coffee/libyuv
1
0
Fork 0
mirror of https://chromium.googlesource.com/libyuv/libyuv synced 2026-02-09 11:16:43 +08:00
Code Issues Packages Projects Releases Wiki Activity

HalfFloat neon armv7 fix for destination pointer.

Browse Source 
Improved unittests detect different in arm64 rounding.

TEST=util/android/test_runner.py gtest -s libyuv_unittest -t 7200 --verbose --release --gtest_filter=*Half* -a "--libyuv_width=640 --libyuv_height=360"
BUG=libyuv:560
R=wangcheng@google.com

Review URL: https://codereview.chromium.org/2478313004 .
This commit is contained in:
Frank Barchard 2016-11-07 12:13:04 -08:00
parent eca08525cb
commit f2c27dafa2
5 changed files with 73 additions and 30 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
README.chromium
View File

@ -1,6 +1,6 @@
Name: libyuv Name: libyuv
URL: http://code.google.com/p/libyuv/ URL: http://code.google.com/p/libyuv/
Version: 1633 Version: 1634
License: BSD License: BSD
License File: LICENSE License File: LICENSE

2
include/libyuv/version.h
View File

@ -11,6 +11,6 @@
#ifndef INCLUDE_LIBYUV_VERSION_H_ #ifndef INCLUDE_LIBYUV_VERSION_H_
#define INCLUDE_LIBYUV_VERSION_H_ #define INCLUDE_LIBYUV_VERSION_H_
#define LIBYUV_VERSION 1633 #define LIBYUV_VERSION 1634
#endif // INCLUDE_LIBYUV_VERSION_H_ #endif // INCLUDE_LIBYUV_VERSION_H_

16
source/row_neon.cc
View File

@ -2742,16 +2742,16 @@ void HalfFloat1Row_NEON(const uint16* src, uint16* dst, float, int width) {
MEMACCESS(0) MEMACCESS(0)
"vld1.8 {q1}, [%0]! \n" // load 8 shorts "vld1.8 {q1}, [%0]! \n" // load 8 shorts
"subs %2, %2, #8 \n" // 8 pixels per loop "subs %2, %2, #8 \n" // 8 pixels per loop
"vmovl.u8 q2, d2 \n" // 8 int's "vmovl.u16 q2, d2 \n" // 8 int's
"vmovl.u8 q3, d3 \n" "vmovl.u16 q3, d3 \n"
"vcvt.f32.u32 q2, q2 \n" // 8 floats "vcvt.f32.u32 q2, q2 \n" // 8 floats
"vcvt.f32.u32 q3, q3 \n" "vcvt.f32.u32 q3, q3 \n"
"vmul.f32 q2, q2, q0 \n" // adjust exponent "vmul.f32 q2, q2, q0 \n" // adjust exponent
"vmul.f32 q3, q3, q0 \n" "vmul.f32 q3, q3, q0 \n"
"vqshrn.u32 d2, q2, #13 \n" // isolate halffloat "vqshrn.u32 d2, q2, #13 \n" // isolate halffloat
"vqshrn.u32 d3, q3, #13 \n" "vqshrn.u32 d3, q3, #13 \n"
MEMACCESS(1) MEMACCESS(1)
"vst1.8 {q1}, [%0]! \n" "vst1.8 {q1}, [%1]! \n"
"bgt 1b \n" "bgt 1b \n"
: "+r"(src), // %0 : "+r"(src), // %0
"+r"(dst), // %1 "+r"(dst), // %1
@ -2770,16 +2770,16 @@ void HalfFloatRow_NEON(const uint16* src, uint16* dst, float scale, int width) {
MEMACCESS(0) MEMACCESS(0)
"vld1.8 {q1}, [%0]! \n" // load 8 shorts "vld1.8 {q1}, [%0]! \n" // load 8 shorts
"subs %2, %2, #8 \n" // 8 pixels per loop "subs %2, %2, #8 \n" // 8 pixels per loop
"vmovl.u8 q2, d2 \n" // 8 int's "vmovl.u16 q2, d2 \n" // 8 int's
"vmovl.u8 q3, d3 \n" "vmovl.u16 q3, d3 \n"
"vcvt.f32.u32 q2, q2 \n" // 8 floats "vcvt.f32.u32 q2, q2 \n" // 8 floats
"vcvt.f32.u32 q3, q3 \n" "vcvt.f32.u32 q3, q3 \n"
"vmul.f32 q2, q2, q0 \n" // adjust exponent "vmul.f32 q2, q2, q0 \n" // adjust exponent
"vmul.f32 q3, q3, q0 \n" "vmul.f32 q3, q3, q0 \n"
"vqshrn.u32 d2, q2, #13 \n" // isolate halffloat "vqshrn.u32 d2, q2, #13 \n" // isolate halffloat
"vqshrn.u32 d3, q3, #13 \n" "vqshrn.u32 d3, q3, #13 \n"
MEMACCESS(1) MEMACCESS(1)
"vst1.8 {q1}, [%0]! \n" "vst1.8 {q1}, [%1]! \n"
"bgt 1b \n" "bgt 1b \n"
: "+r"(src), // %0 : "+r"(src), // %0
"+r"(dst), // %1 "+r"(dst), // %1

3
source/row_neon64.cc
View File

@ -2711,6 +2711,7 @@ void SobelYRow_NEON(const uint8* src_y0, const uint8* src_y1,
); );
} }
// Caveat - rounds float to half float whereas scaling version truncates.
void HalfFloat1Row_NEON(const uint16* src, uint16* dst, float, int width) { void HalfFloat1Row_NEON(const uint16* src, uint16* dst, float, int width) {
asm volatile ( asm volatile (
"1: \n" "1: \n"
@ -2721,7 +2722,7 @@ void HalfFloat1Row_NEON(const uint16* src, uint16* dst, float, int width) {
"uxtl2 v3.4s, v1.8h \n" "uxtl2 v3.4s, v1.8h \n"
"scvtf v2.4s, v2.4s \n" // 8 floats "scvtf v2.4s, v2.4s \n" // 8 floats
"scvtf v3.4s, v3.4s \n" "scvtf v3.4s, v3.4s \n"
"fcvtn v1.4h, v2.4s \n" // 8 floatsgit "fcvtn v1.4h, v2.4s \n" // 8 half floats
"fcvtn2 v1.8h, v3.4s \n" "fcvtn2 v1.8h, v3.4s \n"
MEMACCESS(1) MEMACCESS(1)
"st1 {v1.16b}, [%1], #16 \n" // store 8 shorts "st1 {v1.16b}, [%1], #16 \n" // store 8 shorts

80
unit_test/planar_test.cc
View File

@ -2120,26 +2120,61 @@ int TestHalfFloatPlane(int benchmark_width, int benchmark_height,
} }
opt_time = (get_time() - opt_time) / benchmark_iterations; opt_time = (get_time() - opt_time) / benchmark_iterations;
int diff = 0; int max_diff = 0;
for (i = 0; i < y_plane_size; ++i) { for (i = 0; i < y_plane_size / 2; ++i) {
diff = dst_c[i] - dst_opt[i]; int abs_diff =
if (diff) break; abs(static_cast<int>(reinterpret_cast<uint16*>(dst_c)[i]) -
static_cast<int>(reinterpret_cast<uint16*>(dst_opt)[i]));
if (abs_diff > max_diff) {
max_diff = abs_diff;
}
} }
free_aligned_buffer_page_end(orig_y); free_aligned_buffer_page_end(orig_y);
return diff; return max_diff;
} }
#if defined(__arm__)
static void EnableFlushDenormalToZero(void) {
uint32_t cw;
__asm__ __volatile__ (
"vmrs %0, fpscr \n"
"orr %0, %0, #0x1000000 \n"
"vmsr fpscr, %0 \n"
: "=r"(cw) :: "memory");
}
#endif
// 5 bit exponent with bias of 15 will underflow to a denormal if scale causes // 5 bit exponent with bias of 15 will underflow to a denormal if scale causes
// exponent to be less than 0. 15 - log2(65536) = -1/ This shouldnt normally // exponent to be less than 0. 15 - log2(65536) = -1/ This shouldnt normally
// happen since scale is 1/(1<<bits) where bits is 9, 10 or 12. // happen since scale is 1/(1<<bits) where bits is 9, 10 or 12.
#define MAXHALFDIFF 0
TEST_F(LibYUVPlanarTest, TestHalfFloatPlane_16bit_denormal) { TEST_F(LibYUVPlanarTest, TestHalfFloatPlane_16bit_denormal) {
// 32 bit arm rounding on denormal case is off by 1 compared to C.
#if defined(__arm__)
EnableFlushDenormalToZero();
#endif
int diff = TestHalfFloatPlane(benchmark_width_, benchmark_height_, int diff = TestHalfFloatPlane(benchmark_width_, benchmark_height_,
benchmark_iterations_, benchmark_iterations_,
disable_cpu_flags_, benchmark_cpu_info_, disable_cpu_flags_, benchmark_cpu_info_,
1.0f / 65536.0f, 65535); 1.0f / 65536.0f, 65535);
EXPECT_LE(diff, MAXHALFDIFF); EXPECT_EQ(0, diff);
}
TEST_F(LibYUVPlanarTest, TestHalfFloatPlane_16bit_One) {
int diff = TestHalfFloatPlane(benchmark_width_, benchmark_height_,
benchmark_iterations_,
disable_cpu_flags_, benchmark_cpu_info_,
1.0f, 65535);
EXPECT_LE(diff, 1);
}
TEST_F(LibYUVPlanarTest, TestHalfFloatPlane_16bit_Opt) {
int diff = TestHalfFloatPlane(benchmark_width_, benchmark_height_,
benchmark_iterations_,
disable_cpu_flags_, benchmark_cpu_info_,
1.0f / 4096.0f, 65535);
EXPECT_EQ(0, diff);
} }
TEST_F(LibYUVPlanarTest, TestHalfFloatPlane_10bit_Opt) { TEST_F(LibYUVPlanarTest, TestHalfFloatPlane_10bit_Opt) {
@ -2147,7 +2182,7 @@ TEST_F(LibYUVPlanarTest, TestHalfFloatPlane_10bit_Opt) {
benchmark_iterations_, benchmark_iterations_,
disable_cpu_flags_, benchmark_cpu_info_, disable_cpu_flags_, benchmark_cpu_info_,
1.0f / 1024.0f, 1023); 1.0f / 1024.0f, 1023);
EXPECT_LE(diff, MAXHALFDIFF); EXPECT_EQ(0, diff);
} }
TEST_F(LibYUVPlanarTest, TestHalfFloatPlane_9bit_Opt) { TEST_F(LibYUVPlanarTest, TestHalfFloatPlane_9bit_Opt) {
@ -2155,7 +2190,7 @@ TEST_F(LibYUVPlanarTest, TestHalfFloatPlane_9bit_Opt) {
benchmark_iterations_, benchmark_iterations_,
disable_cpu_flags_, benchmark_cpu_info_, disable_cpu_flags_, benchmark_cpu_info_,
1.0f / 512.0f, 511); 1.0f / 512.0f, 511);
EXPECT_LE(diff, MAXHALFDIFF); EXPECT_EQ(0, diff);
} }
TEST_F(LibYUVPlanarTest, TestHalfFloatPlane_Opt) { TEST_F(LibYUVPlanarTest, TestHalfFloatPlane_Opt) {
@ -2163,15 +2198,7 @@ TEST_F(LibYUVPlanarTest, TestHalfFloatPlane_Opt) {
benchmark_iterations_, benchmark_iterations_,
disable_cpu_flags_, benchmark_cpu_info_, disable_cpu_flags_, benchmark_cpu_info_,
1.0f / 4096.0f, 4095); 1.0f / 4096.0f, 4095);
EXPECT_LE(diff, MAXHALFDIFF); EXPECT_EQ(0, diff);
}
TEST_F(LibYUVPlanarTest, TestHalfFloatPlane_One) {
int diff = TestHalfFloatPlane(benchmark_width_, benchmark_height_,
benchmark_iterations_,
disable_cpu_flags_, benchmark_cpu_info_,
1.0f, 4095);
EXPECT_LE(diff, MAXHALFDIFF);
} }
TEST_F(LibYUVPlanarTest, TestHalfFloatPlane_Offby1) { TEST_F(LibYUVPlanarTest, TestHalfFloatPlane_Offby1) {
@ -2179,9 +2206,24 @@ TEST_F(LibYUVPlanarTest, TestHalfFloatPlane_Offby1) {
benchmark_iterations_, benchmark_iterations_,
disable_cpu_flags_, benchmark_cpu_info_, disable_cpu_flags_, benchmark_cpu_info_,
1.0f / 4095.0f, 4095); 1.0f / 4095.0f, 4095);
EXPECT_LE(diff, MAXHALFDIFF); EXPECT_EQ(0, diff);
} }
TEST_F(LibYUVPlanarTest, TestHalfFloatPlane_One) {
int diff = TestHalfFloatPlane(benchmark_width_, benchmark_height_,
benchmark_iterations_,
disable_cpu_flags_, benchmark_cpu_info_,
1.0f, 2047);
EXPECT_EQ(0, diff);
}
TEST_F(LibYUVPlanarTest, TestHalfFloatPlane_12bit_One) {
int diff = TestHalfFloatPlane(benchmark_width_, benchmark_height_,
benchmark_iterations_,
disable_cpu_flags_, benchmark_cpu_info_,
1.0f, 4095);
EXPECT_LE(diff, 1);
}
TEST_F(LibYUVPlanarTest, TestARGBLumaColorTable) { TEST_F(LibYUVPlanarTest, TestARGBLumaColorTable) {
SIMD_ALIGNED(uint8 orig_pixels[1280][4]); SIMD_ALIGNED(uint8 orig_pixels[1280][4]);