1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libyuv/util/psnr.cc Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,247 @@
1.4 +/*
1.5 + * Copyright 2013 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 "./psnr.h" // NOLINT
1.15 +
1.16 +#include <math.h>
1.17 +
1.18 +#ifdef _OPENMP
1.19 +#include <omp.h>
1.20 +#endif
1.21 +#ifdef _MSC_VER
1.22 +#include <intrin.h> // For __cpuid()
1.23 +#endif
1.24 +
1.25 +#ifdef __cplusplus
1.26 +extern "C" {
1.27 +#endif
1.28 +
1.29 +typedef unsigned int uint32; // NOLINT
1.30 +#ifdef _MSC_VER
1.31 +typedef unsigned __int64 uint64;
1.32 +#else // COMPILER_MSVC
1.33 +#if defined(__LP64__) && !defined(__OpenBSD__) && !defined(__APPLE__)
1.34 +typedef unsigned long uint64; // NOLINT
1.35 +#else // defined(__LP64__) && !defined(__OpenBSD__) && !defined(__APPLE__)
1.36 +typedef unsigned long long uint64; // NOLINT
1.37 +#endif // __LP64__
1.38 +#endif // _MSC_VER
1.39 +
1.40 +// PSNR formula: psnr = 10 * log10 (Peak Signal^2 * size / sse)
1.41 +double ComputePSNR(double sse, double size) {
1.42 + const double kMINSSE = 255.0 * 255.0 * size / pow(10., kMaxPSNR / 10.);
1.43 + if (sse <= kMINSSE)
1.44 + sse = kMINSSE; // Produces max PSNR of 128
1.45 + return 10.0 * log10(65025.0 * size / sse);
1.46 +}
1.47 +
1.48 +#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__)
1.49 +#define HAS_SUMSQUAREERROR_NEON
1.50 +static uint32 SumSquareError_NEON(const uint8* src_a,
1.51 + const uint8* src_b, int count) {
1.52 + volatile uint32 sse;
1.53 + asm volatile ( // NOLINT
1.54 + "vmov.u8 q7, #0 \n"
1.55 + "vmov.u8 q9, #0 \n"
1.56 + "vmov.u8 q8, #0 \n"
1.57 + "vmov.u8 q10, #0 \n"
1.58 +
1.59 + "1: \n"
1.60 + "vld1.u8 {q0}, [%0]! \n"
1.61 + "vld1.u8 {q1}, [%1]! \n"
1.62 + "vsubl.u8 q2, d0, d2 \n"
1.63 + "vsubl.u8 q3, d1, d3 \n"
1.64 + "vmlal.s16 q7, d4, d4 \n"
1.65 + "vmlal.s16 q8, d6, d6 \n"
1.66 + "vmlal.s16 q8, d5, d5 \n"
1.67 + "vmlal.s16 q10, d7, d7 \n"
1.68 + "subs %2, %2, #16 \n"
1.69 + "bhi 1b \n"
1.70 +
1.71 + "vadd.u32 q7, q7, q8 \n"
1.72 + "vadd.u32 q9, q9, q10 \n"
1.73 + "vadd.u32 q10, q7, q9 \n"
1.74 + "vpaddl.u32 q1, q10 \n"
1.75 + "vadd.u64 d0, d2, d3 \n"
1.76 + "vmov.32 %3, d0[0] \n"
1.77 + : "+r"(src_a),
1.78 + "+r"(src_b),
1.79 + "+r"(count),
1.80 + "=r"(sse)
1.81 + :
1.82 + : "memory", "cc", "q0", "q1", "q2", "q3", "q7", "q8", "q9", "q10");
1.83 + return sse;
1.84 +}
1.85 +#elif !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER)
1.86 +#define HAS_SUMSQUAREERROR_SSE2
1.87 +__declspec(naked)
1.88 +static uint32 SumSquareError_SSE2(const uint8* /*src_a*/,
1.89 + const uint8* /*src_b*/, int /*count*/) {
1.90 + __asm {
1.91 + mov eax, [esp + 4] // src_a
1.92 + mov edx, [esp + 8] // src_b
1.93 + mov ecx, [esp + 12] // count
1.94 + pxor xmm0, xmm0
1.95 + pxor xmm5, xmm5
1.96 + sub edx, eax
1.97 +
1.98 + wloop:
1.99 + movdqu xmm1, [eax]
1.100 + movdqu xmm2, [eax + edx]
1.101 + lea eax, [eax + 16]
1.102 + movdqu xmm3, xmm1
1.103 + psubusb xmm1, xmm2
1.104 + psubusb xmm2, xmm3
1.105 + por xmm1, xmm2
1.106 + movdqu xmm2, xmm1
1.107 + punpcklbw xmm1, xmm5
1.108 + punpckhbw xmm2, xmm5
1.109 + pmaddwd xmm1, xmm1
1.110 + pmaddwd xmm2, xmm2
1.111 + paddd xmm0, xmm1
1.112 + paddd xmm0, xmm2
1.113 + sub ecx, 16
1.114 + ja wloop
1.115 +
1.116 + pshufd xmm1, xmm0, 0EEh
1.117 + paddd xmm0, xmm1
1.118 + pshufd xmm1, xmm0, 01h
1.119 + paddd xmm0, xmm1
1.120 + movd eax, xmm0
1.121 + ret
1.122 + }
1.123 +}
1.124 +#elif !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
1.125 +#define HAS_SUMSQUAREERROR_SSE2
1.126 +static uint32 SumSquareError_SSE2(const uint8* src_a,
1.127 + const uint8* src_b, int count) {
1.128 + uint32 sse;
1.129 + asm volatile ( // NOLINT
1.130 + "pxor %%xmm0,%%xmm0 \n"
1.131 + "pxor %%xmm5,%%xmm5 \n"
1.132 + "sub %0,%1 \n"
1.133 +
1.134 + "1: \n"
1.135 + "movdqu (%0),%%xmm1 \n"
1.136 + "movdqu (%0,%1,1),%%xmm2 \n"
1.137 + "lea 0x10(%0),%0 \n"
1.138 + "movdqu %%xmm1,%%xmm3 \n"
1.139 + "psubusb %%xmm2,%%xmm1 \n"
1.140 + "psubusb %%xmm3,%%xmm2 \n"
1.141 + "por %%xmm2,%%xmm1 \n"
1.142 + "movdqu %%xmm1,%%xmm2 \n"
1.143 + "punpcklbw %%xmm5,%%xmm1 \n"
1.144 + "punpckhbw %%xmm5,%%xmm2 \n"
1.145 + "pmaddwd %%xmm1,%%xmm1 \n"
1.146 + "pmaddwd %%xmm2,%%xmm2 \n"
1.147 + "paddd %%xmm1,%%xmm0 \n"
1.148 + "paddd %%xmm2,%%xmm0 \n"
1.149 + "sub $0x10,%2 \n"
1.150 + "ja 1b \n"
1.151 +
1.152 + "pshufd $0xee,%%xmm0,%%xmm1 \n"
1.153 + "paddd %%xmm1,%%xmm0 \n"
1.154 + "pshufd $0x1,%%xmm0,%%xmm1 \n"
1.155 + "paddd %%xmm1,%%xmm0 \n"
1.156 + "movd %%xmm0,%3 \n"
1.157 +
1.158 + : "+r"(src_a), // %0
1.159 + "+r"(src_b), // %1
1.160 + "+r"(count), // %2
1.161 + "=g"(sse) // %3
1.162 + :
1.163 + : "memory", "cc"
1.164 +#if defined(__SSE2__)
1.165 + , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
1.166 +#endif
1.167 + ); // NOLINT
1.168 + return sse;
1.169 +}
1.170 +#endif // LIBYUV_DISABLE_X86 etc
1.171 +
1.172 +#if defined(HAS_SUMSQUAREERROR_SSE2)
1.173 +#if (defined(__pic__) || defined(__APPLE__)) && defined(__i386__)
1.174 +static __inline void __cpuid(int cpu_info[4], int info_type) {
1.175 + asm volatile ( // NOLINT
1.176 + "mov %%ebx, %%edi \n"
1.177 + "cpuid \n"
1.178 + "xchg %%edi, %%ebx \n"
1.179 + : "=a"(cpu_info[0]), "=D"(cpu_info[1]), "=c"(cpu_info[2]), "=d"(cpu_info[3])
1.180 + : "a"(info_type));
1.181 +}
1.182 +#elif defined(__i386__) || defined(__x86_64__)
1.183 +static __inline void __cpuid(int cpu_info[4], int info_type) {
1.184 + asm volatile ( // NOLINT
1.185 + "cpuid \n"
1.186 + : "=a"(cpu_info[0]), "=b"(cpu_info[1]), "=c"(cpu_info[2]), "=d"(cpu_info[3])
1.187 + : "a"(info_type));
1.188 +}
1.189 +#endif
1.190 +
1.191 +static int CpuHasSSE2() {
1.192 +#if defined(__i386__) || defined(__x86_64__) || defined(_M_IX86)
1.193 + int cpu_info[4];
1.194 + __cpuid(cpu_info, 1);
1.195 + if (cpu_info[3] & 0x04000000) {
1.196 + return 1;
1.197 + }
1.198 +#endif
1.199 + return 0;
1.200 +}
1.201 +#endif // HAS_SUMSQUAREERROR_SSE2
1.202 +
1.203 +static uint32 SumSquareError_C(const uint8* src_a,
1.204 + const uint8* src_b, int count) {
1.205 + uint32 sse = 0u;
1.206 + for (int x = 0; x < count; ++x) {
1.207 + int diff = src_a[x] - src_b[x];
1.208 + sse += static_cast<uint32>(diff * diff);
1.209 + }
1.210 + return sse;
1.211 +}
1.212 +
1.213 +double ComputeSumSquareError(const uint8* src_a,
1.214 + const uint8* src_b, int count) {
1.215 + uint32 (*SumSquareError)(const uint8* src_a,
1.216 + const uint8* src_b, int count) = SumSquareError_C;
1.217 +#if defined(HAS_SUMSQUAREERROR_NEON)
1.218 + SumSquareError = SumSquareError_NEON;
1.219 +#endif
1.220 +#if defined(HAS_SUMSQUAREERROR_SSE2)
1.221 + if (CpuHasSSE2()) {
1.222 + SumSquareError = SumSquareError_SSE2;
1.223 + }
1.224 +#endif
1.225 + const int kBlockSize = 1 << 15;
1.226 + uint64 sse = 0;
1.227 +#ifdef _OPENMP
1.228 +#pragma omp parallel for reduction(+: sse)
1.229 +#endif
1.230 + for (int i = 0; i < (count - (kBlockSize - 1)); i += kBlockSize) {
1.231 + sse += SumSquareError(src_a + i, src_b + i, kBlockSize);
1.232 + }
1.233 + src_a += count & ~(kBlockSize - 1);
1.234 + src_b += count & ~(kBlockSize - 1);
1.235 + int remainder = count & (kBlockSize - 1) & ~15;
1.236 + if (remainder) {
1.237 + sse += SumSquareError(src_a, src_b, remainder);
1.238 + src_a += remainder;
1.239 + src_b += remainder;
1.240 + }
1.241 + remainder = count & 15;
1.242 + if (remainder) {
1.243 + sse += SumSquareError_C(src_a, src_b, remainder);
1.244 + }
1.245 + return static_cast<double>(sse);
1.246 +}
1.247 +
1.248 +#ifdef __cplusplus
1.249 +} // extern "C"
1.250 +#endif
