coffee/libyuv
1
0
Fork 0
mirror of https://chromium.googlesource.com/libyuv/libyuv synced 2026-02-08 10:46:50 +08:00
Code Issues Packages Projects Releases Wiki Activity

Scale by even factor low level row function

Browse Source 
Bug: b/171884264
Change-Id: I6a94bde0aa05e681bb4590ea8beec33a61ddbfc9
Reviewed-on: https://chromium-review.googlesource.com/c/libyuv/libyuv/+/2518361
Reviewed-by: richard winterton <rrwinterton@gmail.com>
Commit-Queue: Frank Barchard <fbarchard@chromium.org>
This commit is contained in:
Frank Barchard 2020-11-03 11:25:56 -08:00 committed by Commit Bot
parent f014dbd87a
commit b7a1c5ee5d
17 changed files with 7698 additions and 7554 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
include/libyuv/scale_row.h
View File

@ -113,6 +113,7 @@ extern "C" {
#define HAS_SCALEROWDOWN38_NEON #define HAS_SCALEROWDOWN38_NEON
#define HAS_SCALEROWDOWN4_NEON #define HAS_SCALEROWDOWN4_NEON
#define HAS_SCALEUVROWDOWN2BOX_NEON #define HAS_SCALEUVROWDOWN2BOX_NEON
#define HAS_SCALEUVROWDOWNEVEN_NEON
#endif #endif
#if !defined(LIBYUV_DISABLE_MSA) && defined(__mips_msa) #if !defined(LIBYUV_DISABLE_MSA) && defined(__mips_msa)

6
source/rotate_neon.cc
View File

@ -266,8 +266,10 @@ void TransposeUVWx8_NEON(const uint8_t* src,
"vst1.8 {d21}, [%0] \n" "vst1.8 {d21}, [%0] \n"
"add %1, #8*2 \n" // src += 8*2 "add %1, #8*2 \n" // src += 8*2
"add %3, %3, %4, lsl #3 \n" // dst_a += 8 * dst_stride_a "add %3, %3, %4, lsl #3 \n" // dst_a += 8 *
"add %5, %5, %6, lsl #3 \n" // dst_b += 8 * dst_stride_b // dst_stride_a
"add %5, %5, %6, lsl #3 \n" // dst_b += 8 *
// dst_stride_b
"subs %7, #8 \n" // w -= 8 "subs %7, #8 \n" // w -= 8
"bge 1b \n" "bge 1b \n"

3
source/row_gcc.cc
View File

@ -6937,7 +6937,8 @@ void NV21ToYUV24Row_AVX2(const uint8_t* src_y,
"vmovdqu %5, %%ymm0 \n" // init blend value "vmovdqu %5, %%ymm0 \n" // init blend value
"vmovdqu %6, %%ymm1 \n" // init blend value "vmovdqu %6, %%ymm1 \n" // init blend value
"vmovdqu %7, %%ymm2 \n" // init blend value "vmovdqu %7, %%ymm2 \n" // init blend value
// "sub $0x20, %3 \n" //sub 32 from width for final loop // "sub $0x20, %3 \n" //sub 32 from
// width for final loop
LABELALIGN LABELALIGN
"1: \n" // label 1 "1: \n" // label 1

27
source/row_neon64.cc
View File

@ -906,7 +906,8 @@ void RGB24ToARGBRow_NEON(const uint8_t* src_rgb24,
asm volatile( asm volatile(
"movi v4.8b, #255 \n" // Alpha "movi v4.8b, #255 \n" // Alpha
"1: \n" "1: \n"
"ld3 {v1.8b,v2.8b,v3.8b}, [%0], #24 \n" // load 8 pixels of RGB24. "ld3 {v1.8b,v2.8b,v3.8b}, [%0], #24 \n" // load 8 pixels of
// RGB24.
"prfm pldl1keep, [%0, 448] \n" "prfm pldl1keep, [%0, 448] \n"
"subs %w2, %w2, #8 \n" // 8 processed per loop. "subs %w2, %w2, #8 \n" // 8 processed per loop.
"st4 {v1.8b,v2.8b,v3.8b,v4.8b}, [%1], #32 \n" // store 8 ARGB "st4 {v1.8b,v2.8b,v3.8b,v4.8b}, [%1], #32 \n" // store 8 ARGB
@ -1106,7 +1107,8 @@ void ARGBToRGB24Row_NEON(const uint8_t* src_argb,
"ld4 {v1.8b,v2.8b,v3.8b,v4.8b}, [%0], #32 \n" // load 8 ARGB "ld4 {v1.8b,v2.8b,v3.8b,v4.8b}, [%0], #32 \n" // load 8 ARGB
"prfm pldl1keep, [%0, 448] \n" "prfm pldl1keep, [%0, 448] \n"
"subs %w2, %w2, #8 \n" // 8 processed per loop. "subs %w2, %w2, #8 \n" // 8 processed per loop.
"st3 {v1.8b,v2.8b,v3.8b}, [%1], #24 \n" // store 8 pixels of RGB24 "st3 {v1.8b,v2.8b,v3.8b}, [%1], #24 \n" // store 8 pixels of
// RGB24
"b.gt 1b \n" "b.gt 1b \n"
: "+r"(src_argb), // %0 : "+r"(src_argb), // %0
"+r"(dst_rgb24), // %1 "+r"(dst_rgb24), // %1
@ -1339,7 +1341,8 @@ void ARGBToRGB565Row_NEON(const uint8_t* src_argb,
int width) { int width) {
asm volatile( asm volatile(
"1: \n" "1: \n"
"ld4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%0], #32 \n" // load 8 pixels "ld4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%0], #32 \n" // load 8
// pixels
"prfm pldl1keep, [%0, 448] \n" "prfm pldl1keep, [%0, 448] \n"
"subs %w2, %w2, #8 \n" // 8 processed per loop. "subs %w2, %w2, #8 \n" // 8 processed per loop.
ARGBTORGB565 ARGBTORGB565
@ -1359,7 +1362,8 @@ void ARGBToRGB565DitherRow_NEON(const uint8_t* src_argb,
asm volatile( asm volatile(
"dup v1.4s, %w2 \n" // dither4 "dup v1.4s, %w2 \n" // dither4
"1: \n" "1: \n"
"ld4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], #32 \n" // load 8 pixels "ld4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], #32 \n" // load 8
// pixels
"prfm pldl1keep, [%0, 448] \n" "prfm pldl1keep, [%0, 448] \n"
"subs %w3, %w3, #8 \n" // 8 processed per loop. "subs %w3, %w3, #8 \n" // 8 processed per loop.
"uqadd v20.8b, v20.8b, v1.8b \n" "uqadd v20.8b, v20.8b, v1.8b \n"
@ -1379,7 +1383,8 @@ void ARGBToARGB1555Row_NEON(const uint8_t* src_argb,
int width) { int width) {
asm volatile( asm volatile(
"1: \n" "1: \n"
"ld4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%0], #32 \n" // load 8 pixels "ld4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%0], #32 \n" // load 8
// pixels
"prfm pldl1keep, [%0, 448] \n" "prfm pldl1keep, [%0, 448] \n"
"subs %w2, %w2, #8 \n" // 8 processed per loop. "subs %w2, %w2, #8 \n" // 8 processed per loop.
ARGBTOARGB1555 ARGBTOARGB1555
@ -1399,7 +1404,8 @@ void ARGBToARGB4444Row_NEON(const uint8_t* src_argb,
"movi v4.16b, #0x0f \n" // bits to clear with "movi v4.16b, #0x0f \n" // bits to clear with
// vbic. // vbic.
"1: \n" "1: \n"
"ld4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%0], #32 \n" // load 8 pixels "ld4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%0], #32 \n" // load 8
// pixels
"prfm pldl1keep, [%0, 448] \n" "prfm pldl1keep, [%0, 448] \n"
"subs %w2, %w2, #8 \n" // 8 processed per loop. "subs %w2, %w2, #8 \n" // 8 processed per loop.
ARGBTOARGB4444 ARGBTOARGB4444
@ -2375,8 +2381,10 @@ void ARGBBlendRow_NEON(const uint8_t* src_argb0,
// Blend 1 pixels. // Blend 1 pixels.
"1: \n" "1: \n"
"ld4 {v0.b,v1.b,v2.b,v3.b}[0], [%0], #4 \n" // load 1 pixel ARGB0. "ld4 {v0.b,v1.b,v2.b,v3.b}[0], [%0], #4 \n" // load 1 pixel
"ld4 {v4.b,v5.b,v6.b,v7.b}[0], [%1], #4 \n" // load 1 pixel ARGB1. // ARGB0.
"ld4 {v4.b,v5.b,v6.b,v7.b}[0], [%1], #4 \n" // load 1 pixel
// ARGB1.
"prfm pldl1keep, [%0, 448] \n" "prfm pldl1keep, [%0, 448] \n"
"prfm pldl1keep, [%1, 448] \n" "prfm pldl1keep, [%1, 448] \n"
"subs %w3, %w3, #1 \n" // 1 processed per loop. "subs %w3, %w3, #1 \n" // 1 processed per loop.
@ -3183,7 +3191,8 @@ void GaussRow_F32_NEON(const float* src, float* dst, int width) {
"ld3r {v6.4s, v7.4s, v8.4s}, [%3] \n" // constants 4, 6, 1/256 "ld3r {v6.4s, v7.4s, v8.4s}, [%3] \n" // constants 4, 6, 1/256
"1: \n" "1: \n"
"ld1 {v0.4s, v1.4s, v2.4s}, [%0], %4\n" // load 12 samples, 5 rows "ld1 {v0.4s, v1.4s, v2.4s}, [%0], %4 \n" // load 12 samples, 5
// rows
"fadd v0.4s, v0.4s, v1.4s \n" // * 1 "fadd v0.4s, v0.4s, v1.4s \n" // * 1
"ld1 {v4.4s, v5.4s}, [%0], %5 \n" "ld1 {v4.4s, v5.4s}, [%0], %5 \n"
"fadd v1.4s, v1.4s, v2.4s \n" "fadd v1.4s, v1.4s, v2.4s \n"

7
source/scale_any.cc
View File

@ -490,6 +490,13 @@ SDAANY(ScaleARGBRowDownEvenBox_Any_MMI,
4, 4,
1) 1)
#endif #endif
#ifdef HAS_SCALEUVROWDOWNEVEN_NEON
SDAANY(ScaleUVRowDownEven_Any_NEON,
ScaleUVRowDownEven_NEON,
ScaleUVRowDownEven_C,
2,
3)
#endif
#ifdef SASIMDONLY #ifdef SASIMDONLY
// This also works and uses memcpy and SIMD instead of C, but is slower on ARM // This also works and uses memcpy and SIMD instead of C, but is slower on ARM

3
source/scale_gcc.cc
View File

@ -1323,7 +1323,8 @@ void ScaleARGBFilterCols_SSSE3(uint8_t* dst_argb,
"packuswb %%xmm0,%%xmm0 \n" "packuswb %%xmm0,%%xmm0 \n"
"movd %%xmm0,(%0) \n" "movd %%xmm0,(%0) \n"
LABELALIGN "99: \n" // clang-format error. LABELALIGN
"99: \n" // clang-format error.
: "+r"(dst_argb), // %0 : "+r"(dst_argb), // %0
"+r"(src_argb), // %1 "+r"(src_argb), // %1

29
source/scale_neon.cc
View File

@ -979,6 +979,35 @@ void ScaleUVRowDown2Box_NEON(const uint8_t* src_ptr,
: "memory", "cc", "q0", "q1", "q8", "q9"); : "memory", "cc", "q0", "q1", "q8", "q9");
} }
// Reads 4 pixels at a time.
void ScaleUVRowDownEven_NEON(const uint8_t* src_ptr,
ptrdiff_t src_stride,
int src_stepx, // pixel step
uint8_t* dst_ptr,
int dst_width) {
const uint8_t* src1_ptr = src_ptr + src_stepx * 2;
const uint8_t* src2_ptr = src_ptr + src_stepx * 4;
const uint8_t* src3_ptr = src_ptr + src_stepx * 6;
(void)src_stride;
asm volatile(
"1: \n"
"vld1.16 {d0[0]}, [%0], %6 \n"
"vld1.16 {d0[1]}, [%1], %6 \n"
"vld1.16 {d0[2]}, [%2], %6 \n"
"vld1.16 {d0[3]}, [%3], %6 \n"
"subs %5, %5, #4 \n" // 4 pixels per loop.
"vst1.8 {d0}, [%4]! \n"
"bgt 1b \n"
: "+r"(src_ptr), // %0
"+r"(src1_ptr), // %1
"+r"(src2_ptr), // %2
"+r"(src3_ptr), // %3
"+r"(dst_ptr), // %4
"+r"(dst_width) // %5
: "r"(src_stepx * 8) // %6
: "memory", "cc", "d0");
}
#endif // defined(__ARM_NEON__) && !defined(__aarch64__) #endif // defined(__ARM_NEON__) && !defined(__aarch64__)
#ifdef __cplusplus #ifdef __cplusplus

29
source/scale_neon64.cc
View File

@ -1115,6 +1115,35 @@ void ScaleUVRowDown2Box_NEON(const uint8_t* src_ptr,
: "memory", "cc", "v0", "v1", "v16", "v17"); : "memory", "cc", "v0", "v1", "v16", "v17");
} }
// Reads 4 pixels at a time.
void ScaleUVRowDownEven_NEON(const uint8_t* src_ptr,
ptrdiff_t src_stride,
int src_stepx, // pixel step
uint8_t* dst_ptr,
int dst_width) {
const uint8_t* src1_ptr = src_ptr + src_stepx * 2;
const uint8_t* src2_ptr = src_ptr + src_stepx * 4;
const uint8_t* src3_ptr = src_ptr + src_stepx * 6;
(void)src_stride;
asm volatile(
"1: \n"
"ld1 {v0.h}[0], [%0], %6 \n"
"ld1 {v1.h}[0], [%1], %6 \n"
"ld1 {v2.h}[0], [%2], %6 \n"
"ld1 {v3.h}[0], [%3], %6 \n"
"subs %w5, %w5, #4 \n" // 4 pixels per loop.
"st4 {v0.h, v1.h, v2.h, v3.h}[0], [%4], #8 \n"
"b.gt 1b \n"
: "+r"(src_ptr), // %0
"+r"(src1_ptr), // %1
"+r"(src2_ptr), // %2
"+r"(src3_ptr), // %3
"+r"(dst_ptr), // %4
"+r"(dst_width) // %5
: "r"((int64_t)(src_stepx * 8)) // %6
: "memory", "cc", "v0", "v1", "v2", "v3");
}
#endif // !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__) #endif // !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
#ifdef __cplusplus #ifdef __cplusplus

8
source/scale_uv.cc
View File

@ -299,6 +299,14 @@ static void ScaleUVDownEven(int src_width,
} }
#endif #endif
#if defined(HAS_SCALEUVROWDOWNEVEN_NEON) #if defined(HAS_SCALEUVROWDOWNEVEN_NEON)
if (TestCpuFlag(kCpuHasNEON) && !filtering) {
ScaleUVRowDownEven = ScaleUVRowDownEven_Any_NEON;
if (IS_ALIGNED(dst_width, 4)) {
ScaleUVRowDownEven = ScaleUVRowDownEven_NEON;
}
}
#endif// TODO(fbarchard): Enable Box filter
#if defined(HAS_SCALEUVROWDOWNEVENBOX_NEON)
if (TestCpuFlag(kCpuHasNEON)) { if (TestCpuFlag(kCpuHasNEON)) {
ScaleUVRowDownEven = filtering ? ScaleUVRowDownEvenBox_Any_NEON ScaleUVRowDownEven = filtering ? ScaleUVRowDownEvenBox_Any_NEON
: ScaleUVRowDownEven_Any_NEON; : ScaleUVRowDownEven_Any_NEON;

57
unit_test/scale_test.cc
View File

@ -1052,4 +1052,61 @@ TEST_FACTOR(3, 1, 3, 0)
#undef TEST_FACTOR #undef TEST_FACTOR
#undef SX #undef SX
#undef DX #undef DX
TEST_F(LibYUVScaleTest, PlaneTest3x) {
const int kSrcStride = 48;
const int kDstStride = 16;
const int kSize = kSrcStride * 3;
align_buffer_page_end(orig_pixels, kSize);
for (int i = 0; i < 48 * 3; ++i) {
orig_pixels[i] = i;
}
align_buffer_page_end(dest_pixels, kDstStride);
int iterations16 =
benchmark_width_ * benchmark_height_ / (16 * 1) * benchmark_iterations_;
for (int i = 0; i < iterations16; ++i) {
ScalePlane(orig_pixels, kSrcStride, 48, 3, dest_pixels, kDstStride, 16, 1,
kFilterBilinear);
}
EXPECT_EQ(49, dest_pixels[0]);
ScalePlane(orig_pixels, kSrcStride, 48, 3, dest_pixels, kDstStride, 16, 1,
kFilterNone);
EXPECT_EQ(49, dest_pixels[0]);
free_aligned_buffer_page_end(dest_pixels);
free_aligned_buffer_page_end(orig_pixels);
}
TEST_F(LibYUVScaleTest, PlaneTest4x) {
const int kSrcStride = 64;
const int kDstStride = 16;
const int kSize = kSrcStride * 4;
align_buffer_page_end(orig_pixels, kSize);
for (int i = 0; i < 64 * 4; ++i) {
orig_pixels[i] = i;
}
align_buffer_page_end(dest_pixels, kDstStride);
int iterations16 =
benchmark_width_ * benchmark_height_ / (16 * 1) * benchmark_iterations_;
for (int i = 0; i < iterations16; ++i) {
ScalePlane(orig_pixels, kSrcStride, 64, 4, dest_pixels, kDstStride, 16, 1,
kFilterBilinear);
}
EXPECT_EQ((65 + 66 + 129 + 130 + 2) / 4, dest_pixels[0]);
ScalePlane(orig_pixels, kSrcStride, 64, 4, dest_pixels, kDstStride, 16, 1,
kFilterNone);
EXPECT_EQ(130, dest_pixels[0]); // expect the 3rd pixel of the 3rd row
free_aligned_buffer_page_end(dest_pixels);
free_aligned_buffer_page_end(orig_pixels);
}
} // namespace libyuv } // namespace libyuv