_stride, crop_width,
+                           inv_crop_height, rotation);
+      break;
+    case FOURCC_NV21:
+      src = sample + (src_width * crop_y + crop_x);
+      src_uv = sample + (src_width * src_height) +
+               ((crop_y / 2) * aligned_src_width) + ((crop_x / 2) * 2);
+      // Call NV12 but with u and v parameters swapped.
+      r = NV12ToI420Rotate(src, src_width, src_uv, aligned_src_width, y,
+                           y_stride, v, v_stride, u, u_stride, crop_width,
+                           inv_crop_height, rotation);
+      break;
+    case FOURCC_M420:
+      src = sample + (src_width * crop_y) * 12 / 8 + crop_x;
+      r = M420ToI420(src, src_width, y, y_stride, u, u_stride, v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    // Triplanar formats
+    case FOURCC_I420:
+    case FOURCC_YV12: {
+      const uint8* src_y = sample + (src_width * crop_y + crop_x);
+      const uint8* src_u;
+      const uint8* src_v;
+      int halfwidth = (src_width + 1) / 2;
+      int halfheight = (abs_src_height + 1) / 2;
+      if (format == FOURCC_YV12) {
+        src_v = sample + src_width * abs_src_height +
+                (halfwidth * crop_y + crop_x) / 2;
+        src_u = sample + src_width * abs_src_height +
+                halfwidth * (halfheight + crop_y / 2) + crop_x / 2;
+      } else {
+        src_u = sample + src_width * abs_src_height +
+                (halfwidth * crop_y + crop_x) / 2;
+        src_v = sample + src_width * abs_src_height +
+                halfwidth * (halfheight + crop_y / 2) + crop_x / 2;
+      }
+      r = I420Rotate(src_y, src_width, src_u, halfwidth, src_v, halfwidth, y,
+                     y_stride, u, u_stride, v, v_stride, crop_width,
+                     inv_crop_height, rotation);
+      break;
+    }
+    case FOURCC_I422:
+    case FOURCC_YV16: {
+      const uint8* src_y = sample + src_width * crop_y + crop_x;
+      const uint8* src_u;
+      const uint8* src_v;
+      int halfwidth = (src_width + 1) / 2;
+      if (format == FOURCC_YV16) {
+        src_v = sample + src_width * abs_src_height + halfwidth * crop_y +
+                crop_x / 2;
+        src_u = sample + src_width * abs_src_height +
+                halfwidth * (abs_src_height + crop_y) + crop_x / 2;
+      } else {
+        src_u = sample + src_width * abs_src_height + halfwidth * crop_y +
+                crop_x / 2;
+        src_v = sample + src_width * abs_src_height +
+                halfwidth * (abs_src_height + crop_y) + crop_x / 2;
+      }
+      r = I422ToI420(src_y, src_width, src_u, halfwidth, src_v, halfwidth, y,
+                     y_stride, u, u_stride, v, v_stride, crop_width,
+                     inv_crop_height);
+      break;
+    }
+    case FOURCC_I444:
+    case FOURCC_YV24: {
+      const uint8* src_y = sample + src_width * crop_y + crop_x;
+      const uint8* src_u;
+      const uint8* src_v;
+      if (format == FOURCC_YV24) {
+        src_v = sample + src_width * (abs_src_height + crop_y) + crop_x;
+        src_u = sample + src_width * (abs_src_height * 2 + crop_y) + crop_x;
+      } else {
+        src_u = sample + src_width * (abs_src_height + crop_y) + crop_x;
+        src_v = sample + src_width * (abs_src_height * 2 + crop_y) + crop_x;
+      }
+      r = I444ToI420(src_y, src_width, src_u, src_width, src_v, src_width, y,
+                     y_stride, u, u_stride, v, v_stride, crop_width,
+                     inv_crop_height);
+      break;
+    }
+#ifdef HAVE_JPEG
+    case FOURCC_MJPG:
+      r = MJPGToI420(sample, sample_size, y, y_stride, u, u_stride, v, v_stride,
+                     src_width, abs_src_height, crop_width, inv_crop_height);
+      break;
+#endif
+    default:
+      r = -1;  // unknown fourcc - return failure code.
+  }
+
+  if (need_buf) {
+    if (!r) {
+      r = I420Rotate(y, y_stride, u, u_stride, v, v_stride, tmp_y, tmp_y_stride,
+                     tmp_u, tmp_u_stride, tmp_v, tmp_v_stride, crop_width,
+                     abs_crop_height, rotation);
+    }
+    free(rotate_buffer);
+  }
+
+  return r;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif

pjproject-2.11.1/third_party/yuv/source/cpu_id.cc

@@ -0,0 +1,355 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/cpu_id.h"
+
+#if defined(_MSC_VER)
+#include <intrin.h>  // For __cpuidex()
+#endif
+#if !defined(__pnacl__) && !defined(__CLR_VER) &&                           \
+    !defined(__native_client__) && (defined(_M_IX86) || defined(_M_X64)) && \
+    defined(_MSC_FULL_VER) && (_MSC_FULL_VER >= 160040219)
+#include <immintrin.h>  // For _xgetbv()
+#endif
+
+#if !defined(__native_client__)
+#include <stdlib.h>  // For getenv()
+#endif
+
+// For ArmCpuCaps() but unittested on all platforms
+#include <stdio.h>
+#include <string.h>
+
+#include "libyuv/basic_types.h"  // For CPU_X86
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// For functions that use the stack and have runtime checks for overflow,
+// use SAFEBUFFERS to avoid additional check.
+#if defined(_MSC_FULL_VER) && (_MSC_FULL_VER >= 160040219) && \
+    !defined(__clang__)
+#define SAFEBUFFERS __declspec(safebuffers)
+#else
+#define SAFEBUFFERS
+#endif
+
+// cpu_info_ variable for SIMD instruction sets detected.
+LIBYUV_API int cpu_info_ = 0;
+
+// TODO(fbarchard): Consider using int for cpuid so casting is not needed.
+// Low level cpuid for X86.
+#if (defined(_M_IX86) || defined(_M_X64) || defined(__i386__) || \
+     defined(__x86_64__)) &&                                     \
+    !defined(__pnacl__) && !defined(__CLR_VER)
+LIBYUV_API
+void CpuId(int info_eax, int info_ecx, int* cpu_info) {
+#if defined(_MSC_VER)
+// Visual C version uses intrinsic or inline x86 assembly.
+#if defined(_MSC_FULL_VER) && (_MSC_FULL_VER >= 160040219)
+  __cpuidex(cpu_info, info_eax, info_ecx);
+#elif defined(_M_IX86)
+  __asm {
+    mov        eax, info_eax
+    mov        ecx, info_ecx
+    mov        edi, cpu_info
+    cpuid
+    mov        [edi], eax
+    mov        [edi + 4], ebx
+    mov        [edi + 8], ecx
+    mov        [edi + 12], edx
+  }
+#else  // Visual C but not x86
+  if (info_ecx == 0) {
+    __cpuid(cpu_info, info_eax);
+  } else {
+    cpu_info[3] = cpu_info[2] = cpu_info[1] = cpu_info[0] = 0u;
+  }
+#endif
+// GCC version uses inline x86 assembly.
+#else  // defined(_MSC_VER)
+  int info_ebx, info_edx;
+  asm volatile(
+#if defined(__i386__) && defined(__PIC__)
+      // Preserve ebx for fpic 32 bit.
+      "mov %%ebx, %%edi                          \n"
+      "cpuid                                     \n"
+      "xchg %%edi, %%ebx                         \n"
+      : "=D"(info_ebx),
+#else
+      "cpuid                                     \n"
+      : "=b"(info_ebx),
+#endif  //  defined( __i386__) && defined(__PIC__)
+        "+a"(info_eax), "+c"(info_ecx), "=d"(info_edx));
+  cpu_info[0] = info_eax;
+  cpu_info[1] = info_ebx;
+  cpu_info[2] = info_ecx;
+  cpu_info[3] = info_edx;
+#endif  // defined(_MSC_VER)
+}
+#else  // (defined(_M_IX86) || defined(_M_X64) ...
+LIBYUV_API
+void CpuId(int eax, int ecx, int* cpu_info) {
+  (void)eax;
+  (void)ecx;
+  cpu_info[0] = cpu_info[1] = cpu_info[2] = cpu_info[3] = 0;
+}
+#endif
+
+// For VS2010 and earlier emit can be used:
+//   _asm _emit 0x0f _asm _emit 0x01 _asm _emit 0xd0  // For VS2010 and earlier.
+//  __asm {
+//    xor        ecx, ecx    // xcr 0
+//    xgetbv
+//    mov        xcr0, eax
+//  }
+// For VS2013 and earlier 32 bit, the _xgetbv(0) optimizer produces bad code.
+// https://code.google.com/p/libyuv/issues/detail?id=529
+#if defined(_M_IX86) && (_MSC_VER < 1900)
+#pragma optimize("g", off)
+#endif
+#if (defined(_M_IX86) || defined(_M_X64) || defined(__i386__) || \
+     defined(__x86_64__)) &&                                     \
+    !defined(__pnacl__) && !defined(__CLR_VER) && !defined(__native_client__)
+// X86 CPUs have xgetbv to detect OS saves high parts of ymm registers.
+int GetXCR0() {
+  int xcr0 = 0;
+#if defined(_MSC_FULL_VER) && (_MSC_FULL_VER >= 160040219)
+  xcr0 = (int)_xgetbv(0);  // VS2010 SP1 required.  NOLINT
+#elif defined(__i386__) || defined(__x86_64__)
+  asm(".byte 0x0f, 0x01, 0xd0" : "=a"(xcr0) : "c"(0) : "%edx");
+#endif  // defined(__i386__) || defined(__x86_64__)
+  return xcr0;
+}
+#else
+// xgetbv unavailable to query for OSSave support.  Return 0.
+#define GetXCR0() 0
+#endif  // defined(_M_IX86) || defined(_M_X64) ..
+// Return optimization to previous setting.
+#if defined(_M_IX86) && (_MSC_VER < 1900)
+#pragma optimize("g", on)
+#endif
+
+// based on libvpx arm_cpudetect.c
+// For Arm, but public to allow testing on any CPU
+LIBYUV_API SAFEBUFFERS int ArmCpuCaps(const char* cpuinfo_name) {
+  char cpuinfo_line[512];
+  FILE* f = fopen(cpuinfo_name, "r");
+  if (!f) {
+    // Assume Neon if /proc/cpuinfo is unavailable.
+    // This will occur for Chrome sandbox for Pepper or Render process.
+    return kCpuHasNEON;
+  }
+  while (fgets(cpuinfo_line, sizeof(cpuinfo_line) - 1, f)) {
+    if (memcmp(cpuinfo_line, "Features", 8) == 0) {
+      char* p = strstr(cpuinfo_line, " neon");
+      if (p && (p[5] == ' ' || p[5] == '\n')) {
+        fclose(f);
+        return kCpuHasNEON;
+      }
+      // aarch64 uses asimd for Neon.
+      p = strstr(cpuinfo_line, " asimd");
+      if (p) {
+        fclose(f);
+        return kCpuHasNEON;
+      }
+    }
+  }
+  fclose(f);
+  return 0;
+}
+
+// TODO(fbarchard): Consider read_msa_ir().
+// TODO(fbarchard): Add unittest.
+LIBYUV_API SAFEBUFFERS int MipsCpuCaps(const char* cpuinfo_name,
+                                       const char ase[]) {
+  char cpuinfo_line[512];
+  FILE* f = fopen(cpuinfo_name, "r");
+  if (!f) {
+    // ase enabled if /proc/cpuinfo is unavailable.
+    if (strcmp(ase, " msa") == 0) {
+      return kCpuHasMSA;
+    }
+    return kCpuHasDSPR2;
+  }
+  while (fgets(cpuinfo_line, sizeof(cpuinfo_line) - 1, f)) {
+    if (memcmp(cpuinfo_line, "ASEs implemented", 16) == 0) {
+      char* p = strstr(cpuinfo_line, ase);
+      if (p) {
+        fclose(f);
+        if (strcmp(ase, " msa") == 0) {
+          return kCpuHasMSA;
+        }
+        return kCpuHasDSPR2;
+      }
+    }
+  }
+  fclose(f);
+  return 0;
+}
+
+// Test environment variable for disabling CPU features. Any non-zero value
+// to disable. Zero ignored to make it easy to set the variable on/off.
+#if !defined(__native_client__) && !defined(_M_ARM)
+
+static LIBYUV_BOOL TestEnv(const char* name) {
+  const char* var = getenv(name);
+  if (var) {
+    if (var[0] != '0') {
+      return LIBYUV_TRUE;
+    }
+  }
+  return LIBYUV_FALSE;
+}
+#else  // nacl does not support getenv().
+static LIBYUV_BOOL TestEnv(const char*) {
+  return LIBYUV_FALSE;
+}
+#endif
+
+static SAFEBUFFERS int GetCpuFlags(void) {
+  int cpu_info = 0;
+#if !defined(__pnacl__) && !defined(__CLR_VER) && defined(CPU_X86)
+  int cpu_info0[4] = {0, 0, 0, 0};
+  int cpu_info1[4] = {0, 0, 0, 0};
+  int cpu_info7[4] = {0, 0, 0, 0};
+  CpuId(0, 0, cpu_info0);
+  CpuId(1, 0, cpu_info1);
+  if (cpu_info0[0] >= 7) {
+    CpuId(7, 0, cpu_info7);
+  }
+  cpu_info = kCpuHasX86 | ((cpu_info1[3] & 0x04000000) ? kCpuHasSSE2 : 0) |
+             ((cpu_info1[2] & 0x00000200) ? kCpuHasSSSE3 : 0) |
+             ((cpu_info1[2] & 0x00080000) ? kCpuHasSSE41 : 0) |
+             ((cpu_info1[2] & 0x00100000) ? kCpuHasSSE42 : 0) |
+             ((cpu_info7[1] & 0x00000200) ? kCpuHasERMS : 0);
+
+  // AVX requires OS saves YMM registers.
+  if (((cpu_info1[2] & 0x1c000000) == 0x1c000000) &&  // AVX and OSXSave
+      ((GetXCR0() & 6) == 6)) {  // Test OS saves YMM registers
+    cpu_info |= kCpuHasAVX | ((cpu_info7[1] & 0x00000020) ? kCpuHasAVX2 : 0) |
+                ((cpu_info1[2] & 0x00001000) ? kCpuHasFMA3 : 0) |
+                ((cpu_info1[2] & 0x20000000) ? kCpuHasF16C : 0);
+
+    // Detect AVX512bw
+    if ((GetXCR0() & 0xe0) == 0xe0) {
+      cpu_info |= (cpu_info7[1] & 0x40000000) ? kCpuHasAVX512BW : 0;
+      cpu_info |= (cpu_info7[1] & 0x80000000) ? kCpuHasAVX512VL : 0;
+      cpu_info |= (cpu_info7[2] & 0x00000002) ? kCpuHasAVX512VBMI : 0;
+      cpu_info |= (cpu_info7[2] & 0x00000040) ? kCpuHasAVX512VBMI2 : 0;
+      cpu_info |= (cpu_info7[2] & 0x00001000) ? kCpuHasAVX512VBITALG : 0;
+      cpu_info |= (cpu_info7[2] & 0x00004000) ? kCpuHasAVX512VPOPCNTDQ : 0;
+      cpu_info |= (cpu_info7[2] & 0x00000100) ? kCpuHasGFNI : 0;
+    }
+  }
+
+  // TODO(fbarchard): Consider moving these to gtest
+  // Environment variable overrides for testing.
+  if (TestEnv("LIBYUV_DISABLE_X86")) {
+    cpu_info &= ~kCpuHasX86;
+  }
+  if (TestEnv("LIBYUV_DISABLE_SSE2")) {
+    cpu_info &= ~kCpuHasSSE2;
+  }
+  if (TestEnv("LIBYUV_DISABLE_SSSE3")) {
+    cpu_info &= ~kCpuHasSSSE3;
+  }
+  if (TestEnv("LIBYUV_DISABLE_SSE41")) {
+    cpu_info &= ~kCpuHasSSE41;
+  }
+  if (TestEnv("LIBYUV_DISABLE_SSE42")) {
+    cpu_info &= ~kCpuHasSSE42;
+  }
+  if (TestEnv("LIBYUV_DISABLE_AVX")) {
+    cpu_info &= ~kCpuHasAVX;
+  }
+  if (TestEnv("LIBYUV_DISABLE_AVX2")) {
+    cpu_info &= ~kCpuHasAVX2;
+  }
+  if (TestEnv("LIBYUV_DISABLE_ERMS")) {
+    cpu_info &= ~kCpuHasERMS;
+  }
+  if (TestEnv("LIBYUV_DISABLE_FMA3")) {
+    cpu_info &= ~kCpuHasFMA3;
+  }
+  if (TestEnv("LIBYUV_DISABLE_F16C")) {
+    cpu_info &= ~kCpuHasF16C;
+  }
+  if (TestEnv("LIBYUV_DISABLE_AVX512BW")) {
+    cpu_info &= ~kCpuHasAVX512BW;
+  }
+
+#endif
+#if defined(__mips__) && defined(__linux__)
+#if defined(__mips_dspr2)
+  cpu_info |= kCpuHasDSPR2;
+#endif
+#if defined(__mips_msa)
+  cpu_info = MipsCpuCaps("/proc/cpuinfo", " msa");
+#endif
+  cpu_info |= kCpuHasMIPS;
+  if (getenv("LIBYUV_DISABLE_DSPR2")) {
+    cpu_info &= ~kCpuHasDSPR2;
+  }
+  if (getenv("LIBYUV_DISABLE_MSA")) {
+    cpu_info &= ~kCpuHasMSA;
+  }
+#endif
+#if defined(__arm__) || defined(__aarch64__)
+// gcc -mfpu=neon defines __ARM_NEON__
+// __ARM_NEON__ generates code that requires Neon.  NaCL also requires Neon.
+// For Linux, /proc/cpuinfo can be tested but without that assume Neon.
+#if defined(__ARM_NEON__) || defined(__native_client__) || !defined(__linux__)
+  cpu_info = kCpuHasNEON;
+// For aarch64(arm64), /proc/cpuinfo's feature is not complete, e.g. no neon
+// flag in it.
+// So for aarch64, neon enabling is hard coded here.
+#endif
+#if defined(__aarch64__)
+  cpu_info = kCpuHasNEON;
+#else
+  // Linux arm parse text file for neon detect.
+  cpu_info = ArmCpuCaps("/proc/cpuinfo");
+#endif
+  cpu_info |= kCpuHasARM;
+  if (TestEnv("LIBYUV_DISABLE_NEON")) {
+    cpu_info &= ~kCpuHasNEON;
+  }
+#endif  // __arm__
+  if (TestEnv("LIBYUV_DISABLE_ASM")) {
+    cpu_info = 0;
+  }
+  cpu_info |= kCpuInitialized;
+  return cpu_info;
+}
+
+// Note that use of this function is not thread safe.
+LIBYUV_API
+int MaskCpuFlags(int enable_flags) {
+  int cpu_info = GetCpuFlags() & enable_flags;
+#ifdef __ATOMIC_RELAXED
+  __atomic_store_n(&cpu_info_, cpu_info, __ATOMIC_RELAXED);
+#else
+  cpu_info_ = cpu_info;
+#endif
+  return cpu_info;
+}
+
+LIBYUV_API
+int InitCpuFlags(void) {
+  return MaskCpuFlags(-1);
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif

pjproject-2.11.1/third_party/yuv/source/mjpeg_decoder.cc

@@ -0,0 +1,573 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/mjpeg_decoder.h"
+
+#ifdef HAVE_JPEG
+#include <assert.h>
+
+#if !defined(__pnacl__) && !defined(__CLR_VER) && \
+    !defined(COVERAGE_ENABLED) && !defined(TARGET_IPHONE_SIMULATOR)
+// Must be included before jpeglib.
+#include <setjmp.h>
+#define HAVE_SETJMP
+
+#if defined(_MSC_VER)
+// disable warning 4324: structure was padded due to __declspec(align())
+#pragma warning(disable : 4324)
+#endif
+
+#endif
+struct FILE;  // For jpeglib.h.
+
+// C++ build requires extern C for jpeg internals.
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include <jpeglib.h>
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#include "libyuv/planar_functions.h"  // For CopyPlane().
+
+namespace libyuv {
+
+#ifdef HAVE_SETJMP
+struct SetJmpErrorMgr {
+  jpeg_error_mgr base;  // Must be at the top
+  jmp_buf setjmp_buffer;
+};
+#endif
+
+const int MJpegDecoder::kColorSpaceUnknown = JCS_UNKNOWN;
+const int MJpegDecoder::kColorSpaceGrayscale = JCS_GRAYSCALE;
+const int MJpegDecoder::kColorSpaceRgb = JCS_RGB;
+const int MJpegDecoder::kColorSpaceYCbCr = JCS_YCbCr;
+const int MJpegDecoder::kColorSpaceCMYK = JCS_CMYK;
+const int MJpegDecoder::kColorSpaceYCCK = JCS_YCCK;
+
+// Methods that are passed to jpeglib.
+boolean fill_input_buffer(jpeg_decompress_struct* cinfo);
+void init_source(jpeg_decompress_struct* cinfo);
+void skip_input_data(jpeg_decompress_struct* cinfo, long num_bytes);  // NOLINT
+void term_source(jpeg_decompress_struct* cinfo);
+void ErrorHandler(jpeg_common_struct* cinfo);
+void OutputHandler(jpeg_common_struct* cinfo);
+
+MJpegDecoder::MJpegDecoder()
+    : has_scanline_padding_(LIBYUV_FALSE),
+      num_outbufs_(0),
+      scanlines_(NULL),
+      scanlines_sizes_(NULL),
+      databuf_(NULL),
+      databuf_strides_(NULL) {
+  decompress_struct_ = new jpeg_decompress_struct;
+  source_mgr_ = new jpeg_source_mgr;
+#ifdef HAVE_SETJMP
+  error_mgr_ = new SetJmpErrorMgr;
+  decompress_struct_->err = jpeg_std_error(&error_mgr_->base);
+  // Override standard exit()-based error handler.
+  error_mgr_->base.error_exit = &ErrorHandler;
+  error_mgr_->base.output_message = &OutputHandler;
+#endif
+  decompress_struct_->client_data = NULL;
+  source_mgr_->init_source = &init_source;
+  source_mgr_->fill_input_buffer = &fill_input_buffer;
+  source_mgr_->skip_input_data = &skip_input_data;
+  source_mgr_->resync_to_restart = &jpeg_resync_to_restart;
+  source_mgr_->term_source = &term_source;
+  jpeg_create_decompress(decompress_struct_);
+  decompress_struct_->src = source_mgr_;
+  buf_vec_.buffers = &buf_;
+  buf_vec_.len = 1;
+}
+
+MJpegDecoder::~MJpegDecoder() {
+  jpeg_destroy_decompress(decompress_struct_);
+  delete decompress_struct_;
+  delete source_mgr_;
+#ifdef HAVE_SETJMP
+  delete error_mgr_;
+#endif
+  DestroyOutputBuffers();
+}
+
+LIBYUV_BOOL MJpegDecoder::LoadFrame(const uint8* src, size_t src_len) {
+  if (!ValidateJpeg(src, src_len)) {
+    return LIBYUV_FALSE;
+  }
+
+  buf_.data = src;
+  buf_.len = static_cast<int>(src_len);
+  buf_vec_.pos = 0;
+  decompress_struct_->client_data = &buf_vec_;
+#ifdef HAVE_SETJMP
+  if (setjmp(error_mgr_->setjmp_buffer)) {
+    // We called jpeg_read_header, it experienced an error, and we called
+    // longjmp() and rewound the stack to here. Return error.
+    return LIBYUV_FALSE;
+  }
+#endif
+  if (jpeg_read_header(decompress_struct_, TRUE) != JPEG_HEADER_OK) {
+    // ERROR: Bad MJPEG header
+    return LIBYUV_FALSE;
+  }
+  AllocOutputBuffers(GetNumComponents());
+  for (int i = 0; i < num_outbufs_; ++i) {
+    int scanlines_size = GetComponentScanlinesPerImcuRow(i);
+    if (scanlines_sizes_[i] != scanlines_size) {
+      if (scanlines_[i]) {
+        delete scanlines_[i];
+      }
+      scanlines_[i] = new uint8*[scanlines_size];
+      scanlines_sizes_[i] = scanlines_size;
+    }
+
+    // We allocate padding for the final scanline to pad it up to DCTSIZE bytes
+    // to avoid memory errors, since jpeglib only reads full MCUs blocks. For
+    // the preceding scanlines, the padding is not needed/wanted because the
+    // following addresses will already be valid (they are the initial bytes of
+    // the next scanline) and will be overwritten when jpeglib writes out that
+    // next scanline.
+    int databuf_stride = GetComponentStride(i);
+    int databuf_size = scanlines_size * databuf_stride;
+    if (databuf_strides_[i] != databuf_stride) {
+      if (databuf_[i]) {
+        delete databuf_[i];
+      }
+      databuf_[i] = new uint8[databuf_size];
+      databuf_strides_[i] = databuf_stride;
+    }
+
+    if (GetComponentStride(i) != GetComponentWidth(i)) {
+      has_scanline_padding_ = LIBYUV_TRUE;
+    }
+  }
+  return LIBYUV_TRUE;
+}
+
+static int DivideAndRoundUp(int numerator, int denominator) {
+  return (numerator + denominator - 1) / denominator;
+}
+
+static int DivideAndRoundDown(int numerator, int denominator) {
+  return numerator / denominator;
+}
+
+// Returns width of the last loaded frame.
+int MJpegDecoder::GetWidth() {
+  return decompress_struct_->image_width;
+}
+
+// Returns height of the last loaded frame.
+int MJpegDecoder::GetHeight() {
+  return decompress_struct_->image_height;
+}
+
+// Returns format of the last loaded frame. The return value is one of the
+// kColorSpace* constants.
+int MJpegDecoder::GetColorSpace() {
+  return decompress_struct_->jpeg_color_space;
+}
+
+// Number of color components in the color space.
+int MJpegDecoder::GetNumComponents() {
+  return decompress_struct_->num_components;
+}
+
+// Sample factors of the n-th component.
+int MJpegDecoder::GetHorizSampFactor(int component) {
+  return decompress_struct_->comp_info[component].h_samp_factor;
+}
+
+int MJpegDecoder::GetVertSampFactor(int component) {
+  return decompress_struct_->comp_info[component].v_samp_factor;
+}
+
+int MJpegDecoder::GetHorizSubSampFactor(int component) {
+  return decompress_struct_->max_h_samp_factor / GetHorizSampFactor(component);
+}
+
+int MJpegDecoder::GetVertSubSampFactor(int component) {
+  return decompress_struct_->max_v_samp_factor / GetVertSampFactor(component);
+}
+
+int MJpegDecoder::GetImageScanlinesPerImcuRow() {
+  return decompress_struct_->max_v_samp_factor * DCTSIZE;
+}
+
+int MJpegDecoder::GetComponentScanlinesPerImcuRow(int component) {
+  int vs = GetVertSubSampFactor(component);
+  return DivideAndRoundUp(GetImageScanlinesPerImcuRow(), vs);
+}
+
+int MJpegDecoder::GetComponentWidth(int component) {
+  int hs = GetHorizSubSampFactor(component);
+  return DivideAndRoundUp(GetWidth(), hs);
+}
+
+int MJpegDecoder::GetComponentHeight(int component) {
+  int vs = GetVertSubSampFactor(component);
+  return DivideAndRoundUp(GetHeight(), vs);
+}
+
+// Get width in bytes padded out to a multiple of DCTSIZE
+int MJpegDecoder::GetComponentStride(int component) {
+  return (GetComponentWidth(component) + DCTSIZE - 1) & ~(DCTSIZE - 1);
+}
+
+int MJpegDecoder::GetComponentSize(int component) {
+  return GetComponentWidth(component) * GetComponentHeight(component);
+}
+
+LIBYUV_BOOL MJpegDecoder::UnloadFrame() {
+#ifdef HAVE_SETJMP
+  if (setjmp(error_mgr_->setjmp_buffer)) {
+    // We called jpeg_abort_decompress, it experienced an error, and we called
+    // longjmp() and rewound the stack to here. Return error.
+    return LIBYUV_FALSE;
+  }
+#endif
+  jpeg_abort_decompress(decompress_struct_);
+  return LIBYUV_TRUE;
+}
+
+// TODO(fbarchard): Allow rectangle to be specified: x, y, width, height.
+LIBYUV_BOOL MJpegDecoder::DecodeToBuffers(uint8** planes,
+                                          int dst_width,
+                                          int dst_height) {
+  if (dst_width != GetWidth() || dst_height > GetHeight()) {
+    // ERROR: Bad dimensions
+    return LIBYUV_FALSE;
+  }
+#ifdef HAVE_SETJMP
+  if (setjmp(error_mgr_->setjmp_buffer)) {
+    // We called into jpeglib, it experienced an error sometime during this
+    // function call, and we called longjmp() and rewound the stack to here.
+    // Return error.
+    return LIBYUV_FALSE;
+  }
+#endif
+  if (!StartDecode()) {
+    return LIBYUV_FALSE;
+  }
+  SetScanlinePointers(databuf_);
+  int lines_left = dst_height;
+  // Compute amount of lines to skip to implement vertical crop.
+  // TODO(fbarchard): Ensure skip is a multiple of maximum component
+  // subsample. ie 2
+  int skip = (GetHeight() - dst_height) / 2;
+  if (skip > 0) {
+    // There is no API to skip lines in the output data, so we read them
+    // into the temp buffer.
+    while (skip >= GetImageScanlinesPerImcuRow()) {
+      if (!DecodeImcuRow()) {
+        FinishDecode();
+        return LIBYUV_FALSE;
+      }
+      skip -= GetImageScanlinesPerImcuRow();
+    }
+    if (skip > 0) {
+      // Have a partial iMCU row left over to skip. Must read it and then
+      // copy the parts we want into the destination.
+      if (!DecodeImcuRow()) {
+        FinishDecode();
+        return LIBYUV_FALSE;
+      }
+      for (int i = 0; i < num_outbufs_; ++i) {
+        // TODO(fbarchard): Compute skip to avoid this
+        assert(skip % GetVertSubSampFactor(i) == 0);
+        int rows_to_skip = DivideAndRoundDown(skip, GetVertSubSampFactor(i));
+        int scanlines_to_copy =
+            GetComponentScanlinesPerImcuRow(i) - rows_to_skip;
+        int data_to_skip = rows_to_skip * GetComponentStride(i);
+        CopyPlane(databuf_[i] + data_to_skip, GetComponentStride(i), planes[i],
+                  GetComponentWidth(i), GetComponentWidth(i),
+                  scanlines_to_copy);
+        planes[i] += scanlines_to_copy * GetComponentWidth(i);
+      }
+      lines_left -= (GetImageScanlinesPerImcuRow() - skip);
+    }
+  }
+
+  // Read full MCUs but cropped horizontally
+  for (; lines_left > GetImageScanlinesPerImcuRow();
+       lines_left -= GetImageScanlinesPerImcuRow()) {
+    if (!DecodeImcuRow()) {
+      FinishDecode();
+      return LIBYUV_FALSE;
+    }
+    for (int i = 0; i < num_outbufs_; ++i) {
+      int scanlines_to_copy = GetComponentScanlinesPerImcuRow(i);
+      CopyPlane(databuf_[i], GetComponentStride(i), planes[i],
+                GetComponentWidth(i), GetComponentWidth(i), scanlines_to_copy);
+      planes[i] += scanlines_to_copy * GetComponentWidth(i);
+    }
+  }
+
+  if (lines_left > 0) {
+    // Have a partial iMCU row left over to decode.
+    if (!DecodeImcuRow()) {
+      FinishDecode();
+      return LIBYUV_FALSE;
+    }
+    for (int i = 0; i < num_outbufs_; ++i) {
+      int scanlines_to_copy =
+          DivideAndRoundUp(lines_left, GetVertSubSampFactor(i));
+      CopyPlane(databuf_[i], GetComponentStride(i), planes[i],
+                GetComponentWidth(i), GetComponentWidth(i), scanlines_to_copy);
+      planes[i] += scanlines_to_copy * GetComponentWidth(i);
+    }
+  }
+  return FinishDecode();
+}
+
+LIBYUV_BOOL MJpegDecoder::DecodeToCallback(CallbackFunction fn,
+                                           void* opaque,
+                                           int dst_width,
+                                           int dst_height) {
+  if (dst_width != GetWidth() || dst_height > GetHeight()) {
+    // ERROR: Bad dimensions
+    return LIBYUV_FALSE;
+  }
+#ifdef HAVE_SETJMP
+  if (setjmp(error_mgr_->setjmp_buffer)) {
+    // We called into jpeglib, it experienced an error sometime during this
+    // function call, and we called longjmp() and rewound the stack to here.
+    // Return error.
+    return LIBYUV_FALSE;
+  }
+#endif
+  if (!StartDecode()) {
+    return LIBYUV_FALSE;
+  }
+  SetScanlinePointers(databuf_);
+  int lines_left = dst_height;
+  // TODO(fbarchard): Compute amount of lines to skip to implement vertical crop
+  int skip = (GetHeight() - dst_height) / 2;
+  if (skip > 0) {
+    while (skip >= GetImageScanlinesPerImcuRow()) {
+      if (!DecodeImcuRow()) {
+        FinishDecode();
+        return LIBYUV_FALSE;
+      }
+      skip -= GetImageScanlinesPerImcuRow();
+    }
+    if (skip > 0) {
+      // Have a partial iMCU row left over to skip.
+      if (!DecodeImcuRow()) {
+        FinishDecode();
+        return LIBYUV_FALSE;
+      }
+      for (int i = 0; i < num_outbufs_; ++i) {
+        // TODO(fbarchard): Compute skip to avoid this
+        assert(skip % GetVertSubSampFactor(i) == 0);
+        int rows_to_skip = DivideAndRoundDown(skip, GetVertSubSampFactor(i));
+        int data_to_skip = rows_to_skip * GetComponentStride(i);
+        // Change our own data buffer pointers so we can pass them to the
+        // callback.
+        databuf_[i] += data_to_skip;
+      }
+      int scanlines_to_copy = GetImageScanlinesPerImcuRow() - skip;
+      (*fn)(opaque, databuf_, databuf_strides_, scanlines_to_copy);
+      // Now change them back.
+      for (int i = 0; i < num_outbufs_; ++i) {
+        int rows_to_skip = DivideAndRoundDown(skip, GetVertSubSampFactor(i));
+        int data_to_skip = rows_to_skip * GetComponentStride(i);
+        databuf_[i] -= data_to_skip;
+      }
+      lines_left -= scanlines_to_copy;
+    }
+  }
+  // Read full MCUs until we get to the crop point.
+  for (; lines_left >= GetImageScanlinesPerImcuRow();
+       lines_left -= GetImageScanlinesPerImcuRow()) {
+    if (!DecodeImcuRow()) {
+      FinishDecode();
+      return LIBYUV_FALSE;
+    }
+    (*fn)(opaque, databuf_, databuf_strides_, GetImageScanlinesPerImcuRow());
+  }
+  if (lines_left > 0) {
+    // Have a partial iMCU row left over to decode.
+    if (!DecodeImcuRow()) {
+      FinishDecode();
+      return LIBYUV_FALSE;
+    }
+    (*fn)(opaque, databuf_, databuf_strides_, lines_left);
+  }
+  return FinishDecode();
+}
+
+void init_source(j_decompress_ptr cinfo) {
+  fill_input_buffer(cinfo);
+}
+
+boolean fill_input_buffer(j_decompress_ptr cinfo) {
+  BufferVector* buf_vec = reinterpret_cast<BufferVector*>(cinfo->client_data);
+  if (buf_vec->pos >= buf_vec->len) {
+    assert(0 && "No more data");
+    // ERROR: No more data
+    return FALSE;
+  }
+  cinfo->src->next_input_byte = buf_vec->buffers[buf_vec->pos].data;
+  cinfo->src->bytes_in_buffer = buf_vec->buffers[buf_vec->pos].len;
+  ++buf_vec->pos;
+  return TRUE;
+}
+
+void skip_input_data(j_decompress_ptr cinfo, long num_bytes) {  // NOLINT
+  cinfo->src->next_input_byte += num_bytes;
+}
+
+void term_source(j_decompress_ptr cinfo) {
+  (void)cinfo;  // Nothing to do.
+}
+
+#ifdef HAVE_SETJMP
+void ErrorHandler(j_common_ptr cinfo) {
+// This is called when a jpeglib command experiences an error. Unfortunately
+// jpeglib's error handling model is not very flexible, because it expects the
+// error handler to not return--i.e., it wants the program to terminate. To
+// recover from errors we use setjmp() as shown in their example. setjmp() is
+// C's implementation for the "call with current continuation" functionality
+// seen in some functional programming languages.
+// A formatted message can be output, but is unsafe for release.
+#ifdef DEBUG
+  char buf[JMSG_LENGTH_MAX];
+  (*cinfo->err->format_message)(cinfo, buf);
+// ERROR: Error in jpeglib: buf
+#endif
+
+  SetJmpErrorMgr* mgr = reinterpret_cast<SetJmpErrorMgr*>(cinfo->err);
+  // This rewinds the call stack to the point of the corresponding setjmp()
+  // and causes it to return (for a second time) with value 1.
+  longjmp(mgr->setjmp_buffer, 1);
+}
+
+// Suppress fprintf warnings.
+void OutputHandler(j_common_ptr cinfo) {
+  (void)cinfo;
+}
+
+#endif  // HAVE_SETJMP
+
+void MJpegDecoder::AllocOutputBuffers(int num_outbufs) {
+  if (num_outbufs != num_outbufs_) {
+    // We could perhaps optimize this case to resize the output buffers without
+    // necessarily having to delete and recreate each one, but it's not worth
+    // it.
+    DestroyOutputBuffers();
+
+    scanlines_ = new uint8**[num_outbufs];
+    scanlines_sizes_ = new int[num_outbufs];
+    databuf_ = new uint8*[num_outbufs];
+    databuf_strides_ = new int[num_outbufs];
+
+    for (int i = 0; i < num_outbufs; ++i) {
+      scanlines_[i] = NULL;
+      scanlines_sizes_[i] = 0;
+      databuf_[i] = NULL;
+      databuf_strides_[i] = 0;
+    }
+
+    num_outbufs_ = num_outbufs;
+  }
+}
+
+void MJpegDecoder::DestroyOutputBuffers() {
+  for (int i = 0; i < num_outbufs_; ++i) {
+    delete[] scanlines_[i];
+    delete[] databuf_[i];
+  }
+  delete[] scanlines_;
+  delete[] databuf_;
+  delete[] scanlines_sizes_;
+  delete[] databuf_strides_;
+  scanlines_ = NULL;
+  databuf_ = NULL;
+  scanlines_sizes_ = NULL;
+  databuf_strides_ = NULL;
+  num_outbufs_ = 0;
+}
+
+// JDCT_IFAST and do_block_smoothing improve performance substantially.
+LIBYUV_BOOL MJpegDecoder::StartDecode() {
+  decompress_struct_->raw_data_out = TRUE;
+  decompress_struct_->dct_method = JDCT_IFAST;  // JDCT_ISLOW is default
+  decompress_struct_->dither_mode = JDITHER_NONE;
+  // Not applicable to 'raw':
+  decompress_struct_->do_fancy_upsampling = (boolean)(LIBYUV_FALSE);
+  // Only for buffered mode:
+  decompress_struct_->enable_2pass_quant = (boolean)(LIBYUV_FALSE);
+  // Blocky but fast:
+  decompress_struct_->do_block_smoothing = (boolean)(LIBYUV_FALSE);
+
+  if (!jpeg_start_decompress(decompress_struct_)) {
+    // ERROR: Couldn't start JPEG decompressor";
+    return LIBYUV_FALSE;
+  }
+  return LIBYUV_TRUE;
+}
+
+LIBYUV_BOOL MJpegDecoder::FinishDecode() {
+  // jpeglib considers it an error if we finish without decoding the whole
+  // image, so we call "abort" rather than "finish".
+  jpeg_abort_decompress(decompress_struct_);
+  return LIBYUV_TRUE;
+}
+
+void MJpegDecoder::SetScanlinePointers(uint8** data) {
+  for (int i = 0; i < num_outbufs_; ++i) {
+    uint8* data_i = data[i];
+    for (int j = 0; j < scanlines_sizes_[i]; ++j) {
+      scanlines_[i][j] = data_i;
+      data_i += GetComponentStride(i);
+    }
+  }
+}
+
+inline LIBYUV_BOOL MJpegDecoder::DecodeImcuRow() {
+  return (unsigned int)(GetImageScanlinesPerImcuRow()) ==
+         jpeg_read_raw_data(decompress_struct_, scanlines_,
+                            GetImageScanlinesPerImcuRow());
+}
+
+// The helper function which recognizes the jpeg sub-sampling type.
+JpegSubsamplingType MJpegDecoder::JpegSubsamplingTypeHelper(
+    int* subsample_x,
+    int* subsample_y,
+    int number_of_components) {
+  if (number_of_components == 3) {  // Color images.
+    if (subsample_x[0] == 1 && subsample_y[0] == 1 && subsample_x[1] == 2 &&
+        subsample_y[1] == 2 && subsample_x[2] == 2 && subsample_y[2] == 2) {
+      return kJpegYuv420;
+    } else if (subsample_x[0] == 1 && subsample_y[0] == 1 &&
+               subsample_x[1] == 2 && subsample_y[1] == 1 &&
+               subsample_x[2] == 2 && subsample_y[2] == 1) {
+      return kJpegYuv422;
+    } else if (subsample_x[0] == 1 && subsample_y[0] == 1 &&
+               subsample_x[1] == 1 && subsample_y[1] == 1 &&
+               subsample_x[2] == 1 && subsample_y[2] == 1) {
+      return kJpegYuv444;
+    }
+  } else if (number_of_components == 1) {  // Grey-scale images.
+    if (subsample_x[0] == 1 && subsample_y[0] == 1) {
+      return kJpegYuv400;
+    }
+  }
+  return kJpegUnknown;
+}
+
+}  // namespace libyuv
+#endif  // HAVE_JPEG

pjproject-2.11.1/third_party/yuv/source/mjpeg_validate.cc

@@ -0,0 +1,70 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/mjpeg_decoder.h"
+
+#include <string.h>  // For memchr.
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Helper function to scan for EOI marker (0xff 0xd9).
+static LIBYUV_BOOL ScanEOI(const uint8* sample, size_t sample_size) {
+  if (sample_size >= 2) {
+    const uint8* end = sample + sample_size - 1;
+    const uint8* it = sample;
+    while (it < end) {
+      // TODO(fbarchard): scan for 0xd9 instead.
+      it = (const uint8*)(memchr(it, 0xff, end - it));
+      if (it == NULL) {
+        break;
+      }
+      if (it[1] == 0xd9) {
+        return LIBYUV_TRUE;  // Success: Valid jpeg.
+      }
+      ++it;  // Skip over current 0xff.
+    }
+  }
+  // ERROR: Invalid jpeg end code not found. Size sample_size
+  return LIBYUV_FALSE;
+}
+
+// Helper function to validate the jpeg appears intact.
+LIBYUV_BOOL ValidateJpeg(const uint8* sample, size_t sample_size) {
+  // Maximum size that ValidateJpeg will consider valid.
+  const size_t kMaxJpegSize = 0x7fffffffull;
+  const size_t kBackSearchSize = 1024;
+  if (sample_size < 64 || sample_size > kMaxJpegSize || !sample) {
+    // ERROR: Invalid jpeg size: sample_size
+    return LIBYUV_FALSE;
+  }
+  if (sample[0] != 0xff || sample[1] != 0xd8) {  // SOI marker
+    // ERROR: Invalid jpeg initial start code
+    return LIBYUV_FALSE;
+  }
+
+  // Look for the End Of Image (EOI) marker near the end of the buffer.
+  if (sample_size > kBackSearchSize) {
+    if (ScanEOI(sample + sample_size - kBackSearchSize, kBackSearchSize)) {
+      return LIBYUV_TRUE;  // Success: Valid jpeg.
+    }
+    // Reduce search size for forward search.
+    sample_size = sample_size - kBackSearchSize + 1;
+  }
+  // Step over SOI marker and scan for EOI.
+  return ScanEOI(sample + 2, sample_size - 2);
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif

pjproject-2.11.1/third_party/yuv/source/rotate.cc

@@ -0,0 +1,549 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/rotate.h"
+
+#include "libyuv/convert.h"
+#include "libyuv/cpu_id.h"
+#include "libyuv/planar_functions.h"
+#include "libyuv/rotate_row.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+LIBYUV_API
+void TransposePlane(const uint8* src,
+                    int src_stride,
+                    uint8* dst,
+                    int dst_stride,
+                    int width,
+                    int height) {
+  int i = height;
+#if defined(HAS_TRANSPOSEWX16_MSA)
+  void (*TransposeWx16)(const uint8* src, int src_stride, uint8* dst,
+                        int dst_stride, int width) = TransposeWx16_C;
+#else
+  void (*TransposeWx8)(const uint8* src, int src_stride, uint8* dst,
+                       int dst_stride, int width) = TransposeWx8_C;
+#endif
+#if defined(HAS_TRANSPOSEWX8_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    TransposeWx8 = TransposeWx8_NEON;
+  }
+#endif
+#if defined(HAS_TRANSPOSEWX8_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    TransposeWx8 = TransposeWx8_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      TransposeWx8 = TransposeWx8_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_TRANSPOSEWX8_FAST_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    TransposeWx8 = TransposeWx8_Fast_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      TransposeWx8 = TransposeWx8_Fast_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_TRANSPOSEWX8_DSPR2)
+  if (TestCpuFlag(kCpuHasDSPR2)) {
+    if (IS_ALIGNED(width, 4) && IS_ALIGNED(src, 4) &&
+        IS_ALIGNED(src_stride, 4)) {
+      TransposeWx8 = TransposeWx8_Fast_DSPR2;
+    } else {
+      TransposeWx8 = TransposeWx8_DSPR2;
+    }
+  }
+#endif
+#if defined(HAS_TRANSPOSEWX16_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    TransposeWx16 = TransposeWx16_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      TransposeWx16 = TransposeWx16_MSA;
+    }
+  }
+#endif
+
+#if defined(HAS_TRANSPOSEWX16_MSA)
+  // Work across the source in 16x16 tiles
+  while (i >= 16) {
+    TransposeWx16(src, src_stride, dst, dst_stride, width);
+    src += 16 * src_stride;  // Go down 16 rows.
+    dst += 16;               // Move over 16 columns.
+    i -= 16;
+  }
+#else
+  // Work across the source in 8x8 tiles
+  while (i >= 8) {
+    TransposeWx8(src, src_stride, dst, dst_stride, width);
+    src += 8 * src_stride;  // Go down 8 rows.
+    dst += 8;               // Move over 8 columns.
+    i -= 8;
+  }
+#endif
+
+  if (i > 0) {
+    TransposeWxH_C(src, src_stride, dst, dst_stride, width, i);
+  }
+}
+
+LIBYUV_API
+void RotatePlane90(const uint8* src,
+                   int src_stride,
+                   uint8* dst,
+                   int dst_stride,
+                   int width,
+                   int height) {
+  // Rotate by 90 is a transpose with the source read
+  // from bottom to top. So set the source pointer to the end
+  // of the buffer and flip the sign of the source stride.
+  src += src_stride * (height - 1);
+  src_stride = -src_stride;
+  TransposePlane(src, src_stride, dst, dst_stride, width, height);
+}
+
+LIBYUV_API
+void RotatePlane270(const uint8* src,
+                    int src_stride,
+                    uint8* dst,
+                    int dst_stride,
+                    int width,
+                    int height) {
+  // Rotate by 270 is a transpose with the destination written
+  // from bottom to top. So set the destination pointer to the end
+  // of the buffer and flip the sign of the destination stride.
+  dst += dst_stride * (width - 1);
+  dst_stride = -dst_stride;
+  TransposePlane(src, src_stride, dst, dst_stride, width, height);
+}
+
+LIBYUV_API
+void RotatePlane180(const uint8* src,
+                    int src_stride,
+                    uint8* dst,
+                    int dst_stride,
+                    int width,
+                    int height) {
+  // Swap first and last row and mirror the content. Uses a temporary row.
+  align_buffer_64(row, width);
+  const uint8* src_bot = src + src_stride * (height - 1);
+  uint8* dst_bot = dst + dst_stride * (height - 1);
+  int half_height = (height + 1) >> 1;
+  int y;
+  void (*MirrorRow)(const uint8* src, uint8* dst, int width) = MirrorRow_C;
+  void (*CopyRow)(const uint8* src, uint8* dst, int width) = CopyRow_C;
+#if defined(HAS_MIRRORROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MirrorRow = MirrorRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      MirrorRow = MirrorRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_MIRRORROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    MirrorRow = MirrorRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      MirrorRow = MirrorRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_MIRRORROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MirrorRow = MirrorRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      MirrorRow = MirrorRow_AVX2;
+    }
+  }
+#endif
+// TODO(fbarchard): Mirror on mips handle unaligned memory.
+#if defined(HAS_MIRRORROW_DSPR2)
+  if (TestCpuFlag(kCpuHasDSPR2) && IS_ALIGNED(src, 4) &&
+      IS_ALIGNED(src_stride, 4) && IS_ALIGNED(dst, 4) &&
+      IS_ALIGNED(dst_stride, 4)) {
+    MirrorRow = MirrorRow_DSPR2;
+  }
+#endif
+#if defined(HAS_MIRRORROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    MirrorRow = MirrorRow_Any_MSA;
+    if (IS_ALIGNED(width, 64)) {
+      MirrorRow = MirrorRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_COPYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    CopyRow = IS_ALIGNED(width, 32) ? CopyRow_SSE2 : CopyRow_Any_SSE2;
+  }
+#endif
+#if defined(HAS_COPYROW_AVX)
+  if (TestCpuFlag(kCpuHasAVX)) {
+    CopyRow = IS_ALIGNED(width, 64) ? CopyRow_AVX : CopyRow_Any_AVX;
+  }
+#endif
+#if defined(HAS_COPYROW_ERMS)
+  if (TestCpuFlag(kCpuHasERMS)) {
+    CopyRow = CopyRow_ERMS;
+  }
+#endif
+#if defined(HAS_COPYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    CopyRow = IS_ALIGNED(width, 32) ? CopyRow_NEON : CopyRow_Any_NEON;
+  }
+#endif
+#if defined(HAS_COPYROW_MIPS)
+  if (TestCpuFlag(kCpuHasMIPS)) {
+    CopyRow = CopyRow_MIPS;
+  }
+#endif
+
+  // Odd height will harmlessly mirror the middle row twice.
+  for (y = 0; y < half_height; ++y) {
+    MirrorRow(src, row, width);  // Mirror first row into a buffer
+    src += src_stride;
+    MirrorRow(src_bot, dst, width);  // Mirror last row into first row
+    dst += dst_stride;
+    CopyRow(row, dst_bot, width);  // Copy first mirrored row into last
+    src_bot -= src_stride;
+    dst_bot -= dst_stride;
+  }
+  free_aligned_buffer_64(row);
+}
+
+LIBYUV_API
+void TransposeUV(const uint8* src,
+                 int src_stride,
+                 uint8* dst_a,
+                 int dst_stride_a,
+                 uint8* dst_b,
+                 int dst_stride_b,
+                 int width,
+                 int height) {
+  int i = height;
+#if defined(HAS_TRANSPOSEUVWX16_MSA)
+  void (*TransposeUVWx16)(const uint8* src, int src_stride, uint8* dst_a,
+                          int dst_stride_a, uint8* dst_b, int dst_stride_b,
+                          int width) = TransposeUVWx16_C;
+#else
+  void (*TransposeUVWx8)(const uint8* src, int src_stride, uint8* dst_a,
+                         int dst_stride_a, uint8* dst_b, int dst_stride_b,
+                         int width) = TransposeUVWx8_C;
+#endif
+#if defined(HAS_TRANSPOSEUVWX8_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    TransposeUVWx8 = TransposeUVWx8_NEON;
+  }
+#endif
+#if defined(HAS_TRANSPOSEUVWX8_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    TransposeUVWx8 = TransposeUVWx8_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      TransposeUVWx8 = TransposeUVWx8_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_TRANSPOSEUVWX8_DSPR2)
+  if (TestCpuFlag(kCpuHasDSPR2) && IS_ALIGNED(width, 2) && IS_ALIGNED(src, 4) &&
+      IS_ALIGNED(src_stride, 4)) {
+    TransposeUVWx8 = TransposeUVWx8_DSPR2;
+  }
+#endif
+#if defined(HAS_TRANSPOSEUVWX16_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    TransposeUVWx16 = TransposeUVWx16_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      TransposeUVWx16 = TransposeUVWx16_MSA;
+    }
+  }
+#endif
+
+#if defined(HAS_TRANSPOSEUVWX16_MSA)
+  // Work through the source in 8x8 tiles.
+  while (i >= 16) {
+    TransposeUVWx16(src, src_stride, dst_a, dst_stride_a, dst_b, dst_stride_b,
+                    width);
+    src += 16 * src_stride;  // Go down 16 rows.
+    dst_a += 16;             // Move over 8 columns.
+    dst_b += 16;             // Move over 8 columns.
+    i -= 16;
+  }
+#else
+  // Work through the source in 8x8 tiles.
+  while (i >= 8) {
+    TransposeUVWx8(src, src_stride, dst_a, dst_stride_a, dst_b, dst_stride_b,
+                   width);
+    src += 8 * src_stride;  // Go down 8 rows.
+    dst_a += 8;             // Move over 8 columns.
+    dst_b += 8;             // Move over 8 columns.
+    i -= 8;
+  }
+#endif
+
+  if (i > 0) {
+    TransposeUVWxH_C(src, src_stride, dst_a, dst_stride_a, dst_b, dst_stride_b,
+                     width, i);
+  }
+}
+
+LIBYUV_API
+void RotateUV90(const uint8* src,
+                int src_stride,
+                uint8* dst_a,
+                int dst_stride_a,
+                uint8* dst_b,
+                int dst_stride_b,
+                int width,
+                int height) {
+  src += src_stride * (height - 1);
+  src_stride = -src_stride;
+
+  TransposeUV(src, src_stride, dst_a, dst_stride_a, dst_b, dst_stride_b, width,
+              height);
+}
+
+LIBYUV_API
+void RotateUV270(const uint8* src,
+                 int src_stride,
+                 uint8* dst_a,
+                 int dst_stride_a,
+                 uint8* dst_b,
+                 int dst_stride_b,
+                 int width,
+                 int height) {
+  dst_a += dst_stride_a * (width - 1);
+  dst_b += dst_stride_b * (width - 1);
+  dst_stride_a = -dst_stride_a;
+  dst_stride_b = -dst_stride_b;
+
+  TransposeUV(src, src_stride, dst_a, dst_stride_a, dst_b, dst_stride_b, width,
+              height);
+}
+
+// Rotate 180 is a horizontal and vertical flip.
+LIBYUV_API
+void RotateUV180(const uint8* src,
+                 int src_stride,
+                 uint8* dst_a,
+                 int dst_stride_a,
+                 uint8* dst_b,
+                 int dst_stride_b,
+                 int width,
+                 int height) {
+  int i;
+  void (*MirrorUVRow)(const uint8* src, uint8* dst_u, uint8* dst_v, int width) =
+      MirrorUVRow_C;
+#if defined(HAS_MIRRORUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    MirrorUVRow = MirrorUVRow_NEON;
+  }
+#endif
+#if defined(HAS_MIRRORUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 16)) {
+    MirrorUVRow = MirrorUVRow_SSSE3;
+  }
+#endif
+#if defined(HAS_MIRRORUVROW_DSPR2)
+  if (TestCpuFlag(kCpuHasDSPR2) && IS_ALIGNED(src, 4) &&
+      IS_ALIGNED(src_stride, 4)) {
+    MirrorUVRow = MirrorUVRow_DSPR2;
+  }
+#endif
+#if defined(HAS_MIRRORUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA) && IS_ALIGNED(width, 32)) {
+    MirrorUVRow = MirrorUVRow_MSA;
+  }
+#endif
+
+  dst_a += dst_stride_a * (height - 1);
+  dst_b += dst_stride_b * (height - 1);
+
+  for (i = 0; i < height; ++i) {
+    MirrorUVRow(src, dst_a, dst_b, width);
+    src += src_stride;
+    dst_a -= dst_stride_a;
+    dst_b -= dst_stride_b;
+  }
+}
+
+LIBYUV_API
+int RotatePlane(const uint8* src,
+                int src_stride,
+                uint8* dst,
+                int dst_stride,
+                int width,
+                int height,
+                enum RotationMode mode) {
+  if (!src || width <= 0 || height == 0 || !dst) {
+    return -1;
+  }
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src = src + (height - 1) * src_stride;
+    src_stride = -src_stride;
+  }
+
+  switch (mode) {
+    case kRotate0:
+      // copy frame
+      CopyPlane(src, src_stride, dst, dst_stride, width, height);
+      return 0;
+    case kRotate90:
+      RotatePlane90(src, src_stride, dst, dst_stride, width, height);
+      return 0;
+    case kRotate270:
+      RotatePlane270(src, src_stride, dst, dst_stride, width, height);
+      return 0;
+    case kRotate180:
+      RotatePlane180(src, src_stride, dst, dst_stride, width, height);
+      return 0;
+    default:
+      break;
+  }
+  return -1;
+}
+
+LIBYUV_API
+int I420Rotate(const uint8* src_y,
+               int src_stride_y,
+               const uint8* src_u,
+               int src_stride_u,
+               const uint8* src_v,
+               int src_stride_v,
+               uint8* dst_y,
+               int dst_stride_y,
+               uint8* dst_u,
+               int dst_stride_u,
+               uint8* dst_v,
+               int dst_stride_v,
+               int width,
+               int height,
+               enum RotationMode mode) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if (!src_y || !src_u || !src_v || width <= 0 || height == 0 || !dst_y ||
+      !dst_u || !dst_v) {
+    return -1;
+  }
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (halfheight - 1) * src_stride_u;
+    src_v = src_v + (halfheight - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  switch (mode) {
+    case kRotate0:
+      // copy frame
+      return I420Copy(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                      src_stride_v, dst_y, dst_stride_y, dst_u, dst_stride_u,
+                      dst_v, dst_stride_v, width, height);
+    case kRotate90:
+      RotatePlane90(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+      RotatePlane90(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth,
+                    halfheight);
+      RotatePlane90(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth,
+                    halfheight);
+      return 0;
+    case kRotate270:
+      RotatePlane270(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+      RotatePlane270(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth,
+                     halfheight);
+      RotatePlane270(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth,
+                     halfheight);
+      return 0;
+    case kRotate180:
+      RotatePlane180(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+      RotatePlane180(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth,
+                     halfheight);
+      RotatePlane180(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth,
+                     halfheight);
+      return 0;
+    default:
+      break;
+  }
+  return -1;
+}
+
+LIBYUV_API
+int NV12ToI420Rotate(const uint8* src_y,
+                     int src_stride_y,
+                     const uint8* src_uv,
+                     int src_stride_uv,
+                     uint8* dst_y,
+                     int dst_stride_y,
+                     uint8* dst_u,
+                     int dst_stride_u,
+                     uint8* dst_v,
+                     int dst_stride_v,
+                     int width,
+                     int height,
+                     enum RotationMode mode) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if (!src_y || !src_uv || width <= 0 || height == 0 || !dst_y || !dst_u ||
+      !dst_v) {
+    return -1;
+  }
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_uv = src_uv + (halfheight - 1) * src_stride_uv;
+    src_stride_y = -src_stride_y;
+    src_stride_uv = -src_stride_uv;
+  }
+
+  switch (mode) {
+    case kRotate0:
+      // copy frame
+      return NV12ToI420(src_y, src_stride_y, src_uv, src_stride_uv, dst_y,
+                        dst_stride_y, dst_u, dst_stride_u, dst_v, dst_stride_v,
+                        width, height);
+    case kRotate90:
+      RotatePlane90(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+      RotateUV90(src_uv, src_stride_uv, dst_u, dst_stride_u, dst_v,
+                 dst_stride_v, halfwidth, halfheight);
+      return 0;
+    case kRotate270:
+      RotatePlane270(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+      RotateUV270(src_uv, src_stride_uv, dst_u, dst_stride_u, dst_v,
+                  dst_stride_v, halfwidth, halfheight);
+      return 0;
+    case kRotate180:
+      RotatePlane180(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+      RotateUV180(src_uv, src_stride_uv, dst_u, dst_stride_u, dst_v,
+                  dst_stride_v, halfwidth, halfheight);
+      return 0;
+    default:
+      break;
+  }
+  return -1;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif

pjproject-2.11.1/third_party/yuv/source/rotate_any.cc

@@ -0,0 +1,78 @@
+/*
+ *  Copyright 2015 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/rotate.h"
+#include "libyuv/rotate_row.h"
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#define TANY(NAMEANY, TPOS_SIMD, MASK)                                        \
+  void NAMEANY(const uint8* src, int src_stride, uint8* dst, int dst_stride,  \
+               int width) {                                                   \
+    int r = width & MASK;                                                     \
+    int n = width - r;                                                        \
+    if (n > 0) {                                                              \
+      TPOS_SIMD(src, src_stride, dst, dst_stride, n);                         \
+    }                                                                         \
+    TransposeWx8_C(src + n, src_stride, dst + n * dst_stride, dst_stride, r); \
+  }
+
+#ifdef HAS_TRANSPOSEWX8_NEON
+TANY(TransposeWx8_Any_NEON, TransposeWx8_NEON, 7)
+#endif
+#ifdef HAS_TRANSPOSEWX8_SSSE3
+TANY(TransposeWx8_Any_SSSE3, TransposeWx8_SSSE3, 7)
+#endif
+#ifdef HAS_TRANSPOSEWX8_FAST_SSSE3
+TANY(TransposeWx8_Fast_Any_SSSE3, TransposeWx8_Fast_SSSE3, 15)
+#endif
+#ifdef HAS_TRANSPOSEWX8_DSPR2
+TANY(TransposeWx8_Any_DSPR2, TransposeWx8_DSPR2, 7)
+#endif
+#ifdef HAS_TRANSPOSEWX16_MSA
+TANY(TransposeWx16_Any_MSA, TransposeWx16_MSA, 15)
+#endif
+#undef TANY
+
+#define TUVANY(NAMEANY, TPOS_SIMD, MASK)                                       \
+  void NAMEANY(const uint8* src, int src_stride, uint8* dst_a,                 \
+               int dst_stride_a, uint8* dst_b, int dst_stride_b, int width) {  \
+    int r = width & MASK;                                                      \
+    int n = width - r;                                                         \
+    if (n > 0) {                                                               \
+      TPOS_SIMD(src, src_stride, dst_a, dst_stride_a, dst_b, dst_stride_b, n); \
+    }                                                                          \
+    TransposeUVWx8_C(src + n * 2, src_stride, dst_a + n * dst_stride_a,        \
+                     dst_stride_a, dst_b + n * dst_stride_b, dst_stride_b, r); \
+  }
+
+#ifdef HAS_TRANSPOSEUVWX8_NEON
+TUVANY(TransposeUVWx8_Any_NEON, TransposeUVWx8_NEON, 7)
+#endif
+#ifdef HAS_TRANSPOSEUVWX8_SSE2
+TUVANY(TransposeUVWx8_Any_SSE2, TransposeUVWx8_SSE2, 7)
+#endif
+#ifdef HAS_TRANSPOSEUVWX8_DSPR2
+TUVANY(TransposeUVWx8_Any_DSPR2, TransposeUVWx8_DSPR2, 7)
+#endif
+#ifdef HAS_TRANSPOSEUVWX16_MSA
+TUVANY(TransposeUVWx16_Any_MSA, TransposeUVWx16_MSA, 7)
+#endif
+#undef TUVANY
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif

pjproject-2.11.1/third_party/yuv/source/rotate_argb.cc

@@ -0,0 +1,240 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/rotate.h"
+
+#include "libyuv/convert.h"
+#include "libyuv/cpu_id.h"
+#include "libyuv/planar_functions.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// ARGBScale has a function to copy pixels to a row, striding each source
+// pixel by a constant.
+#if !defined(LIBYUV_DISABLE_X86) &&                          \
+    (defined(_M_IX86) ||                                     \
+     (defined(__x86_64__) && !defined(__native_client__)) || \
+     defined(__i386__))
+#define HAS_SCALEARGBROWDOWNEVEN_SSE2
+void ScaleARGBRowDownEven_SSE2(const uint8* src_ptr,
+                               int src_stride,
+                               int src_stepx,
+                               uint8* dst_ptr,
+                               int dst_width);
+#endif
+#if !defined(LIBYUV_DISABLE_NEON) && !defined(__native_client__) && \
+    (defined(__ARM_NEON__) || defined(LIBYUV_NEON) || defined(__aarch64__))
+#define HAS_SCALEARGBROWDOWNEVEN_NEON
+void ScaleARGBRowDownEven_NEON(const uint8* src_ptr,
+                               int src_stride,
+                               int src_stepx,
+                               uint8* dst_ptr,
+                               int dst_width);
+#endif
+
+void ScaleARGBRowDownEven_C(const uint8* src_ptr,
+                            int,
+                            int src_stepx,
+                            uint8* dst_ptr,
+                            int dst_width);
+
+static void ARGBTranspose(const uint8* src,
+                          int src_stride,
+                          uint8* dst,
+                          int dst_stride,
+                          int width,
+                          int height) {
+  int i;
+  int src_pixel_step = src_stride >> 2;
+  void (*ScaleARGBRowDownEven)(const uint8* src_ptr, int src_stride,
+                               int src_step, uint8* dst_ptr, int dst_width) =
+      ScaleARGBRowDownEven_C;
+#if defined(HAS_SCALEARGBROWDOWNEVEN_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(height, 4)) {  // Width of dest.
+    ScaleARGBRowDownEven = ScaleARGBRowDownEven_SSE2;
+  }
+#endif
+#if defined(HAS_SCALEARGBROWDOWNEVEN_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(height, 4)) {  // Width of dest.
+    ScaleARGBRowDownEven = ScaleARGBRowDownEven_NEON;
+  }
+#endif
+
+  for (i = 0; i < width; ++i) {  // column of source to row of dest.
+    ScaleARGBRowDownEven(src, 0, src_pixel_step, dst, height);
+    dst += dst_stride;
+    src += 4;
+  }
+}
+
+void ARGBRotate90(const uint8* src,
+                  int src_stride,
+                  uint8* dst,
+                  int dst_stride,
+                  int width,
+                  int height) {
+  // Rotate by 90 is a ARGBTranspose with the source read
+  // from bottom to top. So set the source pointer to the end
+  // of the buffer and flip the sign of the source stride.
+  src += src_stride * (height - 1);
+  src_stride = -src_stride;
+  ARGBTranspose(src, src_stride, dst, dst_stride, width, height);
+}
+
+void ARGBRotate270(const uint8* src,
+                   int src_stride,
+                   uint8* dst,
+                   int dst_stride,
+                   int width,
+                   int height) {
+  // Rotate by 270 is a ARGBTranspose with the destination written
+  // from bottom to top. So set the destination pointer to the end
+  // of the buffer and flip the sign of the destination stride.
+  dst += dst_stride * (width - 1);
+  dst_stride = -dst_stride;
+  ARGBTranspose(src, src_stride, dst, dst_stride, width, height);
+}
+
+void ARGBRotate180(const uint8* src,
+                   int src_stride,
+                   uint8* dst,
+                   int dst_stride,
+                   int width,
+                   int height) {
+  // Swap first and last row and mirror the content. Uses a temporary row.
+  align_buffer_64(row, width * 4);
+  const uint8* src_bot = src + src_stride * (height - 1);
+  uint8* dst_bot = dst + dst_stride * (height - 1);
+  int half_height = (height + 1) >> 1;
+  int y;
+  void (*ARGBMirrorRow)(const uint8* src, uint8* dst, int width) =
+      ARGBMirrorRow_C;
+  void (*CopyRow)(const uint8* src, uint8* dst, int width) = CopyRow_C;
+#if defined(HAS_ARGBMIRRORROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBMirrorRow = ARGBMirrorRow_Any_NEON;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBMirrorRow = ARGBMirrorRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBMIRRORROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBMirrorRow = ARGBMirrorRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBMirrorRow = ARGBMirrorRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBMIRRORROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBMirrorRow = ARGBMirrorRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBMirrorRow = ARGBMirrorRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBMIRRORROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBMirrorRow = ARGBMirrorRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBMirrorRow = ARGBMirrorRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_COPYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    CopyRow = IS_ALIGNED(width * 4, 32) ? CopyRow_SSE2 : CopyRow_Any_SSE2;
+  }
+#endif
+#if defined(HAS_COPYROW_AVX)
+  if (TestCpuFlag(kCpuHasAVX)) {
+    CopyRow = IS_ALIGNED(width * 4, 64) ? CopyRow_AVX : CopyRow_Any_AVX;
+  }
+#endif
+#if defined(HAS_COPYROW_ERMS)
+  if (TestCpuFlag(kCpuHasERMS)) {
+    CopyRow = CopyRow_ERMS;
+  }
+#endif
+#if defined(HAS_COPYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    CopyRow = IS_ALIGNED(width * 4, 32) ? CopyRow_NEON : CopyRow_Any_NEON;
+  }
+#endif
+#if defined(HAS_COPYROW_MIPS)
+  if (TestCpuFlag(kCpuHasMIPS)) {
+    CopyRow = CopyRow_MIPS;
+  }
+#endif
+
+  // Odd height will harmlessly mirror the middle row twice.
+  for (y = 0; y < half_height; ++y) {
+    ARGBMirrorRow(src, row, width);      // Mirror first row into a buffer
+    ARGBMirrorRow(src_bot, dst, width);  // Mirror last row into first row
+    CopyRow(row, dst_bot, width * 4);    // Copy first mirrored row into last
+    src += src_stride;
+    dst += dst_stride;
+    src_bot -= src_stride;
+    dst_bot -= dst_stride;
+  }
+  free_aligned_buffer_64(row);
+}
+
+LIBYUV_API
+int ARGBRotate(const uint8* src_argb,
+               int src_stride_argb,
+               uint8* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height,
+               enum RotationMode mode) {
+  if (!src_argb || width <= 0 || height == 0 || !dst_argb) {
+    return -1;
+  }
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+
+  switch (mode) {
+    case kRotate0:
+      // copy frame
+      return ARGBCopy(src_argb, src_stride_argb, dst_argb, dst_stride_argb,
+                      width, height);
+    case kRotate90:
+      ARGBRotate90(src_argb, src_stride_argb, dst_argb, dst_stride_argb, width,
+                   height);
+      return 0;
+    case kRotate270:
+      ARGBRotate270(src_argb, src_stride_argb, dst_argb, dst_stride_argb, width,
+                    height);
+      return 0;
+    case kRotate180:
+      ARGBRotate180(src_argb, src_stride_argb, dst_argb, dst_stride_argb, width,
+                    height);
+      return 0;
+    default:
+      break;
+  }
+  return -1;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif

pjproject-2.11.1/third_party/yuv/source/rotate_common.cc

@@ -0,0 +1,106 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/rotate_row.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+void TransposeWx8_C(const uint8* src,
+                    int src_stride,
+                    uint8* dst,
+                    int dst_stride,
+                    int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    dst[0] = src[0 * src_stride];
+    dst[1] = src[1 * src_stride];
+    dst[2] = src[2 * src_stride];
+    dst[3] = src[3 * src_stride];
+    dst[4] = src[4 * src_stride];
+    dst[5] = src[5 * src_stride];
+    dst[6] = src[6 * src_stride];
+    dst[7] = src[7 * src_stride];
+    ++src;
+    dst += dst_stride;
+  }
+}
+
+void TransposeUVWx8_C(const uint8* src,
+                      int src_stride,
+                      uint8* dst_a,
+                      int dst_stride_a,
+                      uint8* dst_b,
+                      int dst_stride_b,
+                      int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    dst_a[0] = src[0 * src_stride + 0];
+    dst_b[0] = src[0 * src_stride + 1];
+    dst_a[1] = src[1 * src_stride + 0];
+    dst_b[1] = src[1 * src_stride + 1];
+    dst_a[2] = src[2 * src_stride + 0];
+    dst_b[2] = src[2 * src_stride + 1];
+    dst_a[3] = src[3 * src_stride + 0];
+    dst_b[3] = src[3 * src_stride + 1];
+    dst_a[4] = src[4 * src_stride + 0];
+    dst_b[4] = src[4 * src_stride + 1];
+    dst_a[5] = src[5 * src_stride + 0];
+    dst_b[5] = src[5 * src_stride + 1];
+    dst_a[6] = src[6 * src_stride + 0];
+    dst_b[6] = src[6 * src_stride + 1];
+    dst_a[7] = src[7 * src_stride + 0];
+    dst_b[7] = src[7 * src_stride + 1];
+    src += 2;
+    dst_a += dst_stride_a;
+    dst_b += dst_stride_b;
+  }
+}
+
+void TransposeWxH_C(const uint8* src,
+                    int src_stride,
+                    uint8* dst,
+                    int dst_stride,
+                    int width,
+                    int height) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    int j;
+    for (j = 0; j < height; ++j) {
+      dst[i * dst_stride + j] = src[j * src_stride + i];
+    }
+  }
+}
+
+void TransposeUVWxH_C(const uint8* src,
+                      int src_stride,
+                      uint8* dst_a,
+                      int dst_stride_a,
+                      uint8* dst_b,
+                      int dst_stride_b,
+                      int width,
+                      int height) {
+  int i;
+  for (i = 0; i < width * 2; i += 2) {
+    int j;
+    for (j = 0; j < height; ++j) {
+      dst_a[j + ((i >> 1) * dst_stride_a)] = src[i + (j * src_stride)];
+      dst_b[j + ((i >> 1) * dst_stride_b)] = src[i + (j * src_stride) + 1];
+    }
+  }
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif

pjproject-2.11.1/third_party/yuv/source/rotate_dspr2.cc

@@ -0,0 +1,475 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/rotate_row.h"
+#include "libyuv/row.h"
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if !defined(LIBYUV_DISABLE_DSPR2) && defined(__mips_dsp) && \
+    (__mips_dsp_rev >= 2) && (_MIPS_SIM == _MIPS_SIM_ABI32)
+
+void TransposeWx8_DSPR2(const uint8* src,
+                        int src_stride,
+                        uint8* dst,
+                        int dst_stride,
+                        int width) {
+  __asm__ __volatile__(
+      ".set push                                         \n"
+      ".set noreorder                                    \n"
+      "sll              $t2, %[src_stride], 0x1          \n"  // src_stride x 2
+      "sll              $t4, %[src_stride], 0x2          \n"  // src_stride x 4
+      "sll              $t9, %[src_stride], 0x3          \n"  // src_stride x 8
+      "addu             $t3, $t2, %[src_stride]          \n"
+      "addu             $t5, $t4, %[src_stride]          \n"
+      "addu             $t6, $t2, $t4                    \n"
+      "andi             $t0, %[dst], 0x3                 \n"
+      "andi             $t1, %[dst_stride], 0x3          \n"
+      "or               $t0, $t0, $t1                    \n"
+      "bnez             $t0, 11f                         \n"
+      " subu            $t7, $t9, %[src_stride]          \n"
+      // dst + dst_stride word aligned
+      "1:                                                \n"
+      "lbu              $t0, 0(%[src])                   \n"
+      "lbux             $t1, %[src_stride](%[src])       \n"
+      "lbux             $t8, $t2(%[src])                 \n"
+      "lbux             $t9, $t3(%[src])                 \n"
+      "sll              $t1, $t1, 16                     \n"
+      "sll              $t9, $t9, 16                     \n"
+      "or               $t0, $t0, $t1                    \n"
+      "or               $t8, $t8, $t9                    \n"
+      "precr.qb.ph      $s0, $t8, $t0                    \n"
+      "lbux             $t0, $t4(%[src])                 \n"
+      "lbux             $t1, $t5(%[src])                 \n"
+      "lbux             $t8, $t6(%[src])                 \n"
+      "lbux             $t9, $t7(%[src])                 \n"
+      "sll              $t1, $t1, 16                     \n"
+      "sll              $t9, $t9, 16                     \n"
+      "or               $t0, $t0, $t1                    \n"
+      "or               $t8, $t8, $t9                    \n"
+      "precr.qb.ph      $s1, $t8, $t0                    \n"
+      "sw               $s0, 0(%[dst])                   \n"
+      "addiu            %[width], -1                     \n"
+      "addiu            %[src], 1                        \n"
+      "sw               $s1, 4(%[dst])                   \n"
+      "bnez             %[width], 1b                     \n"
+      " addu            %[dst], %[dst], %[dst_stride]    \n"
+      "b                2f                               \n"
+      // dst + dst_stride unaligned
+      "11:                                               \n"
+      "lbu              $t0, 0(%[src])                   \n"
+      "lbux             $t1, %[src_stride](%[src])       \n"
+      "lbux             $t8, $t2(%[src])                 \n"
+      "lbux             $t9, $t3(%[src])                 \n"
+      "sll              $t1, $t1, 16                     \n"
+      "sll              $t9, $t9, 16                     \n"
+      "or               $t0, $t0, $t1                    \n"
+      "or               $t8, $t8, $t9                    \n"
+      "precr.qb.ph      $s0, $t8, $t0                    \n"
+      "lbux             $t0, $t4(%[src])                 \n"
+      "lbux             $t1, $t5(%[src])                 \n"
+      "lbux             $t8, $t6(%[src])                 \n"
+      "lbux             $t9, $t7(%[src])                 \n"
+      "sll              $t1, $t1, 16                     \n"
+      "sll              $t9, $t9, 16                     \n"
+      "or               $t0, $t0, $t1                    \n"
+      "or               $t8, $t8, $t9                    \n"
+      "precr.qb.ph      $s1, $t8, $t0                    \n"
+      "swr              $s0, 0(%[dst])                   \n"
+      "swl              $s0, 3(%[dst])                   \n"
+      "addiu            %[width], -1                     \n"
+      "addiu            %[src], 1                        \n"
+      "swr              $s1, 4(%[dst])                   \n"
+      "swl              $s1, 7(%[dst])                   \n"
+      "bnez             %[width], 11b                    \n"
+      "addu             %[dst], %[dst], %[dst_stride]    \n"
+      "2:                                                \n"
+      ".set pop                                          \n"
+      : [src] "+r"(src), [dst] "+r"(dst), [width] "+r"(width)
+      : [src_stride] "r"(src_stride), [dst_stride] "r"(dst_stride)
+      : "t0", "t1", "t2", "t3", "t4", "t5", "t6", "t7", "t8", "t9", "s0", "s1");
+}
+
+void TransposeWx8_Fast_DSPR2(const uint8* src,
+                             int src_stride,
+                             uint8* dst,
+                             int dst_stride,
+                             int width) {
+  __asm__ __volatile__(
+      ".set noat                                         \n"
+      ".set push                                         \n"
+      ".set noreorder                                    \n"
+      "beqz             %[width], 2f                     \n"
+      " sll             $t2, %[src_stride], 0x1          \n"  // src_stride x 2
+      "sll              $t4, %[src_stride], 0x2          \n"  // src_stride x 4
+      "sll              $t9, %[src_stride], 0x3          \n"  // src_stride x 8
+      "addu             $t3, $t2, %[src_stride]          \n"
+      "addu             $t5, $t4, %[src_stride]          \n"
+      "addu             $t6, $t2, $t4                    \n"
+
+      "srl              $AT, %[width], 0x2               \n"
+      "andi             $t0, %[dst], 0x3                 \n"
+      "andi             $t1, %[dst_stride], 0x3          \n"
+      "or               $t0, $t0, $t1                    \n"
+      "bnez             $t0, 11f                         \n"
+      " subu            $t7, $t9, %[src_stride]          \n"
+      // dst + dst_stride word aligned
+      "1:                                                \n"
+      "lw               $t0, 0(%[src])                   \n"
+      "lwx              $t1, %[src_stride](%[src])       \n"
+      "lwx              $t8, $t2(%[src])                 \n"
+      "lwx              $t9, $t3(%[src])                 \n"
+
+      // t0 = | 30 | 20 | 10 | 00 |
+      // t1 = | 31 | 21 | 11 | 01 |
+      // t8 = | 32 | 22 | 12 | 02 |
+      // t9 = | 33 | 23 | 13 | 03 |
+
+      "precr.qb.ph     $s0, $t1, $t0                     \n"
+      "precr.qb.ph     $s1, $t9, $t8                     \n"
+      "precrq.qb.ph    $s2, $t1, $t0                     \n"
+      "precrq.qb.ph    $s3, $t9, $t8                     \n"
+
+      // s0 = | 21 | 01 | 20 | 00 |
+      // s1 = | 23 | 03 | 22 | 02 |
+      // s2 = | 31 | 11 | 30 | 10 |
+      // s3 = | 33 | 13 | 32 | 12 |
+
+      "precr.qb.ph     $s4, $s1, $s0                     \n"
+      "precrq.qb.ph    $s5, $s1, $s0                     \n"
+      "precr.qb.ph     $s6, $s3, $s2                     \n"
+      "precrq.qb.ph    $s7, $s3, $s2                     \n"
+
+      // s4 = | 03 | 02 | 01 | 00 |
+      // s5 = | 23 | 22 | 21 | 20 |
+      // s6 = | 13 | 12 | 11 | 10 |
+      // s7 = | 33 | 32 | 31 | 30 |
+
+      "lwx              $t0, $t4(%[src])                 \n"
+      "lwx              $t1, $t5(%[src])                 \n"
+      "lwx              $t8, $t6(%[src])                 \n"
+      "lwx              $t9, $t7(%[src])                 \n"
+
+      // t0 = | 34 | 24 | 14 | 04 |
+      // t1 = | 35 | 25 | 15 | 05 |
+      // t8 = | 36 | 26 | 16 | 06 |
+      // t9 = | 37 | 27 | 17 | 07 |
+
+      "precr.qb.ph     $s0, $t1, $t0                     \n"
+      "precr.qb.ph     $s1, $t9, $t8                     \n"
+      "precrq.qb.ph    $s2, $t1, $t0                     \n"
+      "precrq.qb.ph    $s3, $t9, $t8                     \n"
+
+      // s0 = | 25 | 05 | 24 | 04 |
+      // s1 = | 27 | 07 | 26 | 06 |
+      // s2 = | 35 | 15 | 34 | 14 |
+      // s3 = | 37 | 17 | 36 | 16 |
+
+      "precr.qb.ph     $t0, $s1, $s0                     \n"
+      "precrq.qb.ph    $t1, $s1, $s0                     \n"
+      "precr.qb.ph     $t8, $s3, $s2                     \n"
+      "precrq.qb.ph    $t9, $s3, $s2                     \n"
+
+      // t0 = | 07 | 06 | 05 | 04 |
+      // t1 = | 27 | 26 | 25 | 24 |
+      // t8 = | 17 | 16 | 15 | 14 |
+      // t9 = | 37 | 36 | 35 | 34 |
+
+      "addu            $s0, %[dst], %[dst_stride]        \n"
+      "addu            $s1, $s0, %[dst_stride]           \n"
+      "addu            $s2, $s1, %[dst_stride]           \n"
+
+      "sw              $s4, 0(%[dst])                    \n"
+      "sw              $t0, 4(%[dst])                    \n"
+      "sw              $s6, 0($s0)                       \n"
+      "sw              $t8, 4($s0)                       \n"
+      "sw              $s5, 0($s1)                       \n"
+      "sw              $t1, 4($s1)                       \n"
+      "sw              $s7, 0($s2)                       \n"
+      "sw              $t9, 4($s2)                       \n"
+
+      "addiu            $AT, -1                          \n"
+      "addiu            %[src], 4                        \n"
+
+      "bnez             $AT, 1b                          \n"
+      " addu            %[dst], $s2, %[dst_stride]       \n"
+      "b                2f                               \n"
+      // dst + dst_stride unaligned
+      "11:                                               \n"
+      "lw               $t0, 0(%[src])                   \n"
+      "lwx              $t1, %[src_stride](%[src])       \n"
+      "lwx              $t8, $t2(%[src])                 \n"
+      "lwx              $t9, $t3(%[src])                 \n"
+
+      // t0 = | 30 | 20 | 10 | 00 |
+      // t1 = | 31 | 21 | 11 | 01 |
+      // t8 = | 32 | 22 | 12 | 02 |
+      // t9 = | 33 | 23 | 13 | 03 |
+
+      "precr.qb.ph     $s0, $t1, $t0                     \n"
+      "precr.qb.ph     $s1, $t9, $t8                     \n"
+      "precrq.qb.ph    $s2, $t1, $t0                     \n"
+      "precrq.qb.ph    $s3, $t9, $t8                     \n"
+
+      // s0 = | 21 | 01 | 20 | 00 |
+      // s1 = | 23 | 03 | 22 | 02 |
+      // s2 = | 31 | 11 | 30 | 10 |
+      // s3 = | 33 | 13 | 32 | 12 |
+
+      "precr.qb.ph     $s4, $s1, $s0                     \n"
+      "precrq.qb.ph    $s5, $s1, $s0                     \n"
+      "precr.qb.ph     $s6, $s3, $s2                     \n"
+      "precrq.qb.ph    $s7, $s3, $s2                     \n"
+
+      // s4 = | 03 | 02 | 01 | 00 |
+      // s5 = | 23 | 22 | 21 | 20 |
+      // s6 = | 13 | 12 | 11 | 10 |
+      // s7 = | 33 | 32 | 31 | 30 |
+
+      "lwx              $t0, $t4(%[src])                 \n"
+      "lwx              $t1, $t5(%[src])                 \n"
+      "lwx              $t8, $t6(%[src])                 \n"
+      "lwx              $t9, $t7(%[src])                 \n"
+
+      // t0 = | 34 | 24 | 14 | 04 |
+      // t1 = | 35 | 25 | 15 | 05 |
+      // t8 = | 36 | 26 | 16 | 06 |
+      // t9 = | 37 | 27 | 17 | 07 |
+
+      "precr.qb.ph     $s0, $t1, $t0                     \n"
+      "precr.qb.ph     $s1, $t9, $t8                     \n"
+      "precrq.qb.ph    $s2, $t1, $t0                     \n"
+      "precrq.qb.ph    $s3, $t9, $t8                     \n"
+
+      // s0 = | 25 | 05 | 24 | 04 |
+      // s1 = | 27 | 07 | 26 | 06 |
+      // s2 = | 35 | 15 | 34 | 14 |
+      // s3 = | 37 | 17 | 36 | 16 |
+
+      "precr.qb.ph     $t0, $s1, $s0                     \n"
+      "precrq.qb.ph    $t1, $s1, $s0                     \n"
+      "precr.qb.ph     $t8, $s3, $s2                     \n"
+      "precrq.qb.ph    $t9, $s3, $s2                     \n"
+
+      // t0 = | 07 | 06 | 05 | 04 |
+      // t1 = | 27 | 26 | 25 | 24 |
+      // t8 = | 17 | 16 | 15 | 14 |
+      // t9 = | 37 | 36 | 35 | 34 |
+
+      "addu            $s0, %[dst], %[dst_stride]        \n"
+      "addu            $s1, $s0, %[dst_stride]           \n"
+      "addu            $s2, $s1, %[dst_stride]           \n"
+
+      "swr              $s4, 0(%[dst])                   \n"
+      "swl              $s4, 3(%[dst])                   \n"
+      "swr              $t0, 4(%[dst])                   \n"
+      "swl              $t0, 7(%[dst])                   \n"
+      "swr              $s6, 0($s0)                      \n"
+      "swl              $s6, 3($s0)                      \n"
+      "swr              $t8, 4($s0)                      \n"
+      "swl              $t8, 7($s0)                      \n"
+      "swr              $s5, 0($s1)                      \n"
+      "swl              $s5, 3($s1)                      \n"
+      "swr              $t1, 4($s1)                      \n"
+      "swl              $t1, 7($s1)                      \n"
+      "swr              $s7, 0($s2)                      \n"
+      "swl              $s7, 3($s2)                      \n"
+      "swr              $t9, 4($s2)                      \n"
+      "swl              $t9, 7($s2)                      \n"
+
+      "addiu            $AT, -1                          \n"
+      "addiu            %[src], 4                        \n"
+
+      "bnez             $AT, 11b                         \n"
+      " addu            %[dst], $s2, %[dst_stride]       \n"
+      "2:                                                \n"
+      ".set pop                                          \n"
+      ".set at                                           \n"
+      : [src] "+r"(src), [dst] "+r"(dst), [width] "+r"(width)
+      : [src_stride] "r"(src_stride), [dst_stride] "r"(dst_stride)
+      : "t0", "t1", "t2", "t3", "t4", "t5", "t6", "t7", "t8", "t9", "s0", "s1",
+        "s2", "s3", "s4", "s5", "s6", "s7");
+}
+
+void TransposeUVWx8_DSPR2(const uint8* src,
+                          int src_stride,
+                          uint8* dst_a,
+                          int dst_stride_a,
+                          uint8* dst_b,
+                          int dst_stride_b,
+                          int width) {
+  __asm__ __volatile__(
+      ".set push                                         \n"
+      ".set noreorder                                    \n"
+      "beqz            %[width], 2f                      \n"
+      " sll            $t2, %[src_stride], 0x1           \n"  // src_stride x 2
+      "sll             $t4, %[src_stride], 0x2           \n"  // src_stride x 4
+      "sll             $t9, %[src_stride], 0x3           \n"  // src_stride x 8
+      "addu            $t3, $t2, %[src_stride]           \n"
+      "addu            $t5, $t4, %[src_stride]           \n"
+      "addu            $t6, $t2, $t4                     \n"
+      "subu            $t7, $t9, %[src_stride]           \n"
+      "srl             $t1, %[width], 1                  \n"
+
+      // check word aligment for dst_a, dst_b, dst_stride_a and dst_stride_b
+      "andi            $t0, %[dst_a], 0x3                \n"
+      "andi            $t8, %[dst_b], 0x3                \n"
+      "or              $t0, $t0, $t8                     \n"
+      "andi            $t8, %[dst_stride_a], 0x3         \n"
+      "andi            $s5, %[dst_stride_b], 0x3         \n"
+      "or              $t8, $t8, $s5                     \n"
+      "or              $t0, $t0, $t8                     \n"
+      "bnez            $t0, 11f                          \n"
+      " nop                                              \n"
+      // dst + dst_stride word aligned (both, a & b dst addresses)
+      "1:                                                \n"
+      "lw              $t0, 0(%[src])                    \n"  // |B0|A0|b0|a0|
+      "lwx             $t8, %[src_stride](%[src])        \n"  // |B1|A1|b1|a1|
+      "addu            $s5, %[dst_a], %[dst_stride_a]    \n"
+      "lwx             $t9, $t2(%[src])                  \n"  // |B2|A2|b2|a2|
+      "lwx             $s0, $t3(%[src])                  \n"  // |B3|A3|b3|a3|
+      "addu            $s6, %[dst_b], %[dst_stride_b]    \n"
+
+      "precrq.ph.w     $s1, $t8, $t0                     \n"  // |B1|A1|B0|A0|
+      "precrq.ph.w     $s2, $s0, $t9                     \n"  // |B3|A3|B2|A2|
+      "precr.qb.ph     $s3, $s2, $s1                     \n"  // |A3|A2|A1|A0|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n"  // |B3|B2|B1|B0|
+
+      "sll             $t0, $t0, 16                      \n"
+      "packrl.ph       $s1, $t8, $t0                     \n"  // |b1|a1|b0|a0|
+      "sll             $t9, $t9, 16                      \n"
+      "packrl.ph       $s2, $s0, $t9                     \n"  // |b3|a3|b2|a2|
+
+      "sw              $s3, 0($s5)                       \n"
+      "sw              $s4, 0($s6)                       \n"
+
+      "precr.qb.ph     $s3, $s2, $s1                     \n"  // |a3|a2|a1|a0|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n"  // |b3|b2|b1|b0|
+
+      "lwx             $t0, $t4(%[src])                  \n"  // |B4|A4|b4|a4|
+      "lwx             $t8, $t5(%[src])                  \n"  // |B5|A5|b5|a5|
+      "lwx             $t9, $t6(%[src])                  \n"  // |B6|A6|b6|a6|
+      "lwx             $s0, $t7(%[src])                  \n"  // |B7|A7|b7|a7|
+      "sw              $s3, 0(%[dst_a])                  \n"
+      "sw              $s4, 0(%[dst_b])                  \n"
+
+      "precrq.ph.w     $s1, $t8, $t0                     \n"  // |B5|A5|B4|A4|
+      "precrq.ph.w     $s2, $s0, $t9                     \n"  // |B6|A6|B7|A7|
+      "precr.qb.ph     $s3, $s2, $s1                     \n"  // |A7|A6|A5|A4|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n"  // |B7|B6|B5|B4|
+
+      "sll             $t0, $t0, 16                      \n"
+      "packrl.ph       $s1, $t8, $t0                     \n"  // |b5|a5|b4|a4|
+      "sll             $t9, $t9, 16                      \n"
+      "packrl.ph       $s2, $s0, $t9                     \n"  // |b7|a7|b6|a6|
+      "sw              $s3, 4($s5)                       \n"
+      "sw              $s4, 4($s6)                       \n"
+
+      "precr.qb.ph     $s3, $s2, $s1                     \n"  // |a7|a6|a5|a4|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n"  // |b7|b6|b5|b4|
+
+      "addiu           %[src], 4                         \n"
+      "addiu           $t1, -1                           \n"
+      "sll             $t0, %[dst_stride_a], 1           \n"
+      "sll             $t8, %[dst_stride_b], 1           \n"
+      "sw              $s3, 4(%[dst_a])                  \n"
+      "sw              $s4, 4(%[dst_b])                  \n"
+      "addu            %[dst_a], %[dst_a], $t0           \n"
+      "bnez            $t1, 1b                           \n"
+      " addu           %[dst_b], %[dst_b], $t8           \n"
+      "b               2f                                \n"
+      " nop                                              \n"
+
+      // dst_a or dst_b or dst_stride_a or dst_stride_b not word aligned
+      "11:                                               \n"
+      "lw              $t0, 0(%[src])                    \n"  // |B0|A0|b0|a0|
+      "lwx             $t8, %[src_stride](%[src])        \n"  // |B1|A1|b1|a1|
+      "addu            $s5, %[dst_a], %[dst_stride_a]    \n"
+      "lwx             $t9, $t2(%[src])                  \n"  // |B2|A2|b2|a2|
+      "lwx             $s0, $t3(%[src])                  \n"  // |B3|A3|b3|a3|
+      "addu            $s6, %[dst_b], %[dst_stride_b]    \n"
+
+      "precrq.ph.w     $s1, $t8, $t0                     \n"  // |B1|A1|B0|A0|
+      "precrq.ph.w     $s2, $s0, $t9                     \n"  // |B3|A3|B2|A2|
+      "precr.qb.ph     $s3, $s2, $s1                     \n"  // |A3|A2|A1|A0|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n"  // |B3|B2|B1|B0|
+
+      "sll             $t0, $t0, 16                      \n"
+      "packrl.ph       $s1, $t8, $t0                     \n"  // |b1|a1|b0|a0|
+      "sll             $t9, $t9, 16                      \n"
+      "packrl.ph       $s2, $s0, $t9                     \n"  // |b3|a3|b2|a2|
+
+      "swr             $s3, 0($s5)                       \n"
+      "swl             $s3, 3($s5)                       \n"
+      "swr             $s4, 0($s6)                       \n"
+      "swl             $s4, 3($s6)                       \n"
+
+      "precr.qb.ph     $s3, $s2, $s1                     \n"  // |a3|a2|a1|a0|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n"  // |b3|b2|b1|b0|
+
+      "lwx             $t0, $t4(%[src])                  \n"  // |B4|A4|b4|a4|
+      "lwx             $t8, $t5(%[src])                  \n"  // |B5|A5|b5|a5|
+      "lwx             $t9, $t6(%[src])                  \n"  // |B6|A6|b6|a6|
+      "lwx             $s0, $t7(%[src])                  \n"  // |B7|A7|b7|a7|
+      "swr             $s3, 0(%[dst_a])                  \n"
+      "swl             $s3, 3(%[dst_a])                  \n"
+      "swr             $s4, 0(%[dst_b])                  \n"
+      "swl             $s4, 3(%[dst_b])                  \n"
+
+      "precrq.ph.w     $s1, $t8, $t0                     \n"  // |B5|A5|B4|A4|
+      "precrq.ph.w     $s2, $s0, $t9                     \n"  // |B6|A6|B7|A7|
+      "precr.qb.ph     $s3, $s2, $s1                     \n"  // |A7|A6|A5|A4|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n"  // |B7|B6|B5|B4|
+
+      "sll             $t0, $t0, 16                      \n"
+      "packrl.ph       $s1, $t8, $t0                     \n"  // |b5|a5|b4|a4|
+      "sll             $t9, $t9, 16                      \n"
+      "packrl.ph       $s2, $s0, $t9                     \n"  // |b7|a7|b6|a6|
+
+      "swr             $s3, 4($s5)                       \n"
+      "swl             $s3, 7($s5)                       \n"
+      "swr             $s4, 4($s6)                       \n"
+      "swl             $s4, 7($s6)                       \n"
+
+      "precr.qb.ph     $s3, $s2, $s1                     \n"  // |a7|a6|a5|a4|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n"  // |b7|b6|b5|b4|
+
+      "addiu           %[src], 4                         \n"
+      "addiu           $t1, -1                           \n"
+      "sll             $t0, %[dst_stride_a], 1           \n"
+      "sll             $t8, %[dst_stride_b], 1           \n"
+      "swr             $s3, 4(%[dst_a])                  \n"
+      "swl             $s3, 7(%[dst_a])                  \n"
+      "swr             $s4, 4(%[dst_b])                  \n"
+      "swl             $s4, 7(%[dst_b])                  \n"
+      "addu            %[dst_a], %[dst_a], $t0           \n"
+      "bnez            $t1, 11b                          \n"
+      " addu           %[dst_b], %[dst_b], $t8           \n"
+
+      "2:                                                \n"
+      ".set pop                                          \n"
+      : [src] "+r"(src), [dst_a] "+r"(dst_a), [dst_b] "+r"(dst_b),
+        [width] "+r"(width), [src_stride] "+r"(src_stride)
+      : [dst_stride_a] "r"(dst_stride_a), [dst_stride_b] "r"(dst_stride_b)
+      : "t0", "t1", "t2", "t3", "t4", "t5", "t6", "t7", "t8", "t9", "s0", "s1",
+        "s2", "s3", "s4", "s5", "s6");
+}
+
+#endif  // defined(__mips_dsp) && (__mips_dsp_rev >= 2)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif

pjproject-2.11.1/third_party/yuv/source/rotate_gcc.cc

@@ -0,0 +1,374 @@
+/*
+ *  Copyright 2015 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/rotate_row.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC x86 and x64.
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(__x86_64__) || (defined(__i386__) && !defined(_MSC_VER)))
+
+// Transpose 8x8. 32 or 64 bit, but not NaCL for 64 bit.
+#if defined(HAS_TRANSPOSEWX8_SSSE3)
+void TransposeWx8_SSSE3(const uint8* src,
+                        int src_stride,
+                        uint8* dst,
+                        int dst_stride,
+                        int width) {
+  asm volatile(
+      // Read in the data from the source pointer.
+      // First round of bit swap.
+      LABELALIGN
+      "1:                                          \n"
+      "movq       (%0),%%xmm0                      \n"
+      "movq       (%0,%3),%%xmm1                   \n"
+      "lea        (%0,%3,2),%0                     \n"
+      "punpcklbw  %%xmm1,%%xmm0                    \n"
+      "movq       (%0),%%xmm2                      \n"
+      "movdqa     %%xmm0,%%xmm1                    \n"
+      "palignr    $0x8,%%xmm1,%%xmm1               \n"
+      "movq       (%0,%3),%%xmm3                   \n"
+      "lea        (%0,%3,2),%0                     \n"
+      "punpcklbw  %%xmm3,%%xmm2                    \n"
+      "movdqa     %%xmm2,%%xmm3                    \n"
+      "movq       (%0),%%xmm4                      \n"
+      "palignr    $0x8,%%xmm3,%%xmm3               \n"
+      "movq       (%0,%3),%%xmm5                   \n"
+      "lea        (%0,%3,2),%0                     \n"
+      "punpcklbw  %%xmm5,%%xmm4                    \n"
+      "movdqa     %%xmm4,%%xmm5                    \n"
+      "movq       (%0),%%xmm6                      \n"
+      "palignr    $0x8,%%xmm5,%%xmm5               \n"
+      "movq       (%0,%3),%%xmm7                   \n"
+      "lea        (%0,%3,2),%0                     \n"
+      "punpcklbw  %%xmm7,%%xmm6                    \n"
+      "neg        %3                               \n"
+      "movdqa     %%xmm6,%%xmm7                    \n"
+      "lea        0x8(%0,%3,8),%0                  \n"
+      "palignr    $0x8,%%xmm7,%%xmm7               \n"
+      "neg        %3                               \n"
+      // Second round of bit swap.
+      "punpcklwd  %%xmm2,%%xmm0                    \n"
+      "punpcklwd  %%xmm3,%%xmm1                    \n"
+      "movdqa     %%xmm0,%%xmm2                    \n"
+      "movdqa     %%xmm1,%%xmm3                    \n"
+      "palignr    $0x8,%%xmm2,%%xmm2               \n"
+      "palignr    $0x8,%%xmm3,%%xmm3               \n"
+      "punpcklwd  %%xmm6,%%xmm4                    \n"
+      "punpcklwd  %%xmm7,%%xmm5                    \n"
+      "movdqa     %%xmm4,%%xmm6                    \n"
+      "movdqa     %%xmm5,%%xmm7                    \n"
+      "palignr    $0x8,%%xmm6,%%xmm6               \n"
+      "palignr    $0x8,%%xmm7,%%xmm7               \n"
+      // Third round of bit swap.
+      // Write to the destination pointer.
+      "punpckldq  %%xmm4,%%xmm0                    \n"
+      "movq       %%xmm0,(%1)                      \n"
+      "movdqa     %%xmm0,%%xmm4                    \n"
+      "palignr    $0x8,%%xmm4,%%xmm4               \n"
+      "movq       %%xmm4,(%1,%4)                   \n"
+      "lea        (%1,%4,2),%1                     \n"
+      "punpckldq  %%xmm6,%%xmm2                    \n"
+      "movdqa     %%xmm2,%%xmm6                    \n"
+      "movq       %%xmm2,(%1)                      \n"
+      "palignr    $0x8,%%xmm6,%%xmm6               \n"
+      "punpckldq  %%xmm5,%%xmm1                    \n"
+      "movq       %%xmm6,(%1,%4)                   \n"
+      "lea        (%1,%4,2),%1                     \n"
+      "movdqa     %%xmm1,%%xmm5                    \n"
+      "movq       %%xmm1,(%1)                      \n"
+      "palignr    $0x8,%%xmm5,%%xmm5               \n"
+      "movq       %%xmm5,(%1,%4)                   \n"
+      "lea        (%1,%4,2),%1                     \n"
+      "punpckldq  %%xmm7,%%xmm3                    \n"
+      "movq       %%xmm3,(%1)                      \n"
+      "movdqa     %%xmm3,%%xmm7                    \n"
+      "palignr    $0x8,%%xmm7,%%xmm7               \n"
+      "sub        $0x8,%2                          \n"
+      "movq       %%xmm7,(%1,%4)                   \n"
+      "lea        (%1,%4,2),%1                     \n"
+      "jg         1b                               \n"
+      : "+r"(src),                    // %0
+        "+r"(dst),                    // %1
+        "+r"(width)                   // %2
+      : "r"((intptr_t)(src_stride)),  // %3
+        "r"((intptr_t)(dst_stride))   // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
+}
+#endif  // defined(HAS_TRANSPOSEWX8_SSSE3)
+
+// Transpose 16x8. 64 bit
+#if defined(HAS_TRANSPOSEWX8_FAST_SSSE3)
+void TransposeWx8_Fast_SSSE3(const uint8* src,
+                             int src_stride,
+                             uint8* dst,
+                             int dst_stride,
+                             int width) {
+  asm volatile(
+      // Read in the data from the source pointer.
+      // First round of bit swap.
+      LABELALIGN
+      "1:                                          \n"
+      "movdqu     (%0),%%xmm0                      \n"
+      "movdqu     (%0,%3),%%xmm1                   \n"
+      "lea        (%0,%3,2),%0                     \n"
+      "movdqa     %%xmm0,%%xmm8                    \n"
+      "punpcklbw  %%xmm1,%%xmm0                    \n"
+      "punpckhbw  %%xmm1,%%xmm8                    \n"
+      "movdqu     (%0),%%xmm2                      \n"
+      "movdqa     %%xmm0,%%xmm1                    \n"
+      "movdqa     %%xmm8,%%xmm9                    \n"
+      "palignr    $0x8,%%xmm1,%%xmm1               \n"
+      "palignr    $0x8,%%xmm9,%%xmm9               \n"
+      "movdqu     (%0,%3),%%xmm3                   \n"
+      "lea        (%0,%3,2),%0                     \n"
+      "movdqa     %%xmm2,%%xmm10                   \n"
+      "punpcklbw  %%xmm3,%%xmm2                    \n"
+      "punpckhbw  %%xmm3,%%xmm10                   \n"
+      "movdqa     %%xmm2,%%xmm3                    \n"
+      "movdqa     %%xmm10,%%xmm11                  \n"
+      "movdqu     (%0),%%xmm4                      \n"
+      "palignr    $0x8,%%xmm3,%%xmm3               \n"
+      "palignr    $0x8,%%xmm11,%%xmm11             \n"
+      "movdqu     (%0,%3),%%xmm5                   \n"
+      "lea        (%0,%3,2),%0                     \n"
+      "movdqa     %%xmm4,%%xmm12                   \n"
+      "punpcklbw  %%xmm5,%%xmm4                    \n"
+      "punpckhbw  %%xmm5,%%xmm12                   \n"
+      "movdqa     %%xmm4,%%xmm5                    \n"
+      "movdqa     %%xmm12,%%xmm13                  \n"
+      "movdqu     (%0),%%xmm6                      \n"
+      "palignr    $0x8,%%xmm5,%%xmm5               \n"
+      "palignr    $0x8,%%xmm13,%%xmm13             \n"
+      "movdqu     (%0,%3),%%xmm7                   \n"
+      "lea        (%0,%3,2),%0                     \n"
+      "movdqa     %%xmm6,%%xmm14                   \n"
+      "punpcklbw  %%xmm7,%%xmm6                    \n"
+      "punpckhbw  %%xmm7,%%xmm14                   \n"
+      "neg        %3                               \n"
+      "movdqa     %%xmm6,%%xmm7                    \n"
+      "movdqa     %%xmm14,%%xmm15                  \n"
+      "lea        0x10(%0,%3,8),%0                 \n"
+      "palignr    $0x8,%%xmm7,%%xmm7               \n"
+      "palignr    $0x8,%%xmm15,%%xmm15             \n"
+      "neg        %3                               \n"
+      // Second round of bit swap.
+      "punpcklwd  %%xmm2,%%xmm0                    \n"
+      "punpcklwd  %%xmm3,%%xmm1                    \n"
+      "movdqa     %%xmm0,%%xmm2                    \n"
+      "movdqa     %%xmm1,%%xmm3                    \n"
+      "palignr    $0x8,%%xmm2,%%xmm2               \n"
+      "palignr    $0x8,%%xmm3,%%xmm3               \n"
+      "punpcklwd  %%xmm6,%%xmm4                    \n"
+      "punpcklwd  %%xmm7,%%xmm5                    \n"
+      "movdqa     %%xmm4,%%xmm6                    \n"
+      "movdqa     %%xmm5,%%xmm7                    \n"
+      "palignr    $0x8,%%xmm6,%%xmm6               \n"
+      "palignr    $0x8,%%xmm7,%%xmm7               \n"
+      "punpcklwd  %%xmm10,%%xmm8                   \n"
+      "punpcklwd  %%xmm11,%%xmm9                   \n"
+      "movdqa     %%xmm8,%%xmm10                   \n"
+      "movdqa     %%xmm9,%%xmm11                   \n"
+      "palignr    $0x8,%%xmm10,%%xmm10             \n"
+      "palignr    $0x8,%%xmm11,%%xmm11             \n"
+      "punpcklwd  %%xmm14,%%xmm12                  \n"
+      "punpcklwd  %%xmm15,%%xmm13                  \n"
+      "movdqa     %%xmm12,%%xmm14                  \n"
+      "movdqa     %%xmm13,%%xmm15                  \n"
+      "palignr    $0x8,%%xmm14,%%xmm14             \n"
+      "palignr    $0x8,%%xmm15,%%xmm15             \n"
+      // Third round of bit swap.
+      // Write to the destination pointer.
+      "punpckldq  %%xmm4,%%xmm0                    \n"
+      "movq       %%xmm0,(%1)                      \n"
+      "movdqa     %%xmm0,%%xmm4                    \n"
+      "palignr    $0x8,%%xmm4,%%xmm4               \n"
+      "movq       %%xmm4,(%1,%4)                   \n"
+      "lea        (%1,%4,2),%1                     \n"
+      "punpckldq  %%xmm6,%%xmm2                    \n"
+      "movdqa     %%xmm2,%%xmm6                    \n"
+      "movq       %%xmm2,(%1)                      \n"
+      "palignr    $0x8,%%xmm6,%%xmm6               \n"
+      "punpckldq  %%xmm5,%%xmm1                    \n"
+      "movq       %%xmm6,(%1,%4)                   \n"
+      "lea        (%1,%4,2),%1                     \n"
+      "movdqa     %%xmm1,%%xmm5                    \n"
+      "movq       %%xmm1,(%1)                      \n"
+      "palignr    $0x8,%%xmm5,%%xmm5               \n"
+      "movq       %%xmm5,(%1,%4)                   \n"
+      "lea        (%1,%4,2),%1                     \n"
+      "punpckldq  %%xmm7,%%xmm3                    \n"
+      "movq       %%xmm3,(%1)                      \n"
+      "movdqa     %%xmm3,%%xmm7                    \n"
+      "palignr    $0x8,%%xmm7,%%xmm7               \n"
+      "movq       %%xmm7,(%1,%4)                   \n"
+      "lea        (%1,%4,2),%1                     \n"
+      "punpckldq  %%xmm12,%%xmm8                   \n"
+      "movq       %%xmm8,(%1)                      \n"
+      "movdqa     %%xmm8,%%xmm12                   \n"
+      "palignr    $0x8,%%xmm12,%%xmm12             \n"
+      "movq       %%xmm12,(%1,%4)                  \n"
+      "lea        (%1,%4,2),%1                     \n"
+      "punpckldq  %%xmm14,%%xmm10                  \n"
+      "movdqa     %%xmm10,%%xmm14                  \n"
+      "movq       %%xmm10,(%1)                     \n"
+      "palignr    $0x8,%%xmm14,%%xmm14             \n"
+      "punpckldq  %%xmm13,%%xmm9                   \n"
+      "movq       %%xmm14,(%1,%4)                  \n"
+      "lea        (%1,%4,2),%1                     \n"
+      "movdqa     %%xmm9,%%xmm13                   \n"
+      "movq       %%xmm9,(%1)                      \n"
+      "palignr    $0x8,%%xmm13,%%xmm13             \n"
+      "movq       %%xmm13,(%1,%4)                  \n"
+      "lea        (%1,%4,2),%1                     \n"
+      "punpckldq  %%xmm15,%%xmm11                  \n"
+      "movq       %%xmm11,(%1)                     \n"
+      "movdqa     %%xmm11,%%xmm15                  \n"
+      "palignr    $0x8,%%xmm15,%%xmm15             \n"
+      "sub        $0x10,%2                         \n"
+      "movq       %%xmm15,(%1,%4)                  \n"
+      "lea        (%1,%4,2),%1                     \n"
+      "jg         1b                               \n"
+      : "+r"(src),                    // %0
+        "+r"(dst),                    // %1
+        "+r"(width)                   // %2
+      : "r"((intptr_t)(src_stride)),  // %3
+        "r"((intptr_t)(dst_stride))   // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7", "xmm8", "xmm9", "xmm10", "xmm11", "xmm12", "xmm13", "xmm14",
+        "xmm15");
+}
+#endif  // defined(HAS_TRANSPOSEWX8_FAST_SSSE3)
+
+// Transpose UV 8x8.  64 bit.
+#if defined(HAS_TRANSPOSEUVWX8_SSE2)
+void TransposeUVWx8_SSE2(const uint8* src,
+                         int src_stride,
+                         uint8* dst_a,
+                         int dst_stride_a,
+                         uint8* dst_b,
+                         int dst_stride_b,
+                         int width) {
+  asm volatile(
+      // Read in the data from the source pointer.
+      // First round of bit swap.
+      LABELALIGN
+      "1:                                          \n"
+      "movdqu     (%0),%%xmm0                      \n"
+      "movdqu     (%0,%4),%%xmm1                   \n"
+      "lea        (%0,%4,2),%0                     \n"
+      "movdqa     %%xmm0,%%xmm8                    \n"
+      "punpcklbw  %%xmm1,%%xmm0                    \n"
+      "punpckhbw  %%xmm1,%%xmm8                    \n"
+      "movdqa     %%xmm8,%%xmm1                    \n"
+      "movdqu     (%0),%%xmm2                      \n"
+      "movdqu     (%0,%4),%%xmm3                   \n"
+      "lea        (%0,%4,2),%0                     \n"
+      "movdqa     %%xmm2,%%xmm8                    \n"
+      "punpcklbw  %%xmm3,%%xmm2                    \n"
+      "punpckhbw  %%xmm3,%%xmm8                    \n"
+      "movdqa     %%xmm8,%%xmm3                    \n"
+      "movdqu     (%0),%%xmm4                      \n"
+      "movdqu     (%0,%4),%%xmm5                   \n"
+      "lea        (%0,%4,2),%0                     \n"
+      "movdqa     %%xmm4,%%xmm8                    \n"
+      "punpcklbw  %%xmm5,%%xmm4                    \n"
+      "punpckhbw  %%xmm5,%%xmm8                    \n"
+      "movdqa     %%xmm8,%%xmm5                    \n"
+      "movdqu     (%0),%%xmm6                      \n"
+      "movdqu     (%0,%4),%%xmm7                   \n"
+      "lea        (%0,%4,2),%0                     \n"
+      "movdqa     %%xmm6,%%xmm8                    \n"
+      "punpcklbw  %%xmm7,%%xmm6                    \n"
+      "neg        %4                               \n"
+      "lea        0x10(%0,%4,8),%0                 \n"
+      "punpckhbw  %%xmm7,%%xmm8                    \n"
+      "movdqa     %%xmm8,%%xmm7                    \n"
+      "neg        %4                               \n"
+      // Second round of bit swap.
+      "movdqa     %%xmm0,%%xmm8                    \n"
+      "movdqa     %%xmm1,%%xmm9                    \n"
+      "punpckhwd  %%xmm2,%%xmm8                    \n"
+      "punpckhwd  %%xmm3,%%xmm9                    \n"
+      "punpcklwd  %%xmm2,%%xmm0                    \n"
+      "punpcklwd  %%xmm3,%%xmm1                    \n"
+      "movdqa     %%xmm8,%%xmm2                    \n"
+      "movdqa     %%xmm9,%%xmm3                    \n"
+      "movdqa     %%xmm4,%%xmm8                    \n"
+      "movdqa     %%xmm5,%%xmm9                    \n"
+      "punpckhwd  %%xmm6,%%xmm8                    \n"
+      "punpckhwd  %%xmm7,%%xmm9                    \n"
+      "punpcklwd  %%xmm6,%%xmm4                    \n"
+      "punpcklwd  %%xmm7,%%xmm5                    \n"
+      "movdqa     %%xmm8,%%xmm6                    \n"
+      "movdqa     %%xmm9,%%xmm7                    \n"
+      // Third round of bit swap.
+      // Write to the destination pointer.
+      "movdqa     %%xmm0,%%xmm8                    \n"
+      "punpckldq  %%xmm4,%%xmm0                    \n"
+      "movlpd     %%xmm0,(%1)                      \n"  // Write back U channel
+      "movhpd     %%xmm0,(%2)                      \n"  // Write back V channel
+      "punpckhdq  %%xmm4,%%xmm8                    \n"
+      "movlpd     %%xmm8,(%1,%5)                   \n"
+      "lea        (%1,%5,2),%1                     \n"
+      "movhpd     %%xmm8,(%2,%6)                   \n"
+      "lea        (%2,%6,2),%2                     \n"
+      "movdqa     %%xmm2,%%xmm8                    \n"
+      "punpckldq  %%xmm6,%%xmm2                    \n"
+      "movlpd     %%xmm2,(%1)                      \n"
+      "movhpd     %%xmm2,(%2)                      \n"
+      "punpckhdq  %%xmm6,%%xmm8                    \n"
+      "movlpd     %%xmm8,(%1,%5)                   \n"
+      "lea        (%1,%5,2),%1                     \n"
+      "movhpd     %%xmm8,(%2,%6)                   \n"
+      "lea        (%2,%6,2),%2                     \n"
+      "movdqa     %%xmm1,%%xmm8                    \n"
+      "punpckldq  %%xmm5,%%xmm1                    \n"
+      "movlpd     %%xmm1,(%1)                      \n"
+      "movhpd     %%xmm1,(%2)                      \n"
+      "punpckhdq  %%xmm5,%%xmm8                    \n"
+      "movlpd     %%xmm8,(%1,%5)                   \n"
+      "lea        (%1,%5,2),%1                     \n"
+      "movhpd     %%xmm8,(%2,%6)                   \n"
+      "lea        (%2,%6,2),%2                     \n"
+      "movdqa     %%xmm3,%%xmm8                    \n"
+      "punpckldq  %%xmm7,%%xmm3                    \n"
+      "movlpd     %%xmm3,(%1)                      \n"
+      "movhpd     %%xmm3,(%2)                      \n"
+      "punpckhdq  %%xmm7,%%xmm8                    \n"
+      "sub        $0x8,%3                          \n"
+      "movlpd     %%xmm8,(%1,%5)                   \n"
+      "lea        (%1,%5,2),%1                     \n"
+      "movhpd     %%xmm8,(%2,%6)                   \n"
+      "lea        (%2,%6,2),%2                     \n"
+      "jg         1b                               \n"
+      : "+r"(src),                      // %0
+        "+r"(dst_a),                    // %1
+        "+r"(dst_b),                    // %2
+        "+r"(width)                     // %3
+      : "r"((intptr_t)(src_stride)),    // %4
+        "r"((intptr_t)(dst_stride_a)),  // %5
+        "r"((intptr_t)(dst_stride_b))   // %6
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7", "xmm8", "xmm9");
+}
+#endif  // defined(HAS_TRANSPOSEUVWX8_SSE2)
+#endif  // defined(__x86_64__) || defined(__i386__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif

pjproject-2.11.1/third_party/yuv/source/rotate_msa.cc

@@ -0,0 +1,250 @@
+/*
+ *  Copyright 2016 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/rotate_row.h"
+
+// This module is for GCC MSA
+#if !defined(LIBYUV_DISABLE_MSA) && defined(__mips_msa)
+#include "libyuv/macros_msa.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#define ILVRL_B(in0, in1, in2, in3, out0, out1, out2, out3) \
+  {                                                         \
+    out0 = (v16u8)__msa_ilvr_b((v16i8)in1, (v16i8)in0);     \
+    out1 = (v16u8)__msa_ilvl_b((v16i8)in1, (v16i8)in0);     \
+    out2 = (v16u8)__msa_ilvr_b((v16i8)in3, (v16i8)in2);     \
+    out3 = (v16u8)__msa_ilvl_b((v16i8)in3, (v16i8)in2);     \
+  }
+
+#define ILVRL_H(in0, in1, in2, in3, out0, out1, out2, out3) \
+  {                                                         \
+    out0 = (v16u8)__msa_ilvr_h((v8i16)in1, (v8i16)in0);     \
+    out1 = (v16u8)__msa_ilvl_h((v8i16)in1, (v8i16)in0);     \
+    out2 = (v16u8)__msa_ilvr_h((v8i16)in3, (v8i16)in2);     \
+    out3 = (v16u8)__msa_ilvl_h((v8i16)in3, (v8i16)in2);     \
+  }
+
+#define ILVRL_W(in0, in1, in2, in3, out0, out1, out2, out3) \
+  {                                                         \
+    out0 = (v16u8)__msa_ilvr_w((v4i32)in1, (v4i32)in0);     \
+    out1 = (v16u8)__msa_ilvl_w((v4i32)in1, (v4i32)in0);     \
+    out2 = (v16u8)__msa_ilvr_w((v4i32)in3, (v4i32)in2);     \
+    out3 = (v16u8)__msa_ilvl_w((v4i32)in3, (v4i32)in2);     \
+  }
+
+#define ILVRL_D(in0, in1, in2, in3, out0, out1, out2, out3) \
+  {                                                         \
+    out0 = (v16u8)__msa_ilvr_d((v2i64)in1, (v2i64)in0);     \
+    out1 = (v16u8)__msa_ilvl_d((v2i64)in1, (v2i64)in0);     \
+    out2 = (v16u8)__msa_ilvr_d((v2i64)in3, (v2i64)in2);     \
+    out3 = (v16u8)__msa_ilvl_d((v2i64)in3, (v2i64)in2);     \
+  }
+
+void TransposeWx16_C(const uint8* src,
+                     int src_stride,
+                     uint8* dst,
+                     int dst_stride,
+                     int width) {
+  TransposeWx8_C(src, src_stride, dst, dst_stride, width);
+  TransposeWx8_C((src + 8 * src_stride), src_stride, (dst + 8), dst_stride,
+                 width);
+}
+
+void TransposeUVWx16_C(const uint8* src,
+                       int src_stride,
+                       uint8* dst_a,
+                       int dst_stride_a,
+                       uint8* dst_b,
+                       int dst_stride_b,
+                       int width) {
+  TransposeUVWx8_C(src, src_stride, dst_a, dst_stride_a, dst_b, dst_stride_b,
+                   width);
+  TransposeUVWx8_C((src + 8 * src_stride), src_stride, (dst_a + 8),
+                   dst_stride_a, (dst_b + 8), dst_stride_b, width);
+}
+
+void TransposeWx16_MSA(const uint8* src,
+                       int src_stride,
+                       uint8* dst,
+                       int dst_stride,
+                       int width) {
+  int x;
+  const uint8* s;
+  v16u8 src0, src1, src2, src3, dst0, dst1, dst2, dst3, vec0, vec1, vec2, vec3;
+  v16u8 reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;
+  v16u8 res0, res1, res2, res3, res4, res5, res6, res7, res8, res9;
+
+  for (x = 0; x < width; x += 16) {
+    s = src;
+    src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src1 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src2 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src3 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    ILVRL_B(src0, src1, src2, src3, vec0, vec1, vec2, vec3);
+    ILVRL_H(vec0, vec2, vec1, vec3, reg0, reg1, reg2, reg3);
+    src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src1 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src2 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src3 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    ILVRL_B(src0, src1, src2, src3, vec0, vec1, vec2, vec3);
+    ILVRL_H(vec0, vec2, vec1, vec3, reg4, reg5, reg6, reg7);
+    ILVRL_W(reg0, reg4, reg1, reg5, res0, res1, res2, res3);
+    ILVRL_W(reg2, reg6, reg3, reg7, res4, res5, res6, res7);
+    src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src1 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src2 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src3 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    ILVRL_B(src0, src1, src2, src3, vec0, vec1, vec2, vec3);
+    ILVRL_H(vec0, vec2, vec1, vec3, reg0, reg1, reg2, reg3);
+    src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src1 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src2 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src3 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    ILVRL_B(src0, src1, src2, src3, vec0, vec1, vec2, vec3);
+    ILVRL_H(vec0, vec2, vec1, vec3, reg4, reg5, reg6, reg7);
+    res8 = (v16u8)__msa_ilvr_w((v4i32)reg4, (v4i32)reg0);
+    res9 = (v16u8)__msa_ilvl_w((v4i32)reg4, (v4i32)reg0);
+    ILVRL_D(res0, res8, res1, res9, dst0, dst1, dst2, dst3);
+    ST_UB4(dst0, dst1, dst2, dst3, dst, dst_stride);
+    dst += dst_stride * 4;
+    res8 = (v16u8)__msa_ilvr_w((v4i32)reg5, (v4i32)reg1);
+    res9 = (v16u8)__msa_ilvl_w((v4i32)reg5, (v4i32)reg1);
+    ILVRL_D(res2, res8, res3, res9, dst0, dst1, dst2, dst3);
+    ST_UB4(dst0, dst1, dst2, dst3, dst, dst_stride);
+    dst += dst_stride * 4;
+    res8 = (v16u8)__msa_ilvr_w((v4i32)reg6, (v4i32)reg2);
+    res9 = (v16u8)__msa_ilvl_w((v4i32)reg6, (v4i32)reg2);
+    ILVRL_D(res4, res8, res5, res9, dst0, dst1, dst2, dst3);
+    ST_UB4(dst0, dst1, dst2, dst3, dst, dst_stride);
+    dst += dst_stride * 4;
+    res8 = (v16u8)__msa_ilvr_w((v4i32)reg7, (v4i32)reg3);
+    res9 = (v16u8)__msa_ilvl_w((v4i32)reg7, (v4i32)reg3);
+    ILVRL_D(res6, res8, res7, res9, dst0, dst1, dst2, dst3);
+    ST_UB4(dst0, dst1, dst2, dst3, dst, dst_stride);
+    src += 16;
+    dst += dst_stride * 4;
+  }
+}
+
+void TransposeUVWx16_MSA(const uint8* src,
+                         int src_stride,
+                         uint8* dst_a,
+                         int dst_stride_a,
+                         uint8* dst_b,
+                         int dst_stride_b,
+                         int width) {
+  int x;
+  const uint8* s;
+  v16u8 src0, src1, src2, src3, dst0, dst1, dst2, dst3, vec0, vec1, vec2, vec3;
+  v16u8 reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;
+  v16u8 res0, res1, res2, res3, res4, res5, res6, res7, res8, res9;
+
+  for (x = 0; x < width; x += 8) {
+    s = src;
+    src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src1 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src2 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src3 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    ILVRL_B(src0, src1, src2, src3, vec0, vec1, vec2, vec3);
+    ILVRL_H(vec0, vec2, vec1, vec3, reg0, reg1, reg2, reg3);
+    src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src1 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src2 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src3 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    ILVRL_B(src0, src1, src2, src3, vec0, vec1, vec2, vec3);
+    ILVRL_H(vec0, vec2, vec1, vec3, reg4, reg5, reg6, reg7);
+    ILVRL_W(reg0, reg4, reg1, reg5, res0, res1, res2, res3);
+    ILVRL_W(reg2, reg6, reg3, reg7, res4, res5, res6, res7);
+    src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src1 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src2 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src3 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    ILVRL_B(src0, src1, src2, src3, vec0, vec1, vec2, vec3);
+    ILVRL_H(vec0, vec2, vec1, vec3, reg0, reg1, reg2, reg3);
+    src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src1 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src2 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    src3 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    s += src_stride;
+    ILVRL_B(src0, src1, src2, src3, vec0, vec1, vec2, vec3);
+    ILVRL_H(vec0, vec2, vec1, vec3, reg4, reg5, reg6, reg7);
+    res8 = (v16u8)__msa_ilvr_w((v4i32)reg4, (v4i32)reg0);
+    res9 = (v16u8)__msa_ilvl_w((v4i32)reg4, (v4i32)reg0);
+    ILVRL_D(res0, res8, res1, res9, dst0, dst1, dst2, dst3);
+    ST_UB2(dst0, dst2, dst_a, dst_stride_a);
+    ST_UB2(dst1, dst3, dst_b, dst_stride_b);
+    dst_a += dst_stride_a * 2;
+    dst_b += dst_stride_b * 2;
+    res8 = (v16u8)__msa_ilvr_w((v4i32)reg5, (v4i32)reg1);
+    res9 = (v16u8)__msa_ilvl_w((v4i32)reg5, (v4i32)reg1);
+    ILVRL_D(res2, res8, res3, res9, dst0, dst1, dst2, dst3);
+    ST_UB2(dst0, dst2, dst_a, dst_stride_a);
+    ST_UB2(dst1, dst3, dst_b, dst_stride_b);
+    dst_a += dst_stride_a * 2;
+    dst_b += dst_stride_b * 2;
+    res8 = (v16u8)__msa_ilvr_w((v4i32)reg6, (v4i32)reg2);
+    res9 = (v16u8)__msa_ilvl_w((v4i32)reg6, (v4i32)reg2);
+    ILVRL_D(res4, res8, res5, res9, dst0, dst1, dst2, dst3);
+    ST_UB2(dst0, dst2, dst_a, dst_stride_a);
+    ST_UB2(dst1, dst3, dst_b, dst_stride_b);
+    dst_a += dst_stride_a * 2;
+    dst_b += dst_stride_b * 2;
+    res8 = (v16u8)__msa_ilvr_w((v4i32)reg7, (v4i32)reg3);
+    res9 = (v16u8)__msa_ilvl_w((v4i32)reg7, (v4i32)reg3);
+    ILVRL_D(res6, res8, res7, res9, dst0, dst1, dst2, dst3);
+    ST_UB2(dst0, dst2, dst_a, dst_stride_a);
+    ST_UB2(dst1, dst3, dst_b, dst_stride_b);
+    src += 16;
+    dst_a += dst_stride_a * 2;
+    dst_b += dst_stride_b * 2;
+  }
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // !defined(LIBYUV_DISABLE_MSA) && defined(__mips_msa)

pjproject-2.11.1/third_party/yuv/source/rotate_neon.cc

@@ -0,0 +1,416 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/rotate_row.h"
+#include "libyuv/row.h"
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__) && \
+    !defined(__aarch64__)
+
+static uvec8 kVTbl4x4Transpose = {0, 4, 8,  12, 1, 5, 9,  13,
+                                  2, 6, 10, 14, 3, 7, 11, 15};
+
+void TransposeWx8_NEON(const uint8* src,
+                       int src_stride,
+                       uint8* dst,
+                       int dst_stride,
+                       int width) {
+  const uint8* src_temp;
+  asm volatile(
+      // loops are on blocks of 8. loop will stop when
+      // counter gets to or below 0. starting the counter
+      // at w-8 allow for this
+      "sub         %5, #8                        \n"
+
+      // handle 8x8 blocks. this should be the majority of the plane
+      "1:                                        \n"
+      "mov         %0, %1                      \n"
+
+      "vld1.8      {d0}, [%0], %2              \n"
+      "vld1.8      {d1}, [%0], %2              \n"
+      "vld1.8      {d2}, [%0], %2              \n"
+      "vld1.8      {d3}, [%0], %2              \n"
+      "vld1.8      {d4}, [%0], %2              \n"
+      "vld1.8      {d5}, [%0], %2              \n"
+      "vld1.8      {d6}, [%0], %2              \n"
+      "vld1.8      {d7}, [%0]                  \n"
+
+      "vtrn.8      d1, d0                      \n"
+      "vtrn.8      d3, d2                      \n"
+      "vtrn.8      d5, d4                      \n"
+      "vtrn.8      d7, d6                      \n"
+
+      "vtrn.16     d1, d3                      \n"
+      "vtrn.16     d0, d2                      \n"
+      "vtrn.16     d5, d7                      \n"
+      "vtrn.16     d4, d6                      \n"
+
+      "vtrn.32     d1, d5                      \n"
+      "vtrn.32     d0, d4                      \n"
+      "vtrn.32     d3, d7                      \n"
+      "vtrn.32     d2, d6                      \n"
+
+      "vrev16.8    q0, q0                      \n"
+      "vrev16.8    q1, q1                      \n"
+      "vrev16.8    q2, q2                      \n"
+      "vrev16.8    q3, q3                      \n"
+
+      "mov         %0, %3                      \n"
+
+      "vst1.8      {d1}, [%0], %4              \n"
+      "vst1.8      {d0}, [%0], %4              \n"
+      "vst1.8      {d3}, [%0], %4              \n"
+      "vst1.8      {d2}, [%0], %4              \n"
+      "vst1.8      {d5}, [%0], %4              \n"
+      "vst1.8      {d4}, [%0], %4              \n"
+      "vst1.8      {d7}, [%0], %4              \n"
+      "vst1.8      {d6}, [%0]                  \n"
+
+      "add         %1, #8                      \n"  // src += 8
+      "add         %3, %3, %4, lsl #3          \n"  // dst += 8 * dst_stride
+      "subs        %5,  #8                     \n"  // w   -= 8
+      "bge         1b                          \n"
+
+      // add 8 back to counter. if the result is 0 there are
+      // no residuals.
+      "adds        %5, #8                        \n"
+      "beq         4f                            \n"
+
+      // some residual, so between 1 and 7 lines left to transpose
+      "cmp         %5, #2                        \n"
+      "blt         3f                            \n"
+
+      "cmp         %5, #4                        \n"
+      "blt         2f                            \n"
+
+      // 4x8 block
+      "mov         %0, %1                        \n"
+      "vld1.32     {d0[0]}, [%0], %2             \n"
+      "vld1.32     {d0[1]}, [%0], %2             \n"
+      "vld1.32     {d1[0]}, [%0], %2             \n"
+      "vld1.32     {d1[1]}, [%0], %2             \n"
+      "vld1.32     {d2[0]}, [%0], %2             \n"
+      "vld1.32     {d2[1]}, [%0], %2             \n"
+      "vld1.32     {d3[0]}, [%0], %2             \n"
+      "vld1.32     {d3[1]}, [%0]                 \n"
+
+      "mov         %0, %3                        \n"
+
+      "vld1.8      {q3}, [%6]                    \n"
+
+      "vtbl.8      d4, {d0, d1}, d6              \n"
+      "vtbl.8      d5, {d0, d1}, d7              \n"
+      "vtbl.8      d0, {d2, d3}, d6              \n"
+      "vtbl.8      d1, {d2, d3}, d7              \n"
+
+      // TODO(frkoenig): Rework shuffle above to
+      // write out with 4 instead of 8 writes.
+      "vst1.32     {d4[0]}, [%0], %4             \n"
+      "vst1.32     {d4[1]}, [%0], %4             \n"
+      "vst1.32     {d5[0]}, [%0], %4             \n"
+      "vst1.32     {d5[1]}, [%0]                 \n"
+
+      "add         %0, %3, #4                    \n"
+      "vst1.32     {d0[0]}, [%0], %4             \n"
+      "vst1.32     {d0[1]}, [%0], %4             \n"
+      "vst1.32     {d1[0]}, [%0], %4             \n"
+      "vst1.32     {d1[1]}, [%0]                 \n"
+
+      "add         %1, #4                        \n"  // src += 4
+      "add         %3, %3, %4, lsl #2            \n"  // dst += 4 * dst_stride
+      "subs        %5,  #4                       \n"  // w   -= 4
+      "beq         4f                            \n"
+
+      // some residual, check to see if it includes a 2x8 block,
+      // or less
+      "cmp         %5, #2                        \n"
+      "blt         3f                            \n"
+
+      // 2x8 block
+      "2:                                        \n"
+      "mov         %0, %1                        \n"
+      "vld1.16     {d0[0]}, [%0], %2             \n"
+      "vld1.16     {d1[0]}, [%0], %2             \n"
+      "vld1.16     {d0[1]}, [%0], %2             \n"
+      "vld1.16     {d1[1]}, [%0], %2             \n"
+      "vld1.16     {d0[2]}, [%0], %2             \n"
+      "vld1.16     {d1[2]}, [%0], %2             \n"
+      "vld1.16     {d0[3]}, [%0], %2             \n"
+      "vld1.16     {d1[3]}, [%0]                 \n"
+
+      "vtrn.8      d0, d1                        \n"
+
+      "mov         %0, %3                        \n"
+
+      "vst1.64     {d0}, [%0], %4                \n"
+      "vst1.64     {d1}, [%0]                    \n"
+
+      "add         %1, #2                        \n"  // src += 2
+      "add         %3, %3, %4, lsl #1            \n"  // dst += 2 * dst_stride
+      "subs        %5,  #2                       \n"  // w   -= 2
+      "beq         4f                            \n"
+
+      // 1x8 block
+      "3:                                        \n"
+      "vld1.8      {d0[0]}, [%1], %2             \n"
+      "vld1.8      {d0[1]}, [%1], %2             \n"
+      "vld1.8      {d0[2]}, [%1], %2             \n"
+      "vld1.8      {d0[3]}, [%1], %2             \n"
+      "vld1.8      {d0[4]}, [%1], %2             \n"
+      "vld1.8      {d0[5]}, [%1], %2             \n"
+      "vld1.8      {d0[6]}, [%1], %2             \n"
+      "vld1.8      {d0[7]}, [%1]                 \n"
+
+      "vst1.64     {d0}, [%3]                    \n"
+
+      "4:                                        \n"
+
+      : "=&r"(src_temp),         // %0
+        "+r"(src),               // %1
+        "+r"(src_stride),        // %2
+        "+r"(dst),               // %3
+        "+r"(dst_stride),        // %4
+        "+r"(width)              // %5
+      : "r"(&kVTbl4x4Transpose)  // %6
+      : "memory", "cc", "q0", "q1", "q2", "q3");
+}
+
+static uvec8 kVTbl4x4TransposeDi = {0, 8,  1, 9,  2, 10, 3, 11,
+                                    4, 12, 5, 13, 6, 14, 7, 15};
+
+void TransposeUVWx8_NEON(const uint8* src,
+                         int src_stride,
+                         uint8* dst_a,
+                         int dst_stride_a,
+                         uint8* dst_b,
+                         int dst_stride_b,
+                         int width) {
+  const uint8* src_temp;
+  asm volatile(
+      // loops are on blocks of 8. loop will stop when
+      // counter gets to or below 0. starting the counter
+      // at w-8 allow for this
+      "sub         %7, #8                        \n"
+
+      // handle 8x8 blocks. this should be the majority of the plane
+      "1:                                        \n"
+      "mov         %0, %1                      \n"
+
+      "vld2.8      {d0,  d1},  [%0], %2        \n"
+      "vld2.8      {d2,  d3},  [%0], %2        \n"
+      "vld2.8      {d4,  d5},  [%0], %2        \n"
+      "vld2.8      {d6,  d7},  [%0], %2        \n"
+      "vld2.8      {d16, d17}, [%0], %2        \n"
+      "vld2.8      {d18, d19}, [%0], %2        \n"
+      "vld2.8      {d20, d21}, [%0], %2        \n"
+      "vld2.8      {d22, d23}, [%0]            \n"
+
+      "vtrn.8      q1, q0                      \n"
+      "vtrn.8      q3, q2                      \n"
+      "vtrn.8      q9, q8                      \n"
+      "vtrn.8      q11, q10                    \n"
+
+      "vtrn.16     q1, q3                      \n"
+      "vtrn.16     q0, q2                      \n"
+      "vtrn.16     q9, q11                     \n"
+      "vtrn.16     q8, q10                     \n"
+
+      "vtrn.32     q1, q9                      \n"
+      "vtrn.32     q0, q8                      \n"
+      "vtrn.32     q3, q11                     \n"
+      "vtrn.32     q2, q10                     \n"
+
+      "vrev16.8    q0, q0                      \n"
+      "vrev16.8    q1, q1                      \n"
+      "vrev16.8    q2, q2                      \n"
+      "vrev16.8    q3, q3                      \n"
+      "vrev16.8    q8, q8                      \n"
+      "vrev16.8    q9, q9                      \n"
+      "vrev16.8    q10, q10                    \n"
+      "vrev16.8    q11, q11                    \n"
+
+      "mov         %0, %3                      \n"
+
+      "vst1.8      {d2},  [%0], %4             \n"
+      "vst1.8      {d0},  [%0], %4             \n"
+      "vst1.8      {d6},  [%0], %4             \n"
+      "vst1.8      {d4},  [%0], %4             \n"
+      "vst1.8      {d18}, [%0], %4             \n"
+      "vst1.8      {d16}, [%0], %4             \n"
+      "vst1.8      {d22}, [%0], %4             \n"
+      "vst1.8      {d20}, [%0]                 \n"
+
+      "mov         %0, %5                      \n"
+
+      "vst1.8      {d3},  [%0], %6             \n"
+      "vst1.8      {d1},  [%0], %6             \n"
+      "vst1.8      {d7},  [%0], %6             \n"
+      "vst1.8      {d5},  [%0], %6             \n"
+      "vst1.8      {d19}, [%0], %6             \n"
+      "vst1.8      {d17}, [%0], %6             \n"
+      "vst1.8      {d23}, [%0], %6             \n"
+      "vst1.8      {d21}, [%0]                 \n"
+
+      "add         %1, #8*2                    \n"  // src   += 8*2
+      "add         %3, %3, %4, lsl #3          \n"  // dst_a += 8 * dst_stride_a
+      "add         %5, %5, %6, lsl #3          \n"  // dst_b += 8 * dst_stride_b
+      "subs        %7,  #8                     \n"  // w     -= 8
+      "bge         1b                          \n"
+
+      // add 8 back to counter. if the result is 0 there are
+      // no residuals.
+      "adds        %7, #8                        \n"
+      "beq         4f                            \n"
+
+      // some residual, so between 1 and 7 lines left to transpose
+      "cmp         %7, #2                        \n"
+      "blt         3f                            \n"
+
+      "cmp         %7, #4                        \n"
+      "blt         2f                            \n"
+
+      // TODO(frkoenig): Clean this up
+      // 4x8 block
+      "mov         %0, %1                        \n"
+      "vld1.64     {d0}, [%0], %2                \n"
+      "vld1.64     {d1}, [%0], %2                \n"
+      "vld1.64     {d2}, [%0], %2                \n"
+      "vld1.64     {d3}, [%0], %2                \n"
+      "vld1.64     {d4}, [%0], %2                \n"
+      "vld1.64     {d5}, [%0], %2                \n"
+      "vld1.64     {d6}, [%0], %2                \n"
+      "vld1.64     {d7}, [%0]                    \n"
+
+      "vld1.8      {q15}, [%8]                   \n"
+
+      "vtrn.8      q0, q1                        \n"
+      "vtrn.8      q2, q3                        \n"
+
+      "vtbl.8      d16, {d0, d1}, d30            \n"
+      "vtbl.8      d17, {d0, d1}, d31            \n"
+      "vtbl.8      d18, {d2, d3}, d30            \n"
+      "vtbl.8      d19, {d2, d3}, d31            \n"
+      "vtbl.8      d20, {d4, d5}, d30            \n"
+      "vtbl.8      d21, {d4, d5}, d31            \n"
+      "vtbl.8      d22, {d6, d7}, d30            \n"
+      "vtbl.8      d23, {d6, d7}, d31            \n"
+
+      "mov         %0, %3                        \n"
+
+      "vst1.32     {d16[0]},  [%0], %4           \n"
+      "vst1.32     {d16[1]},  [%0], %4           \n"
+      "vst1.32     {d17[0]},  [%0], %4           \n"
+      "vst1.32     {d17[1]},  [%0], %4           \n"
+
+      "add         %0, %3, #4                    \n"
+      "vst1.32     {d20[0]}, [%0], %4            \n"
+      "vst1.32     {d20[1]}, [%0], %4            \n"
+      "vst1.32     {d21[0]}, [%0], %4            \n"
+      "vst1.32     {d21[1]}, [%0]                \n"
+
+      "mov         %0, %5                        \n"
+
+      "vst1.32     {d18[0]}, [%0], %6            \n"
+      "vst1.32     {d18[1]}, [%0], %6            \n"
+      "vst1.32     {d19[0]}, [%0], %6            \n"
+      "vst1.32     {d19[1]}, [%0], %6            \n"
+
+      "add         %0, %5, #4                    \n"
+      "vst1.32     {d22[0]},  [%0], %6           \n"
+      "vst1.32     {d22[1]},  [%0], %6           \n"
+      "vst1.32     {d23[0]},  [%0], %6           \n"
+      "vst1.32     {d23[1]},  [%0]               \n"
+
+      "add         %1, #4*2                      \n"  // src   += 4 * 2
+      "add         %3, %3, %4, lsl #2            \n"  // dst_a += 4 *
+                                                      // dst_stride_a
+      "add         %5, %5, %6, lsl #2            \n"  // dst_b += 4 *
+                                                      // dst_stride_b
+      "subs        %7,  #4                       \n"  // w     -= 4
+      "beq         4f                            \n"
+
+      // some residual, check to see if it includes a 2x8 block,
+      // or less
+      "cmp         %7, #2                        \n"
+      "blt         3f                            \n"
+
+      // 2x8 block
+      "2:                                        \n"
+      "mov         %0, %1                        \n"
+      "vld2.16     {d0[0], d2[0]}, [%0], %2      \n"
+      "vld2.16     {d1[0], d3[0]}, [%0], %2      \n"
+      "vld2.16     {d0[1], d2[1]}, [%0], %2      \n"
+      "vld2.16     {d1[1], d3[1]}, [%0], %2      \n"
+      "vld2.16     {d0[2], d2[2]}, [%0], %2      \n"
+      "vld2.16     {d1[2], d3[2]}, [%0], %2      \n"
+      "vld2.16     {d0[3], d2[3]}, [%0], %2      \n"
+      "vld2.16     {d1[3], d3[3]}, [%0]          \n"
+
+      "vtrn.8      d0, d1                        \n"
+      "vtrn.8      d2, d3                        \n"
+
+      "mov         %0, %3                        \n"
+
+      "vst1.64     {d0}, [%0], %4                \n"
+      "vst1.64     {d2}, [%0]                    \n"
+
+      "mov         %0, %5                        \n"
+
+      "vst1.64     {d1}, [%0], %6                \n"
+      "vst1.64     {d3}, [%0]                    \n"
+
+      "add         %1, #2*2                      \n"  // src   += 2 * 2
+      "add         %3, %3, %4, lsl #1            \n"  // dst_a += 2 *
+                                                      // dst_stride_a
+      "add         %5, %5, %6, lsl #1            \n"  // dst_b += 2 *
+                                                      // dst_stride_b
+      "subs        %7,  #2                       \n"  // w     -= 2
+      "beq         4f                            \n"
+
+      // 1x8 block
+      "3:                                        \n"
+      "vld2.8      {d0[0], d1[0]}, [%1], %2      \n"
+      "vld2.8      {d0[1], d1[1]}, [%1], %2      \n"
+      "vld2.8      {d0[2], d1[2]}, [%1], %2      \n"
+      "vld2.8      {d0[3], d1[3]}, [%1], %2      \n"
+      "vld2.8      {d0[4], d1[4]}, [%1], %2      \n"
+      "vld2.8      {d0[5], d1[5]}, [%1], %2      \n"
+      "vld2.8      {d0[6], d1[6]}, [%1], %2      \n"
+      "vld2.8      {d0[7], d1[7]}, [%1]          \n"
+
+      "vst1.64     {d0}, [%3]                    \n"
+      "vst1.64     {d1}, [%5]                    \n"
+
+      "4:                                        \n"
+
+      : "=&r"(src_temp),           // %0
+        "+r"(src),                 // %1
+        "+r"(src_stride),          // %2
+        "+r"(dst_a),               // %3
+        "+r"(dst_stride_a),        // %4
+        "+r"(dst_b),               // %5
+        "+r"(dst_stride_b),        // %6
+        "+r"(width)                // %7
+      : "r"(&kVTbl4x4TransposeDi)  // %8
+      : "memory", "cc", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11");
+}
+#endif  // defined(__ARM_NEON__) && !defined(__aarch64__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif

pjproject-2.11.1/third_party/yuv/source/rotate_neon64.cc

@@ -0,0 +1,426 @@
+/*
+ *  Copyright 2014 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/rotate_row.h"
+#include "libyuv/row.h"
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC Neon armv8 64 bit.
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
+
+static uvec8 kVTbl4x4Transpose = {0, 4, 8,  12, 1, 5, 9,  13,
+                                  2, 6, 10, 14, 3, 7, 11, 15};
+
+void TransposeWx8_NEON(const uint8* src,
+                       int src_stride,
+                       uint8* dst,
+                       int dst_stride,
+                       int width) {
+  const uint8* src_temp;
+  asm volatile(
+      // loops are on blocks of 8. loop will stop when
+      // counter gets to or below 0. starting the counter
+      // at w-8 allow for this
+      "sub         %w3, %w3, #8                     \n"
+
+      // handle 8x8 blocks. this should be the majority of the plane
+      "1:                                          \n"
+      "mov         %0, %1                        \n"
+
+      "ld1        {v0.8b}, [%0], %5              \n"
+      "ld1        {v1.8b}, [%0], %5              \n"
+      "ld1        {v2.8b}, [%0], %5              \n"
+      "ld1        {v3.8b}, [%0], %5              \n"
+      "ld1        {v4.8b}, [%0], %5              \n"
+      "ld1        {v5.8b}, [%0], %5              \n"
+      "ld1        {v6.8b}, [%0], %5              \n"
+      "ld1        {v7.8b}, [%0]                  \n"
+
+      "trn2     v16.8b, v0.8b, v1.8b             \n"
+      "trn1     v17.8b, v0.8b, v1.8b             \n"
+      "trn2     v18.8b, v2.8b, v3.8b             \n"
+      "trn1     v19.8b, v2.8b, v3.8b             \n"
+      "trn2     v20.8b, v4.8b, v5.8b             \n"
+      "trn1     v21.8b, v4.8b, v5.8b             \n"
+      "trn2     v22.8b, v6.8b, v7.8b             \n"
+      "trn1     v23.8b, v6.8b, v7.8b             \n"
+
+      "trn2     v3.4h, v17.4h, v19.4h            \n"
+      "trn1     v1.4h, v17.4h, v19.4h            \n"
+      "trn2     v2.4h, v16.4h, v18.4h            \n"
+      "trn1     v0.4h, v16.4h, v18.4h            \n"
+      "trn2     v7.4h, v21.4h, v23.4h            \n"
+      "trn1     v5.4h, v21.4h, v23.4h            \n"
+      "trn2     v6.4h, v20.4h, v22.4h            \n"
+      "trn1     v4.4h, v20.4h, v22.4h            \n"
+
+      "trn2     v21.2s, v1.2s, v5.2s             \n"
+      "trn1     v17.2s, v1.2s, v5.2s             \n"
+      "trn2     v20.2s, v0.2s, v4.2s             \n"
+      "trn1     v16.2s, v0.2s, v4.2s             \n"
+      "trn2     v23.2s, v3.2s, v7.2s             \n"
+      "trn1     v19.2s, v3.2s, v7.2s             \n"
+      "trn2     v22.2s, v2.2s, v6.2s             \n"
+      "trn1     v18.2s, v2.2s, v6.2s             \n"
+
+      "mov         %0, %2                        \n"
+
+      "st1      {v17.8b}, [%0], %6               \n"
+      "st1      {v16.8b}, [%0], %6               \n"
+      "st1      {v19.8b}, [%0], %6               \n"
+      "st1      {v18.8b}, [%0], %6               \n"
+      "st1      {v21.8b}, [%0], %6               \n"
+      "st1      {v20.8b}, [%0], %6               \n"
+      "st1      {v23.8b}, [%0], %6               \n"
+      "st1      {v22.8b}, [%0]                   \n"
+
+      "add         %1, %1, #8                    \n"  // src += 8
+      "add         %2, %2, %6, lsl #3            \n"  // dst += 8 * dst_stride
+      "subs        %w3, %w3, #8                  \n"  // w   -= 8
+      "b.ge        1b                            \n"
+
+      // add 8 back to counter. if the result is 0 there are
+      // no residuals.
+      "adds        %w3, %w3, #8                    \n"
+      "b.eq        4f                              \n"
+
+      // some residual, so between 1 and 7 lines left to transpose
+      "cmp         %w3, #2                          \n"
+      "b.lt        3f                              \n"
+
+      "cmp         %w3, #4                          \n"
+      "b.lt        2f                              \n"
+
+      // 4x8 block
+      "mov         %0, %1                          \n"
+      "ld1     {v0.s}[0], [%0], %5                 \n"
+      "ld1     {v0.s}[1], [%0], %5                 \n"
+      "ld1     {v0.s}[2], [%0], %5                 \n"
+      "ld1     {v0.s}[3], [%0], %5                 \n"
+      "ld1     {v1.s}[0], [%0], %5                 \n"
+      "ld1     {v1.s}[1], [%0], %5                 \n"
+      "ld1     {v1.s}[2], [%0], %5                 \n"
+      "ld1     {v1.s}[3], [%0]                     \n"
+
+      "mov         %0, %2                          \n"
+
+      "ld1      {v2.16b}, [%4]                     \n"
+
+      "tbl      v3.16b, {v0.16b}, v2.16b           \n"
+      "tbl      v0.16b, {v1.16b}, v2.16b           \n"
+
+      // TODO(frkoenig): Rework shuffle above to
+      // write out with 4 instead of 8 writes.
+      "st1 {v3.s}[0], [%0], %6                     \n"
+      "st1 {v3.s}[1], [%0], %6                     \n"
+      "st1 {v3.s}[2], [%0], %6                     \n"
+      "st1 {v3.s}[3], [%0]                         \n"
+
+      "add         %0, %2, #4                      \n"
+      "st1 {v0.s}[0], [%0], %6                     \n"
+      "st1 {v0.s}[1], [%0], %6                     \n"
+      "st1 {v0.s}[2], [%0], %6                     \n"
+      "st1 {v0.s}[3], [%0]                         \n"
+
+      "add         %1, %1, #4                      \n"  // src += 4
+      "add         %2, %2, %6, lsl #2              \n"  // dst += 4 * dst_stride
+      "subs        %w3, %w3, #4                    \n"  // w   -= 4
+      "b.eq        4f                              \n"
+
+      // some residual, check to see if it includes a 2x8 block,
+      // or less
+      "cmp         %w3, #2                         \n"
+      "b.lt        3f                              \n"
+
+      // 2x8 block
+      "2:                                          \n"
+      "mov         %0, %1                          \n"
+      "ld1     {v0.h}[0], [%0], %5                 \n"
+      "ld1     {v1.h}[0], [%0], %5                 \n"
+      "ld1     {v0.h}[1], [%0], %5                 \n"
+      "ld1     {v1.h}[1], [%0], %5                 \n"
+      "ld1     {v0.h}[2], [%0], %5                 \n"
+      "ld1     {v1.h}[2], [%0], %5                 \n"
+      "ld1     {v0.h}[3], [%0], %5                 \n"
+      "ld1     {v1.h}[3], [%0]                     \n"
+
+      "trn2    v2.8b, v0.8b, v1.8b                 \n"
+      "trn1    v3.8b, v0.8b, v1.8b                 \n"
+
+      "mov         %0, %2                          \n"
+
+      "st1     {v3.8b}, [%0], %6                   \n"
+      "st1     {v2.8b}, [%0]                       \n"
+
+      "add         %1, %1, #2                      \n"  // src += 2
+      "add         %2, %2, %6, lsl #1              \n"  // dst += 2 * dst_stride
+      "subs        %w3, %w3,  #2                   \n"  // w   -= 2
+      "b.eq        4f                              \n"
+
+      // 1x8 block
+      "3:                                          \n"
+      "ld1         {v0.b}[0], [%1], %5             \n"
+      "ld1         {v0.b}[1], [%1], %5             \n"
+      "ld1         {v0.b}[2], [%1], %5             \n"
+      "ld1         {v0.b}[3], [%1], %5             \n"
+      "ld1         {v0.b}[4], [%1], %5             \n"
+      "ld1         {v0.b}[5], [%1], %5             \n"
+      "ld1         {v0.b}[6], [%1], %5             \n"
+      "ld1         {v0.b}[7], [%1]                 \n"
+
+      "st1         {v0.8b}, [%2]                   \n"
+
+      "4:                                          \n"
+
+      : "=&r"(src_temp),                          // %0
+        "+r"(src),                                // %1
+        "+r"(dst),                                // %2
+        "+r"(width)                               // %3
+      : "r"(&kVTbl4x4Transpose),                  // %4
+        "r"(static_cast<ptrdiff_t>(src_stride)),  // %5
+        "r"(static_cast<ptrdiff_t>(dst_stride))   // %6
+      : "memory", "cc", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16",
+        "v17", "v18", "v19", "v20", "v21", "v22", "v23");
+}
+
+static uint8 kVTbl4x4TransposeDi[32] = {
+    0, 16, 32, 48, 2, 18, 34, 50, 4, 20, 36, 52, 6, 22, 38, 54,
+    1, 17, 33, 49, 3, 19, 35, 51, 5, 21, 37, 53, 7, 23, 39, 55};
+
+void TransposeUVWx8_NEON(const uint8* src,
+                         int src_stride,
+                         uint8* dst_a,
+                         int dst_stride_a,
+                         uint8* dst_b,
+                         int dst_stride_b,
+                         int width) {
+  const uint8* src_temp;
+  asm volatile(
+      // loops are on blocks of 8. loop will stop when
+      // counter gets to or below 0. starting the counter
+      // at w-8 allow for this
+      "sub       %w4, %w4, #8                    \n"
+
+      // handle 8x8 blocks. this should be the majority of the plane
+      "1:                                        \n"
+      "mov       %0, %1                          \n"
+
+      "ld1       {v0.16b}, [%0], %5              \n"
+      "ld1       {v1.16b}, [%0], %5              \n"
+      "ld1       {v2.16b}, [%0], %5              \n"
+      "ld1       {v3.16b}, [%0], %5              \n"
+      "ld1       {v4.16b}, [%0], %5              \n"
+      "ld1       {v5.16b}, [%0], %5              \n"
+      "ld1       {v6.16b}, [%0], %5              \n"
+      "ld1       {v7.16b}, [%0]                  \n"
+
+      "trn1      v16.16b, v0.16b, v1.16b         \n"
+      "trn2      v17.16b, v0.16b, v1.16b         \n"
+      "trn1      v18.16b, v2.16b, v3.16b         \n"
+      "trn2      v19.16b, v2.16b, v3.16b         \n"
+      "trn1      v20.16b, v4.16b, v5.16b         \n"
+      "trn2      v21.16b, v4.16b, v5.16b         \n"
+      "trn1      v22.16b, v6.16b, v7.16b         \n"
+      "trn2      v23.16b, v6.16b, v7.16b         \n"
+
+      "trn1      v0.8h, v16.8h, v18.8h           \n"
+      "trn2      v1.8h, v16.8h, v18.8h           \n"
+      "trn1      v2.8h, v20.8h, v22.8h           \n"
+      "trn2      v3.8h, v20.8h, v22.8h           \n"
+      "trn1      v4.8h, v17.8h, v19.8h           \n"
+      "trn2      v5.8h, v17.8h, v19.8h           \n"
+      "trn1      v6.8h, v21.8h, v23.8h           \n"
+      "trn2      v7.8h, v21.8h, v23.8h           \n"
+
+      "trn1      v16.4s, v0.4s, v2.4s            \n"
+      "trn2      v17.4s, v0.4s, v2.4s            \n"
+      "trn1      v18.4s, v1.4s, v3.4s            \n"
+      "trn2      v19.4s, v1.4s, v3.4s            \n"
+      "trn1      v20.4s, v4.4s, v6.4s            \n"
+      "trn2      v21.4s, v4.4s, v6.4s            \n"
+      "trn1      v22.4s, v5.4s, v7.4s            \n"
+      "trn2      v23.4s, v5.4s, v7.4s            \n"
+
+      "mov       %0, %2                          \n"
+
+      "st1       {v16.d}[0], [%0], %6            \n"
+      "st1       {v18.d}[0], [%0], %6            \n"
+      "st1       {v17.d}[0], [%0], %6            \n"
+      "st1       {v19.d}[0], [%0], %6            \n"
+      "st1       {v16.d}[1], [%0], %6            \n"
+      "st1       {v18.d}[1], [%0], %6            \n"
+      "st1       {v17.d}[1], [%0], %6            \n"
+      "st1       {v19.d}[1], [%0]                \n"
+
+      "mov       %0, %3                          \n"
+
+      "st1       {v20.d}[0], [%0], %7            \n"
+      "st1       {v22.d}[0], [%0], %7            \n"
+      "st1       {v21.d}[0], [%0], %7            \n"
+      "st1       {v23.d}[0], [%0], %7            \n"
+      "st1       {v20.d}[1], [%0], %7            \n"
+      "st1       {v22.d}[1], [%0], %7            \n"
+      "st1       {v21.d}[1], [%0], %7            \n"
+      "st1       {v23.d}[1], [%0]                \n"
+
+      "add       %1, %1, #16                     \n"  // src   += 8*2
+      "add       %2, %2, %6, lsl #3              \n"  // dst_a += 8 *
+                                                      // dst_stride_a
+      "add       %3, %3, %7, lsl #3              \n"  // dst_b += 8 *
+                                                      // dst_stride_b
+      "subs      %w4, %w4,  #8                   \n"  // w     -= 8
+      "b.ge      1b                              \n"
+
+      // add 8 back to counter. if the result is 0 there are
+      // no residuals.
+      "adds      %w4, %w4, #8                    \n"
+      "b.eq      4f                              \n"
+
+      // some residual, so between 1 and 7 lines left to transpose
+      "cmp       %w4, #2                         \n"
+      "b.lt      3f                              \n"
+
+      "cmp       %w4, #4                         \n"
+      "b.lt      2f                              \n"
+
+      // TODO(frkoenig): Clean this up
+      // 4x8 block
+      "mov       %0, %1                          \n"
+      "ld1       {v0.8b}, [%0], %5               \n"
+      "ld1       {v1.8b}, [%0], %5               \n"
+      "ld1       {v2.8b}, [%0], %5               \n"
+      "ld1       {v3.8b}, [%0], %5               \n"
+      "ld1       {v4.8b}, [%0], %5               \n"
+      "ld1       {v5.8b}, [%0], %5               \n"
+      "ld1       {v6.8b}, [%0], %5               \n"
+      "ld1       {v7.8b}, [%0]                   \n"
+
+      "ld1       {v30.16b}, [%8], #16            \n"
+      "ld1       {v31.16b}, [%8]                 \n"
+
+      "tbl       v16.16b, {v0.16b, v1.16b, v2.16b, v3.16b}, v30.16b  \n"
+      "tbl       v17.16b, {v0.16b, v1.16b, v2.16b, v3.16b}, v31.16b  \n"
+      "tbl       v18.16b, {v4.16b, v5.16b, v6.16b, v7.16b}, v30.16b  \n"
+      "tbl       v19.16b, {v4.16b, v5.16b, v6.16b, v7.16b}, v31.16b  \n"
+
+      "mov       %0, %2                          \n"
+
+      "st1       {v16.s}[0],  [%0], %6           \n"
+      "st1       {v16.s}[1],  [%0], %6           \n"
+      "st1       {v16.s}[2],  [%0], %6           \n"
+      "st1       {v16.s}[3],  [%0], %6           \n"
+
+      "add       %0, %2, #4                      \n"
+      "st1       {v18.s}[0], [%0], %6            \n"
+      "st1       {v18.s}[1], [%0], %6            \n"
+      "st1       {v18.s}[2], [%0], %6            \n"
+      "st1       {v18.s}[3], [%0]                \n"
+
+      "mov       %0, %3                          \n"
+
+      "st1       {v17.s}[0], [%0], %7            \n"
+      "st1       {v17.s}[1], [%0], %7            \n"
+      "st1       {v17.s}[2], [%0], %7            \n"
+      "st1       {v17.s}[3], [%0], %7            \n"
+
+      "add       %0, %3, #4                      \n"
+      "st1       {v19.s}[0],  [%0], %7           \n"
+      "st1       {v19.s}[1],  [%0], %7           \n"
+      "st1       {v19.s}[2],  [%0], %7           \n"
+      "st1       {v19.s}[3],  [%0]               \n"
+
+      "add       %1, %1, #8                      \n"  // src   += 4 * 2
+      "add       %2, %2, %6, lsl #2              \n"  // dst_a += 4 *
+                                                      // dst_stride_a
+      "add       %3, %3, %7, lsl #2              \n"  // dst_b += 4 *
+                                                      // dst_stride_b
+      "subs      %w4,  %w4,  #4                  \n"  // w     -= 4
+      "b.eq      4f                              \n"
+
+      // some residual, check to see if it includes a 2x8 block,
+      // or less
+      "cmp       %w4, #2                         \n"
+      "b.lt      3f                              \n"
+
+      // 2x8 block
+      "2:                                        \n"
+      "mov       %0, %1                          \n"
+      "ld2       {v0.h, v1.h}[0], [%0], %5       \n"
+      "ld2       {v2.h, v3.h}[0], [%0], %5       \n"
+      "ld2       {v0.h, v1.h}[1], [%0], %5       \n"
+      "ld2       {v2.h, v3.h}[1], [%0], %5       \n"
+      "ld2       {v0.h, v1.h}[2], [%0], %5       \n"
+      "ld2       {v2.h, v3.h}[2], [%0], %5       \n"
+      "ld2       {v0.h, v1.h}[3], [%0], %5       \n"
+      "ld2       {v2.h, v3.h}[3], [%0]           \n"
+
+      "trn1      v4.8b, v0.8b, v2.8b             \n"
+      "trn2      v5.8b, v0.8b, v2.8b             \n"
+      "trn1      v6.8b, v1.8b, v3.8b             \n"
+      "trn2      v7.8b, v1.8b, v3.8b             \n"
+
+      "mov       %0, %2                          \n"
+
+      "st1       {v4.d}[0], [%0], %6             \n"
+      "st1       {v6.d}[0], [%0]                 \n"
+
+      "mov       %0, %3                          \n"
+
+      "st1       {v5.d}[0], [%0], %7             \n"
+      "st1       {v7.d}[0], [%0]                 \n"
+
+      "add       %1, %1, #4                      \n"  // src   += 2 * 2
+      "add       %2, %2, %6, lsl #1              \n"  // dst_a += 2 *
+                                                      // dst_stride_a
+      "add       %3, %3, %7, lsl #1              \n"  // dst_b += 2 *
+                                                      // dst_stride_b
+      "subs      %w4,  %w4,  #2                  \n"  // w     -= 2
+      "b.eq      4f                              \n"
+
+      // 1x8 block
+      "3:                                        \n"
+      "ld2       {v0.b, v1.b}[0], [%1], %5       \n"
+      "ld2       {v0.b, v1.b}[1], [%1], %5       \n"
+      "ld2       {v0.b, v1.b}[2], [%1], %5       \n"
+      "ld2       {v0.b, v1.b}[3], [%1], %5       \n"
+      "ld2       {v0.b, v1.b}[4], [%1], %5       \n"
+      "ld2       {v0.b, v1.b}[5], [%1], %5       \n"
+      "ld2       {v0.b, v1.b}[6], [%1], %5       \n"
+      "ld2       {v0.b, v1.b}[7], [%1]           \n"
+
+      "st1       {v0.d}[0], [%2]                 \n"
+      "st1       {v1.d}[0], [%3]                 \n"
+
+      "4:                                        \n"
+
+      : "=&r"(src_temp),                            // %0
+        "+r"(src),                                  // %1
+        "+r"(dst_a),                                // %2
+        "+r"(dst_b),                                // %3
+        "+r"(width)                                 // %4
+      : "r"(static_cast<ptrdiff_t>(src_stride)),    // %5
+        "r"(static_cast<ptrdiff_t>(dst_stride_a)),  // %6
+        "r"(static_cast<ptrdiff_t>(dst_stride_b)),  // %7
+        "r"(&kVTbl4x4TransposeDi)                   // %8
+      : "memory", "cc", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16",
+        "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v30", "v31");
+}
+#endif  // !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif

pjproject-2.11.1/third_party/yuv/source/rotate_win.cc

@@ -0,0 +1,252 @@
+/*
+ *  Copyright 2013 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/rotate_row.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for 32 bit Visual C x86 and clangcl
+#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER)
+
+__declspec(naked) void TransposeWx8_SSSE3(const uint8* src,
+                                          int src_stride,
+                                          uint8* dst,
+                                          int dst_stride,
+                                          int width) {
+  __asm {
+    push      edi
+    push      esi
+    push      ebp
+    mov       eax, [esp + 12 + 4]  // src
+    mov       edi, [esp + 12 + 8]  // src_stride
+    mov       edx, [esp + 12 + 12]  // dst
+    mov       esi, [esp + 12 + 16]  // dst_stride
+    mov       ecx, [esp + 12 + 20]  // width
+
+    // Read in the data from the source pointer.
+    // First round of bit swap.
+    align      4
+ convertloop:
+    movq      xmm0, qword ptr [eax]
+    lea       ebp, [eax + 8]
+    movq      xmm1, qword ptr [eax + edi]
+    lea       eax, [eax + 2 * edi]
+    punpcklbw xmm0, xmm1
+    movq      xmm2, qword ptr [eax]
+    movdqa    xmm1, xmm0
+    palignr   xmm1, xmm1, 8
+    movq      xmm3, qword ptr [eax + edi]
+    lea       eax, [eax + 2 * edi]
+    punpcklbw xmm2, xmm3
+    movdqa    xmm3, xmm2
+    movq      xmm4, qword ptr [eax]
+    palignr   xmm3, xmm3, 8
+    movq      xmm5, qword ptr [eax + edi]
+    punpcklbw xmm4, xmm5
+    lea       eax, [eax + 2 * edi]
+    movdqa    xmm5, xmm4
+    movq      xmm6, qword ptr [eax]
+    palignr   xmm5, xmm5, 8
+    movq      xmm7, qword ptr [eax + edi]
+    punpcklbw xmm6, xmm7
+    mov       eax, ebp
+    movdqa    xmm7, xmm6
+    palignr   xmm7, xmm7, 8
+    // Second round of bit swap.
+    punpcklwd xmm0, xmm2
+    punpcklwd xmm1, xmm3
+    movdqa    xmm2, xmm0
+    movdqa    xmm3, xmm1
+    palignr   xmm2, xmm2, 8
+    palignr   xmm3, xmm3, 8
+    punpcklwd xmm4, xmm6
+    punpcklwd xmm5, xmm7
+    movdqa    xmm6, xmm4
+    movdqa    xmm7, xmm5
+    palignr   xmm6, xmm6, 8
+    palignr   xmm7, xmm7, 8
+    // Third round of bit swap.
+    // Write to the destination pointer.
+    punpckldq xmm0, xmm4
+    movq      qword ptr [edx], xmm0
+    movdqa    xmm4, xmm0
+    palignr   xmm4, xmm4, 8
+    movq      qword ptr [edx + esi], xmm4
+    lea       edx, [edx + 2 * esi]
+    punpckldq xmm2, xmm6
+    movdqa    xmm6, xmm2
+    palignr   xmm6, xmm6, 8
+    movq      qword ptr [edx], xmm2
+    punpckldq xmm1, xmm5
+    movq      qword ptr [edx + esi], xmm6
+    lea       edx, [edx + 2 * esi]
+    movdqa    xmm5, xmm1
+    movq      qword ptr [edx], xmm1
+    palignr   xmm5, xmm5, 8
+    punpckldq xmm3, xmm7
+    movq      qword ptr [edx + esi], xmm5
+    lea       edx, [edx + 2 * esi]
+    movq      qword ptr [edx], xmm3
+    movdqa    xmm7, xmm3
+    palignr   xmm7, xmm7, 8
+    sub       ecx, 8
+    movq      qword ptr [edx + esi], xmm7
+    lea       edx, [edx + 2 * esi]
+    jg        convertloop
+
+    pop       ebp
+    pop       esi
+    pop       edi
+    ret
+  }
+}
+
+__declspec(naked) void TransposeUVWx8_SSE2(const uint8* src,
+                                           int src_stride,
+                                           uint8* dst_a,
+                                           int dst_stride_a,
+                                           uint8* dst_b,
+                                           int dst_stride_b,
+                                           int w) {
+  __asm {
+    push      ebx
+    push      esi
+    push      edi
+    push      ebp
+    mov       eax, [esp + 16 + 4]  // src
+    mov       edi, [esp + 16 + 8]  // src_stride
+    mov       edx, [esp + 16 + 12]  // dst_a
+    mov       esi, [esp + 16 + 16]  // dst_stride_a
+    mov       ebx, [esp + 16 + 20]  // dst_b
+    mov       ebp, [esp + 16 + 24]  // dst_stride_b
+    mov       ecx, esp
+    sub       esp, 4 + 16
+    and       esp, ~15
+    mov       [esp + 16], ecx
+    mov       ecx, [ecx + 16 + 28]  // w
+
+    align      4
+    // Read in the data from the source pointer.
+    // First round of bit swap.
+  convertloop:
+    movdqu    xmm0, [eax]
+    movdqu    xmm1, [eax + edi]
+    lea       eax, [eax + 2 * edi]
+    movdqa    xmm7, xmm0  // use xmm7 as temp register.
+    punpcklbw xmm0, xmm1
+    punpckhbw xmm7, xmm1
+    movdqa    xmm1, xmm7
+    movdqu    xmm2, [eax]
+    movdqu    xmm3, [eax + edi]
+    lea       eax, [eax + 2 * edi]
+    movdqa    xmm7, xmm2
+    punpcklbw xmm2, xmm3
+    punpckhbw xmm7, xmm3
+    movdqa    xmm3, xmm7
+    movdqu    xmm4, [eax]
+    movdqu    xmm5, [eax + edi]
+    lea       eax, [eax + 2 * edi]
+    movdqa    xmm7, xmm4
+    punpcklbw xmm4, xmm5
+    punpckhbw xmm7, xmm5
+    movdqa    xmm5, xmm7
+    movdqu    xmm6, [eax]
+    movdqu    xmm7, [eax + edi]
+    lea       eax, [eax + 2 * edi]
+    movdqu    [esp], xmm5  // backup xmm5
+    neg       edi
+    movdqa    xmm5, xmm6  // use xmm5 as temp register.
+    punpcklbw xmm6, xmm7
+    punpckhbw xmm5, xmm7
+    movdqa    xmm7, xmm5
+    lea       eax, [eax + 8 * edi + 16]
+    neg       edi
+        // Second round of bit swap.
+    movdqa    xmm5, xmm0
+    punpcklwd xmm0, xmm2
+    punpckhwd xmm5, xmm2
+    movdqa    xmm2, xmm5
+    movdqa    xmm5, xmm1
+    punpcklwd xmm1, xmm3
+    punpckhwd xmm5, xmm3
+    movdqa    xmm3, xmm5
+    movdqa    xmm5, xmm4
+    punpcklwd xmm4, xmm6
+    punpckhwd xmm5, xmm6
+    movdqa    xmm6, xmm5
+    movdqu    xmm5, [esp]  // restore xmm5
+    movdqu    [esp], xmm6  // backup xmm6
+    movdqa    xmm6, xmm5  // use xmm6 as temp register.
+    punpcklwd xmm5, xmm7
+    punpckhwd xmm6, xmm7
+    movdqa    xmm7, xmm6
+
+        // Third round of bit swap.
+        // Write to the destination pointer.
+    movdqa    xmm6, xmm0
+    punpckldq xmm0, xmm4
+    punpckhdq xmm6, xmm4
+    movdqa    xmm4, xmm6
+    movdqu    xmm6, [esp]  // restore xmm6
+    movlpd    qword ptr [edx], xmm0
+    movhpd    qword ptr [ebx], xmm0
+    movlpd    qword ptr [edx + esi], xmm4
+    lea       edx, [edx + 2 * esi]
+    movhpd    qword ptr [ebx + ebp], xmm4
+    lea       ebx, [ebx + 2 * ebp]
+    movdqa    xmm0, xmm2  // use xmm0 as the temp register.
+    punpckldq xmm2, xmm6
+    movlpd    qword ptr [edx], xmm2
+    movhpd    qword ptr [ebx], xmm2
+    punpckhdq xmm0, xmm6
+    movlpd    qword ptr [edx + esi], xmm0
+    lea       edx, [edx + 2 * esi]
+    movhpd    qword ptr [ebx + ebp], xmm0
+    lea       ebx, [ebx + 2 * ebp]
+    movdqa    xmm0, xmm1  // use xmm0 as the temp register.
+    punpckldq xmm1, xmm5
+    movlpd    qword ptr [edx], xmm1
+    movhpd    qword ptr [ebx], xmm1
+    punpckhdq xmm0, xmm5
+    movlpd    qword ptr [edx + esi], xmm0
+    lea       edx, [edx + 2 * esi]
+    movhpd    qword ptr [ebx + ebp], xmm0
+    lea       ebx, [ebx + 2 * ebp]
+    movdqa    xmm0, xmm3  // use xmm0 as the temp register.
+    punpckldq xmm3, xmm7
+    movlpd    qword ptr [edx], xmm3
+    movhpd    qword ptr [ebx], xmm3
+    punpckhdq xmm0, xmm7
+    sub       ecx, 8
+    movlpd    qword ptr [edx + esi], xmm0
+    lea       edx, [edx + 2 * esi]
+    movhpd    qword ptr [ebx + ebp], xmm0
+    lea       ebx, [ebx + 2 * ebp]
+    jg        convertloop
+
+    mov       esp, [esp + 16]
+    pop       ebp
+    pop       edi
+    pop       esi
+    pop       ebx
+    ret
+  }
+}
+
+#endif  // !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif

pjproject-2.11.1/third_party/yuv/source/scale_any.cc

@@ -0,0 +1,467 @@
+/*
+ *  Copyright 2015 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/scale.h"
+#include "libyuv/scale_row.h"
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Definition for ScaleFilterCols, ScaleARGBCols and ScaleARGBFilterCols
+#define CANY(NAMEANY, TERP_SIMD, TERP_C, BPP, MASK)                        \
+  void NAMEANY(uint8* dst_ptr, const uint8* src_ptr, int dst_width, int x, \
+               int dx) {                                                   \
+    int r = dst_width & MASK;                                              \
+    int n = dst_width & ~MASK;                                             \
+    if (n > 0) {                                                           \
+      TERP_SIMD(dst_ptr, src_ptr, n, x, dx);                               \
+    }                                                                      \
+    TERP_C(dst_ptr + n * BPP, src_ptr, r, x + n * dx, dx);                 \
+  }
+
+#ifdef HAS_SCALEFILTERCOLS_NEON
+CANY(ScaleFilterCols_Any_NEON, ScaleFilterCols_NEON, ScaleFilterCols_C, 1, 7)
+#endif
+#ifdef HAS_SCALEFILTERCOLS_MSA
+CANY(ScaleFilterCols_Any_MSA, ScaleFilterCols_MSA, ScaleFilterCols_C, 1, 15)
+#endif
+#ifdef HAS_SCALEARGBCOLS_NEON
+CANY(ScaleARGBCols_Any_NEON, ScaleARGBCols_NEON, ScaleARGBCols_C, 4, 7)
+#endif
+#ifdef HAS_SCALEARGBCOLS_MSA
+CANY(ScaleARGBCols_Any_MSA, ScaleARGBCols_MSA, ScaleARGBCols_C, 4, 3)
+#endif
+#ifdef HAS_SCALEARGBFILTERCOLS_NEON
+CANY(ScaleARGBFilterCols_Any_NEON,
+     ScaleARGBFilterCols_NEON,
+     ScaleARGBFilterCols_C,
+     4,
+     3)
+#endif
+#ifdef HAS_SCALEARGBFILTERCOLS_MSA
+CANY(ScaleARGBFilterCols_Any_MSA,
+     ScaleARGBFilterCols_MSA,
+     ScaleARGBFilterCols_C,
+     4,
+     7)
+#endif
+#undef CANY
+
+// Fixed scale down.
+// Mask may be non-power of 2, so use MOD
+#define SDANY(NAMEANY, SCALEROWDOWN_SIMD, SCALEROWDOWN_C, FACTOR, BPP, MASK) \
+  void NAMEANY(const uint8* src_ptr, ptrdiff_t src_stride, uint8* dst_ptr,   \
+               int dst_width) {                                              \
+    int r = (int)((unsigned int)dst_width % (MASK + 1)); /* NOLINT */        \
+    int n = dst_width - r;                                                   \
+    if (n > 0) {                                                             \
+      SCALEROWDOWN_SIMD(src_ptr, src_stride, dst_ptr, n);                    \
+    }                                                                        \
+    SCALEROWDOWN_C(src_ptr + (n * FACTOR) * BPP, src_stride,                 \
+                   dst_ptr + n * BPP, r);                                    \
+  }
+
+// Fixed scale down for odd source width.  Used by I420Blend subsampling.
+// Since dst_width is (width + 1) / 2, this function scales one less pixel
+// and copies the last pixel.
+#define SDODD(NAMEANY, SCALEROWDOWN_SIMD, SCALEROWDOWN_C, FACTOR, BPP, MASK) \
+  void NAMEANY(const uint8* src_ptr, ptrdiff_t src_stride, uint8* dst_ptr,   \
+               int dst_width) {                                              \
+    int r = (int)((unsigned int)(dst_width - 1) % (MASK + 1)); /* NOLINT */  \
+    int n = (dst_width - 1) - r;                                             \
+    if (n > 0) {                                                             \
+      SCALEROWDOWN_SIMD(src_ptr, src_stride, dst_ptr, n);                    \
+    }                                                                        \
+    SCALEROWDOWN_C(src_ptr + (n * FACTOR) * BPP, src_stride,                 \
+                   dst_ptr + n * BPP, r + 1);                                \
+  }
+
+#ifdef HAS_SCALEROWDOWN2_SSSE3
+SDANY(ScaleRowDown2_Any_SSSE3, ScaleRowDown2_SSSE3, ScaleRowDown2_C, 2, 1, 15)
+SDANY(ScaleRowDown2Linear_Any_SSSE3,
+      ScaleRowDown2Linear_SSSE3,
+      ScaleRowDown2Linear_C,
+      2,
+      1,
+      15)
+SDANY(ScaleRowDown2Box_Any_SSSE3,
+      ScaleRowDown2Box_SSSE3,
+      ScaleRowDown2Box_C,
+      2,
+      1,
+      15)
+SDODD(ScaleRowDown2Box_Odd_SSSE3,
+      ScaleRowDown2Box_SSSE3,
+      ScaleRowDown2Box_Odd_C,
+      2,
+      1,
+      15)
+#endif
+#ifdef HAS_SCALEROWDOWN2_AVX2
+SDANY(ScaleRowDown2_Any_AVX2, ScaleRowDown2_AVX2, ScaleRowDown2_C, 2, 1, 31)
+SDANY(ScaleRowDown2Linear_Any_AVX2,
+      ScaleRowDown2Linear_AVX2,
+      ScaleRowDown2Linear_C,
+      2,
+      1,
+      31)
+SDANY(ScaleRowDown2Box_Any_AVX2,
+      ScaleRowDown2Box_AVX2,
+      ScaleRowDown2Box_C,
+      2,
+      1,
+      31)
+SDODD(ScaleRowDown2Box_Odd_AVX2,
+      ScaleRowDown2Box_AVX2,
+      ScaleRowDown2Box_Odd_C,
+      2,
+      1,
+      31)
+#endif
+#ifdef HAS_SCALEROWDOWN2_NEON
+SDANY(ScaleRowDown2_Any_NEON, ScaleRowDown2_NEON, ScaleRowDown2_C, 2, 1, 15)
+SDANY(ScaleRowDown2Linear_Any_NEON,
+      ScaleRowDown2Linear_NEON,
+      ScaleRowDown2Linear_C,
+      2,
+      1,
+      15)
+SDANY(ScaleRowDown2Box_Any_NEON,
+      ScaleRowDown2Box_NEON,
+      ScaleRowDown2Box_C,
+      2,
+      1,
+      15)
+SDODD(ScaleRowDown2Box_Odd_NEON,
+      ScaleRowDown2Box_NEON,
+      ScaleRowDown2Box_Odd_C,
+      2,
+      1,
+      15)
+#endif
+#ifdef HAS_SCALEROWDOWN2_MSA
+SDANY(ScaleRowDown2_Any_MSA, ScaleRowDown2_MSA, ScaleRowDown2_C, 2, 1, 31)
+SDANY(ScaleRowDown2Linear_Any_MSA,
+      ScaleRowDown2Linear_MSA,
+      ScaleRowDown2Linear_C,
+      2,
+      1,
+      31)
+SDANY(ScaleRowDown2Box_Any_MSA,
+      ScaleRowDown2Box_MSA,
+      ScaleRowDown2Box_C,
+      2,
+      1,
+      31)
+#endif
+#ifdef HAS_SCALEROWDOWN4_SSSE3
+SDANY(ScaleRowDown4_Any_SSSE3, ScaleRowDown4_SSSE3, ScaleRowDown4_C, 4, 1, 7)
+SDANY(ScaleRowDown4Box_Any_SSSE3,
+      ScaleRowDown4Box_SSSE3,
+      ScaleRowDown4Box_C,
+      4,
+      1,
+      7)
+#endif
+#ifdef HAS_SCALEROWDOWN4_AVX2
+SDANY(ScaleRowDown4_Any_AVX2, ScaleRowDown4_AVX2, ScaleRowDown4_C, 4, 1, 15)
+SDANY(ScaleRowDown4Box_Any_AVX2,
+      ScaleRowDown4Box_AVX2,
+      ScaleRowDown4Box_C,
+      4,
+      1,
+      15)
+#endif
+#ifdef HAS_SCALEROWDOWN4_NEON
+SDANY(ScaleRowDown4_Any_NEON, ScaleRowDown4_NEON, ScaleRowDown4_C, 4, 1, 7)
+SDANY(ScaleRowDown4Box_Any_NEON,
+      ScaleRowDown4Box_NEON,
+      ScaleRowDown4Box_C,
+      4,
+      1,
+      7)
+#endif
+#ifdef HAS_SCALEROWDOWN4_MSA
+SDANY(ScaleRowDown4_Any_MSA, ScaleRowDown4_MSA, ScaleRowDown4_C, 4, 1, 15)
+SDANY(ScaleRowDown4Box_Any_MSA,
+      ScaleRowDown4Box_MSA,
+      ScaleRowDown4Box_C,
+      4,
+      1,
+      15)
+#endif
+#ifdef HAS_SCALEROWDOWN34_SSSE3
+SDANY(ScaleRowDown34_Any_SSSE3,
+      ScaleRowDown34_SSSE3,
+      ScaleRowDown34_C,
+      4 / 3,
+      1,
+      23)
+SDANY(ScaleRowDown34_0_Box_Any_SSSE3,
+      ScaleRowDown34_0_Box_SSSE3,
+      ScaleRowDown34_0_Box_C,
+      4 / 3,
+      1,
+      23)
+SDANY(ScaleRowDown34_1_Box_Any_SSSE3,
+      ScaleRowDown34_1_Box_SSSE3,
+      ScaleRowDown34_1_Box_C,
+      4 / 3,
+      1,
+      23)
+#endif
+#ifdef HAS_SCALEROWDOWN34_NEON
+SDANY(ScaleRowDown34_Any_NEON,
+      ScaleRowDown34_NEON,
+      ScaleRowDown34_C,
+      4 / 3,
+      1,
+      23)
+SDANY(ScaleRowDown34_0_Box_Any_NEON,
+      ScaleRowDown34_0_Box_NEON,
+      ScaleRowDown34_0_Box_C,
+      4 / 3,
+      1,
+      23)
+SDANY(ScaleRowDown34_1_Box_Any_NEON,
+      ScaleRowDown34_1_Box_NEON,
+      ScaleRowDown34_1_Box_C,
+      4 / 3,
+      1,
+      23)
+#endif
+#ifdef HAS_SCALEROWDOWN34_MSA
+SDANY(ScaleRowDown34_Any_MSA,
+      ScaleRowDown34_MSA,
+      ScaleRowDown34_C,
+      4 / 3,
+      1,
+      47)
+SDANY(ScaleRowDown34_0_Box_Any_MSA,
+      ScaleRowDown34_0_Box_MSA,
+      ScaleRowDown34_0_Box_C,
+      4 / 3,
+      1,
+      47)
+SDANY(ScaleRowDown34_1_Box_Any_MSA,
+      ScaleRowDown34_1_Box_MSA,
+      ScaleRowDown34_1_Box_C,
+      4 / 3,
+      1,
+      47)
+#endif
+#ifdef HAS_SCALEROWDOWN38_SSSE3
+SDANY(ScaleRowDown38_Any_SSSE3,
+      ScaleRowDown38_SSSE3,
+      ScaleRowDown38_C,
+      8 / 3,
+      1,
+      11)
+SDANY(ScaleRowDown38_3_Box_Any_SSSE3,
+      ScaleRowDown38_3_Box_SSSE3,
+      ScaleRowDown38_3_Box_C,
+      8 / 3,
+      1,
+      5)
+SDANY(ScaleRowDown38_2_Box_Any_SSSE3,
+      ScaleRowDown38_2_Box_SSSE3,
+      ScaleRowDown38_2_Box_C,
+      8 / 3,
+      1,
+      5)
+#endif
+#ifdef HAS_SCALEROWDOWN38_NEON
+SDANY(ScaleRowDown38_Any_NEON,
+      ScaleRowDown38_NEON,
+      ScaleRowDown38_C,
+      8 / 3,
+      1,
+      11)
+SDANY(ScaleRowDown38_3_Box_Any_NEON,
+      ScaleRowDown38_3_Box_NEON,
+      ScaleRowDown38_3_Box_C,
+      8 / 3,
+      1,
+      11)
+SDANY(ScaleRowDown38_2_Box_Any_NEON,
+      ScaleRowDown38_2_Box_NEON,
+      ScaleRowDown38_2_Box_C,
+      8 / 3,
+      1,
+      11)
+#endif
+#ifdef HAS_SCALEROWDOWN38_MSA
+SDANY(ScaleRowDown38_Any_MSA,
+      ScaleRowDown38_MSA,
+      ScaleRowDown38_C,
+      8 / 3,
+      1,
+      11)
+SDANY(ScaleRowDown38_3_Box_Any_MSA,
+      ScaleRowDown38_3_Box_MSA,
+      ScaleRowDown38_3_Box_C,
+      8 / 3,
+      1,
+      11)
+SDANY(ScaleRowDown38_2_Box_Any_MSA,
+      ScaleRowDown38_2_Box_MSA,
+      ScaleRowDown38_2_Box_C,
+      8 / 3,
+      1,
+      11)
+#endif
+
+#ifdef HAS_SCALEARGBROWDOWN2_SSE2
+SDANY(ScaleARGBRowDown2_Any_SSE2,
+      ScaleARGBRowDown2_SSE2,
+      ScaleARGBRowDown2_C,
+      2,
+      4,
+      3)
+SDANY(ScaleARGBRowDown2Linear_Any_SSE2,
+      ScaleARGBRowDown2Linear_SSE2,
+      ScaleARGBRowDown2Linear_C,
+      2,
+      4,
+      3)
+SDANY(ScaleARGBRowDown2Box_Any_SSE2,
+      ScaleARGBRowDown2Box_SSE2,
+      ScaleARGBRowDown2Box_C,
+      2,
+      4,
+      3)
+#endif
+#ifdef HAS_SCALEARGBROWDOWN2_NEON
+SDANY(ScaleARGBRowDown2_Any_NEON,
+      ScaleARGBRowDown2_NEON,
+      ScaleARGBRowDown2_C,
+      2,
+      4,
+      7)
+SDANY(ScaleARGBRowDown2Linear_Any_NEON,
+      ScaleARGBRowDown2Linear_NEON,
+      ScaleARGBRowDown2Linear_C,
+      2,
+      4,
+      7)
+SDANY(ScaleARGBRowDown2Box_Any_NEON,
+      ScaleARGBRowDown2Box_NEON,
+      ScaleARGBRowDown2Box_C,
+      2,
+      4,
+      7)
+#endif
+#ifdef HAS_SCALEARGBROWDOWN2_MSA
+SDANY(ScaleARGBRowDown2_Any_MSA,
+      ScaleARGBRowDown2_MSA,
+      ScaleARGBRowDown2_C,
+      2,
+      4,
+      3)
+SDANY(ScaleARGBRowDown2Linear_Any_MSA,
+      ScaleARGBRowDown2Linear_MSA,
+      ScaleARGBRowDown2Linear_C,
+      2,
+      4,
+      3)
+SDANY(ScaleARGBRowDown2Box_Any_MSA,
+      ScaleARGBRowDown2Box_MSA,
+      ScaleARGBRowDown2Box_C,
+      2,
+      4,
+      3)
+#endif
+#undef SDANY
+
+// Scale down by even scale factor.
+#define SDAANY(NAMEANY, SCALEROWDOWN_SIMD, SCALEROWDOWN_C, BPP, MASK)      \
+  void NAMEANY(const uint8* src_ptr, ptrdiff_t src_stride, int src_stepx,  \
+               uint8* dst_ptr, int dst_width) {                            \
+    int r = dst_width & MASK;                                              \
+    int n = dst_width & ~MASK;                                             \
+    if (n > 0) {                                                           \
+      SCALEROWDOWN_SIMD(src_ptr, src_stride, src_stepx, dst_ptr, n);       \
+    }                                                                      \
+    SCALEROWDOWN_C(src_ptr + (n * src_stepx) * BPP, src_stride, src_stepx, \
+                   dst_ptr + n * BPP, r);                                  \
+  }
+
+#ifdef HAS_SCALEARGBROWDOWNEVEN_SSE2
+SDAANY(ScaleARGBRowDownEven_Any_SSE2,
+       ScaleARGBRowDownEven_SSE2,
+       ScaleARGBRowDownEven_C,
+       4,
+       3)
+SDAANY(ScaleARGBRowDownEvenBox_Any_SSE2,
+       ScaleARGBRowDownEvenBox_SSE2,
+       ScaleARGBRowDownEvenBox_C,
+       4,
+       3)
+#endif
+#ifdef HAS_SCALEARGBROWDOWNEVEN_NEON
+SDAANY(ScaleARGBRowDownEven_Any_NEON,
+       ScaleARGBRowDownEven_NEON,
+       ScaleARGBRowDownEven_C,
+       4,
+       3)
+SDAANY(ScaleARGBRowDownEvenBox_Any_NEON,
+       ScaleARGBRowDownEvenBox_NEON,
+       ScaleARGBRowDownEvenBox_C,
+       4,
+       3)
+#endif
+#ifdef HAS_SCALEARGBROWDOWNEVEN_MSA
+SDAANY(ScaleARGBRowDownEven_Any_MSA,
+       ScaleARGBRowDownEven_MSA,
+       ScaleARGBRowDownEven_C,
+       4,
+       3)
+SDAANY(ScaleARGBRowDownEvenBox_Any_MSA,
+       ScaleARGBRowDownEvenBox_MSA,
+       ScaleARGBRowDownEvenBox_C,
+       4,
+       3)
+#endif
+
+// Add rows box filter scale down.
+#define SAANY(NAMEANY, SCALEADDROW_SIMD, SCALEADDROW_C, MASK)          \
+  void NAMEANY(const uint8* src_ptr, uint16* dst_ptr, int src_width) { \
+    int n = src_width & ~MASK;                                         \
+    if (n > 0) {                                                       \
+      SCALEADDROW_SIMD(src_ptr, dst_ptr, n);                           \
+    }                                                                  \
+    SCALEADDROW_C(src_ptr + n, dst_ptr + n, src_width & MASK);         \
+  }
+
+#ifdef HAS_SCALEADDROW_SSE2
+SAANY(ScaleAddRow_Any_SSE2, ScaleAddRow_SSE2, ScaleAddRow_C, 15)
+#endif
+#ifdef HAS_SCALEADDROW_AVX2
+SAANY(ScaleAddRow_Any_AVX2, ScaleAddRow_AVX2, ScaleAddRow_C, 31)
+#endif
+#ifdef HAS_SCALEADDROW_NEON
+SAANY(ScaleAddRow_Any_NEON, ScaleAddRow_NEON, ScaleAddRow_C, 15)
+#endif
+#ifdef HAS_SCALEADDROW_MSA
+SAANY(ScaleAddRow_Any_MSA, ScaleAddRow_MSA, ScaleAddRow_C, 15)
+#endif
+#ifdef HAS_SCALEADDROW_DSPR2
+SAANY(ScaleAddRow_Any_DSPR2, ScaleAddRow_DSPR2, ScaleAddRow_C, 15)
+#endif
+#undef SAANY
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif

pjproject-2.11.1/third_party/yuv/source/scale_dspr2.cc

@@ -0,0 +1,668 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/basic_types.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC MIPS DSPR2
+#if !defined(LIBYUV_DISABLE_DSPR2) && defined(__mips_dsp) && \
+    (__mips_dsp_rev >= 2) && (_MIPS_SIM == _MIPS_SIM_ABI32)
+
+void ScaleRowDown2_DSPR2(const uint8* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint8* dst,
+                         int dst_width) {
+  __asm__ __volatile__(
+      ".set push                                     \n"
+      ".set noreorder                                \n"
+
+      "srl            $t9, %[dst_width], 4           \n"  // iterations -> by 16
+      "beqz           $t9, 2f                        \n"
+      " nop                                          \n"
+
+      "1:                                            \n"
+      "lw             $t0, 0(%[src_ptr])             \n"  // |3|2|1|0|
+      "lw             $t1, 4(%[src_ptr])             \n"  // |7|6|5|4|
+      "lw             $t2, 8(%[src_ptr])             \n"  // |11|10|9|8|
+      "lw             $t3, 12(%[src_ptr])            \n"  // |15|14|13|12|
+      "lw             $t4, 16(%[src_ptr])            \n"  // |19|18|17|16|
+      "lw             $t5, 20(%[src_ptr])            \n"  // |23|22|21|20|
+      "lw             $t6, 24(%[src_ptr])            \n"  // |27|26|25|24|
+      "lw             $t7, 28(%[src_ptr])            \n"  // |31|30|29|28|
+      // TODO(fbarchard): Use odd pixels instead of even.
+      "precrq.qb.ph   $t8, $t1, $t0                  \n"  // |7|5|3|1|
+      "precrq.qb.ph   $t0, $t3, $t2                  \n"  // |15|13|11|9|
+      "precrq.qb.ph   $t1, $t5, $t4                  \n"  // |23|21|19|17|
+      "precrq.qb.ph   $t2, $t7, $t6                  \n"  // |31|29|27|25|
+      "addiu          %[src_ptr], %[src_ptr], 32     \n"
+      "addiu          $t9, $t9, -1                   \n"
+      "sw             $t8, 0(%[dst])                 \n"
+      "sw             $t0, 4(%[dst])                 \n"
+      "sw             $t1, 8(%[dst])                 \n"
+      "sw             $t2, 12(%[dst])                \n"
+      "bgtz           $t9, 1b                        \n"
+      " addiu         %[dst], %[dst], 16             \n"
+
+      "2:                                            \n"
+      "andi           $t9, %[dst_width], 0xf         \n"  // residue
+      "beqz           $t9, 3f                        \n"
+      " nop                                          \n"
+
+      "21:                                           \n"
+      "lbu            $t0, 1(%[src_ptr])             \n"
+      "addiu          %[src_ptr], %[src_ptr], 2      \n"
+      "addiu          $t9, $t9, -1                   \n"
+      "sb             $t0, 0(%[dst])                 \n"
+      "bgtz           $t9, 21b                       \n"
+      " addiu         %[dst], %[dst], 1              \n"
+
+      "3:                                            \n"
+      ".set pop                                      \n"
+      : [src_ptr] "+r"(src_ptr), [dst] "+r"(dst)
+      : [dst_width] "r"(dst_width)
+      : "t0", "t1", "t2", "t3", "t4", "t5", "t6", "t7", "t8", "t9");
+}
+
+void ScaleRowDown2Box_DSPR2(const uint8* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8* dst,
+                            int dst_width) {
+  const uint8* t = src_ptr + src_stride;
+
+  __asm__ __volatile__(
+      ".set push                                    \n"
+      ".set noreorder                               \n"
+
+      "srl            $t9, %[dst_width], 3          \n"  // iterations -> step 8
+      "bltz           $t9, 2f                       \n"
+      " nop                                         \n"
+
+      "1:                                           \n"
+      "lw             $t0, 0(%[src_ptr])            \n"  // |3|2|1|0|
+      "lw             $t1, 4(%[src_ptr])            \n"  // |7|6|5|4|
+      "lw             $t2, 8(%[src_ptr])            \n"  // |11|10|9|8|
+      "lw             $t3, 12(%[src_ptr])           \n"  // |15|14|13|12|
+      "lw             $t4, 0(%[t])                  \n"  // |19|18|17|16|
+      "lw             $t5, 4(%[t])                  \n"  // |23|22|21|20|
+      "lw             $t6, 8(%[t])                  \n"  // |27|26|25|24|
+      "lw             $t7, 12(%[t])                 \n"  // |31|30|29|28|
+      "addiu          $t9, $t9, -1                  \n"
+      "srl            $t8, $t0, 16                  \n"  // |X|X|3|2|
+      "ins            $t0, $t4, 16, 16              \n"  // |17|16|1|0|
+      "ins            $t4, $t8, 0, 16               \n"  // |19|18|3|2|
+      "raddu.w.qb     $t0, $t0                      \n"  // |17+16+1+0|
+      "raddu.w.qb     $t4, $t4                      \n"  // |19+18+3+2|
+      "shra_r.w       $t0, $t0, 2                   \n"  // |t0+2|>>2
+      "shra_r.w       $t4, $t4, 2                   \n"  // |t4+2|>>2
+      "srl            $t8, $t1, 16                  \n"  // |X|X|7|6|
+      "ins            $t1, $t5, 16, 16              \n"  // |21|20|5|4|
+      "ins            $t5, $t8, 0, 16               \n"  // |22|23|7|6|
+      "raddu.w.qb     $t1, $t1                      \n"  // |21+20+5+4|
+      "raddu.w.qb     $t5, $t5                      \n"  // |23+22+7+6|
+      "shra_r.w       $t1, $t1, 2                   \n"  // |t1+2|>>2
+      "shra_r.w       $t5, $t5, 2                   \n"  // |t5+2|>>2
+      "srl            $t8, $t2, 16                  \n"  // |X|X|11|10|
+      "ins            $t2, $t6, 16, 16              \n"  // |25|24|9|8|
+      "ins            $t6, $t8, 0, 16               \n"  // |27|26|11|10|
+      "raddu.w.qb     $t2, $t2                      \n"  // |25+24+9+8|
+      "raddu.w.qb     $t6, $t6                      \n"  // |27+26+11+10|
+      "shra_r.w       $t2, $t2, 2                   \n"  // |t2+2|>>2
+      "shra_r.w       $t6, $t6, 2                   \n"  // |t5+2|>>2
+      "srl            $t8, $t3, 16                  \n"  // |X|X|15|14|
+      "ins            $t3, $t7, 16, 16              \n"  // |29|28|13|12|
+      "ins            $t7, $t8, 0, 16               \n"  // |31|30|15|14|
+      "raddu.w.qb     $t3, $t3                      \n"  // |29+28+13+12|
+      "raddu.w.qb     $t7, $t7                      \n"  // |31+30+15+14|
+      "shra_r.w       $t3, $t3, 2                   \n"  // |t3+2|>>2
+      "shra_r.w       $t7, $t7, 2                   \n"  // |t7+2|>>2
+      "addiu          %[src_ptr], %[src_ptr], 16    \n"
+      "addiu          %[t], %[t], 16                \n"
+      "sb             $t0, 0(%[dst])                \n"
+      "sb             $t4, 1(%[dst])                \n"
+      "sb             $t1, 2(%[dst])                \n"
+      "sb             $t5, 3(%[dst])                \n"
+      "sb             $t2, 4(%[dst])                \n"
+      "sb             $t6, 5(%[dst])                \n"
+      "sb             $t3, 6(%[dst])                \n"
+      "sb             $t7, 7(%[dst])                \n"
+      "bgtz           $t9, 1b                       \n"
+      " addiu         %[dst], %[dst], 8             \n"
+
+      "2:                                           \n"
+      "andi           $t9, %[dst_width], 0x7        \n"  // x = residue
+      "beqz           $t9, 3f                       \n"
+      " nop                                         \n"
+
+      "21:                                          \n"
+      "lwr            $t1, 0(%[src_ptr])            \n"
+      "lwl            $t1, 3(%[src_ptr])            \n"
+      "lwr            $t2, 0(%[t])                  \n"
+      "lwl            $t2, 3(%[t])                  \n"
+      "srl            $t8, $t1, 16                  \n"
+      "ins            $t1, $t2, 16, 16              \n"
+      "ins            $t2, $t8, 0, 16               \n"
+      "raddu.w.qb     $t1, $t1                      \n"
+      "raddu.w.qb     $t2, $t2                      \n"
+      "shra_r.w       $t1, $t1, 2                   \n"
+      "shra_r.w       $t2, $t2, 2                   \n"
+      "sb             $t1, 0(%[dst])                \n"
+      "sb             $t2, 1(%[dst])                \n"
+      "addiu          %[src_ptr], %[src_ptr], 4     \n"
+      "addiu          $t9, $t9, -2                  \n"
+      "addiu          %[t], %[t], 4                 \n"
+      "bgtz           $t9, 21b                      \n"
+      " addiu         %[dst], %[dst], 2             \n"
+
+      "3:                                           \n"
+      ".set pop                                     \n"
+
+      : [src_ptr] "+r"(src_ptr), [dst] "+r"(dst), [t] "+r"(t)
+      : [dst_width] "r"(dst_width)
+      : "t0", "t1", "t2", "t3", "t4", "t5", "t6", "t7", "t8", "t9");
+}
+
+void ScaleRowDown4_DSPR2(const uint8* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint8* dst,
+                         int dst_width) {
+  __asm__ __volatile__(
+      ".set push                                    \n"
+      ".set noreorder                               \n"
+
+      "srl            $t9, %[dst_width], 3          \n"
+      "beqz           $t9, 2f                       \n"
+      " nop                                         \n"
+
+      "1:                                           \n"
+      "lw             $t1, 0(%[src_ptr])            \n"  // |3|2|1|0|
+      "lw             $t2, 4(%[src_ptr])            \n"  // |7|6|5|4|
+      "lw             $t3, 8(%[src_ptr])            \n"  // |11|10|9|8|
+      "lw             $t4, 12(%[src_ptr])           \n"  // |15|14|13|12|
+      "lw             $t5, 16(%[src_ptr])           \n"  // |19|18|17|16|
+      "lw             $t6, 20(%[src_ptr])           \n"  // |23|22|21|20|
+      "lw             $t7, 24(%[src_ptr])           \n"  // |27|26|25|24|
+      "lw             $t8, 28(%[src_ptr])           \n"  // |31|30|29|28|
+      "precr.qb.ph    $t1, $t2, $t1                 \n"  // |6|4|2|0|
+      "precr.qb.ph    $t2, $t4, $t3                 \n"  // |14|12|10|8|
+      "precr.qb.ph    $t5, $t6, $t5                 \n"  // |22|20|18|16|
+      "precr.qb.ph    $t6, $t8, $t7                 \n"  // |30|28|26|24|
+      "precrq.qb.ph   $t1, $t2, $t1                 \n"  // |14|10|6|2|
+      "precrq.qb.ph   $t5, $t6, $t5                 \n"  // |30|26|22|18|
+      "addiu          %[src_ptr], %[src_ptr], 32    \n"
+      "addiu          $t9, $t9, -1                  \n"
+      "sw             $t1, 0(%[dst])                \n"
+      "sw             $t5, 4(%[dst])                \n"
+      "bgtz           $t9, 1b                       \n"
+      " addiu         %[dst], %[dst], 8             \n"
+
+      "2:                                           \n"
+      "andi           $t9, %[dst_width], 7          \n"  // residue
+      "beqz           $t9, 3f                       \n"
+      " nop                                         \n"
+
+      "21:                                          \n"
+      "lbu            $t1, 2(%[src_ptr])            \n"
+      "addiu          %[src_ptr], %[src_ptr], 4     \n"
+      "addiu          $t9, $t9, -1                  \n"
+      "sb             $t1, 0(%[dst])                \n"
+      "bgtz           $t9, 21b                      \n"
+      " addiu         %[dst], %[dst], 1             \n"
+
+      "3:                                           \n"
+      ".set pop                                     \n"
+      : [src_ptr] "+r"(src_ptr), [dst] "+r"(dst)
+      : [dst_width] "r"(dst_width)
+      : "t1", "t2", "t3", "t4", "t5", "t6", "t7", "t8", "t9");
+}
+
+void ScaleRowDown4Box_DSPR2(const uint8* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8* dst,
+                            int dst_width) {
+  intptr_t stride = src_stride;
+  const uint8* s1 = src_ptr + stride;
+  const uint8* s2 = s1 + stride;
+  const uint8* s3 = s2 + stride;
+
+  __asm__ __volatile__(
+      ".set push                                  \n"
+      ".set noreorder                             \n"
+
+      "srl           $t9, %[dst_width], 1         \n"
+      "andi          $t8, %[dst_width], 1         \n"
+
+      "1:                                         \n"
+      "lw            $t0, 0(%[src_ptr])           \n"  // |3|2|1|0|
+      "lw            $t1, 0(%[s1])                \n"  // |7|6|5|4|
+      "lw            $t2, 0(%[s2])                \n"  // |11|10|9|8|
+      "lw            $t3, 0(%[s3])                \n"  // |15|14|13|12|
+      "lw            $t4, 4(%[src_ptr])           \n"  // |19|18|17|16|
+      "lw            $t5, 4(%[s1])                \n"  // |23|22|21|20|
+      "lw            $t6, 4(%[s2])                \n"  // |27|26|25|24|
+      "lw            $t7, 4(%[s3])                \n"  // |31|30|29|28|
+      "raddu.w.qb    $t0, $t0                     \n"  // |3 + 2 + 1 + 0|
+      "raddu.w.qb    $t1, $t1                     \n"  // |7 + 6 + 5 + 4|
+      "raddu.w.qb    $t2, $t2                     \n"  // |11 + 10 + 9 + 8|
+      "raddu.w.qb    $t3, $t3                     \n"  // |15 + 14 + 13 + 12|
+      "raddu.w.qb    $t4, $t4                     \n"  // |19 + 18 + 17 + 16|
+      "raddu.w.qb    $t5, $t5                     \n"  // |23 + 22 + 21 + 20|
+      "raddu.w.qb    $t6, $t6                     \n"  // |27 + 26 + 25 + 24|
+      "raddu.w.qb    $t7, $t7                     \n"  // |31 + 30 + 29 + 28|
+      "add           $t0, $t0, $t1                \n"
+      "add           $t1, $t2, $t3                \n"
+      "add           $t0, $t0, $t1                \n"
+      "add           $t4, $t4, $t5                \n"
+      "add           $t6, $t6, $t7                \n"
+      "add           $t4, $t4, $t6                \n"
+      "shra_r.w      $t0, $t0, 4                  \n"
+      "shra_r.w      $t4, $t4, 4                  \n"
+      "sb            $t0, 0(%[dst])               \n"
+      "sb            $t4, 1(%[dst])               \n"
+      "addiu         %[src_ptr], %[src_ptr], 8    \n"
+      "addiu         %[s1], %[s1], 8              \n"
+      "addiu         %[s2], %[s2], 8              \n"
+      "addiu         %[s3], %[s3], 8              \n"
+      "addiu         $t9, $t9, -1                 \n"
+      "bgtz          $t9, 1b                      \n"
+      " addiu        %[dst], %[dst], 2            \n"
+      "beqz          $t8, 2f                      \n"
+      " nop                                       \n"
+
+      "lw            $t0, 0(%[src_ptr])           \n"  // |3|2|1|0|
+      "lw            $t1, 0(%[s1])                \n"  // |7|6|5|4|
+      "lw            $t2, 0(%[s2])                \n"  // |11|10|9|8|
+      "lw            $t3, 0(%[s3])                \n"  // |15|14|13|12|
+      "raddu.w.qb    $t0, $t0                     \n"  // |3 + 2 + 1 + 0|
+      "raddu.w.qb    $t1, $t1                     \n"  // |7 + 6 + 5 + 4|
+      "raddu.w.qb    $t2, $t2                     \n"  // |11 + 10 + 9 + 8|
+      "raddu.w.qb    $t3, $t3                     \n"  // |15 + 14 + 13 + 12|
+      "add           $t0, $t0, $t1                \n"
+      "add           $t1, $t2, $t3                \n"
+      "add           $t0, $t0, $t1                \n"
+      "shra_r.w      $t0, $t0, 4                  \n"
+      "sb            $t0, 0(%[dst])               \n"
+
+      "2:                                         \n"
+      ".set pop                                   \n"
+
+      : [src_ptr] "+r"(src_ptr), [dst] "+r"(dst), [s1] "+r"(s1), [s2] "+r"(s2),
+        [s3] "+r"(s3)
+      : [dst_width] "r"(dst_width)
+      : "t0", "t1", "t2", "t3", "t4", "t5", "t6", "t7", "t8", "t9");
+}
+
+void ScaleRowDown34_DSPR2(const uint8* src_ptr,
+                          ptrdiff_t src_stride,
+                          uint8* dst,
+                          int dst_width) {
+  __asm__ __volatile__(
+      ".set push                                          \n"
+      ".set noreorder                                     \n"
+      "1:                                                 \n"
+      "lw              $t1, 0(%[src_ptr])                 \n"  // |3|2|1|0|
+      "lw              $t2, 4(%[src_ptr])                 \n"  // |7|6|5|4|
+      "lw              $t3, 8(%[src_ptr])                 \n"  // |11|10|9|8|
+      "lw              $t4, 12(%[src_ptr])                \n"  // |15|14|13|12|
+      "lw              $t5, 16(%[src_ptr])                \n"  // |19|18|17|16|
+      "lw              $t6, 20(%[src_ptr])                \n"  // |23|22|21|20|
+      "lw              $t7, 24(%[src_ptr])                \n"  // |27|26|25|24|
+      "lw              $t8, 28(%[src_ptr])                \n"  // |31|30|29|28|
+      "precrq.qb.ph    $t0, $t2, $t4                      \n"  // |7|5|15|13|
+      "precrq.qb.ph    $t9, $t6, $t8                      \n"  // |23|21|31|30|
+      "addiu           %[dst_width], %[dst_width], -24    \n"
+      "ins             $t1, $t1, 8, 16                    \n"  // |3|1|0|X|
+      "ins             $t4, $t0, 8, 16                    \n"  // |X|15|13|12|
+      "ins             $t5, $t5, 8, 16                    \n"  // |19|17|16|X|
+      "ins             $t8, $t9, 8, 16                    \n"  // |X|31|29|28|
+      "addiu           %[src_ptr], %[src_ptr], 32         \n"
+      "packrl.ph       $t0, $t3, $t0                      \n"  // |9|8|7|5|
+      "packrl.ph       $t9, $t7, $t9                      \n"  // |25|24|23|21|
+      "prepend         $t1, $t2, 8                        \n"  // |4|3|1|0|
+      "prepend         $t3, $t4, 24                       \n"  // |15|13|12|11|
+      "prepend         $t5, $t6, 8                        \n"  // |20|19|17|16|
+      "prepend         $t7, $t8, 24                       \n"  // |31|29|28|27|
+      "sw              $t1, 0(%[dst])                     \n"
+      "sw              $t0, 4(%[dst])                     \n"
+      "sw              $t3, 8(%[dst])                     \n"
+      "sw              $t5, 12(%[dst])                    \n"
+      "sw              $t9, 16(%[dst])                    \n"
+      "sw              $t7, 20(%[dst])                    \n"
+      "bnez            %[dst_width], 1b                   \n"
+      " addiu          %[dst], %[dst], 24                 \n"
+      ".set pop                                           \n"
+      : [src_ptr] "+r"(src_ptr), [dst] "+r"(dst), [dst_width] "+r"(dst_width)
+      :
+      : "t0", "t1", "t2", "t3", "t4", "t5", "t6", "t7", "t8", "t9");
+}
+
+void ScaleRowDown34_0_Box_DSPR2(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* d,
+                                int dst_width) {
+  __asm__ __volatile__(
+      ".set push                                         \n"
+      ".set noreorder                                    \n"
+      "repl.ph           $t3, 3                          \n"  // 0x00030003
+
+      "1:                                                \n"
+      "lw                $t0, 0(%[src_ptr])              \n"  // |S3|S2|S1|S0|
+      "lwx               $t1, %[src_stride](%[src_ptr])  \n"  // |T3|T2|T1|T0|
+      "rotr              $t2, $t0, 8                     \n"  // |S0|S3|S2|S1|
+      "rotr              $t6, $t1, 8                     \n"  // |T0|T3|T2|T1|
+      "muleu_s.ph.qbl    $t4, $t2, $t3                   \n"  // |S0*3|S3*3|
+      "muleu_s.ph.qbl    $t5, $t6, $t3                   \n"  // |T0*3|T3*3|
+      "andi              $t0, $t2, 0xFFFF                \n"  // |0|0|S2|S1|
+      "andi              $t1, $t6, 0xFFFF                \n"  // |0|0|T2|T1|
+      "raddu.w.qb        $t0, $t0                        \n"
+      "raddu.w.qb        $t1, $t1                        \n"
+      "shra_r.w          $t0, $t0, 1                     \n"
+      "shra_r.w          $t1, $t1, 1                     \n"
+      "preceu.ph.qbr     $t2, $t2                        \n"  // |0|S2|0|S1|
+      "preceu.ph.qbr     $t6, $t6                        \n"  // |0|T2|0|T1|
+      "rotr              $t2, $t2, 16                    \n"  // |0|S1|0|S2|
+      "rotr              $t6, $t6, 16                    \n"  // |0|T1|0|T2|
+      "addu.ph           $t2, $t2, $t4                   \n"
+      "addu.ph           $t6, $t6, $t5                   \n"
+      "sll               $t5, $t0, 1                     \n"
+      "add               $t0, $t5, $t0                   \n"
+      "shra_r.ph         $t2, $t2, 2                     \n"
+      "shra_r.ph         $t6, $t6, 2                     \n"
+      "shll.ph           $t4, $t2, 1                     \n"
+      "addq.ph           $t4, $t4, $t2                   \n"
+      "addu              $t0, $t0, $t1                   \n"
+      "addiu             %[src_ptr], %[src_ptr], 4       \n"
+      "shra_r.w          $t0, $t0, 2                     \n"
+      "addu.ph           $t6, $t6, $t4                   \n"
+      "shra_r.ph         $t6, $t6, 2                     \n"
+      "srl               $t1, $t6, 16                    \n"
+      "addiu             %[dst_width], %[dst_width], -3  \n"
+      "sb                $t1, 0(%[d])                    \n"
+      "sb                $t0, 1(%[d])                    \n"
+      "sb                $t6, 2(%[d])                    \n"
+      "bgtz              %[dst_width], 1b                \n"
+      " addiu            %[d], %[d], 3                   \n"
+      "3:                                                \n"
+      ".set pop                                          \n"
+      : [src_ptr] "+r"(src_ptr), [src_stride] "+r"(src_stride), [d] "+r"(d),
+        [dst_width] "+r"(dst_width)
+      :
+      : "t0", "t1", "t2", "t3", "t4", "t5", "t6");
+}
+
+void ScaleRowDown34_1_Box_DSPR2(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* d,
+                                int dst_width) {
+  __asm__ __volatile__(
+      ".set push                                           \n"
+      ".set noreorder                                      \n"
+      "repl.ph           $t2, 3                            \n"  // 0x00030003
+
+      "1:                                                  \n"
+      "lw                $t0, 0(%[src_ptr])                \n"  // |S3|S2|S1|S0|
+      "lwx               $t1, %[src_stride](%[src_ptr])    \n"  // |T3|T2|T1|T0|
+      "rotr              $t4, $t0, 8                       \n"  // |S0|S3|S2|S1|
+      "rotr              $t6, $t1, 8                       \n"  // |T0|T3|T2|T1|
+      "muleu_s.ph.qbl    $t3, $t4, $t2                     \n"  // |S0*3|S3*3|
+      "muleu_s.ph.qbl    $t5, $t6, $t2                     \n"  // |T0*3|T3*3|
+      "andi              $t0, $t4, 0xFFFF                  \n"  // |0|0|S2|S1|
+      "andi              $t1, $t6, 0xFFFF                  \n"  // |0|0|T2|T1|
+      "raddu.w.qb        $t0, $t0                          \n"
+      "raddu.w.qb        $t1, $t1                          \n"
+      "shra_r.w          $t0, $t0, 1                       \n"
+      "shra_r.w          $t1, $t1, 1                       \n"
+      "preceu.ph.qbr     $t4, $t4                          \n"  // |0|S2|0|S1|
+      "preceu.ph.qbr     $t6, $t6                          \n"  // |0|T2|0|T1|
+      "rotr              $t4, $t4, 16                      \n"  // |0|S1|0|S2|
+      "rotr              $t6, $t6, 16                      \n"  // |0|T1|0|T2|
+      "addu.ph           $t4, $t4, $t3                     \n"
+      "addu.ph           $t6, $t6, $t5                     \n"
+      "shra_r.ph         $t6, $t6, 2                       \n"
+      "shra_r.ph         $t4, $t4, 2                       \n"
+      "addu.ph           $t6, $t6, $t4                     \n"
+      "addiu             %[src_ptr], %[src_ptr], 4         \n"
+      "shra_r.ph         $t6, $t6, 1                       \n"
+      "addu              $t0, $t0, $t1                     \n"
+      "addiu             %[dst_width], %[dst_width], -3    \n"
+      "shra_r.w          $t0, $t0, 1                       \n"
+      "srl               $t1, $t6, 16                      \n"
+      "sb                $t1, 0(%[d])                      \n"
+      "sb                $t0, 1(%[d])                      \n"
+      "sb                $t6, 2(%[d])                      \n"
+      "bgtz              %[dst_width], 1b                  \n"
+      " addiu            %[d], %[d], 3                     \n"
+      "3:                                                  \n"
+      ".set pop                                            \n"
+      : [src_ptr] "+r"(src_ptr), [src_stride] "+r"(src_stride), [d] "+r"(d),
+        [dst_width] "+r"(dst_width)
+      :
+      : "t0", "t1", "t2", "t3", "t4", "t5", "t6");
+}
+
+void ScaleRowDown38_DSPR2(const uint8* src_ptr,
+                          ptrdiff_t src_stride,
+                          uint8* dst,
+                          int dst_width) {
+  __asm__ __volatile__(
+      ".set push                                     \n"
+      ".set noreorder                                \n"
+
+      "1:                                            \n"
+      "lw         $t0, 0(%[src_ptr])                 \n"  // |3|2|1|0|
+      "lw         $t1, 4(%[src_ptr])                 \n"  // |7|6|5|4|
+      "lw         $t2, 8(%[src_ptr])                 \n"  // |11|10|9|8|
+      "lw         $t3, 12(%[src_ptr])                \n"  // |15|14|13|12|
+      "lw         $t4, 16(%[src_ptr])                \n"  // |19|18|17|16|
+      "lw         $t5, 20(%[src_ptr])                \n"  // |23|22|21|20|
+      "lw         $t6, 24(%[src_ptr])                \n"  // |27|26|25|24|
+      "lw         $t7, 28(%[src_ptr])                \n"  // |31|30|29|28|
+      "wsbh       $t0, $t0                           \n"  // |2|3|0|1|
+      "wsbh       $t6, $t6                           \n"  // |26|27|24|25|
+      "srl        $t0, $t0, 8                        \n"  // |X|2|3|0|
+      "srl        $t3, $t3, 16                       \n"  // |X|X|15|14|
+      "srl        $t5, $t5, 16                       \n"  // |X|X|23|22|
+      "srl        $t7, $t7, 16                       \n"  // |X|X|31|30|
+      "ins        $t1, $t2, 24, 8                    \n"  // |8|6|5|4|
+      "ins        $t6, $t5, 0, 8                     \n"  // |26|27|24|22|
+      "ins        $t1, $t0, 0, 16                    \n"  // |8|6|3|0|
+      "ins        $t6, $t7, 24, 8                    \n"  // |30|27|24|22|
+      "prepend    $t2, $t3, 24                       \n"  // |X|15|14|11|
+      "ins        $t4, $t4, 16, 8                    \n"  // |19|16|17|X|
+      "ins        $t4, $t2, 0, 16                    \n"  // |19|16|14|11|
+      "addiu      %[src_ptr], %[src_ptr], 32         \n"
+      "addiu      %[dst_width], %[dst_width], -12    \n"
+      "addiu      $t8,%[dst_width], -12              \n"
+      "sw         $t1, 0(%[dst])                     \n"
+      "sw         $t4, 4(%[dst])                     \n"
+      "sw         $t6, 8(%[dst])                     \n"
+      "bgez       $t8, 1b                            \n"
+      " addiu     %[dst], %[dst], 12                 \n"
+      ".set pop                                      \n"
+      : [src_ptr] "+r"(src_ptr), [dst] "+r"(dst), [dst_width] "+r"(dst_width)
+      :
+      : "t0", "t1", "t2", "t3", "t4", "t5", "t6", "t7", "t8");
+}
+
+void ScaleRowDown38_2_Box_DSPR2(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr,
+                                int dst_width) {
+  intptr_t stride = src_stride;
+  const uint8* t = src_ptr + stride;
+  const int c = 0x2AAA;
+
+  __asm__ __volatile__(
+      ".set push                                         \n"
+      ".set noreorder                                    \n"
+
+      "1:                                                \n"
+      "lw              $t0, 0(%[src_ptr])                \n"  // |S3|S2|S1|S0|
+      "lw              $t1, 4(%[src_ptr])                \n"  // |S7|S6|S5|S4|
+      "lw              $t2, 0(%[t])                      \n"  // |T3|T2|T1|T0|
+      "lw              $t3, 4(%[t])                      \n"  // |T7|T6|T5|T4|
+      "rotr            $t1, $t1, 16                      \n"  // |S5|S4|S7|S6|
+      "packrl.ph       $t4, $t1, $t3                     \n"  // |S7|S6|T7|T6|
+      "packrl.ph       $t5, $t3, $t1                     \n"  // |T5|T4|S5|S4|
+      "raddu.w.qb      $t4, $t4                          \n"  // S7+S6+T7+T6
+      "raddu.w.qb      $t5, $t5                          \n"  // T5+T4+S5+S4
+      "precrq.qb.ph    $t6, $t0, $t2                     \n"  // |S3|S1|T3|T1|
+      "precrq.qb.ph    $t6, $t6, $t6                     \n"  // |S3|T3|S3|T3|
+      "srl             $t4, $t4, 2                       \n"  // t4 / 4
+      "srl             $t6, $t6, 16                      \n"  // |0|0|S3|T3|
+      "raddu.w.qb      $t6, $t6                          \n"  // 0+0+S3+T3
+      "addu            $t6, $t5, $t6                     \n"
+      "mul             $t6, $t6, %[c]                    \n"  // t6 * 0x2AAA
+      "sll             $t0, $t0, 8                       \n"  // |S2|S1|S0|0|
+      "sll             $t2, $t2, 8                       \n"  // |T2|T1|T0|0|
+      "raddu.w.qb      $t0, $t0                          \n"  // S2+S1+S0+0
+      "raddu.w.qb      $t2, $t2                          \n"  // T2+T1+T0+0
+      "addu            $t0, $t0, $t2                     \n"
+      "mul             $t0, $t0, %[c]                    \n"  // t0 * 0x2AAA
+      "addiu           %[src_ptr], %[src_ptr], 8         \n"
+      "addiu           %[t], %[t], 8                     \n"
+      "addiu           %[dst_width], %[dst_width], -3    \n"
+      "addiu           %[dst_ptr], %[dst_ptr], 3         \n"
+      "srl             $t6, $t6, 16                      \n"
+      "srl             $t0, $t0, 16                      \n"
+      "sb              $t4, -1(%[dst_ptr])               \n"
+      "sb              $t6, -2(%[dst_ptr])               \n"
+      "bgtz            %[dst_width], 1b                  \n"
+      " sb             $t0, -3(%[dst_ptr])               \n"
+      ".set pop                                          \n"
+      : [src_ptr] "+r"(src_ptr), [dst_ptr] "+r"(dst_ptr), [t] "+r"(t),
+        [dst_width] "+r"(dst_width)
+      : [c] "r"(c)
+      : "t0", "t1", "t2", "t3", "t4", "t5", "t6");
+}
+
+void ScaleRowDown38_3_Box_DSPR2(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr,
+                                int dst_width) {
+  intptr_t stride = src_stride;
+  const uint8* s1 = src_ptr + stride;
+  stride += stride;
+  const uint8* s2 = src_ptr + stride;
+  const int c1 = 0x1C71;
+  const int c2 = 0x2AAA;
+
+  __asm__ __volatile__(
+      ".set push                                         \n"
+      ".set noreorder                                    \n"
+
+      "1:                                                \n"
+      "lw              $t0, 0(%[src_ptr])                \n"  // |S3|S2|S1|S0|
+      "lw              $t1, 4(%[src_ptr])                \n"  // |S7|S6|S5|S4|
+      "lw              $t2, 0(%[s1])                     \n"  // |T3|T2|T1|T0|
+      "lw              $t3, 4(%[s1])                     \n"  // |T7|T6|T5|T4|
+      "lw              $t4, 0(%[s2])                     \n"  // |R3|R2|R1|R0|
+      "lw              $t5, 4(%[s2])                     \n"  // |R7|R6|R5|R4|
+      "rotr            $t1, $t1, 16                      \n"  // |S5|S4|S7|S6|
+      "packrl.ph       $t6, $t1, $t3                     \n"  // |S7|S6|T7|T6|
+      "raddu.w.qb      $t6, $t6                          \n"  // S7+S6+T7+T6
+      "packrl.ph       $t7, $t3, $t1                     \n"  // |T5|T4|S5|S4|
+      "raddu.w.qb      $t7, $t7                          \n"  // T5+T4+S5+S4
+      "sll             $t8, $t5, 16                      \n"  // |R5|R4|0|0|
+      "raddu.w.qb      $t8, $t8                          \n"  // R5+R4
+      "addu            $t7, $t7, $t8                     \n"
+      "srl             $t8, $t5, 16                      \n"  // |0|0|R7|R6|
+      "raddu.w.qb      $t8, $t8                          \n"  // R7 + R6
+      "addu            $t6, $t6, $t8                     \n"
+      "mul             $t6, $t6, %[c2]                   \n"  // t6 * 0x2AAA
+      "precrq.qb.ph    $t8, $t0, $t2                     \n"  // |S3|S1|T3|T1|
+      "precrq.qb.ph    $t8, $t8, $t4                     \n"  // |S3|T3|R3|R1|
+      "srl             $t8, $t8, 8                       \n"  // |0|S3|T3|R3|
+      "raddu.w.qb      $t8, $t8                          \n"  // S3 + T3 + R3
+      "addu            $t7, $t7, $t8                     \n"
+      "mul             $t7, $t7, %[c1]                   \n"  // t7 * 0x1C71
+      "sll             $t0, $t0, 8                       \n"  // |S2|S1|S0|0|
+      "sll             $t2, $t2, 8                       \n"  // |T2|T1|T0|0|
+      "sll             $t4, $t4, 8                       \n"  // |R2|R1|R0|0|
+      "raddu.w.qb      $t0, $t0                          \n"
+      "raddu.w.qb      $t2, $t2                          \n"
+      "raddu.w.qb      $t4, $t4                          \n"
+      "addu            $t0, $t0, $t2                     \n"
+      "addu            $t0, $t0, $t4                     \n"
+      "mul             $t0, $t0, %[c1]                   \n"  // t0 * 0x1C71
+      "addiu           %[src_ptr], %[src_ptr], 8         \n"
+      "addiu           %[s1], %[s1], 8                   \n"
+      "addiu           %[s2], %[s2], 8                   \n"
+      "addiu           %[dst_width], %[dst_width], -3    \n"
+      "addiu           %[dst_ptr], %[dst_ptr], 3         \n"
+      "srl             $t6, $t6, 16                      \n"
+      "srl             $t7, $t7, 16                      \n"
+      "srl             $t0, $t0, 16                      \n"
+      "sb              $t6, -1(%[dst_ptr])               \n"
+      "sb              $t7, -2(%[dst_ptr])               \n"
+      "bgtz            %[dst_width], 1b                  \n"
+      " sb             $t0, -3(%[dst_ptr])               \n"
+      ".set pop                                          \n"
+      : [src_ptr] "+r"(src_ptr), [dst_ptr] "+r"(dst_ptr), [s1] "+r"(s1),
+        [s2] "+r"(s2), [dst_width] "+r"(dst_width)
+      : [c1] "r"(c1), [c2] "r"(c2)
+      : "t0", "t1", "t2", "t3", "t4", "t5", "t6", "t7", "t8");
+}
+
+void ScaleAddRow_DSPR2(const uint8* src_ptr, uint16* dst_ptr, int src_width) {
+  int x;
+  for (x = 0; x < ((src_width - 1)); x += 8) {
+    uint32 tmp_t1, tmp_t2, tmp_t3, tmp_t4;
+    uint32 tmp_t5, tmp_t6, tmp_t7, tmp_t8;
+    __asm__ __volatile__(
+        ".set push                                                \n"
+        ".set noreorder                                           \n"
+        "lw                %[tmp_t5],   0(%[src_ptr])             \n"
+        "lw                %[tmp_t6],   4(%[src_ptr])             \n"
+        "lw                %[tmp_t1],   0(%[dst_ptr])             \n"
+        "lw                %[tmp_t2],   4(%[dst_ptr])             \n"
+        "lw                %[tmp_t3],   8(%[dst_ptr])             \n"
+        "lw                %[tmp_t4],   12(%[dst_ptr])            \n"
+        "preceu.ph.qbr     %[tmp_t7],   %[tmp_t5]                 \n"
+        "preceu.ph.qbl     %[tmp_t8],   %[tmp_t5]                 \n"
+        "addu.ph           %[tmp_t1],   %[tmp_t1],     %[tmp_t7]  \n"
+        "addu.ph           %[tmp_t2],   %[tmp_t2],     %[tmp_t8]  \n"
+        "preceu.ph.qbr     %[tmp_t7],   %[tmp_t6]                 \n"
+        "preceu.ph.qbl     %[tmp_t8],   %[tmp_t6]                 \n"
+        "addu.ph           %[tmp_t3],   %[tmp_t3],     %[tmp_t7]  \n"
+        "addu.ph           %[tmp_t4],   %[tmp_t4],     %[tmp_t8]  \n"
+        "sw                %[tmp_t1],   0(%[dst_ptr])             \n"
+        "sw                %[tmp_t2],   4(%[dst_ptr])             \n"
+        "sw                %[tmp_t3],   8(%[dst_ptr])             \n"
+        "sw                %[tmp_t4],   12(%[dst_ptr])            \n"
+        ".set pop                                                 \n"
+        :
+        [tmp_t1] "=&r"(tmp_t1), [tmp_t2] "=&r"(tmp_t2), [tmp_t3] "=&r"(tmp_t3),
+        [tmp_t4] "=&r"(tmp_t4), [tmp_t5] "=&r"(tmp_t5), [tmp_t6] "=&r"(tmp_t6),
+        [tmp_t7] "=&r"(tmp_t7), [tmp_t8] "=&r"(tmp_t8), [src_ptr] "+r"(src_ptr)
+        : [dst_ptr] "r"(dst_ptr));
+    src_ptr += 8;
+    dst_ptr += 8;
+  }
+
+  if ((src_width)&7) {
+    for (x = 0; x < ((src_width - 1) & 7); x += 1) {
+      dst_ptr[0] += src_ptr[0];
+      src_ptr += 1;
+      dst_ptr += 1;
+    }
+  }
+}
+
+#endif  // defined(__mips_dsp) && (__mips_dsp_rev >= 2)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif

pjproject-2.11.1/third_party/yuv/source/scale_msa.cc

@@ -0,0 +1,949 @@
+/*
+ *  Copyright 2016 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+
+#include "libyuv/scale_row.h"
+
+// This module is for GCC MSA
+#if !defined(LIBYUV_DISABLE_MSA) && defined(__mips_msa)
+#include "libyuv/macros_msa.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#define LOAD_INDEXED_DATA(srcp, indx0, out0) \
+  {                                          \
+    out0[0] = srcp[indx0[0]];                \
+    out0[1] = srcp[indx0[1]];                \
+    out0[2] = srcp[indx0[2]];                \
+    out0[3] = srcp[indx0[3]];                \
+  }
+
+void ScaleARGBRowDown2_MSA(const uint8_t* src_argb,
+                           ptrdiff_t src_stride,
+                           uint8_t* dst_argb,
+                           int dst_width) {
+  int x;
+  v16u8 src0, src1, dst0;
+  (void)src_stride;
+
+  for (x = 0; x < dst_width; x += 4) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)src_argb, 16);
+    dst0 = (v16u8)__msa_pckod_w((v4i32)src1, (v4i32)src0);
+    ST_UB(dst0, dst_argb);
+    src_argb += 32;
+    dst_argb += 16;
+  }
+}
+
+void ScaleARGBRowDown2Linear_MSA(const uint8_t* src_argb,
+                                 ptrdiff_t src_stride,
+                                 uint8_t* dst_argb,
+                                 int dst_width) {
+  int x;
+  v16u8 src0, src1, vec0, vec1, dst0;
+  (void)src_stride;
+
+  for (x = 0; x < dst_width; x += 4) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)src_argb, 16);
+    vec0 = (v16u8)__msa_pckev_w((v4i32)src1, (v4i32)src0);
+    vec1 = (v16u8)__msa_pckod_w((v4i32)src1, (v4i32)src0);
+    dst0 = (v16u8)__msa_aver_u_b((v16u8)vec0, (v16u8)vec1);
+    ST_UB(dst0, dst_argb);
+    src_argb += 32;
+    dst_argb += 16;
+  }
+}
+
+void ScaleARGBRowDown2Box_MSA(const uint8_t* src_argb,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_argb,
+                              int dst_width) {
+  int x;
+  const uint8_t* s = src_argb;
+  const uint8_t* t = src_argb + src_stride;
+  v16u8 src0, src1, src2, src3, vec0, vec1, vec2, vec3, dst0;
+  v8u16 reg0, reg1, reg2, reg3;
+  v16i8 shuffler = {0, 4, 1, 5, 2, 6, 3, 7, 8, 12, 9, 13, 10, 14, 11, 15};
+
+  for (x = 0; x < dst_width; x += 4) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)s, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)t, 0);
+    src3 = (v16u8)__msa_ld_b((v16i8*)t, 16);
+    vec0 = (v16u8)__msa_vshf_b(shuffler, (v16i8)src0, (v16i8)src0);
+    vec1 = (v16u8)__msa_vshf_b(shuffler, (v16i8)src1, (v16i8)src1);
+    vec2 = (v16u8)__msa_vshf_b(shuffler, (v16i8)src2, (v16i8)src2);
+    vec3 = (v16u8)__msa_vshf_b(shuffler, (v16i8)src3, (v16i8)src3);
+    reg0 = __msa_hadd_u_h(vec0, vec0);
+    reg1 = __msa_hadd_u_h(vec1, vec1);
+    reg2 = __msa_hadd_u_h(vec2, vec2);
+    reg3 = __msa_hadd_u_h(vec3, vec3);
+    reg0 += reg2;
+    reg1 += reg3;
+    reg0 = (v8u16)__msa_srari_h((v8i16)reg0, 2);
+    reg1 = (v8u16)__msa_srari_h((v8i16)reg1, 2);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)reg1, (v16i8)reg0);
+    ST_UB(dst0, dst_argb);
+    s += 32;
+    t += 32;
+    dst_argb += 16;
+  }
+}
+
+void ScaleARGBRowDownEven_MSA(const uint8_t* src_argb,
+                              ptrdiff_t src_stride,
+                              int32_t src_stepx,
+                              uint8_t* dst_argb,
+                              int dst_width) {
+  int x;
+  int32_t stepx = src_stepx * 4;
+  int32_t data0, data1, data2, data3;
+  (void)src_stride;
+
+  for (x = 0; x < dst_width; x += 4) {
+    data0 = LW(src_argb);
+    data1 = LW(src_argb + stepx);
+    data2 = LW(src_argb + stepx * 2);
+    data3 = LW(src_argb + stepx * 3);
+    SW(data0, dst_argb);
+    SW(data1, dst_argb + 4);
+    SW(data2, dst_argb + 8);
+    SW(data3, dst_argb + 12);
+    src_argb += stepx * 4;
+    dst_argb += 16;
+  }
+}
+
+void ScaleARGBRowDownEvenBox_MSA(const uint8* src_argb,
+                                 ptrdiff_t src_stride,
+                                 int src_stepx,
+                                 uint8* dst_argb,
+                                 int dst_width) {
+  int x;
+  const uint8* nxt_argb = src_argb + src_stride;
+  int32_t stepx = src_stepx * 4;
+  int64_t data0, data1, data2, data3;
+  v16u8 src0 = {0}, src1 = {0}, src2 = {0}, src3 = {0};
+  v16u8 vec0, vec1, vec2, vec3;
+  v8u16 reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;
+  v16u8 dst0;
+
+  for (x = 0; x < dst_width; x += 4) {
+    data0 = LD(src_argb);
+    data1 = LD(src_argb + stepx);
+    data2 = LD(src_argb + stepx * 2);
+    data3 = LD(src_argb + stepx * 3);
+    src0 = (v16u8)__msa_insert_d((v2i64)src0, 0, data0);
+    src0 = (v16u8)__msa_insert_d((v2i64)src0, 1, data1);
+    src1 = (v16u8)__msa_insert_d((v2i64)src1, 0, data2);
+    src1 = (v16u8)__msa_insert_d((v2i64)src1, 1, data3);
+    data0 = LD(nxt_argb);
+    data1 = LD(nxt_argb + stepx);
+    data2 = LD(nxt_argb + stepx * 2);
+    data3 = LD(nxt_argb + stepx * 3);
+    src2 = (v16u8)__msa_insert_d((v2i64)src2, 0, data0);
+    src2 = (v16u8)__msa_insert_d((v2i64)src2, 1, data1);
+    src3 = (v16u8)__msa_insert_d((v2i64)src3, 0, data2);
+    src3 = (v16u8)__msa_insert_d((v2i64)src3, 1, data3);
+    vec0 = (v16u8)__msa_ilvr_b((v16i8)src2, (v16i8)src0);
+    vec1 = (v16u8)__msa_ilvr_b((v16i8)src3, (v16i8)src1);
+    vec2 = (v16u8)__msa_ilvl_b((v16i8)src2, (v16i8)src0);
+    vec3 = (v16u8)__msa_ilvl_b((v16i8)src3, (v16i8)src1);
+    reg0 = __msa_hadd_u_h(vec0, vec0);
+    reg1 = __msa_hadd_u_h(vec1, vec1);
+    reg2 = __msa_hadd_u_h(vec2, vec2);
+    reg3 = __msa_hadd_u_h(vec3, vec3);
+    reg4 = (v8u16)__msa_pckev_d((v2i64)reg2, (v2i64)reg0);
+    reg5 = (v8u16)__msa_pckev_d((v2i64)reg3, (v2i64)reg1);
+    reg6 = (v8u16)__msa_pckod_d((v2i64)reg2, (v2i64)reg0);
+    reg7 = (v8u16)__msa_pckod_d((v2i64)reg3, (v2i64)reg1);
+    reg4 += reg6;
+    reg5 += reg7;
+    reg4 = (v8u16)__msa_srari_h((v8i16)reg4, 2);
+    reg5 = (v8u16)__msa_srari_h((v8i16)reg5, 2);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)reg5, (v16i8)reg4);
+    ST_UB(dst0, dst_argb);
+    src_argb += stepx * 4;
+    nxt_argb += stepx * 4;
+    dst_argb += 16;
+  }
+}
+
+void ScaleRowDown2_MSA(const uint8_t* src_ptr,
+                       ptrdiff_t src_stride,
+                       uint8_t* dst,
+                       int dst_width) {
+  int x;
+  v16u8 src0, src1, src2, src3, dst0, dst1;
+  (void)src_stride;
+
+  for (x = 0; x < dst_width; x += 32) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 32);
+    src3 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 48);
+    dst0 = (v16u8)__msa_pckod_b((v16i8)src1, (v16i8)src0);
+    dst1 = (v16u8)__msa_pckod_b((v16i8)src3, (v16i8)src2);
+    ST_UB2(dst0, dst1, dst, 16);
+    src_ptr += 64;
+    dst += 32;
+  }
+}
+
+void ScaleRowDown2Linear_MSA(const uint8_t* src_ptr,
+                             ptrdiff_t src_stride,
+                             uint8_t* dst,
+                             int dst_width) {
+  int x;
+  v16u8 src0, src1, src2, src3, vec0, vec1, vec2, vec3, dst0, dst1;
+  (void)src_stride;
+
+  for (x = 0; x < dst_width; x += 32) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 32);
+    src3 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 48);
+    vec0 = (v16u8)__msa_pckev_b((v16i8)src1, (v16i8)src0);
+    vec2 = (v16u8)__msa_pckev_b((v16i8)src3, (v16i8)src2);
+    vec1 = (v16u8)__msa_pckod_b((v16i8)src1, (v16i8)src0);
+    vec3 = (v16u8)__msa_pckod_b((v16i8)src3, (v16i8)src2);
+    dst0 = __msa_aver_u_b(vec1, vec0);
+    dst1 = __msa_aver_u_b(vec3, vec2);
+    ST_UB2(dst0, dst1, dst, 16);
+    src_ptr += 64;
+    dst += 32;
+  }
+}
+
+void ScaleRowDown2Box_MSA(const uint8_t* src_ptr,
+                          ptrdiff_t src_stride,
+                          uint8_t* dst,
+                          int dst_width) {
+  int x;
+  const uint8_t* s = src_ptr;
+  const uint8_t* t = src_ptr + src_stride;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7, dst0, dst1;
+  v8u16 vec0, vec1, vec2, vec3;
+
+  for (x = 0; x < dst_width; x += 32) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)s, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)s, 32);
+    src3 = (v16u8)__msa_ld_b((v16i8*)s, 48);
+    src4 = (v16u8)__msa_ld_b((v16i8*)t, 0);
+    src5 = (v16u8)__msa_ld_b((v16i8*)t, 16);
+    src6 = (v16u8)__msa_ld_b((v16i8*)t, 32);
+    src7 = (v16u8)__msa_ld_b((v16i8*)t, 48);
+    vec0 = __msa_hadd_u_h(src0, src0);
+    vec1 = __msa_hadd_u_h(src1, src1);
+    vec2 = __msa_hadd_u_h(src2, src2);
+    vec3 = __msa_hadd_u_h(src3, src3);
+    vec0 += __msa_hadd_u_h(src4, src4);
+    vec1 += __msa_hadd_u_h(src5, src5);
+    vec2 += __msa_hadd_u_h(src6, src6);
+    vec3 += __msa_hadd_u_h(src7, src7);
+    vec0 = (v8u16)__msa_srari_h((v8i16)vec0, 2);
+    vec1 = (v8u16)__msa_srari_h((v8i16)vec1, 2);
+    vec2 = (v8u16)__msa_srari_h((v8i16)vec2, 2);
+    vec3 = (v8u16)__msa_srari_h((v8i16)vec3, 2);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0);
+    dst1 = (v16u8)__msa_pckev_b((v16i8)vec3, (v16i8)vec2);
+    ST_UB2(dst0, dst1, dst, 16);
+    s += 64;
+    t += 64;
+    dst += 32;
+  }
+}
+
+void ScaleRowDown4_MSA(const uint8_t* src_ptr,
+                       ptrdiff_t src_stride,
+                       uint8_t* dst,
+                       int dst_width) {
+  int x;
+  v16u8 src0, src1, src2, src3, vec0, vec1, dst0;
+  (void)src_stride;
+
+  for (x = 0; x < dst_width; x += 16) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 32);
+    src3 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 48);
+    vec0 = (v16u8)__msa_pckev_b((v16i8)src1, (v16i8)src0);
+    vec1 = (v16u8)__msa_pckev_b((v16i8)src3, (v16i8)src2);
+    dst0 = (v16u8)__msa_pckod_b((v16i8)vec1, (v16i8)vec0);
+    ST_UB(dst0, dst);
+    src_ptr += 64;
+    dst += 16;
+  }
+}
+
+void ScaleRowDown4Box_MSA(const uint8_t* src_ptr,
+                          ptrdiff_t src_stride,
+                          uint8_t* dst,
+                          int dst_width) {
+  int x;
+  const uint8_t* s = src_ptr;
+  const uint8_t* t0 = s + src_stride;
+  const uint8_t* t1 = s + src_stride * 2;
+  const uint8_t* t2 = s + src_stride * 3;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7, dst0;
+  v8u16 vec0, vec1, vec2, vec3;
+  v4u32 reg0, reg1, reg2, reg3;
+
+  for (x = 0; x < dst_width; x += 16) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)s, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)s, 32);
+    src3 = (v16u8)__msa_ld_b((v16i8*)s, 48);
+    src4 = (v16u8)__msa_ld_b((v16i8*)t0, 0);
+    src5 = (v16u8)__msa_ld_b((v16i8*)t0, 16);
+    src6 = (v16u8)__msa_ld_b((v16i8*)t0, 32);
+    src7 = (v16u8)__msa_ld_b((v16i8*)t0, 48);
+    vec0 = __msa_hadd_u_h(src0, src0);
+    vec1 = __msa_hadd_u_h(src1, src1);
+    vec2 = __msa_hadd_u_h(src2, src2);
+    vec3 = __msa_hadd_u_h(src3, src3);
+    vec0 += __msa_hadd_u_h(src4, src4);
+    vec1 += __msa_hadd_u_h(src5, src5);
+    vec2 += __msa_hadd_u_h(src6, src6);
+    vec3 += __msa_hadd_u_h(src7, src7);
+    src0 = (v16u8)__msa_ld_b((v16i8*)t1, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)t1, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)t1, 32);
+    src3 = (v16u8)__msa_ld_b((v16i8*)t1, 48);
+    src4 = (v16u8)__msa_ld_b((v16i8*)t2, 0);
+    src5 = (v16u8)__msa_ld_b((v16i8*)t2, 16);
+    src6 = (v16u8)__msa_ld_b((v16i8*)t2, 32);
+    src7 = (v16u8)__msa_ld_b((v16i8*)t2, 48);
+    vec0 += __msa_hadd_u_h(src0, src0);
+    vec1 += __msa_hadd_u_h(src1, src1);
+    vec2 += __msa_hadd_u_h(src2, src2);
+    vec3 += __msa_hadd_u_h(src3, src3);
+    vec0 += __msa_hadd_u_h(src4, src4);
+    vec1 += __msa_hadd_u_h(src5, src5);
+    vec2 += __msa_hadd_u_h(src6, src6);
+    vec3 += __msa_hadd_u_h(src7, src7);
+    reg0 = __msa_hadd_u_w(vec0, vec0);
+    reg1 = __msa_hadd_u_w(vec1, vec1);
+    reg2 = __msa_hadd_u_w(vec2, vec2);
+    reg3 = __msa_hadd_u_w(vec3, vec3);
+    reg0 = (v4u32)__msa_srari_w((v4i32)reg0, 4);
+    reg1 = (v4u32)__msa_srari_w((v4i32)reg1, 4);
+    reg2 = (v4u32)__msa_srari_w((v4i32)reg2, 4);
+    reg3 = (v4u32)__msa_srari_w((v4i32)reg3, 4);
+    vec0 = (v8u16)__msa_pckev_h((v8i16)reg1, (v8i16)reg0);
+    vec1 = (v8u16)__msa_pckev_h((v8i16)reg3, (v8i16)reg2);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0);
+    ST_UB(dst0, dst);
+    s += 64;
+    t0 += 64;
+    t1 += 64;
+    t2 += 64;
+    dst += 16;
+  }
+}
+
+void ScaleRowDown38_MSA(const uint8_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint8_t* dst,
+                        int dst_width) {
+  int x, width;
+  uint64_t dst0;
+  uint32_t dst1;
+  v16u8 src0, src1, vec0;
+  v16i8 mask = {0, 3, 6, 8, 11, 14, 16, 19, 22, 24, 27, 30, 0, 0, 0, 0};
+  (void)src_stride;
+
+  assert(dst_width % 3 == 0);
+  width = dst_width / 3;
+
+  for (x = 0; x < width; x += 4) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 16);
+    vec0 = (v16u8)__msa_vshf_b(mask, (v16i8)src1, (v16i8)src0);
+    dst0 = __msa_copy_u_d((v2i64)vec0, 0);
+    dst1 = __msa_copy_u_w((v4i32)vec0, 2);
+    SD(dst0, dst);
+    SW(dst1, dst + 8);
+    src_ptr += 32;
+    dst += 12;
+  }
+}
+
+void ScaleRowDown38_2_Box_MSA(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_ptr,
+                              int dst_width) {
+  int x, width;
+  const uint8_t* s = src_ptr;
+  const uint8_t* t = src_ptr + src_stride;
+  uint64_t dst0;
+  uint32_t dst1;
+  v16u8 src0, src1, src2, src3, out;
+  v8u16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v4u32 tmp0, tmp1, tmp2, tmp3, tmp4;
+  v8i16 zero = {0};
+  v8i16 mask = {0, 1, 2, 8, 3, 4, 5, 9};
+  v16i8 dst_mask = {0, 2, 16, 4, 6, 18, 8, 10, 20, 12, 14, 22, 0, 0, 0, 0};
+  v4u32 const_0x2AAA = (v4u32)__msa_fill_w(0x2AAA);
+  v4u32 const_0x4000 = (v4u32)__msa_fill_w(0x4000);
+
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  width = dst_width / 3;
+
+  for (x = 0; x < width; x += 4) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)s, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)t, 0);
+    src3 = (v16u8)__msa_ld_b((v16i8*)t, 16);
+    vec0 = (v8u16)__msa_ilvr_b((v16i8)src2, (v16i8)src0);
+    vec1 = (v8u16)__msa_ilvl_b((v16i8)src2, (v16i8)src0);
+    vec2 = (v8u16)__msa_ilvr_b((v16i8)src3, (v16i8)src1);
+    vec3 = (v8u16)__msa_ilvl_b((v16i8)src3, (v16i8)src1);
+    vec0 = __msa_hadd_u_h((v16u8)vec0, (v16u8)vec0);
+    vec1 = __msa_hadd_u_h((v16u8)vec1, (v16u8)vec1);
+    vec2 = __msa_hadd_u_h((v16u8)vec2, (v16u8)vec2);
+    vec3 = __msa_hadd_u_h((v16u8)vec3, (v16u8)vec3);
+    vec4 = (v8u16)__msa_vshf_h(mask, zero, (v8i16)vec0);
+    vec5 = (v8u16)__msa_vshf_h(mask, zero, (v8i16)vec1);
+    vec6 = (v8u16)__msa_vshf_h(mask, zero, (v8i16)vec2);
+    vec7 = (v8u16)__msa_vshf_h(mask, zero, (v8i16)vec3);
+    vec0 = (v8u16)__msa_pckod_w((v4i32)vec1, (v4i32)vec0);
+    vec1 = (v8u16)__msa_pckod_w((v4i32)vec3, (v4i32)vec2);
+    vec0 = (v8u16)__msa_pckod_w((v4i32)vec1, (v4i32)vec0);
+    tmp0 = __msa_hadd_u_w(vec4, vec4);
+    tmp1 = __msa_hadd_u_w(vec5, vec5);
+    tmp2 = __msa_hadd_u_w(vec6, vec6);
+    tmp3 = __msa_hadd_u_w(vec7, vec7);
+    tmp4 = __msa_hadd_u_w(vec0, vec0);
+    vec0 = (v8u16)__msa_pckev_h((v8i16)tmp1, (v8i16)tmp0);
+    vec1 = (v8u16)__msa_pckev_h((v8i16)tmp3, (v8i16)tmp2);
+    tmp0 = __msa_hadd_u_w(vec0, vec0);
+    tmp1 = __msa_hadd_u_w(vec1, vec1);
+    tmp0 *= const_0x2AAA;
+    tmp1 *= const_0x2AAA;
+    tmp4 *= const_0x4000;
+    tmp0 = (v4u32)__msa_srai_w((v4i32)tmp0, 16);
+    tmp1 = (v4u32)__msa_srai_w((v4i32)tmp1, 16);
+    tmp4 = (v4u32)__msa_srai_w((v4i32)tmp4, 16);
+    vec0 = (v8u16)__msa_pckev_h((v8i16)tmp1, (v8i16)tmp0);
+    vec1 = (v8u16)__msa_pckev_h((v8i16)tmp4, (v8i16)tmp4);
+    out = (v16u8)__msa_vshf_b(dst_mask, (v16i8)vec1, (v16i8)vec0);
+    dst0 = __msa_copy_u_d((v2i64)out, 0);
+    dst1 = __msa_copy_u_w((v4i32)out, 2);
+    SD(dst0, dst_ptr);
+    SW(dst1, dst_ptr + 8);
+    s += 32;
+    t += 32;
+    dst_ptr += 12;
+  }
+}
+
+void ScaleRowDown38_3_Box_MSA(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_ptr,
+                              int dst_width) {
+  int x, width;
+  const uint8_t* s = src_ptr;
+  const uint8_t* t0 = s + src_stride;
+  const uint8_t* t1 = s + src_stride * 2;
+  uint64_t dst0;
+  uint32_t dst1;
+  v16u8 src0, src1, src2, src3, src4, src5, out;
+  v8u16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v4u32 tmp0, tmp1, tmp2, tmp3, tmp4;
+  v8u16 zero = {0};
+  v8i16 mask = {0, 1, 2, 8, 3, 4, 5, 9};
+  v16i8 dst_mask = {0, 2, 16, 4, 6, 18, 8, 10, 20, 12, 14, 22, 0, 0, 0, 0};
+  v4u32 const_0x1C71 = (v4u32)__msa_fill_w(0x1C71);
+  v4u32 const_0x2AAA = (v4u32)__msa_fill_w(0x2AAA);
+
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  width = dst_width / 3;
+
+  for (x = 0; x < width; x += 4) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)s, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)t0, 0);
+    src3 = (v16u8)__msa_ld_b((v16i8*)t0, 16);
+    src4 = (v16u8)__msa_ld_b((v16i8*)t1, 0);
+    src5 = (v16u8)__msa_ld_b((v16i8*)t1, 16);
+    vec0 = (v8u16)__msa_ilvr_b((v16i8)src2, (v16i8)src0);
+    vec1 = (v8u16)__msa_ilvl_b((v16i8)src2, (v16i8)src0);
+    vec2 = (v8u16)__msa_ilvr_b((v16i8)src3, (v16i8)src1);
+    vec3 = (v8u16)__msa_ilvl_b((v16i8)src3, (v16i8)src1);
+    vec4 = (v8u16)__msa_ilvr_b((v16i8)zero, (v16i8)src4);
+    vec5 = (v8u16)__msa_ilvl_b((v16i8)zero, (v16i8)src4);
+    vec6 = (v8u16)__msa_ilvr_b((v16i8)zero, (v16i8)src5);
+    vec7 = (v8u16)__msa_ilvl_b((v16i8)zero, (v16i8)src5);
+    vec0 = __msa_hadd_u_h((v16u8)vec0, (v16u8)vec0);
+    vec1 = __msa_hadd_u_h((v16u8)vec1, (v16u8)vec1);
+    vec2 = __msa_hadd_u_h((v16u8)vec2, (v16u8)vec2);
+    vec3 = __msa_hadd_u_h((v16u8)vec3, (v16u8)vec3);
+    vec0 += __msa_hadd_u_h((v16u8)vec4, (v16u8)vec4);
+    vec1 += __msa_hadd_u_h((v16u8)vec5, (v16u8)vec5);
+    vec2 += __msa_hadd_u_h((v16u8)vec6, (v16u8)vec6);
+    vec3 += __msa_hadd_u_h((v16u8)vec7, (v16u8)vec7);
+    vec4 = (v8u16)__msa_vshf_h(mask, (v8i16)zero, (v8i16)vec0);
+    vec5 = (v8u16)__msa_vshf_h(mask, (v8i16)zero, (v8i16)vec1);
+    vec6 = (v8u16)__msa_vshf_h(mask, (v8i16)zero, (v8i16)vec2);
+    vec7 = (v8u16)__msa_vshf_h(mask, (v8i16)zero, (v8i16)vec3);
+    vec0 = (v8u16)__msa_pckod_w((v4i32)vec1, (v4i32)vec0);
+    vec1 = (v8u16)__msa_pckod_w((v4i32)vec3, (v4i32)vec2);
+    vec0 = (v8u16)__msa_pckod_w((v4i32)vec1, (v4i32)vec0);
+    tmp0 = __msa_hadd_u_w(vec4, vec4);
+    tmp1 = __msa_hadd_u_w(vec5, vec5);
+    tmp2 = __msa_hadd_u_w(vec6, vec6);
+    tmp3 = __msa_hadd_u_w(vec7, vec7);
+    tmp4 = __msa_hadd_u_w(vec0, vec0);
+    vec0 = (v8u16)__msa_pckev_h((v8i16)tmp1, (v8i16)tmp0);
+    vec1 = (v8u16)__msa_pckev_h((v8i16)tmp3, (v8i16)tmp2);
+    tmp0 = __msa_hadd_u_w(vec0, vec0);
+    tmp1 = __msa_hadd_u_w(vec1, vec1);
+    tmp0 *= const_0x1C71;
+    tmp1 *= const_0x1C71;
+    tmp4 *= const_0x2AAA;
+    tmp0 = (v4u32)__msa_srai_w((v4i32)tmp0, 16);
+    tmp1 = (v4u32)__msa_srai_w((v4i32)tmp1, 16);
+    tmp4 = (v4u32)__msa_srai_w((v4i32)tmp4, 16);
+    vec0 = (v8u16)__msa_pckev_h((v8i16)tmp1, (v8i16)tmp0);
+    vec1 = (v8u16)__msa_pckev_h((v8i16)tmp4, (v8i16)tmp4);
+    out = (v16u8)__msa_vshf_b(dst_mask, (v16i8)vec1, (v16i8)vec0);
+    dst0 = __msa_copy_u_d((v2i64)out, 0);
+    dst1 = __msa_copy_u_w((v4i32)out, 2);
+    SD(dst0, dst_ptr);
+    SW(dst1, dst_ptr + 8);
+    s += 32;
+    t0 += 32;
+    t1 += 32;
+    dst_ptr += 12;
+  }
+}
+
+void ScaleAddRow_MSA(const uint8_t* src_ptr, uint16_t* dst_ptr, int src_width) {
+  int x;
+  v16u8 src0;
+  v8u16 dst0, dst1;
+  v16i8 zero = {0};
+
+  assert(src_width > 0);
+
+  for (x = 0; x < src_width; x += 16) {
+    src0 = LD_UB(src_ptr);
+    dst0 = (v8u16)__msa_ld_h((v8i16*)dst_ptr, 0);
+    dst1 = (v8u16)__msa_ld_h((v8i16*)dst_ptr, 16);
+    dst0 += (v8u16)__msa_ilvr_b(zero, (v16i8)src0);
+    dst1 += (v8u16)__msa_ilvl_b(zero, (v16i8)src0);
+    ST_UH2(dst0, dst1, dst_ptr, 8);
+    src_ptr += 16;
+    dst_ptr += 16;
+  }
+}
+
+void ScaleFilterCols_MSA(uint8* dst_ptr,
+                         const uint8* src_ptr,
+                         int dst_width,
+                         int x,
+                         int dx) {
+  int j;
+  v4i32 vec_x = __msa_fill_w(x);
+  v4i32 vec_dx = __msa_fill_w(dx);
+  v4i32 vec_const = {0, 1, 2, 3};
+  v4i32 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, vec8, vec9;
+  v4i32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  v8u16 reg0, reg1;
+  v16u8 dst0;
+  v4i32 const_0xFFFF = __msa_fill_w(0xFFFF);
+  v4i32 const_0x40 = __msa_fill_w(0x40);
+
+  vec0 = vec_dx * vec_const;
+  vec1 = vec_dx * 4;
+  vec_x += vec0;
+
+  for (j = 0; j < dst_width - 1; j += 16) {
+    vec2 = vec_x >> 16;
+    vec6 = vec_x & const_0xFFFF;
+    vec_x += vec1;
+    vec3 = vec_x >> 16;
+    vec7 = vec_x & const_0xFFFF;
+    vec_x += vec1;
+    vec4 = vec_x >> 16;
+    vec8 = vec_x & const_0xFFFF;
+    vec_x += vec1;
+    vec5 = vec_x >> 16;
+    vec9 = vec_x & const_0xFFFF;
+    vec_x += vec1;
+    vec6 >>= 9;
+    vec7 >>= 9;
+    vec8 >>= 9;
+    vec9 >>= 9;
+    LOAD_INDEXED_DATA(src_ptr, vec2, tmp0);
+    LOAD_INDEXED_DATA(src_ptr, vec3, tmp1);
+    LOAD_INDEXED_DATA(src_ptr, vec4, tmp2);
+    LOAD_INDEXED_DATA(src_ptr, vec5, tmp3);
+    vec2 += 1;
+    vec3 += 1;
+    vec4 += 1;
+    vec5 += 1;
+    LOAD_INDEXED_DATA(src_ptr, vec2, tmp4);
+    LOAD_INDEXED_DATA(src_ptr, vec3, tmp5);
+    LOAD_INDEXED_DATA(src_ptr, vec4, tmp6);
+    LOAD_INDEXED_DATA(src_ptr, vec5, tmp7);
+    tmp4 -= tmp0;
+    tmp5 -= tmp1;
+    tmp6 -= tmp2;
+    tmp7 -= tmp3;
+    tmp4 *= vec6;
+    tmp5 *= vec7;
+    tmp6 *= vec8;
+    tmp7 *= vec9;
+    tmp4 += const_0x40;
+    tmp5 += const_0x40;
+    tmp6 += const_0x40;
+    tmp7 += const_0x40;
+    tmp4 >>= 7;
+    tmp5 >>= 7;
+    tmp6 >>= 7;
+    tmp7 >>= 7;
+    tmp0 += tmp4;
+    tmp1 += tmp5;
+    tmp2 += tmp6;
+    tmp3 += tmp7;
+    reg0 = (v8u16)__msa_pckev_h((v8i16)tmp1, (v8i16)tmp0);
+    reg1 = (v8u16)__msa_pckev_h((v8i16)tmp3, (v8i16)tmp2);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)reg1, (v16i8)reg0);
+    __msa_st_b(dst0, dst_ptr, 0);
+    dst_ptr += 16;
+  }
+}
+
+void ScaleARGBCols_MSA(uint8* dst_argb,
+                       const uint8* src_argb,
+                       int dst_width,
+                       int x,
+                       int dx) {
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+  int j;
+  v4i32 x_vec = __msa_fill_w(x);
+  v4i32 dx_vec = __msa_fill_w(dx);
+  v4i32 const_vec = {0, 1, 2, 3};
+  v4i32 vec0, vec1, vec2;
+  v4i32 dst0;
+
+  vec0 = dx_vec * const_vec;
+  vec1 = dx_vec * 4;
+  x_vec += vec0;
+
+  for (j = 0; j < dst_width; j += 4) {
+    vec2 = x_vec >> 16;
+    x_vec += vec1;
+    LOAD_INDEXED_DATA(src, vec2, dst0);
+    __msa_st_w(dst0, dst, 0);
+    dst += 4;
+  }
+}
+
+void ScaleARGBFilterCols_MSA(uint8* dst_argb,
+                             const uint8* src_argb,
+                             int dst_width,
+                             int x,
+                             int dx) {
+  const uint32* src = (const uint32*)(src_argb);
+  int j;
+  v4u32 src0, src1, src2, src3;
+  v4u32 vec0, vec1, vec2, vec3;
+  v16u8 reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;
+  v16u8 mult0, mult1, mult2, mult3;
+  v8u16 tmp0, tmp1, tmp2, tmp3;
+  v16u8 dst0, dst1;
+  v4u32 vec_x = (v4u32)__msa_fill_w(x);
+  v4u32 vec_dx = (v4u32)__msa_fill_w(dx);
+  v4u32 vec_const = {0, 1, 2, 3};
+  v16u8 const_0x7f = (v16u8)__msa_fill_b(0x7f);
+
+  vec0 = vec_dx * vec_const;
+  vec1 = vec_dx * 4;
+  vec_x += vec0;
+
+  for (j = 0; j < dst_width - 1; j += 8) {
+    vec2 = vec_x >> 16;
+    reg0 = (v16u8)(vec_x >> 9);
+    vec_x += vec1;
+    vec3 = vec_x >> 16;
+    reg1 = (v16u8)(vec_x >> 9);
+    vec_x += vec1;
+    reg0 = reg0 & const_0x7f;
+    reg1 = reg1 & const_0x7f;
+    reg0 = (v16u8)__msa_shf_b((v16i8)reg0, 0);
+    reg1 = (v16u8)__msa_shf_b((v16i8)reg1, 0);
+    reg2 = reg0 ^ const_0x7f;
+    reg3 = reg1 ^ const_0x7f;
+    mult0 = (v16u8)__msa_ilvr_b((v16i8)reg0, (v16i8)reg2);
+    mult1 = (v16u8)__msa_ilvl_b((v16i8)reg0, (v16i8)reg2);
+    mult2 = (v16u8)__msa_ilvr_b((v16i8)reg1, (v16i8)reg3);
+    mult3 = (v16u8)__msa_ilvl_b((v16i8)reg1, (v16i8)reg3);
+    LOAD_INDEXED_DATA(src, vec2, src0);
+    LOAD_INDEXED_DATA(src, vec3, src1);
+    vec2 += 1;
+    vec3 += 1;
+    LOAD_INDEXED_DATA(src, vec2, src2);
+    LOAD_INDEXED_DATA(src, vec3, src3);
+    reg4 = (v16u8)__msa_ilvr_b((v16i8)src2, (v16i8)src0);
+    reg5 = (v16u8)__msa_ilvl_b((v16i8)src2, (v16i8)src0);
+    reg6 = (v16u8)__msa_ilvr_b((v16i8)src3, (v16i8)src1);
+    reg7 = (v16u8)__msa_ilvl_b((v16i8)src3, (v16i8)src1);
+    tmp0 = __msa_dotp_u_h(reg4, mult0);
+    tmp1 = __msa_dotp_u_h(reg5, mult1);
+    tmp2 = __msa_dotp_u_h(reg6, mult2);
+    tmp3 = __msa_dotp_u_h(reg7, mult3);
+    tmp0 >>= 7;
+    tmp1 >>= 7;
+    tmp2 >>= 7;
+    tmp3 >>= 7;
+    dst0 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+    dst1 = (v16u8)__msa_pckev_b((v16i8)tmp3, (v16i8)tmp2);
+    __msa_st_b(dst0, dst_argb, 0);
+    __msa_st_b(dst1, dst_argb, 16);
+    dst_argb += 32;
+  }
+}
+
+void ScaleRowDown34_MSA(const uint8* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint8* dst,
+                        int dst_width) {
+  int x;
+  (void)src_stride;
+  v16u8 src0, src1, src2, src3;
+  v16u8 vec0, vec1, vec2;
+  v16i8 mask0 = {0, 1, 3, 4, 5, 7, 8, 9, 11, 12, 13, 15, 16, 17, 19, 20};
+  v16i8 mask1 = {5, 7, 8, 9, 11, 12, 13, 15, 16, 17, 19, 20, 21, 23, 24, 25};
+  v16i8 mask2 = {11, 12, 13, 15, 16, 17, 19, 20,
+                 21, 23, 24, 25, 27, 28, 29, 31};
+
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+
+  for (x = 0; x < dst_width; x += 48) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 32);
+    src3 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 48);
+    vec0 = (v16u8)__msa_vshf_b(mask0, (v16i8)src1, (v16i8)src0);
+    vec1 = (v16u8)__msa_vshf_b(mask1, (v16i8)src2, (v16i8)src1);
+    vec2 = (v16u8)__msa_vshf_b(mask2, (v16i8)src3, (v16i8)src2);
+    __msa_st_b((v16i8)vec0, dst, 0);
+    __msa_st_b((v16i8)vec1, dst, 16);
+    __msa_st_b((v16i8)vec2, dst, 32);
+    src_ptr += 64;
+    dst += 48;
+  }
+}
+
+void ScaleRowDown34_0_Box_MSA(const uint8* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8* d,
+                              int dst_width) {
+  const uint8* s = src_ptr;
+  const uint8* t = src_ptr + src_stride;
+  int x;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7, dst0, dst1, dst2;
+  v16u8 vec0, vec1, vec2, vec3, vec4, vec5;
+  v16u8 vec6, vec7, vec8, vec9, vec10, vec11;
+  v8i16 reg0, reg1, reg2, reg3, reg4, reg5;
+  v8i16 reg6, reg7, reg8, reg9, reg10, reg11;
+  v16u8 const0 = {3, 1, 1, 1, 1, 3, 3, 1, 1, 1, 1, 3, 3, 1, 1, 1};
+  v16u8 const1 = {1, 3, 3, 1, 1, 1, 1, 3, 3, 1, 1, 1, 1, 3, 3, 1};
+  v16u8 const2 = {1, 1, 1, 3, 3, 1, 1, 1, 1, 3, 3, 1, 1, 1, 1, 3};
+  v16i8 mask0 = {0, 1, 1, 2, 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10};
+  v16i8 mask1 = {10, 11, 12, 13, 13, 14, 14, 15,
+                 16, 17, 17, 18, 18, 19, 20, 21};
+  v16i8 mask2 = {5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13, 13, 14, 14, 15};
+  v8i16 shft0 = {2, 1, 2, 2, 1, 2, 2, 1};
+  v8i16 shft1 = {2, 2, 1, 2, 2, 1, 2, 2};
+  v8i16 shft2 = {1, 2, 2, 1, 2, 2, 1, 2};
+
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+
+  for (x = 0; x < dst_width; x += 48) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)s, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)s, 32);
+    src3 = (v16u8)__msa_ld_b((v16i8*)s, 48);
+    src4 = (v16u8)__msa_ld_b((v16i8*)t, 0);
+    src5 = (v16u8)__msa_ld_b((v16i8*)t, 16);
+    src6 = (v16u8)__msa_ld_b((v16i8*)t, 32);
+    src7 = (v16u8)__msa_ld_b((v16i8*)t, 48);
+    vec0 = (v16u8)__msa_vshf_b(mask0, (v16i8)src0, (v16i8)src0);
+    vec1 = (v16u8)__msa_vshf_b(mask1, (v16i8)src1, (v16i8)src0);
+    vec2 = (v16u8)__msa_vshf_b(mask2, (v16i8)src1, (v16i8)src1);
+    vec3 = (v16u8)__msa_vshf_b(mask0, (v16i8)src2, (v16i8)src2);
+    vec4 = (v16u8)__msa_vshf_b(mask1, (v16i8)src3, (v16i8)src2);
+    vec5 = (v16u8)__msa_vshf_b(mask2, (v16i8)src3, (v16i8)src3);
+    vec6 = (v16u8)__msa_vshf_b(mask0, (v16i8)src4, (v16i8)src4);
+    vec7 = (v16u8)__msa_vshf_b(mask1, (v16i8)src5, (v16i8)src4);
+    vec8 = (v16u8)__msa_vshf_b(mask2, (v16i8)src5, (v16i8)src5);
+    vec9 = (v16u8)__msa_vshf_b(mask0, (v16i8)src6, (v16i8)src6);
+    vec10 = (v16u8)__msa_vshf_b(mask1, (v16i8)src7, (v16i8)src6);
+    vec11 = (v16u8)__msa_vshf_b(mask2, (v16i8)src7, (v16i8)src7);
+    reg0 = (v8i16)__msa_dotp_u_h(vec0, const0);
+    reg1 = (v8i16)__msa_dotp_u_h(vec1, const1);
+    reg2 = (v8i16)__msa_dotp_u_h(vec2, const2);
+    reg3 = (v8i16)__msa_dotp_u_h(vec3, const0);
+    reg4 = (v8i16)__msa_dotp_u_h(vec4, const1);
+    reg5 = (v8i16)__msa_dotp_u_h(vec5, const2);
+    reg6 = (v8i16)__msa_dotp_u_h(vec6, const0);
+    reg7 = (v8i16)__msa_dotp_u_h(vec7, const1);
+    reg8 = (v8i16)__msa_dotp_u_h(vec8, const2);
+    reg9 = (v8i16)__msa_dotp_u_h(vec9, const0);
+    reg10 = (v8i16)__msa_dotp_u_h(vec10, const1);
+    reg11 = (v8i16)__msa_dotp_u_h(vec11, const2);
+    reg0 = __msa_srar_h(reg0, shft0);
+    reg1 = __msa_srar_h(reg1, shft1);
+    reg2 = __msa_srar_h(reg2, shft2);
+    reg3 = __msa_srar_h(reg3, shft0);
+    reg4 = __msa_srar_h(reg4, shft1);
+    reg5 = __msa_srar_h(reg5, shft2);
+    reg6 = __msa_srar_h(reg6, shft0);
+    reg7 = __msa_srar_h(reg7, shft1);
+    reg8 = __msa_srar_h(reg8, shft2);
+    reg9 = __msa_srar_h(reg9, shft0);
+    reg10 = __msa_srar_h(reg10, shft1);
+    reg11 = __msa_srar_h(reg11, shft2);
+    reg0 = reg0 * 3 + reg6;
+    reg1 = reg1 * 3 + reg7;
+    reg2 = reg2 * 3 + reg8;
+    reg3 = reg3 * 3 + reg9;
+    reg4 = reg4 * 3 + reg10;
+    reg5 = reg5 * 3 + reg11;
+    reg0 = __msa_srari_h(reg0, 2);
+    reg1 = __msa_srari_h(reg1, 2);
+    reg2 = __msa_srari_h(reg2, 2);
+    reg3 = __msa_srari_h(reg3, 2);
+    reg4 = __msa_srari_h(reg4, 2);
+    reg5 = __msa_srari_h(reg5, 2);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)reg1, (v16i8)reg0);
+    dst1 = (v16u8)__msa_pckev_b((v16i8)reg3, (v16i8)reg2);
+    dst2 = (v16u8)__msa_pckev_b((v16i8)reg5, (v16i8)reg4);
+    __msa_st_b((v16i8)dst0, d, 0);
+    __msa_st_b((v16i8)dst1, d, 16);
+    __msa_st_b((v16i8)dst2, d, 32);
+    s += 64;
+    t += 64;
+    d += 48;
+  }
+}
+
+void ScaleRowDown34_1_Box_MSA(const uint8* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8* d,
+                              int dst_width) {
+  const uint8* s = src_ptr;
+  const uint8* t = src_ptr + src_stride;
+  int x;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7, dst0, dst1, dst2;
+  v16u8 vec0, vec1, vec2, vec3, vec4, vec5;
+  v16u8 vec6, vec7, vec8, vec9, vec10, vec11;
+  v8i16 reg0, reg1, reg2, reg3, reg4, reg5;
+  v8i16 reg6, reg7, reg8, reg9, reg10, reg11;
+  v16u8 const0 = {3, 1, 1, 1, 1, 3, 3, 1, 1, 1, 1, 3, 3, 1, 1, 1};
+  v16u8 const1 = {1, 3, 3, 1, 1, 1, 1, 3, 3, 1, 1, 1, 1, 3, 3, 1};
+  v16u8 const2 = {1, 1, 1, 3, 3, 1, 1, 1, 1, 3, 3, 1, 1, 1, 1, 3};
+  v16i8 mask0 = {0, 1, 1, 2, 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10};
+  v16i8 mask1 = {10, 11, 12, 13, 13, 14, 14, 15,
+                 16, 17, 17, 18, 18, 19, 20, 21};
+  v16i8 mask2 = {5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13, 13, 14, 14, 15};
+  v8i16 shft0 = {2, 1, 2, 2, 1, 2, 2, 1};
+  v8i16 shft1 = {2, 2, 1, 2, 2, 1, 2, 2};
+  v8i16 shft2 = {1, 2, 2, 1, 2, 2, 1, 2};
+
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+
+  for (x = 0; x < dst_width; x += 48) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)s, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)s, 32);
+    src3 = (v16u8)__msa_ld_b((v16i8*)s, 48);
+    src4 = (v16u8)__msa_ld_b((v16i8*)t, 0);
+    src5 = (v16u8)__msa_ld_b((v16i8*)t, 16);
+    src6 = (v16u8)__msa_ld_b((v16i8*)t, 32);
+    src7 = (v16u8)__msa_ld_b((v16i8*)t, 48);
+    vec0 = (v16u8)__msa_vshf_b(mask0, (v16i8)src0, (v16i8)src0);
+    vec1 = (v16u8)__msa_vshf_b(mask1, (v16i8)src1, (v16i8)src0);
+    vec2 = (v16u8)__msa_vshf_b(mask2, (v16i8)src1, (v16i8)src1);
+    vec3 = (v16u8)__msa_vshf_b(mask0, (v16i8)src2, (v16i8)src2);
+    vec4 = (v16u8)__msa_vshf_b(mask1, (v16i8)src3, (v16i8)src2);
+    vec5 = (v16u8)__msa_vshf_b(mask2, (v16i8)src3, (v16i8)src3);
+    vec6 = (v16u8)__msa_vshf_b(mask0, (v16i8)src4, (v16i8)src4);
+    vec7 = (v16u8)__msa_vshf_b(mask1, (v16i8)src5, (v16i8)src4);
+    vec8 = (v16u8)__msa_vshf_b(mask2, (v16i8)src5, (v16i8)src5);
+    vec9 = (v16u8)__msa_vshf_b(mask0, (v16i8)src6, (v16i8)src6);
+    vec10 = (v16u8)__msa_vshf_b(mask1, (v16i8)src7, (v16i8)src6);
+    vec11 = (v16u8)__msa_vshf_b(mask2, (v16i8)src7, (v16i8)src7);
+    reg0 = (v8i16)__msa_dotp_u_h(vec0, const0);
+    reg1 = (v8i16)__msa_dotp_u_h(vec1, const1);
+    reg2 = (v8i16)__msa_dotp_u_h(vec2, const2);
+    reg3 = (v8i16)__msa_dotp_u_h(vec3, const0);
+    reg4 = (v8i16)__msa_dotp_u_h(vec4, const1);
+    reg5 = (v8i16)__msa_dotp_u_h(vec5, const2);
+    reg6 = (v8i16)__msa_dotp_u_h(vec6, const0);
+    reg7 = (v8i16)__msa_dotp_u_h(vec7, const1);
+    reg8 = (v8i16)__msa_dotp_u_h(vec8, const2);
+    reg9 = (v8i16)__msa_dotp_u_h(vec9, const0);
+    reg10 = (v8i16)__msa_dotp_u_h(vec10, const1);
+    reg11 = (v8i16)__msa_dotp_u_h(vec11, const2);
+    reg0 = __msa_srar_h(reg0, shft0);
+    reg1 = __msa_srar_h(reg1, shft1);
+    reg2 = __msa_srar_h(reg2, shft2);
+    reg3 = __msa_srar_h(reg3, shft0);
+    reg4 = __msa_srar_h(reg4, shft1);
+    reg5 = __msa_srar_h(reg5, shft2);
+    reg6 = __msa_srar_h(reg6, shft0);
+    reg7 = __msa_srar_h(reg7, shft1);
+    reg8 = __msa_srar_h(reg8, shft2);
+    reg9 = __msa_srar_h(reg9, shft0);
+    reg10 = __msa_srar_h(reg10, shft1);
+    reg11 = __msa_srar_h(reg11, shft2);
+    reg0 += reg6;
+    reg1 += reg7;
+    reg2 += reg8;
+    reg3 += reg9;
+    reg4 += reg10;
+    reg5 += reg11;
+    reg0 = __msa_srari_h(reg0, 1);
+    reg1 = __msa_srari_h(reg1, 1);
+    reg2 = __msa_srari_h(reg2, 1);
+    reg3 = __msa_srari_h(reg3, 1);
+    reg4 = __msa_srari_h(reg4, 1);
+    reg5 = __msa_srari_h(reg5, 1);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)reg1, (v16i8)reg0);
+    dst1 = (v16u8)__msa_pckev_b((v16i8)reg3, (v16i8)reg2);
+    dst2 = (v16u8)__msa_pckev_b((v16i8)reg5, (v16i8)reg4);
+    __msa_st_b((v16i8)dst0, d, 0);
+    __msa_st_b((v16i8)dst1, d, 16);
+    __msa_st_b((v16i8)dst2, d, 32);
+    s += 64;
+    t += 64;
+    d += 48;
+  }
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // !defined(LIBYUV_DISABLE_MSA) && defined(__mips_msa)

pjproject-2.11.1/third_party/yuv/source/scale_neon.cc

@@ -0,0 +1,967 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC Neon.
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__) && \
+    !defined(__aarch64__)
+
+// NEON downscalers with interpolation.
+// Provided by Fritz Koenig
+
+// Read 32x1 throw away even pixels, and write 16x1.
+void ScaleRowDown2_NEON(const uint8* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint8* dst,
+                        int dst_width) {
+  (void)src_stride;
+  asm volatile(
+      "1:                                        \n"
+      // load even pixels into q0, odd into q1
+      "vld2.8     {q0, q1}, [%0]!                \n"
+      "subs       %2, %2, #16                    \n"  // 16 processed per loop
+      "vst1.8     {q1}, [%1]!                    \n"  // store odd pixels
+      "bgt        1b                             \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(dst),       // %1
+        "+r"(dst_width)  // %2
+      :
+      : "q0", "q1"  // Clobber List
+      );
+}
+
+// Read 32x1 average down and write 16x1.
+void ScaleRowDown2Linear_NEON(const uint8* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8* dst,
+                              int dst_width) {
+  (void)src_stride;
+  asm volatile(
+      "1:                                        \n"
+      "vld2.8     {q0, q1}, [%0]!                \n"  // load 32 pixels
+      "subs       %2, %2, #16                    \n"  // 16 processed per loop
+      "vrhadd.u8  q0, q0, q1                     \n"  // rounding half add
+      "vst1.8     {q0}, [%1]!                    \n"
+      "bgt        1b                             \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(dst),       // %1
+        "+r"(dst_width)  // %2
+      :
+      : "q0", "q1"  // Clobber List
+      );
+}
+
+// Read 32x2 average down and write 16x1.
+void ScaleRowDown2Box_NEON(const uint8* src_ptr,
+                           ptrdiff_t src_stride,
+                           uint8* dst,
+                           int dst_width) {
+  asm volatile(
+      // change the stride to row 2 pointer
+      "add        %1, %0                         \n"
+      "1:                                        \n"
+      "vld1.8     {q0, q1}, [%0]!                \n"  // load row 1 and post inc
+      "vld1.8     {q2, q3}, [%1]!                \n"  // load row 2 and post inc
+      "subs       %3, %3, #16                    \n"  // 16 processed per loop
+      "vpaddl.u8  q0, q0                         \n"  // row 1 add adjacent
+      "vpaddl.u8  q1, q1                         \n"
+      "vpadal.u8  q0, q2                         \n"  // row 2 add adjacent +
+                                                      // row1
+      "vpadal.u8  q1, q3                         \n"
+      "vrshrn.u16 d0, q0, #2                     \n"  // downshift, round and
+                                                      // pack
+      "vrshrn.u16 d1, q1, #2                     \n"
+      "vst1.8     {q0}, [%2]!                    \n"
+      "bgt        1b                             \n"
+      : "+r"(src_ptr),     // %0
+        "+r"(src_stride),  // %1
+        "+r"(dst),         // %2
+        "+r"(dst_width)    // %3
+      :
+      : "q0", "q1", "q2", "q3"  // Clobber List
+      );
+}
+
+void ScaleRowDown4_NEON(const uint8* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint8* dst_ptr,
+                        int dst_width) {
+  (void)src_stride;
+  asm volatile(
+      "1:                                        \n"
+      "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // src line 0
+      "subs       %2, %2, #8                     \n"  // 8 processed per loop
+      "vst1.8     {d2}, [%1]!                    \n"
+      "bgt        1b                             \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(dst_ptr),   // %1
+        "+r"(dst_width)  // %2
+      :
+      : "q0", "q1", "memory", "cc");
+}
+
+void ScaleRowDown4Box_NEON(const uint8* src_ptr,
+                           ptrdiff_t src_stride,
+                           uint8* dst_ptr,
+                           int dst_width) {
+  const uint8* src_ptr1 = src_ptr + src_stride;
+  const uint8* src_ptr2 = src_ptr + src_stride * 2;
+  const uint8* src_ptr3 = src_ptr + src_stride * 3;
+  asm volatile(
+      "1:                                        \n"
+      "vld1.8     {q0}, [%0]!                    \n"  // load up 16x4
+      "vld1.8     {q1}, [%3]!                    \n"
+      "vld1.8     {q2}, [%4]!                    \n"
+      "vld1.8     {q3}, [%5]!                    \n"
+      "subs       %2, %2, #4                     \n"
+      "vpaddl.u8  q0, q0                         \n"
+      "vpadal.u8  q0, q1                         \n"
+      "vpadal.u8  q0, q2                         \n"
+      "vpadal.u8  q0, q3                         \n"
+      "vpaddl.u16 q0, q0                         \n"
+      "vrshrn.u32 d0, q0, #4                     \n"  // divide by 16 w/rounding
+      "vmovn.u16  d0, q0                         \n"
+      "vst1.32    {d0[0]}, [%1]!                 \n"
+      "bgt        1b                             \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(dst_ptr),    // %1
+        "+r"(dst_width),  // %2
+        "+r"(src_ptr1),   // %3
+        "+r"(src_ptr2),   // %4
+        "+r"(src_ptr3)    // %5
+      :
+      : "q0", "q1", "q2", "q3", "memory", "cc");
+}
+
+// Down scale from 4 to 3 pixels. Use the neon multilane read/write
+// to load up the every 4th pixel into a 4 different registers.
+// Point samples 32 pixels to 24 pixels.
+void ScaleRowDown34_NEON(const uint8* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint8* dst_ptr,
+                         int dst_width) {
+  (void)src_stride;
+  asm volatile(
+      "1:                                        \n"
+      "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // src line 0
+      "subs       %2, %2, #24                    \n"
+      "vmov       d2, d3                         \n"  // order d0, d1, d2
+      "vst3.8     {d0, d1, d2}, [%1]!            \n"
+      "bgt        1b                             \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(dst_ptr),   // %1
+        "+r"(dst_width)  // %2
+      :
+      : "d0", "d1", "d2", "d3", "memory", "cc");
+}
+
+void ScaleRowDown34_0_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr,
+                               int dst_width) {
+  asm volatile(
+      "vmov.u8    d24, #3                        \n"
+      "add        %3, %0                         \n"
+      "1:                                        \n"
+      "vld4.8       {d0, d1, d2, d3}, [%0]!      \n"  // src line 0
+      "vld4.8       {d4, d5, d6, d7}, [%3]!      \n"  // src line 1
+      "subs         %2, %2, #24                  \n"
+
+      // filter src line 0 with src line 1
+      // expand chars to shorts to allow for room
+      // when adding lines together
+      "vmovl.u8     q8, d4                       \n"
+      "vmovl.u8     q9, d5                       \n"
+      "vmovl.u8     q10, d6                      \n"
+      "vmovl.u8     q11, d7                      \n"
+
+      // 3 * line_0 + line_1
+      "vmlal.u8     q8, d0, d24                  \n"
+      "vmlal.u8     q9, d1, d24                  \n"
+      "vmlal.u8     q10, d2, d24                 \n"
+      "vmlal.u8     q11, d3, d24                 \n"
+
+      // (3 * line_0 + line_1) >> 2
+      "vqrshrn.u16  d0, q8, #2                   \n"
+      "vqrshrn.u16  d1, q9, #2                   \n"
+      "vqrshrn.u16  d2, q10, #2                  \n"
+      "vqrshrn.u16  d3, q11, #2                  \n"
+
+      // a0 = (src[0] * 3 + s[1] * 1) >> 2
+      "vmovl.u8     q8, d1                       \n"
+      "vmlal.u8     q8, d0, d24                  \n"
+      "vqrshrn.u16  d0, q8, #2                   \n"
+
+      // a1 = (src[1] * 1 + s[2] * 1) >> 1
+      "vrhadd.u8    d1, d1, d2                   \n"
+
+      // a2 = (src[2] * 1 + s[3] * 3) >> 2
+      "vmovl.u8     q8, d2                       \n"
+      "vmlal.u8     q8, d3, d24                  \n"
+      "vqrshrn.u16  d2, q8, #2                   \n"
+
+      "vst3.8       {d0, d1, d2}, [%1]!          \n"
+
+      "bgt          1b                           \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(dst_ptr),    // %1
+        "+r"(dst_width),  // %2
+        "+r"(src_stride)  // %3
+      :
+      : "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11", "d24", "memory",
+        "cc");
+}
+
+void ScaleRowDown34_1_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr,
+                               int dst_width) {
+  asm volatile(
+      "vmov.u8    d24, #3                        \n"
+      "add        %3, %0                         \n"
+      "1:                                        \n"
+      "vld4.8       {d0, d1, d2, d3}, [%0]!      \n"  // src line 0
+      "vld4.8       {d4, d5, d6, d7}, [%3]!      \n"  // src line 1
+      "subs         %2, %2, #24                  \n"
+      // average src line 0 with src line 1
+      "vrhadd.u8    q0, q0, q2                   \n"
+      "vrhadd.u8    q1, q1, q3                   \n"
+
+      // a0 = (src[0] * 3 + s[1] * 1) >> 2
+      "vmovl.u8     q3, d1                       \n"
+      "vmlal.u8     q3, d0, d24                  \n"
+      "vqrshrn.u16  d0, q3, #2                   \n"
+
+      // a1 = (src[1] * 1 + s[2] * 1) >> 1
+      "vrhadd.u8    d1, d1, d2                   \n"
+
+      // a2 = (src[2] * 1 + s[3] * 3) >> 2
+      "vmovl.u8     q3, d2                       \n"
+      "vmlal.u8     q3, d3, d24                  \n"
+      "vqrshrn.u16  d2, q3, #2                   \n"
+
+      "vst3.8       {d0, d1, d2}, [%1]!          \n"
+      "bgt          1b                           \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(dst_ptr),    // %1
+        "+r"(dst_width),  // %2
+        "+r"(src_stride)  // %3
+      :
+      : "r4", "q0", "q1", "q2", "q3", "d24", "memory", "cc");
+}
+
+#define HAS_SCALEROWDOWN38_NEON
+static uvec8 kShuf38 = {0, 3, 6, 8, 11, 14, 16, 19, 22, 24, 27, 30, 0, 0, 0, 0};
+static uvec8 kShuf38_2 = {0,  8, 16, 2,  10, 17, 4, 12,
+                          18, 6, 14, 19, 0,  0,  0, 0};
+static vec16 kMult38_Div6 = {65536 / 12, 65536 / 12, 65536 / 12, 65536 / 12,
+                             65536 / 12, 65536 / 12, 65536 / 12, 65536 / 12};
+static vec16 kMult38_Div9 = {65536 / 18, 65536 / 18, 65536 / 18, 65536 / 18,
+                             65536 / 18, 65536 / 18, 65536 / 18, 65536 / 18};
+
+// 32 -> 12
+void ScaleRowDown38_NEON(const uint8* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint8* dst_ptr,
+                         int dst_width) {
+  (void)src_stride;
+  asm volatile(
+      "vld1.8     {q3}, [%3]                     \n"
+      "1:                                        \n"
+      "vld1.8     {d0, d1, d2, d3}, [%0]!        \n"
+      "subs       %2, %2, #12                    \n"
+      "vtbl.u8    d4, {d0, d1, d2, d3}, d6       \n"
+      "vtbl.u8    d5, {d0, d1, d2, d3}, d7       \n"
+      "vst1.8     {d4}, [%1]!                    \n"
+      "vst1.32    {d5[0]}, [%1]!                 \n"
+      "bgt        1b                             \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(dst_ptr),   // %1
+        "+r"(dst_width)  // %2
+      : "r"(&kShuf38)    // %3
+      : "d0", "d1", "d2", "d3", "d4", "d5", "memory", "cc");
+}
+
+// 32x3 -> 12x1
+void OMITFP ScaleRowDown38_3_Box_NEON(const uint8* src_ptr,
+                                      ptrdiff_t src_stride,
+                                      uint8* dst_ptr,
+                                      int dst_width) {
+  const uint8* src_ptr1 = src_ptr + src_stride * 2;
+
+  asm volatile(
+      "vld1.16    {q13}, [%5]                    \n"
+      "vld1.8     {q14}, [%6]                    \n"
+      "vld1.8     {q15}, [%7]                    \n"
+      "add        %3, %0                         \n"
+      "1:                                        \n"
+
+      // d0 = 00 40 01 41 02 42 03 43
+      // d1 = 10 50 11 51 12 52 13 53
+      // d2 = 20 60 21 61 22 62 23 63
+      // d3 = 30 70 31 71 32 72 33 73
+      "vld4.8       {d0, d1, d2, d3}, [%0]!      \n"
+      "vld4.8       {d4, d5, d6, d7}, [%3]!      \n"
+      "vld4.8       {d16, d17, d18, d19}, [%4]!  \n"
+      "subs         %2, %2, #12                  \n"
+
+      // Shuffle the input data around to get align the data
+      //  so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7
+      // d0 = 00 10 01 11 02 12 03 13
+      // d1 = 40 50 41 51 42 52 43 53
+      "vtrn.u8      d0, d1                       \n"
+      "vtrn.u8      d4, d5                       \n"
+      "vtrn.u8      d16, d17                     \n"
+
+      // d2 = 20 30 21 31 22 32 23 33
+      // d3 = 60 70 61 71 62 72 63 73
+      "vtrn.u8      d2, d3                       \n"
+      "vtrn.u8      d6, d7                       \n"
+      "vtrn.u8      d18, d19                     \n"
+
+      // d0 = 00+10 01+11 02+12 03+13
+      // d2 = 40+50 41+51 42+52 43+53
+      "vpaddl.u8    q0, q0                       \n"
+      "vpaddl.u8    q2, q2                       \n"
+      "vpaddl.u8    q8, q8                       \n"
+
+      // d3 = 60+70 61+71 62+72 63+73
+      "vpaddl.u8    d3, d3                       \n"
+      "vpaddl.u8    d7, d7                       \n"
+      "vpaddl.u8    d19, d19                     \n"
+
+      // combine source lines
+      "vadd.u16     q0, q2                       \n"
+      "vadd.u16     q0, q8                       \n"
+      "vadd.u16     d4, d3, d7                   \n"
+      "vadd.u16     d4, d19                      \n"
+
+      // dst_ptr[3] = (s[6 + st * 0] + s[7 + st * 0]
+      //             + s[6 + st * 1] + s[7 + st * 1]
+      //             + s[6 + st * 2] + s[7 + st * 2]) / 6
+      "vqrdmulh.s16 q2, q2, q13                  \n"
+      "vmovn.u16    d4, q2                       \n"
+
+      // Shuffle 2,3 reg around so that 2 can be added to the
+      //  0,1 reg and 3 can be added to the 4,5 reg. This
+      //  requires expanding from u8 to u16 as the 0,1 and 4,5
+      //  registers are already expanded. Then do transposes
+      //  to get aligned.
+      // q2 = xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
+      "vmovl.u8     q1, d2                       \n"
+      "vmovl.u8     q3, d6                       \n"
+      "vmovl.u8     q9, d18                      \n"
+
+      // combine source lines
+      "vadd.u16     q1, q3                       \n"
+      "vadd.u16     q1, q9                       \n"
+
+      // d4 = xx 20 xx 30 xx 22 xx 32
+      // d5 = xx 21 xx 31 xx 23 xx 33
+      "vtrn.u32     d2, d3                       \n"
+
+      // d4 = xx 20 xx 21 xx 22 xx 23
+      // d5 = xx 30 xx 31 xx 32 xx 33
+      "vtrn.u16     d2, d3                       \n"
+
+      // 0+1+2, 3+4+5
+      "vadd.u16     q0, q1                       \n"
+
+      // Need to divide, but can't downshift as the the value
+      //  isn't a power of 2. So multiply by 65536 / n
+      //  and take the upper 16 bits.
+      "vqrdmulh.s16 q0, q0, q15                  \n"
+
+      // Align for table lookup, vtbl requires registers to
+      //  be adjacent
+      "vmov.u8      d2, d4                       \n"
+
+      "vtbl.u8      d3, {d0, d1, d2}, d28        \n"
+      "vtbl.u8      d4, {d0, d1, d2}, d29        \n"
+
+      "vst1.8       {d3}, [%1]!                  \n"
+      "vst1.32      {d4[0]}, [%1]!               \n"
+      "bgt          1b                           \n"
+      : "+r"(src_ptr),       // %0
+        "+r"(dst_ptr),       // %1
+        "+r"(dst_width),     // %2
+        "+r"(src_stride),    // %3
+        "+r"(src_ptr1)       // %4
+      : "r"(&kMult38_Div6),  // %5
+        "r"(&kShuf38_2),     // %6
+        "r"(&kMult38_Div9)   // %7
+      : "q0", "q1", "q2", "q3", "q8", "q9", "q13", "q14", "q15", "memory",
+        "cc");
+}
+
+// 32x2 -> 12x1
+void ScaleRowDown38_2_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr,
+                               int dst_width) {
+  asm volatile(
+      "vld1.16    {q13}, [%4]                    \n"
+      "vld1.8     {q14}, [%5]                    \n"
+      "add        %3, %0                         \n"
+      "1:                                        \n"
+
+      // d0 = 00 40 01 41 02 42 03 43
+      // d1 = 10 50 11 51 12 52 13 53
+      // d2 = 20 60 21 61 22 62 23 63
+      // d3 = 30 70 31 71 32 72 33 73
+      "vld4.8       {d0, d1, d2, d3}, [%0]!      \n"
+      "vld4.8       {d4, d5, d6, d7}, [%3]!      \n"
+      "subs         %2, %2, #12                  \n"
+
+      // Shuffle the input data around to get align the data
+      //  so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7
+      // d0 = 00 10 01 11 02 12 03 13
+      // d1 = 40 50 41 51 42 52 43 53
+      "vtrn.u8      d0, d1                       \n"
+      "vtrn.u8      d4, d5                       \n"
+
+      // d2 = 20 30 21 31 22 32 23 33
+      // d3 = 60 70 61 71 62 72 63 73
+      "vtrn.u8      d2, d3                       \n"
+      "vtrn.u8      d6, d7                       \n"
+
+      // d0 = 00+10 01+11 02+12 03+13
+      // d2 = 40+50 41+51 42+52 43+53
+      "vpaddl.u8    q0, q0                       \n"
+      "vpaddl.u8    q2, q2                       \n"
+
+      // d3 = 60+70 61+71 62+72 63+73
+      "vpaddl.u8    d3, d3                       \n"
+      "vpaddl.u8    d7, d7                       \n"
+
+      // combine source lines
+      "vadd.u16     q0, q2                       \n"
+      "vadd.u16     d4, d3, d7                   \n"
+
+      // dst_ptr[3] = (s[6] + s[7] + s[6+st] + s[7+st]) / 4
+      "vqrshrn.u16  d4, q2, #2                   \n"
+
+      // Shuffle 2,3 reg around so that 2 can be added to the
+      //  0,1 reg and 3 can be added to the 4,5 reg. This
+      //  requires expanding from u8 to u16 as the 0,1 and 4,5
+      //  registers are already expanded. Then do transposes
+      //  to get aligned.
+      // q2 = xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
+      "vmovl.u8     q1, d2                       \n"
+      "vmovl.u8     q3, d6                       \n"
+
+      // combine source lines
+      "vadd.u16     q1, q3                       \n"
+
+      // d4 = xx 20 xx 30 xx 22 xx 32
+      // d5 = xx 21 xx 31 xx 23 xx 33
+      "vtrn.u32     d2, d3                       \n"
+
+      // d4 = xx 20 xx 21 xx 22 xx 23
+      // d5 = xx 30 xx 31 xx 32 xx 33
+      "vtrn.u16     d2, d3                       \n"
+
+      // 0+1+2, 3+4+5
+      "vadd.u16     q0, q1                       \n"
+
+      // Need to divide, but can't downshift as the the value
+      //  isn't a power of 2. So multiply by 65536 / n
+      //  and take the upper 16 bits.
+      "vqrdmulh.s16 q0, q0, q13                  \n"
+
+      // Align for table lookup, vtbl requires registers to
+      //  be adjacent
+      "vmov.u8      d2, d4                       \n"
+
+      "vtbl.u8      d3, {d0, d1, d2}, d28        \n"
+      "vtbl.u8      d4, {d0, d1, d2}, d29        \n"
+
+      "vst1.8       {d3}, [%1]!                  \n"
+      "vst1.32      {d4[0]}, [%1]!               \n"
+      "bgt          1b                           \n"
+      : "+r"(src_ptr),       // %0
+        "+r"(dst_ptr),       // %1
+        "+r"(dst_width),     // %2
+        "+r"(src_stride)     // %3
+      : "r"(&kMult38_Div6),  // %4
+        "r"(&kShuf38_2)      // %5
+      : "q0", "q1", "q2", "q3", "q13", "q14", "memory", "cc");
+}
+
+void ScaleAddRows_NEON(const uint8* src_ptr,
+                       ptrdiff_t src_stride,
+                       uint16* dst_ptr,
+                       int src_width,
+                       int src_height) {
+  const uint8* src_tmp;
+  asm volatile(
+      "1:                                        \n"
+      "mov       %0, %1                          \n"
+      "mov       r12, %5                         \n"
+      "veor      q2, q2, q2                      \n"
+      "veor      q3, q3, q3                      \n"
+      "2:                                        \n"
+      // load 16 pixels into q0
+      "vld1.8     {q0}, [%0], %3                 \n"
+      "vaddw.u8   q3, q3, d1                     \n"
+      "vaddw.u8   q2, q2, d0                     \n"
+      "subs       r12, r12, #1                   \n"
+      "bgt        2b                             \n"
+      "vst1.16    {q2, q3}, [%2]!                \n"  // store pixels
+      "add        %1, %1, #16                    \n"
+      "subs       %4, %4, #16                    \n"  // 16 processed per loop
+      "bgt        1b                             \n"
+      : "=&r"(src_tmp),    // %0
+        "+r"(src_ptr),     // %1
+        "+r"(dst_ptr),     // %2
+        "+r"(src_stride),  // %3
+        "+r"(src_width),   // %4
+        "+r"(src_height)   // %5
+      :
+      : "memory", "cc", "r12", "q0", "q1", "q2", "q3"  // Clobber List
+      );
+}
+
+// TODO(Yang Zhang): Investigate less load instructions for
+// the x/dx stepping
+#define LOAD2_DATA8_LANE(n)                      \
+  "lsr        %5, %3, #16                    \n" \
+  "add        %6, %1, %5                     \n" \
+  "add        %3, %3, %4                     \n" \
+  "vld2.8     {d6[" #n "], d7[" #n "]}, [%6] \n"
+
+// The NEON version mimics this formula (from row_common.cc):
+// #define BLENDER(a, b, f) (uint8)((int)(a) +
+//    ((((int)((f)) * ((int)(b) - (int)(a))) + 0x8000) >> 16))
+
+void ScaleFilterCols_NEON(uint8* dst_ptr,
+                          const uint8* src_ptr,
+                          int dst_width,
+                          int x,
+                          int dx) {
+  int dx_offset[4] = {0, 1, 2, 3};
+  int* tmp = dx_offset;
+  const uint8* src_tmp = src_ptr;
+  asm volatile (
+    "vdup.32    q0, %3                         \n"  // x
+    "vdup.32    q1, %4                         \n"  // dx
+    "vld1.32    {q2}, [%5]                     \n"  // 0 1 2 3
+    "vshl.i32   q3, q1, #2                     \n"  // 4 * dx
+    "vmul.s32   q1, q1, q2                     \n"
+    // x         , x + 1 * dx, x + 2 * dx, x + 3 * dx
+    "vadd.s32   q1, q1, q0                     \n"
+    // x + 4 * dx, x + 5 * dx, x + 6 * dx, x + 7 * dx
+    "vadd.s32   q2, q1, q3                     \n"
+    "vshl.i32   q0, q3, #1                     \n"  // 8 * dx
+  "1:                                          \n"
+    LOAD2_DATA8_LANE(0)
+    LOAD2_DATA8_LANE(1)
+    LOAD2_DATA8_LANE(2)
+    LOAD2_DATA8_LANE(3)
+    LOAD2_DATA8_LANE(4)
+    LOAD2_DATA8_LANE(5)
+    LOAD2_DATA8_LANE(6)
+    LOAD2_DATA8_LANE(7)
+    "vmov       q10, q1                        \n"
+    "vmov       q11, q2                        \n"
+    "vuzp.16    q10, q11                       \n"
+    "vmovl.u8   q8, d6                         \n"
+    "vmovl.u8   q9, d7                         \n"
+    "vsubl.s16  q11, d18, d16                  \n"
+    "vsubl.s16  q12, d19, d17                  \n"
+    "vmovl.u16  q13, d20                       \n"
+    "vmovl.u16  q10, d21                       \n"
+    "vmul.s32   q11, q11, q13                  \n"
+    "vmul.s32   q12, q12, q10                  \n"
+    "vrshrn.s32  d18, q11, #16                 \n"
+    "vrshrn.s32  d19, q12, #16                 \n"
+    "vadd.s16   q8, q8, q9                     \n"
+    "vmovn.s16  d6, q8                         \n"
+
+    "vst1.8     {d6}, [%0]!                    \n"  // store pixels
+    "vadd.s32   q1, q1, q0                     \n"
+    "vadd.s32   q2, q2, q0                     \n"
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop
+    "bgt        1b                             \n"
+  : "+r"(dst_ptr),          // %0
+    "+r"(src_ptr),          // %1
+    "+r"(dst_width),        // %2
+    "+r"(x),                // %3
+    "+r"(dx),               // %4
+    "+r"(tmp),              // %5
+    "+r"(src_tmp)           // %6
+  :
+  : "memory", "cc", "q0", "q1", "q2", "q3",
+    "q8", "q9", "q10", "q11", "q12", "q13"
+  );
+}
+
+#undef LOAD2_DATA8_LANE
+
+// 16x2 -> 16x1
+void ScaleFilterRows_NEON(uint8* dst_ptr,
+                          const uint8* src_ptr,
+                          ptrdiff_t src_stride,
+                          int dst_width,
+                          int source_y_fraction) {
+  asm volatile(
+      "cmp          %4, #0                       \n"
+      "beq          100f                         \n"
+      "add          %2, %1                       \n"
+      "cmp          %4, #64                      \n"
+      "beq          75f                          \n"
+      "cmp          %4, #128                     \n"
+      "beq          50f                          \n"
+      "cmp          %4, #192                     \n"
+      "beq          25f                          \n"
+
+      "vdup.8       d5, %4                       \n"
+      "rsb          %4, #256                     \n"
+      "vdup.8       d4, %4                       \n"
+      // General purpose row blend.
+      "1:                                        \n"
+      "vld1.8       {q0}, [%1]!                  \n"
+      "vld1.8       {q1}, [%2]!                  \n"
+      "subs         %3, %3, #16                  \n"
+      "vmull.u8     q13, d0, d4                  \n"
+      "vmull.u8     q14, d1, d4                  \n"
+      "vmlal.u8     q13, d2, d5                  \n"
+      "vmlal.u8     q14, d3, d5                  \n"
+      "vrshrn.u16   d0, q13, #8                  \n"
+      "vrshrn.u16   d1, q14, #8                  \n"
+      "vst1.8       {q0}, [%0]!                  \n"
+      "bgt          1b                           \n"
+      "b            99f                          \n"
+
+      // Blend 25 / 75.
+      "25:                                       \n"
+      "vld1.8       {q0}, [%1]!                  \n"
+      "vld1.8       {q1}, [%2]!                  \n"
+      "subs         %3, %3, #16                  \n"
+      "vrhadd.u8    q0, q1                       \n"
+      "vrhadd.u8    q0, q1                       \n"
+      "vst1.8       {q0}, [%0]!                  \n"
+      "bgt          25b                          \n"
+      "b            99f                          \n"
+
+      // Blend 50 / 50.
+      "50:                                       \n"
+      "vld1.8       {q0}, [%1]!                  \n"
+      "vld1.8       {q1}, [%2]!                  \n"
+      "subs         %3, %3, #16                  \n"
+      "vrhadd.u8    q0, q1                       \n"
+      "vst1.8       {q0}, [%0]!                  \n"
+      "bgt          50b                          \n"
+      "b            99f                          \n"
+
+      // Blend 75 / 25.
+      "75:                                       \n"
+      "vld1.8       {q1}, [%1]!                  \n"
+      "vld1.8       {q0}, [%2]!                  \n"
+      "subs         %3, %3, #16                  \n"
+      "vrhadd.u8    q0, q1                       \n"
+      "vrhadd.u8    q0, q1                       \n"
+      "vst1.8       {q0}, [%0]!                  \n"
+      "bgt          75b                          \n"
+      "b            99f                          \n"
+
+      // Blend 100 / 0 - Copy row unchanged.
+      "100:                                      \n"
+      "vld1.8       {q0}, [%1]!                  \n"
+      "subs         %3, %3, #16                  \n"
+      "vst1.8       {q0}, [%0]!                  \n"
+      "bgt          100b                         \n"
+
+      "99:                                       \n"
+      "vst1.8       {d1[7]}, [%0]                \n"
+      : "+r"(dst_ptr),           // %0
+        "+r"(src_ptr),           // %1
+        "+r"(src_stride),        // %2
+        "+r"(dst_width),         // %3
+        "+r"(source_y_fraction)  // %4
+      :
+      : "q0", "q1", "d4", "d5", "q13", "q14", "memory", "cc");
+}
+
+void ScaleARGBRowDown2_NEON(const uint8* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8* dst,
+                            int dst_width) {
+  (void)src_stride;
+  asm volatile(
+      "1:                                        \n"
+      "vld4.32    {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+      "vld4.32    {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB
+      "subs       %2, %2, #8                     \n"  // 8 processed per loop
+      "vmov       q2, q1                         \n"  // load next 8 ARGB
+      "vst2.32    {q2, q3}, [%1]!                \n"  // store odd pixels
+      "bgt        1b                             \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(dst),       // %1
+        "+r"(dst_width)  // %2
+      :
+      : "memory", "cc", "q0", "q1", "q2", "q3"  // Clobber List
+      );
+}
+
+//  46:  f964 018d   vld4.32  {d16,d18,d20,d22}, [r4]!
+//  4a:  3e04        subs  r6, #4
+//  4c:  f964 118d   vld4.32  {d17,d19,d21,d23}, [r4]!
+//  50:  ef64 21f4   vorr  q9, q10, q10
+//  54:  f942 038d   vst2.32  {d16-d19}, [r2]!
+//  58:  d1f5        bne.n  46 <ScaleARGBRowDown2_C+0x46>
+
+void ScaleARGBRowDown2Linear_NEON(const uint8* src_argb,
+                                  ptrdiff_t src_stride,
+                                  uint8* dst_argb,
+                                  int dst_width) {
+  (void)src_stride;
+  asm volatile(
+      "1:                                        \n"
+      "vld4.32    {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+      "vld4.32    {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB
+      "subs       %2, %2, #8                     \n"  // 8 processed per loop
+      "vrhadd.u8  q0, q0, q1                     \n"  // rounding half add
+      "vrhadd.u8  q1, q2, q3                     \n"  // rounding half add
+      "vst2.32    {q0, q1}, [%1]!                \n"
+      "bgt       1b                              \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_argb),  // %1
+        "+r"(dst_width)  // %2
+      :
+      : "memory", "cc", "q0", "q1", "q2", "q3"  // Clobber List
+      );
+}
+
+void ScaleARGBRowDown2Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst,
+                               int dst_width) {
+  asm volatile(
+      // change the stride to row 2 pointer
+      "add        %1, %1, %0                     \n"
+      "1:                                        \n"
+      "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+      "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB
+      "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+      "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+      "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+      "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+      "vpaddl.u8  q3, q3                         \n"  // A 16 bytes -> 8 shorts.
+      "vld4.8     {d16, d18, d20, d22}, [%1]!    \n"  // load 8 more ARGB
+      "vld4.8     {d17, d19, d21, d23}, [%1]!    \n"  // load last 8 ARGB
+      "vpadal.u8  q0, q8                         \n"  // B 16 bytes -> 8 shorts.
+      "vpadal.u8  q1, q9                         \n"  // G 16 bytes -> 8 shorts.
+      "vpadal.u8  q2, q10                        \n"  // R 16 bytes -> 8 shorts.
+      "vpadal.u8  q3, q11                        \n"  // A 16 bytes -> 8 shorts.
+      "vrshrn.u16 d0, q0, #2                     \n"  // round and pack to bytes
+      "vrshrn.u16 d1, q1, #2                     \n"
+      "vrshrn.u16 d2, q2, #2                     \n"
+      "vrshrn.u16 d3, q3, #2                     \n"
+      "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"
+      "bgt        1b                             \n"
+      : "+r"(src_ptr),     // %0
+        "+r"(src_stride),  // %1
+        "+r"(dst),         // %2
+        "+r"(dst_width)    // %3
+      :
+      : "memory", "cc", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11");
+}
+
+// Reads 4 pixels at a time.
+// Alignment requirement: src_argb 4 byte aligned.
+void ScaleARGBRowDownEven_NEON(const uint8* src_argb,
+                               ptrdiff_t src_stride,
+                               int src_stepx,
+                               uint8* dst_argb,
+                               int dst_width) {
+  (void)src_stride;
+  asm volatile(
+      "mov        r12, %3, lsl #2                \n"
+      "1:                                        \n"
+      "vld1.32    {d0[0]}, [%0], r12             \n"
+      "vld1.32    {d0[1]}, [%0], r12             \n"
+      "vld1.32    {d1[0]}, [%0], r12             \n"
+      "vld1.32    {d1[1]}, [%0], r12             \n"
+      "subs       %2, %2, #4                     \n"  // 4 pixels per loop.
+      "vst1.8     {q0}, [%1]!                    \n"
+      "bgt        1b                             \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_argb),  // %1
+        "+r"(dst_width)  // %2
+      : "r"(src_stepx)   // %3
+      : "memory", "cc", "r12", "q0");
+}
+
+// Reads 4 pixels at a time.
+// Alignment requirement: src_argb 4 byte aligned.
+void ScaleARGBRowDownEvenBox_NEON(const uint8* src_argb,
+                                  ptrdiff_t src_stride,
+                                  int src_stepx,
+                                  uint8* dst_argb,
+                                  int dst_width) {
+  asm volatile(
+      "mov        r12, %4, lsl #2                \n"
+      "add        %1, %1, %0                     \n"
+      "1:                                        \n"
+      "vld1.8     {d0}, [%0], r12                \n"  // 4 2x2 blocks -> 2x1
+      "vld1.8     {d1}, [%1], r12                \n"
+      "vld1.8     {d2}, [%0], r12                \n"
+      "vld1.8     {d3}, [%1], r12                \n"
+      "vld1.8     {d4}, [%0], r12                \n"
+      "vld1.8     {d5}, [%1], r12                \n"
+      "vld1.8     {d6}, [%0], r12                \n"
+      "vld1.8     {d7}, [%1], r12                \n"
+      "vaddl.u8   q0, d0, d1                     \n"
+      "vaddl.u8   q1, d2, d3                     \n"
+      "vaddl.u8   q2, d4, d5                     \n"
+      "vaddl.u8   q3, d6, d7                     \n"
+      "vswp.8     d1, d2                         \n"  // ab_cd -> ac_bd
+      "vswp.8     d5, d6                         \n"  // ef_gh -> eg_fh
+      "vadd.u16   q0, q0, q1                     \n"  // (a+b)_(c+d)
+      "vadd.u16   q2, q2, q3                     \n"  // (e+f)_(g+h)
+      "vrshrn.u16 d0, q0, #2                     \n"  // first 2 pixels.
+      "vrshrn.u16 d1, q2, #2                     \n"  // next 2 pixels.
+      "subs       %3, %3, #4                     \n"  // 4 pixels per loop.
+      "vst1.8     {q0}, [%2]!                    \n"
+      "bgt        1b                             \n"
+      : "+r"(src_argb),    // %0
+        "+r"(src_stride),  // %1
+        "+r"(dst_argb),    // %2
+        "+r"(dst_width)    // %3
+      : "r"(src_stepx)     // %4
+      : "memory", "cc", "r12", "q0", "q1", "q2", "q3");
+}
+
+// TODO(Yang Zhang): Investigate less load instructions for
+// the x/dx stepping
+#define LOAD1_DATA32_LANE(dn, n)                 \
+  "lsr        %5, %3, #16                    \n" \
+  "add        %6, %1, %5, lsl #2             \n" \
+  "add        %3, %3, %4                     \n" \
+  "vld1.32    {" #dn "[" #n "]}, [%6]        \n"
+
+void ScaleARGBCols_NEON(uint8* dst_argb,
+                        const uint8* src_argb,
+                        int dst_width,
+                        int x,
+                        int dx) {
+  int tmp;
+  const uint8* src_tmp = src_argb;
+  asm volatile(
+      "1:                                        \n"
+      // clang-format off
+      LOAD1_DATA32_LANE(d0, 0)
+      LOAD1_DATA32_LANE(d0, 1)
+      LOAD1_DATA32_LANE(d1, 0)
+      LOAD1_DATA32_LANE(d1, 1)
+      LOAD1_DATA32_LANE(d2, 0)
+      LOAD1_DATA32_LANE(d2, 1)
+      LOAD1_DATA32_LANE(d3, 0)
+      LOAD1_DATA32_LANE(d3, 1)
+      // clang-format on
+      "vst1.32     {q0, q1}, [%0]!               \n"  // store pixels
+      "subs       %2, %2, #8                     \n"  // 8 processed per loop
+      "bgt        1b                             \n"
+      : "+r"(dst_argb),   // %0
+        "+r"(src_argb),   // %1
+        "+r"(dst_width),  // %2
+        "+r"(x),          // %3
+        "+r"(dx),         // %4
+        "=&r"(tmp),       // %5
+        "+r"(src_tmp)     // %6
+      :
+      : "memory", "cc", "q0", "q1");
+}
+
+#undef LOAD1_DATA32_LANE
+
+// TODO(Yang Zhang): Investigate less load instructions for
+// the x/dx stepping
+#define LOAD2_DATA32_LANE(dn1, dn2, n)                       \
+  "lsr        %5, %3, #16                                \n" \
+  "add        %6, %1, %5, lsl #2                         \n" \
+  "add        %3, %3, %4                                 \n" \
+  "vld2.32    {" #dn1 "[" #n "], " #dn2 "[" #n "]}, [%6] \n"
+
+void ScaleARGBFilterCols_NEON(uint8* dst_argb,
+                              const uint8* src_argb,
+                              int dst_width,
+                              int x,
+                              int dx) {
+  int dx_offset[4] = {0, 1, 2, 3};
+  int* tmp = dx_offset;
+  const uint8* src_tmp = src_argb;
+  asm volatile (
+    "vdup.32    q0, %3                         \n"  // x
+    "vdup.32    q1, %4                         \n"  // dx
+    "vld1.32    {q2}, [%5]                     \n"  // 0 1 2 3
+    "vshl.i32   q9, q1, #2                     \n"  // 4 * dx
+    "vmul.s32   q1, q1, q2                     \n"
+    "vmov.i8    q3, #0x7f                      \n"  // 0x7F
+    "vmov.i16   q15, #0x7f                     \n"  // 0x7F
+    // x         , x + 1 * dx, x + 2 * dx, x + 3 * dx
+    "vadd.s32   q8, q1, q0                     \n"
+  "1:                                          \n"
+    // d0, d1: a
+    // d2, d3: b
+    LOAD2_DATA32_LANE(d0, d2, 0)
+    LOAD2_DATA32_LANE(d0, d2, 1)
+    LOAD2_DATA32_LANE(d1, d3, 0)
+    LOAD2_DATA32_LANE(d1, d3, 1)
+    "vshrn.i32   d22, q8, #9                   \n"
+    "vand.16     d22, d22, d30                 \n"
+    "vdup.8      d24, d22[0]                   \n"
+    "vdup.8      d25, d22[2]                   \n"
+    "vdup.8      d26, d22[4]                   \n"
+    "vdup.8      d27, d22[6]                   \n"
+    "vext.8      d4, d24, d25, #4              \n"
+    "vext.8      d5, d26, d27, #4              \n"  // f
+    "veor.8      q10, q2, q3                   \n"  // 0x7f ^ f
+    "vmull.u8    q11, d0, d20                  \n"
+    "vmull.u8    q12, d1, d21                  \n"
+    "vmull.u8    q13, d2, d4                   \n"
+    "vmull.u8    q14, d3, d5                   \n"
+    "vadd.i16    q11, q11, q13                 \n"
+    "vadd.i16    q12, q12, q14                 \n"
+    "vshrn.i16   d0, q11, #7                   \n"
+    "vshrn.i16   d1, q12, #7                   \n"
+
+    "vst1.32     {d0, d1}, [%0]!               \n"  // store pixels
+    "vadd.s32    q8, q8, q9                    \n"
+    "subs        %2, %2, #4                    \n"  // 4 processed per loop
+    "bgt         1b                            \n"
+  : "+r"(dst_argb),         // %0
+    "+r"(src_argb),         // %1
+    "+r"(dst_width),        // %2
+    "+r"(x),                // %3
+    "+r"(dx),               // %4
+    "+r"(tmp),              // %5
+    "+r"(src_tmp)           // %6
+  :
+  : "memory", "cc", "q0", "q1", "q2", "q3", "q8", "q9",
+    "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+#undef LOAD2_DATA32_LANE
+
+#endif  // defined(__ARM_NEON__) && !defined(__aarch64__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif

pjproject-2.11.1/third_party/yuv/source/video_common.cc

@@ -0,0 +1,63 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/video_common.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#define ARRAY_SIZE(x) (int)(sizeof(x) / sizeof(x[0]))
+
+struct FourCCAliasEntry {
+  uint32 alias;
+  uint32 canonical;
+};
+
+static const struct FourCCAliasEntry kFourCCAliases[] = {
+    {FOURCC_IYUV, FOURCC_I420},
+    {FOURCC_YU12, FOURCC_I420},
+    {FOURCC_YU16, FOURCC_I422},
+    {FOURCC_YU24, FOURCC_I444},
+    {FOURCC_YUYV, FOURCC_YUY2},
+    {FOURCC_YUVS, FOURCC_YUY2},  // kCMPixelFormat_422YpCbCr8_yuvs
+    {FOURCC_HDYC, FOURCC_UYVY},
+    {FOURCC_2VUY, FOURCC_UYVY},  // kCMPixelFormat_422YpCbCr8
+    {FOURCC_JPEG, FOURCC_MJPG},  // Note: JPEG has DHT while MJPG does not.
+    {FOURCC_DMB1, FOURCC_MJPG},
+    {FOURCC_BA81, FOURCC_BGGR},  // deprecated.
+    {FOURCC_RGB3, FOURCC_RAW},
+    {FOURCC_BGR3, FOURCC_24BG},
+    {FOURCC_CM32, FOURCC_BGRA},  // kCMPixelFormat_32ARGB
+    {FOURCC_CM24, FOURCC_RAW},   // kCMPixelFormat_24RGB
+    {FOURCC_L555, FOURCC_RGBO},  // kCMPixelFormat_16LE555
+    {FOURCC_L565, FOURCC_RGBP},  // kCMPixelFormat_16LE565
+    {FOURCC_5551, FOURCC_RGBO},  // kCMPixelFormat_16LE5551
+};
+// TODO(fbarchard): Consider mapping kCMPixelFormat_32BGRA to FOURCC_ARGB.
+//  {FOURCC_BGRA, FOURCC_ARGB},  // kCMPixelFormat_32BGRA
+
+LIBYUV_API
+uint32 CanonicalFourCC(uint32 fourcc) {
+  int i;
+  for (i = 0; i < ARRAY_SIZE(kFourCCAliases); ++i) {
+    if (kFourCCAliases[i].alias == fourcc) {
+      return kFourCCAliases[i].canonical;
+    }
+  }
+  // Not an alias, so return it as-is.
+  return fourcc;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif