1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libyuv/source/compare.cc Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,325 @@
1.4 +/*
1.5 + * Copyright 2011 The LibYuv Project Authors. All rights reserved.
1.6 + *
1.7 + * Use of this source code is governed by a BSD-style license
1.8 + * that can be found in the LICENSE file in the root of the source
1.9 + * tree. An additional intellectual property rights grant can be found
1.10 + * in the file PATENTS. All contributing project authors may
1.11 + * be found in the AUTHORS file in the root of the source tree.
1.12 + */
1.13 +
1.14 +#include "libyuv/compare.h"
1.15 +
1.16 +#include <float.h>
1.17 +#include <math.h>
1.18 +#ifdef _OPENMP
1.19 +#include <omp.h>
1.20 +#endif
1.21 +
1.22 +#include "libyuv/basic_types.h"
1.23 +#include "libyuv/cpu_id.h"
1.24 +#include "libyuv/row.h"
1.25 +
1.26 +#ifdef __cplusplus
1.27 +namespace libyuv {
1.28 +extern "C" {
1.29 +#endif
1.30 +
1.31 +// hash seed of 5381 recommended.
1.32 +// Internal C version of HashDjb2 with int sized count for efficiency.
1.33 +uint32 HashDjb2_C(const uint8* src, int count, uint32 seed);
1.34 +
1.35 +// This module is for Visual C x86
1.36 +#if !defined(LIBYUV_DISABLE_X86) && \
1.37 + (defined(_M_IX86) || \
1.38 + (defined(__x86_64__) || (defined(__i386__) && !defined(__pic__))))
1.39 +#define HAS_HASHDJB2_SSE41
1.40 +uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed);
1.41 +
1.42 +#if _MSC_VER >= 1700
1.43 +#define HAS_HASHDJB2_AVX2
1.44 +uint32 HashDjb2_AVX2(const uint8* src, int count, uint32 seed);
1.45 +#endif
1.46 +
1.47 +#endif // HAS_HASHDJB2_SSE41
1.48 +
1.49 +// hash seed of 5381 recommended.
1.50 +LIBYUV_API
1.51 +uint32 HashDjb2(const uint8* src, uint64 count, uint32 seed) {
1.52 + const int kBlockSize = 1 << 15; // 32768;
1.53 + int remainder;
1.54 + uint32 (*HashDjb2_SSE)(const uint8* src, int count, uint32 seed) = HashDjb2_C;
1.55 +#if defined(HAS_HASHDJB2_SSE41)
1.56 + if (TestCpuFlag(kCpuHasSSE41)) {
1.57 + HashDjb2_SSE = HashDjb2_SSE41;
1.58 + }
1.59 +#endif
1.60 +#if defined(HAS_HASHDJB2_AVX2)
1.61 + if (TestCpuFlag(kCpuHasAVX2)) {
1.62 + HashDjb2_SSE = HashDjb2_AVX2;
1.63 + }
1.64 +#endif
1.65 +
1.66 + while (count >= (uint64)(kBlockSize)) {
1.67 + seed = HashDjb2_SSE(src, kBlockSize, seed);
1.68 + src += kBlockSize;
1.69 + count -= kBlockSize;
1.70 + }
1.71 + remainder = (int)(count) & ~15;
1.72 + if (remainder) {
1.73 + seed = HashDjb2_SSE(src, remainder, seed);
1.74 + src += remainder;
1.75 + count -= remainder;
1.76 + }
1.77 + remainder = (int)(count) & 15;
1.78 + if (remainder) {
1.79 + seed = HashDjb2_C(src, remainder, seed);
1.80 + }
1.81 + return seed;
1.82 +}
1.83 +
1.84 +uint32 SumSquareError_C(const uint8* src_a, const uint8* src_b, int count);
1.85 +#if !defined(LIBYUV_DISABLE_NEON) && \
1.86 + (defined(__ARM_NEON__) || defined(LIBYUV_NEON))
1.87 +#define HAS_SUMSQUAREERROR_NEON
1.88 +uint32 SumSquareError_NEON(const uint8* src_a, const uint8* src_b, int count);
1.89 +#endif
1.90 +#if !defined(LIBYUV_DISABLE_X86) && \
1.91 + (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__))
1.92 +#define HAS_SUMSQUAREERROR_SSE2
1.93 +uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count);
1.94 +#endif
1.95 +// Visual C 2012 required for AVX2.
1.96 +#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && _MSC_VER >= 1700
1.97 +#define HAS_SUMSQUAREERROR_AVX2
1.98 +uint32 SumSquareError_AVX2(const uint8* src_a, const uint8* src_b, int count);
1.99 +#endif
1.100 +
1.101 +// TODO(fbarchard): Refactor into row function.
1.102 +LIBYUV_API
1.103 +uint64 ComputeSumSquareError(const uint8* src_a, const uint8* src_b,
1.104 + int count) {
1.105 + // SumSquareError returns values 0 to 65535 for each squared difference.
1.106 + // Up to 65536 of those can be summed and remain within a uint32.
1.107 + // After each block of 65536 pixels, accumulate into a uint64.
1.108 + const int kBlockSize = 65536;
1.109 + int remainder = count & (kBlockSize - 1) & ~31;
1.110 + uint64 sse = 0;
1.111 + int i;
1.112 + uint32 (*SumSquareError)(const uint8* src_a, const uint8* src_b, int count) =
1.113 + SumSquareError_C;
1.114 +#if defined(HAS_SUMSQUAREERROR_NEON)
1.115 + if (TestCpuFlag(kCpuHasNEON)) {
1.116 + SumSquareError = SumSquareError_NEON;
1.117 + }
1.118 +#endif
1.119 +#if defined(HAS_SUMSQUAREERROR_SSE2)
1.120 + if (TestCpuFlag(kCpuHasSSE2) &&
1.121 + IS_ALIGNED(src_a, 16) && IS_ALIGNED(src_b, 16)) {
1.122 + // Note only used for multiples of 16 so count is not checked.
1.123 + SumSquareError = SumSquareError_SSE2;
1.124 + }
1.125 +#endif
1.126 +#if defined(HAS_SUMSQUAREERROR_AVX2)
1.127 + if (TestCpuFlag(kCpuHasAVX2)) {
1.128 + // Note only used for multiples of 32 so count is not checked.
1.129 + SumSquareError = SumSquareError_AVX2;
1.130 + }
1.131 +#endif
1.132 +#ifdef _OPENMP
1.133 +#pragma omp parallel for reduction(+: sse)
1.134 +#endif
1.135 + for (i = 0; i < (count - (kBlockSize - 1)); i += kBlockSize) {
1.136 + sse += SumSquareError(src_a + i, src_b + i, kBlockSize);
1.137 + }
1.138 + src_a += count & ~(kBlockSize - 1);
1.139 + src_b += count & ~(kBlockSize - 1);
1.140 + if (remainder) {
1.141 + sse += SumSquareError(src_a, src_b, remainder);
1.142 + src_a += remainder;
1.143 + src_b += remainder;
1.144 + }
1.145 + remainder = count & 31;
1.146 + if (remainder) {
1.147 + sse += SumSquareError_C(src_a, src_b, remainder);
1.148 + }
1.149 + return sse;
1.150 +}
1.151 +
1.152 +LIBYUV_API
1.153 +uint64 ComputeSumSquareErrorPlane(const uint8* src_a, int stride_a,
1.154 + const uint8* src_b, int stride_b,
1.155 + int width, int height) {
1.156 + uint64 sse = 0;
1.157 + int h;
1.158 + // Coalesce rows.
1.159 + if (stride_a == width &&
1.160 + stride_b == width) {
1.161 + width *= height;
1.162 + height = 1;
1.163 + stride_a = stride_b = 0;
1.164 + }
1.165 + for (h = 0; h < height; ++h) {
1.166 + sse += ComputeSumSquareError(src_a, src_b, width);
1.167 + src_a += stride_a;
1.168 + src_b += stride_b;
1.169 + }
1.170 + return sse;
1.171 +}
1.172 +
1.173 +LIBYUV_API
1.174 +double SumSquareErrorToPsnr(uint64 sse, uint64 count) {
1.175 + double psnr;
1.176 + if (sse > 0) {
1.177 + double mse = (double)(count) / (double)(sse);
1.178 + psnr = 10.0 * log10(255.0 * 255.0 * mse);
1.179 + } else {
1.180 + psnr = kMaxPsnr; // Limit to prevent divide by 0
1.181 + }
1.182 +
1.183 + if (psnr > kMaxPsnr)
1.184 + psnr = kMaxPsnr;
1.185 +
1.186 + return psnr;
1.187 +}
1.188 +
1.189 +LIBYUV_API
1.190 +double CalcFramePsnr(const uint8* src_a, int stride_a,
1.191 + const uint8* src_b, int stride_b,
1.192 + int width, int height) {
1.193 + const uint64 samples = width * height;
1.194 + const uint64 sse = ComputeSumSquareErrorPlane(src_a, stride_a,
1.195 + src_b, stride_b,
1.196 + width, height);
1.197 + return SumSquareErrorToPsnr(sse, samples);
1.198 +}
1.199 +
1.200 +LIBYUV_API
1.201 +double I420Psnr(const uint8* src_y_a, int stride_y_a,
1.202 + const uint8* src_u_a, int stride_u_a,
1.203 + const uint8* src_v_a, int stride_v_a,
1.204 + const uint8* src_y_b, int stride_y_b,
1.205 + const uint8* src_u_b, int stride_u_b,
1.206 + const uint8* src_v_b, int stride_v_b,
1.207 + int width, int height) {
1.208 + const uint64 sse_y = ComputeSumSquareErrorPlane(src_y_a, stride_y_a,
1.209 + src_y_b, stride_y_b,
1.210 + width, height);
1.211 + const int width_uv = (width + 1) >> 1;
1.212 + const int height_uv = (height + 1) >> 1;
1.213 + const uint64 sse_u = ComputeSumSquareErrorPlane(src_u_a, stride_u_a,
1.214 + src_u_b, stride_u_b,
1.215 + width_uv, height_uv);
1.216 + const uint64 sse_v = ComputeSumSquareErrorPlane(src_v_a, stride_v_a,
1.217 + src_v_b, stride_v_b,
1.218 + width_uv, height_uv);
1.219 + const uint64 samples = width * height + 2 * (width_uv * height_uv);
1.220 + const uint64 sse = sse_y + sse_u + sse_v;
1.221 + return SumSquareErrorToPsnr(sse, samples);
1.222 +}
1.223 +
1.224 +static const int64 cc1 = 26634; // (64^2*(.01*255)^2
1.225 +static const int64 cc2 = 239708; // (64^2*(.03*255)^2
1.226 +
1.227 +static double Ssim8x8_C(const uint8* src_a, int stride_a,
1.228 + const uint8* src_b, int stride_b) {
1.229 + int64 sum_a = 0;
1.230 + int64 sum_b = 0;
1.231 + int64 sum_sq_a = 0;
1.232 + int64 sum_sq_b = 0;
1.233 + int64 sum_axb = 0;
1.234 +
1.235 + int i;
1.236 + for (i = 0; i < 8; ++i) {
1.237 + int j;
1.238 + for (j = 0; j < 8; ++j) {
1.239 + sum_a += src_a[j];
1.240 + sum_b += src_b[j];
1.241 + sum_sq_a += src_a[j] * src_a[j];
1.242 + sum_sq_b += src_b[j] * src_b[j];
1.243 + sum_axb += src_a[j] * src_b[j];
1.244 + }
1.245 +
1.246 + src_a += stride_a;
1.247 + src_b += stride_b;
1.248 + }
1.249 +
1.250 + {
1.251 + const int64 count = 64;
1.252 + // scale the constants by number of pixels
1.253 + const int64 c1 = (cc1 * count * count) >> 12;
1.254 + const int64 c2 = (cc2 * count * count) >> 12;
1.255 +
1.256 + const int64 sum_a_x_sum_b = sum_a * sum_b;
1.257 +
1.258 + const int64 ssim_n = (2 * sum_a_x_sum_b + c1) *
1.259 + (2 * count * sum_axb - 2 * sum_a_x_sum_b + c2);
1.260 +
1.261 + const int64 sum_a_sq = sum_a*sum_a;
1.262 + const int64 sum_b_sq = sum_b*sum_b;
1.263 +
1.264 + const int64 ssim_d = (sum_a_sq + sum_b_sq + c1) *
1.265 + (count * sum_sq_a - sum_a_sq +
1.266 + count * sum_sq_b - sum_b_sq + c2);
1.267 +
1.268 + if (ssim_d == 0.0) {
1.269 + return DBL_MAX;
1.270 + }
1.271 + return ssim_n * 1.0 / ssim_d;
1.272 + }
1.273 +}
1.274 +
1.275 +// We are using a 8x8 moving window with starting location of each 8x8 window
1.276 +// on the 4x4 pixel grid. Such arrangement allows the windows to overlap
1.277 +// block boundaries to penalize blocking artifacts.
1.278 +LIBYUV_API
1.279 +double CalcFrameSsim(const uint8* src_a, int stride_a,
1.280 + const uint8* src_b, int stride_b,
1.281 + int width, int height) {
1.282 + int samples = 0;
1.283 + double ssim_total = 0;
1.284 + double (*Ssim8x8)(const uint8* src_a, int stride_a,
1.285 + const uint8* src_b, int stride_b) = Ssim8x8_C;
1.286 +
1.287 + // sample point start with each 4x4 location
1.288 + int i;
1.289 + for (i = 0; i < height - 8; i += 4) {
1.290 + int j;
1.291 + for (j = 0; j < width - 8; j += 4) {
1.292 + ssim_total += Ssim8x8(src_a + j, stride_a, src_b + j, stride_b);
1.293 + samples++;
1.294 + }
1.295 +
1.296 + src_a += stride_a * 4;
1.297 + src_b += stride_b * 4;
1.298 + }
1.299 +
1.300 + ssim_total /= samples;
1.301 + return ssim_total;
1.302 +}
1.303 +
1.304 +LIBYUV_API
1.305 +double I420Ssim(const uint8* src_y_a, int stride_y_a,
1.306 + const uint8* src_u_a, int stride_u_a,
1.307 + const uint8* src_v_a, int stride_v_a,
1.308 + const uint8* src_y_b, int stride_y_b,
1.309 + const uint8* src_u_b, int stride_u_b,
1.310 + const uint8* src_v_b, int stride_v_b,
1.311 + int width, int height) {
1.312 + const double ssim_y = CalcFrameSsim(src_y_a, stride_y_a,
1.313 + src_y_b, stride_y_b, width, height);
1.314 + const int width_uv = (width + 1) >> 1;
1.315 + const int height_uv = (height + 1) >> 1;
1.316 + const double ssim_u = CalcFrameSsim(src_u_a, stride_u_a,
1.317 + src_u_b, stride_u_b,
1.318 + width_uv, height_uv);
1.319 + const double ssim_v = CalcFrameSsim(src_v_a, stride_v_a,
1.320 + src_v_b, stride_v_b,
1.321 + width_uv, height_uv);
1.322 + return ssim_y * 0.8 + 0.1 * (ssim_u + ssim_v);
1.323 +}
1.324 +
1.325 +#ifdef __cplusplus
1.326 +} // extern "C"
1.327 +} // namespace libyuv
1.328 +#endif
