coffee/libyuv
1
0
Fork 0
mirror of https://chromium.googlesource.com/libyuv/libyuv synced 2026-10-28 18:07:12 +08:00
Code Issues Packages Projects Releases Wiki Activity
libyuv/source/compare.cc
Frank Barchard 700f153cb8 [libyuv] Add Qualcomm Hexagon HVX SIMD support 
Implement Qualcomm Hexagon Vector Extensions (HVX) 128-byte SIMD
optimizations and runtime CPU detection for Hexagon DSP targets:
- CPU detection: Add kCpuHasHexagon and kCpuHasHVX flags,
  LIBYUV_DISABLE_HVX environment variable support, and Hexagon
  unit_test mmap/munmap shims.
- Rotate/Transpose (rotate_hvx.cc): Implement TransposeWxH_HVX,
  TransposeUVWxH_HVX, TransposeWxH_16_HVX, and TransposeWxH_32_HVX
  using power-of-2 vdeal/vshuff butterfly networks.
- Compare (compare_hvx.cc): Implement HammingDistance_HVX,
  SumSquareError_HVX, HashDjb2_HVX, and Ssim4x4Row_HVX.
- Scale (scale_hvx.cc): Implement 16 scale row kernels including
  ScaleAddRow_HVX, ScaleRowDown2*_HVX (8-bit and 16-bit),
  ScaleRowDown4*_HVX, ScaleARGBRowDown2*_HVX, and
  ScaleUVRowDown2*_HVX.
- Row functions (row_hvx.cc): Implement 58 row kernels covering
  copy/set/mirror, split/merge, 8/16-bit conversions, YUV-to-ARGB,
  ARGB-to-YUV, ARGB channel swizzles, blend, attenuate, and
  interpolate using hardware-predicated partial loads/stores.

R=rrwinterton@gmail.com

Bug: None
Test: libyuv_unittest on Hexagon v81 simulator and Snapdragon 8 Elite
Change-Id: Ib6713de51f6246d95e58c95e629ff03f18e3b91e
Reviewed-on: https://chromium-review.googlesource.com/c/libyuv/libyuv/+/8604586
Reviewed-by: richard winterton <rrwinterton@gmail.com>
2026-10-19 14:16:18 -07:00

627 lines
18 KiB
C++
Raw Blame History

/*
* 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 <limits.h>
#include <math.h>
#ifdef _OPENMP
#include <omp.h>
#endif
#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_t HashDjb2(const uint8_t* src, uint64_t count, uint32_t seed) {
const int kBlockSize = 1 << 15; // 32768;
int simd_size = 16; // SIMD for multiple of 16, and C for remainder.
int remainder;
uint32_t (*HashDjb2_SSE)(const uint8_t* src, int count, uint32_t 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;
simd_size = 32;
}
#endif
#if defined(HAS_HASHDJB2_NEON)
if (TestCpuFlag(kCpuHasNEON)) {
HashDjb2_SSE = HashDjb2_NEON;
}
#endif
#if defined(HAS_HASHDJB2_AVX512BW)
if (TestCpuFlag(kCpuHasAVX512BW)) {
HashDjb2_SSE = HashDjb2_AVX512BW;
simd_size = 1; // Any count.
}
#endif
#if defined(HAS_HASHDJB2_WASMSIMD)
if (TestCpuFlag(kCpuHasWASMSIMD)) {
HashDjb2_SSE = HashDjb2_WASMSIMD;
}
#endif
#if defined(HAS_HASHDJB2_HVX)
if (TestCpuFlag(kCpuHasHVX)) {
HashDjb2_SSE = HashDjb2_HVX;
simd_size = 1; // Any count.
}
#endif
while (count >= (uint64_t)kBlockSize) {
seed = HashDjb2_SSE(src, kBlockSize, seed);
src += kBlockSize;
count -= kBlockSize;
}
remainder = (int)count & ~(simd_size - 1);
if (remainder) {
seed = HashDjb2_SSE(src, remainder, seed);
src += remainder;
count -= remainder;
}
remainder = (int)count & (simd_size - 1);
if (remainder) {
seed = HashDjb2_C(src, remainder, seed);
}
return seed;
}
static uint32_t ARGBDetectRow_C(const uint8_t* 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) { // Fourth 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 fourth 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_t ARGBDetect(const uint8_t* argb,
int stride_argb,
int width,
int height) {
uint32_t fourcc = 0;
int h;
if (!argb || width <= 0 || height <= 0) {
return fourcc;
}
// Coalesce rows.
if (stride_argb == width * 4 && (ptrdiff_t)width * height <= INT_MAX) {
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_t ComputeHammingDistance(const uint8_t* src_a,
const uint8_t* src_b,
int count) {
const int kBlockSize = 1 << 15; // 32768;
int simd_size = 64; // SIMD for multiple of 64, and C for remainder.
int remainder;
uint64_t diff = 0;
int i;
uint32_t (*HammingDistance)(const uint8_t* src_a, const uint8_t* src_b,
int count) = HammingDistance_C;
#if defined(HAS_HAMMINGDISTANCE_NEON)
if (TestCpuFlag(kCpuHasNEON)) {
HammingDistance = HammingDistance_NEON;
}
#endif
#if defined(HAS_HAMMINGDISTANCE_NEON_DOTPROD)
if (TestCpuFlag(kCpuHasNeonDotProd)) {
HammingDistance = HammingDistance_NEON_DotProd;
}
#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_AVX512BW)
if (TestCpuFlag(kCpuHasAVX512BW)) {
HammingDistance = HammingDistance_AVX512BW;
simd_size = 1; // Any count.
}
#endif
#if defined(HAS_HAMMINGDISTANCE_WASMSIMD)
if (TestCpuFlag(kCpuHasWASMSIMD)) {
HammingDistance = HammingDistance_WASMSIMD;
}
#endif
#if defined(HAS_HAMMINGDISTANCE_HVX)
if (TestCpuFlag(kCpuHasHVX)) {
HammingDistance = HammingDistance_HVX;
simd_size = 1; // Any count.
}
#endif
remainder = count & (kBlockSize - 1) & ~(simd_size - 1);
#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 & (simd_size - 1);
if (remainder) {
diff += HammingDistance_C(src_a, src_b, remainder);
}
return diff;
}
// TODO(fbarchard): Refactor into row function.
LIBYUV_API
uint64_t ComputeSumSquareError(const uint8_t* src_a,
const uint8_t* 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_t.
// After each block of 65536 pixels, accumulate into a uint64_t.
const int kBlockSize = 65536;
int simd_size = 32; // SIMD for multiple of 32, and C for remainder.
int remainder;
uint64_t sse = 0;
int i;
uint32_t (*SumSquareError)(const uint8_t* src_a, const uint8_t* src_b,
int count) = SumSquareError_C;
#if defined(HAS_SUMSQUAREERROR_NEON)
if (TestCpuFlag(kCpuHasNEON)) {
SumSquareError = SumSquareError_NEON;
}
#endif
#if defined(HAS_SUMSQUAREERROR_NEON_DOTPROD)
if (TestCpuFlag(kCpuHasNeonDotProd)) {
SumSquareError = SumSquareError_NEON_DotProd;
}
#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_AVX512BW)
if (TestCpuFlag(kCpuHasAVX512BW)) {
SumSquareError = SumSquareError_AVX512BW;
simd_size = 1; // Any count.
}
#endif
#if defined(HAS_SUMSQUAREERROR_WASMSIMD)
if (TestCpuFlag(kCpuHasWASMSIMD)) {
SumSquareError = SumSquareError_WASMSIMD;
}
#endif
#if defined(HAS_SUMSQUAREERROR_HVX)
if (TestCpuFlag(kCpuHasHVX)) {
SumSquareError = SumSquareError_HVX;
simd_size = 1; // Any count.
}
#endif
remainder = count & (kBlockSize - 1) & ~(simd_size - 1);
#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 & (simd_size - 1);
if (remainder) {
sse += SumSquareError_C(src_a, src_b, remainder);
}
return sse;
}
LIBYUV_API
uint64_t ComputeSumSquareErrorPlane(const uint8_t* src_a,
int stride_a,
const uint8_t* src_b,
int stride_b,
int width,
int height) {
uint64_t sse = 0;
int h;
if (!src_a || !src_b || width <= 0 || height <= 0) {
return sse;
}
// Coalesce rows.
if (stride_a == width && stride_b == width &&
(ptrdiff_t)width * height <= INT_MAX) {
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_t sse, uint64_t 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_t* src_a,
int stride_a,
const uint8_t* src_b,
int stride_b,
int width,
int height) {
const uint64_t samples = (uint64_t)width * (uint64_t)height;
const uint64_t sse = ComputeSumSquareErrorPlane(src_a, stride_a, src_b,
stride_b, width, height);
return SumSquareErrorToPsnr(sse, samples);
}
LIBYUV_API
double I420Psnr(const uint8_t* src_y_a,
int stride_y_a,
const uint8_t* src_u_a,
int stride_u_a,
const uint8_t* src_v_a,
int stride_v_a,
const uint8_t* src_y_b,
int stride_y_b,
const uint8_t* src_u_b,
int stride_u_b,
const uint8_t* src_v_b,
int stride_v_b,
int width,
int height) {
const uint64_t 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_t sse_u = ComputeSumSquareErrorPlane(
src_u_a, stride_u_a, src_u_b, stride_u_b, width_uv, height_uv);
const uint64_t sse_v = ComputeSumSquareErrorPlane(
src_v_a, stride_v_a, src_v_b, stride_v_b, width_uv, height_uv);
const uint64_t samples = (uint64_t)width * (uint64_t)height +
2 * ((uint64_t)width_uv * (uint64_t)height_uv);
const uint64_t sse = sse_y + sse_u + sse_v;
return SumSquareErrorToPsnr(sse, samples);
}
static const int64_t cc1 = 26634; // (64^2*(.01*255)^2
static const int64_t cc2 = 239708; // (64^2*(.03*255)^2
static double Ssim8x8_C(const uint8_t* src_a,
int stride_a,
const uint8_t* src_b,
int stride_b) {
int64_t sum_a = 0;
int64_t sum_b = 0;
int64_t sum_sq_a = 0;
int64_t sum_sq_b = 0;
int64_t 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_t count = 64;
// scale the constants by number of pixels
const int64_t c1 = (cc1 * count * count) >> 12;
const int64_t c2 = (cc2 * count * count) >> 12;
const int64_t sum_a_x_sum_b = sum_a * sum_b;
const int64_t ssim_n = (2 * sum_a_x_sum_b + c1) *
(2 * count * sum_axb - 2 * sum_a_x_sum_b + c2);
const int64_t sum_a_sq = sum_a * sum_a;
const int64_t sum_b_sq = sum_b * sum_b;
const int64_t 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) {
return DBL_MAX;
}
return (double)ssim_n / (double)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_t* src_a,
int stride_a,
const uint8_t* src_b,
int stride_b,
int width,
int height) {
int samples = 0;
double ssim_total = 0;
if (width <= 8 || height <= 8) {
return ssim_total / samples;
}
int simd_size = 4;
void (*Ssim4x4Row)(const uint8_t* src_a, int stride_a, const uint8_t* src_b,
int stride_b, uint32_t* dst, int width) = Ssim4x4Row_C;
#if defined(HAS_SSIM4X4ROW_SSE2)
if (TestCpuFlag(kCpuHasSSE2)) {
Ssim4x4Row = Ssim4x4Row_SSE2;
simd_size = 16;
}
#endif
#if defined(HAS_SSIM4X4ROW_AVX2)
if (TestCpuFlag(kCpuHasAVX2)) {
Ssim4x4Row = Ssim4x4Row_AVX2;
simd_size = 32;
}
#endif
#if defined(HAS_SSIM4X4ROW_AVX512BW)
if (TestCpuFlag(kCpuHasAVX512BW)) {
Ssim4x4Row = Ssim4x4Row_AVX512BW;
simd_size = 4;
}
#endif
#if defined(HAS_SSIM4X4ROW_NEON)
if (TestCpuFlag(kCpuHasNEON)) {
Ssim4x4Row = Ssim4x4Row_NEON;
simd_size = 16;
}
#endif
#if defined(HAS_SSIM4X4ROW_NEON_DOTPROD)
if (TestCpuFlag(kCpuHasNeonDotProd)) {
Ssim4x4Row = Ssim4x4Row_NEON_DotProd;
simd_size = 16;
}
#endif
#if defined(HAS_SSIM4X4ROW_SVE2)
if (TestCpuFlag(kCpuHasSVE2)) {
Ssim4x4Row = Ssim4x4Row_SVE2;
simd_size = 4;
}
#endif
#if defined(HAS_SSIM4X4ROW_SME)
if (TestCpuFlag(kCpuHasSME)) {
Ssim4x4Row = Ssim4x4Row_SME;
simd_size = 4;
}
#endif
#if defined(HAS_SSIM4X4ROW_RVV)
if (TestCpuFlag(kCpuHasRVV)) {
Ssim4x4Row = Ssim4x4Row_RVV;
simd_size = 4;
}
#endif
#if defined(HAS_SSIM4X4ROW_LSX)
if (TestCpuFlag(kCpuHasLSX)) {
Ssim4x4Row = Ssim4x4Row_LSX;
simd_size = 16;
}
#endif
#if defined(HAS_SSIM4X4ROW_LASX)
if (TestCpuFlag(kCpuHasLASX)) {
Ssim4x4Row = Ssim4x4Row_LASX;
simd_size = 32;
}
#endif
#if defined(HAS_SSIM4X4ROW_WASMSIMD)
if (TestCpuFlag(kCpuHasWASMSIMD)) {
Ssim4x4Row = Ssim4x4Row_WASMSIMD;
simd_size = 16;
}
#endif
#if defined(HAS_SSIM4X4ROW_HVX)
if (TestCpuFlag(kCpuHasHVX)) {
Ssim4x4Row = Ssim4x4Row_HVX;
simd_size = 4;
}
#endif
const int num_windows = (width - 8 + 3) >> 2;
const int num_tiles = num_windows + 1;
const int tile_width = num_tiles * 4;
const int main_width = tile_width & ~(simd_size - 1);
const int rem_width = tile_width & (simd_size - 1);
#ifdef _OPENMP
#pragma omp parallel reduction(+ : ssim_total, samples)
#endif
{
align_buffer_64(row_buf, (size_t)tile_width * 2 * sizeof(uint32_t));
if (row_buf) {
uint32_t* row0 = (uint32_t*)row_buf;
uint32_t* row1 = row0 + tile_width;
int prev_i = -8;
int i;
#ifdef _OPENMP
#pragma omp for schedule(static)
#endif
for (i = 0; i < height - 8; i += 4) {
const uint8_t* a = src_a + (ptrdiff_t)i * stride_a;
const uint8_t* b = src_b + (ptrdiff_t)i * stride_b;
if (prev_i + 4 != i) {
if (main_width > 0) {
Ssim4x4Row(a, stride_a, b, stride_b, row0, main_width);
}
if (rem_width > 0) {
Ssim4x4Row_C(a + main_width, stride_a, b + main_width, stride_b,
row0 + main_width, rem_width);
}
}
a += (ptrdiff_t)stride_a * 4;
b += (ptrdiff_t)stride_b * 4;
if (main_width > 0) {
Ssim4x4Row(a, stride_a, b, stride_b, row1, main_width);
}
if (rem_width > 0) {
Ssim4x4Row_C(a + main_width, stride_a, b + main_width, stride_b,
row1 + main_width, rem_width);
}
ssim_total += SsimWindowRow_C(row0, row1, num_windows);
samples += num_windows;
uint32_t* tmp = row0;
row0 = row1;
row1 = tmp;
prev_i = i;
}
free_aligned_buffer_64(row_buf);
} else {
int i;
#ifdef _OPENMP
#pragma omp for schedule(static)
#endif
for (i = 0; i < height - 8; i += 4) {
const uint8_t* a = src_a + (ptrdiff_t)i * stride_a;
const uint8_t* b = src_b + (ptrdiff_t)i * stride_b;
int j;
for (j = 0; j < width - 8; j += 4) {
ssim_total += Ssim8x8_C(a + j, stride_a, b + j, stride_b);
samples++;
}
}
}
}
ssim_total /= samples;
return ssim_total;
}
LIBYUV_API
double I420Ssim(const uint8_t* src_y_a,
int stride_y_a,
const uint8_t* src_u_a,
int stride_u_a,
const uint8_t* src_v_a,
int stride_v_a,
const uint8_t* src_y_b,
int stride_y_b,
const uint8_t* src_u_b,
int stride_u_b,
const uint8_t* 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
Reference in New Issue View Git Blame Copy Permalink