michael@0: /*
michael@0:  *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
michael@0:  *
michael@0:  *  Use of this source code is governed by a BSD-style license
michael@0:  *  that can be found in the LICENSE file in the root of the source
michael@0:  *  tree. An additional intellectual property rights grant can be found
michael@0:  *  in the file PATENTS. All contributing project authors may
michael@0:  *  be found in the AUTHORS file in the root of the source tree.
michael@0:  */
michael@0: 
michael@0: #include "libyuv/compare.h"
michael@0: 
michael@0: #include <float.h>
michael@0: #include <math.h>
michael@0: #ifdef _OPENMP
michael@0: #include <omp.h>
michael@0: #endif
michael@0: 
michael@0: #include "libyuv/basic_types.h"
michael@0: #include "libyuv/cpu_id.h"
michael@0: #include "libyuv/row.h"
michael@0: 
michael@0: #ifdef __cplusplus
michael@0: namespace libyuv {
michael@0: extern "C" {
michael@0: #endif
michael@0: 
michael@0: // hash seed of 5381 recommended.
michael@0: // Internal C version of HashDjb2 with int sized count for efficiency.
michael@0: uint32 HashDjb2_C(const uint8* src, int count, uint32 seed);
michael@0: 
michael@0: // This module is for Visual C x86
michael@0: #if !defined(LIBYUV_DISABLE_X86) && \
michael@0:     (defined(_M_IX86) || \
michael@0:     (defined(__x86_64__) || (defined(__i386__) && !defined(__pic__))))
michael@0: #define HAS_HASHDJB2_SSE41
michael@0: uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed);
michael@0: 
michael@0: #if _MSC_VER >= 1700
michael@0: #define HAS_HASHDJB2_AVX2
michael@0: uint32 HashDjb2_AVX2(const uint8* src, int count, uint32 seed);
michael@0: #endif
michael@0: 
michael@0: #endif  // HAS_HASHDJB2_SSE41
michael@0: 
michael@0: // hash seed of 5381 recommended.
michael@0: LIBYUV_API
michael@0: uint32 HashDjb2(const uint8* src, uint64 count, uint32 seed) {
michael@0:   const int kBlockSize = 1 << 15;  // 32768;
michael@0:   int remainder;
michael@0:   uint32 (*HashDjb2_SSE)(const uint8* src, int count, uint32 seed) = HashDjb2_C;
michael@0: #if defined(HAS_HASHDJB2_SSE41)
michael@0:   if (TestCpuFlag(kCpuHasSSE41)) {
michael@0:     HashDjb2_SSE = HashDjb2_SSE41;
michael@0:   }
michael@0: #endif
michael@0: #if defined(HAS_HASHDJB2_AVX2)
michael@0:   if (TestCpuFlag(kCpuHasAVX2)) {
michael@0:     HashDjb2_SSE = HashDjb2_AVX2;
michael@0:   }
michael@0: #endif
michael@0: 
michael@0:   while (count >= (uint64)(kBlockSize)) {
michael@0:     seed = HashDjb2_SSE(src, kBlockSize, seed);
michael@0:     src += kBlockSize;
michael@0:     count -= kBlockSize;
michael@0:   }
michael@0:   remainder = (int)(count) & ~15;
michael@0:   if (remainder) {
michael@0:     seed = HashDjb2_SSE(src, remainder, seed);
michael@0:     src += remainder;
michael@0:     count -= remainder;
michael@0:   }
michael@0:   remainder = (int)(count) & 15;
michael@0:   if (remainder) {
michael@0:     seed = HashDjb2_C(src, remainder, seed);
michael@0:   }
michael@0:   return seed;
michael@0: }
michael@0: 
michael@0: uint32 SumSquareError_C(const uint8* src_a, const uint8* src_b, int count);
michael@0: #if !defined(LIBYUV_DISABLE_NEON) && \
michael@0:     (defined(__ARM_NEON__) || defined(LIBYUV_NEON))
michael@0: #define HAS_SUMSQUAREERROR_NEON
michael@0: uint32 SumSquareError_NEON(const uint8* src_a, const uint8* src_b, int count);
michael@0: #endif
michael@0: #if !defined(LIBYUV_DISABLE_X86) && \
michael@0:     (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__))
michael@0: #define HAS_SUMSQUAREERROR_SSE2
michael@0: uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count);
michael@0: #endif
michael@0: // Visual C 2012 required for AVX2.
michael@0: #if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && _MSC_VER >= 1700
michael@0: #define HAS_SUMSQUAREERROR_AVX2
michael@0: uint32 SumSquareError_AVX2(const uint8* src_a, const uint8* src_b, int count);
michael@0: #endif
michael@0: 
michael@0: // TODO(fbarchard): Refactor into row function.
michael@0: LIBYUV_API
michael@0: uint64 ComputeSumSquareError(const uint8* src_a, const uint8* src_b,
michael@0:                              int count) {
michael@0:   // SumSquareError returns values 0 to 65535 for each squared difference.
michael@0:   // Up to 65536 of those can be summed and remain within a uint32.
michael@0:   // After each block of 65536 pixels, accumulate into a uint64.
michael@0:   const int kBlockSize = 65536;
michael@0:   int remainder = count & (kBlockSize - 1) & ~31;
michael@0:   uint64 sse = 0;
michael@0:   int i;
michael@0:   uint32 (*SumSquareError)(const uint8* src_a, const uint8* src_b, int count) =
michael@0:       SumSquareError_C;
michael@0: #if defined(HAS_SUMSQUAREERROR_NEON)
michael@0:   if (TestCpuFlag(kCpuHasNEON)) {
michael@0:     SumSquareError = SumSquareError_NEON;
michael@0:   }
michael@0: #endif
michael@0: #if defined(HAS_SUMSQUAREERROR_SSE2)
michael@0:   if (TestCpuFlag(kCpuHasSSE2) &&
michael@0:       IS_ALIGNED(src_a, 16) && IS_ALIGNED(src_b, 16)) {
michael@0:     // Note only used for multiples of 16 so count is not checked.
michael@0:     SumSquareError = SumSquareError_SSE2;
michael@0:   }
michael@0: #endif
michael@0: #if defined(HAS_SUMSQUAREERROR_AVX2)
michael@0:   if (TestCpuFlag(kCpuHasAVX2)) {
michael@0:     // Note only used for multiples of 32 so count is not checked.
michael@0:     SumSquareError = SumSquareError_AVX2;
michael@0:   }
michael@0: #endif
michael@0: #ifdef _OPENMP
michael@0: #pragma omp parallel for reduction(+: sse)
michael@0: #endif
michael@0:   for (i = 0; i < (count - (kBlockSize - 1)); i += kBlockSize) {
michael@0:     sse += SumSquareError(src_a + i, src_b + i, kBlockSize);
michael@0:   }
michael@0:   src_a += count & ~(kBlockSize - 1);
michael@0:   src_b += count & ~(kBlockSize - 1);
michael@0:   if (remainder) {
michael@0:     sse += SumSquareError(src_a, src_b, remainder);
michael@0:     src_a += remainder;
michael@0:     src_b += remainder;
michael@0:   }
michael@0:   remainder = count & 31;
michael@0:   if (remainder) {
michael@0:     sse += SumSquareError_C(src_a, src_b, remainder);
michael@0:   }
michael@0:   return sse;
michael@0: }
michael@0: 
michael@0: LIBYUV_API
michael@0: uint64 ComputeSumSquareErrorPlane(const uint8* src_a, int stride_a,
michael@0:                                   const uint8* src_b, int stride_b,
michael@0:                                   int width, int height) {
michael@0:   uint64 sse = 0;
michael@0:   int h;
michael@0:   // Coalesce rows.
michael@0:   if (stride_a == width &&
michael@0:       stride_b == width) {
michael@0:     width *= height;
michael@0:     height = 1;
michael@0:     stride_a = stride_b = 0;
michael@0:   }
michael@0:   for (h = 0; h < height; ++h) {
michael@0:     sse += ComputeSumSquareError(src_a, src_b, width);
michael@0:     src_a += stride_a;
michael@0:     src_b += stride_b;
michael@0:   }
michael@0:   return sse;
michael@0: }
michael@0: 
michael@0: LIBYUV_API
michael@0: double SumSquareErrorToPsnr(uint64 sse, uint64 count) {
michael@0:   double psnr;
michael@0:   if (sse > 0) {
michael@0:     double mse = (double)(count) / (double)(sse);
michael@0:     psnr = 10.0 * log10(255.0 * 255.0 * mse);
michael@0:   } else {
michael@0:     psnr = kMaxPsnr;      // Limit to prevent divide by 0
michael@0:   }
michael@0: 
michael@0:   if (psnr > kMaxPsnr)
michael@0:     psnr = kMaxPsnr;
michael@0: 
michael@0:   return psnr;
michael@0: }
michael@0: 
michael@0: LIBYUV_API
michael@0: double CalcFramePsnr(const uint8* src_a, int stride_a,
michael@0:                      const uint8* src_b, int stride_b,
michael@0:                      int width, int height) {
michael@0:   const uint64 samples = width * height;
michael@0:   const uint64 sse = ComputeSumSquareErrorPlane(src_a, stride_a,
michael@0:                                                 src_b, stride_b,
michael@0:                                                 width, height);
michael@0:   return SumSquareErrorToPsnr(sse, samples);
michael@0: }
michael@0: 
michael@0: LIBYUV_API
michael@0: double I420Psnr(const uint8* src_y_a, int stride_y_a,
michael@0:                 const uint8* src_u_a, int stride_u_a,
michael@0:                 const uint8* src_v_a, int stride_v_a,
michael@0:                 const uint8* src_y_b, int stride_y_b,
michael@0:                 const uint8* src_u_b, int stride_u_b,
michael@0:                 const uint8* src_v_b, int stride_v_b,
michael@0:                 int width, int height) {
michael@0:   const uint64 sse_y = ComputeSumSquareErrorPlane(src_y_a, stride_y_a,
michael@0:                                                   src_y_b, stride_y_b,
michael@0:                                                   width, height);
michael@0:   const int width_uv = (width + 1) >> 1;
michael@0:   const int height_uv = (height + 1) >> 1;
michael@0:   const uint64 sse_u = ComputeSumSquareErrorPlane(src_u_a, stride_u_a,
michael@0:                                                   src_u_b, stride_u_b,
michael@0:                                                   width_uv, height_uv);
michael@0:   const uint64 sse_v = ComputeSumSquareErrorPlane(src_v_a, stride_v_a,
michael@0:                                                   src_v_b, stride_v_b,
michael@0:                                                   width_uv, height_uv);
michael@0:   const uint64 samples = width * height + 2 * (width_uv * height_uv);
michael@0:   const uint64 sse = sse_y + sse_u + sse_v;
michael@0:   return SumSquareErrorToPsnr(sse, samples);
michael@0: }
michael@0: 
michael@0: static const int64 cc1 =  26634;  // (64^2*(.01*255)^2
michael@0: static const int64 cc2 = 239708;  // (64^2*(.03*255)^2
michael@0: 
michael@0: static double Ssim8x8_C(const uint8* src_a, int stride_a,
michael@0:                         const uint8* src_b, int stride_b) {
michael@0:   int64 sum_a = 0;
michael@0:   int64 sum_b = 0;
michael@0:   int64 sum_sq_a = 0;
michael@0:   int64 sum_sq_b = 0;
michael@0:   int64 sum_axb = 0;
michael@0: 
michael@0:   int i;
michael@0:   for (i = 0; i < 8; ++i) {
michael@0:     int j;
michael@0:     for (j = 0; j < 8; ++j) {
michael@0:       sum_a += src_a[j];
michael@0:       sum_b += src_b[j];
michael@0:       sum_sq_a += src_a[j] * src_a[j];
michael@0:       sum_sq_b += src_b[j] * src_b[j];
michael@0:       sum_axb += src_a[j] * src_b[j];
michael@0:     }
michael@0: 
michael@0:     src_a += stride_a;
michael@0:     src_b += stride_b;
michael@0:   }
michael@0: 
michael@0:   {
michael@0:     const int64 count = 64;
michael@0:     // scale the constants by number of pixels
michael@0:     const int64 c1 = (cc1 * count * count) >> 12;
michael@0:     const int64 c2 = (cc2 * count * count) >> 12;
michael@0: 
michael@0:     const int64 sum_a_x_sum_b = sum_a * sum_b;
michael@0: 
michael@0:     const int64 ssim_n = (2 * sum_a_x_sum_b + c1) *
michael@0:                          (2 * count * sum_axb - 2 * sum_a_x_sum_b + c2);
michael@0: 
michael@0:     const int64 sum_a_sq = sum_a*sum_a;
michael@0:     const int64 sum_b_sq = sum_b*sum_b;
michael@0: 
michael@0:     const int64 ssim_d = (sum_a_sq + sum_b_sq + c1) *
michael@0:                          (count * sum_sq_a - sum_a_sq +
michael@0:                           count * sum_sq_b - sum_b_sq + c2);
michael@0: 
michael@0:     if (ssim_d == 0.0) {
michael@0:       return DBL_MAX;
michael@0:     }
michael@0:     return ssim_n * 1.0 / ssim_d;
michael@0:   }
michael@0: }
michael@0: 
michael@0: // We are using a 8x8 moving window with starting location of each 8x8 window
michael@0: // on the 4x4 pixel grid. Such arrangement allows the windows to overlap
michael@0: // block boundaries to penalize blocking artifacts.
michael@0: LIBYUV_API
michael@0: double CalcFrameSsim(const uint8* src_a, int stride_a,
michael@0:                      const uint8* src_b, int stride_b,
michael@0:                      int width, int height) {
michael@0:   int samples = 0;
michael@0:   double ssim_total = 0;
michael@0:   double (*Ssim8x8)(const uint8* src_a, int stride_a,
michael@0:                     const uint8* src_b, int stride_b) = Ssim8x8_C;
michael@0: 
michael@0:   // sample point start with each 4x4 location
michael@0:   int i;
michael@0:   for (i = 0; i < height - 8; i += 4) {
michael@0:     int j;
michael@0:     for (j = 0; j < width - 8; j += 4) {
michael@0:       ssim_total += Ssim8x8(src_a + j, stride_a, src_b + j, stride_b);
michael@0:       samples++;
michael@0:     }
michael@0: 
michael@0:     src_a += stride_a * 4;
michael@0:     src_b += stride_b * 4;
michael@0:   }
michael@0: 
michael@0:   ssim_total /= samples;
michael@0:   return ssim_total;
michael@0: }
michael@0: 
michael@0: LIBYUV_API
michael@0: double I420Ssim(const uint8* src_y_a, int stride_y_a,
michael@0:                 const uint8* src_u_a, int stride_u_a,
michael@0:                 const uint8* src_v_a, int stride_v_a,
michael@0:                 const uint8* src_y_b, int stride_y_b,
michael@0:                 const uint8* src_u_b, int stride_u_b,
michael@0:                 const uint8* src_v_b, int stride_v_b,
michael@0:                 int width, int height) {
michael@0:   const double ssim_y = CalcFrameSsim(src_y_a, stride_y_a,
michael@0:                                       src_y_b, stride_y_b, width, height);
michael@0:   const int width_uv = (width + 1) >> 1;
michael@0:   const int height_uv = (height + 1) >> 1;
michael@0:   const double ssim_u = CalcFrameSsim(src_u_a, stride_u_a,
michael@0:                                       src_u_b, stride_u_b,
michael@0:                                       width_uv, height_uv);
michael@0:   const double ssim_v = CalcFrameSsim(src_v_a, stride_v_a,
michael@0:                                       src_v_b, stride_v_b,
michael@0:                                       width_uv, height_uv);
michael@0:   return ssim_y * 0.8 + 0.1 * (ssim_u + ssim_v);
michael@0: }
michael@0: 
michael@0: #ifdef __cplusplus
michael@0: }  // extern "C"
michael@0: }  // namespace libyuv
michael@0: #endif