Skip to content

L
libyuv
Commit b14f46fa authored by fbarchard@google.com's avatar fbarchard@google.com
Browse files
Options

NaCL pepper_33 port of scale and compare using lock/unlock. Remove less useful… 

NaCL pepper_33 port of scale and compare using lock/unlock.  Remove less useful scaling tests and change default size to a multiple of 16 for better assembly coverage.
BUG=none
TESTED=ncval
R=nfullagar@google.com

Review URL: https://webrtc-codereview.appspot.com/5939005

git-svn-id: http://libyuv.googlecode.com/svn/trunk@917 16f28f9a-4ce2-e073-06de-1de4eb20be90
parent f2bd3153
Hide whitespace changes
Inline Side-by-side
Showing with 113 additions and 186 deletions
include/libyuv/row.h
View file @ b14f46fa
...@@ -369,6 +369,67 @@ typedef uint8 uvec8[16]; ...@@ -369,6 +369,67 @@ typedef uint8 uvec8[16];
#define OMITFP __attribute__((optimize("omit-frame-pointer"))) #define OMITFP __attribute__((optimize("omit-frame-pointer")))
#endif #endif
// NaCL macros for GCC x86 and x64.
// TODO(nfullagar): When pepper_33 toolchain is distributed, default to
// NEW_BINUTILS and remove all BUNDLEALIGN occurances.
#if defined(__native_client__) && defined(__x86_64__)
#if defined(NEW_BINUTILS)
#define BUNDLELOCK ".bundle_lock\n"
#define BUNDLEUNLOCK ".bundle_unlock\n"
#define BUNDLEALIGN "\n"
#else
#define BUNDLELOCK "\n"
#define BUNDLEUNLOCK "\n"
#define BUNDLEALIGN ".p2align 5\n"
#endif
#define LABELALIGN ".p2align 5\n"
#define MEMACCESS(base) "%%nacl:(%%r15,%q" #base ")"
#define MEMACCESS2(offset, base) "%%nacl:" #offset "(%%r15,%q" #base ")"
#define MEMLEA(offset, base) #offset "(%q" #base ")"
#define MEMLEA3(offset, index, scale) \
#offset "(,%q" #index "," #scale ")"
#define MEMLEA4(offset, base, index, scale) \
#offset "(%q" #base ",%q" #index "," #scale ")"
#define MEMMOVESTRING(s, d) "%%nacl:(%q" #s "),%%nacl:(%q" #d "), %%r15"
#define MEMSTORESTRING(reg, d) "%%" #reg ",%%nacl:(%q" #d "), %%r15"
#define MEMOPREG(opcode, offset, base, index, scale, reg) \
BUNDLELOCK \
"lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
#opcode " (%%r15,%%r14),%%" #reg "\n" \
BUNDLEUNLOCK
#define MEMOPMEM(opcode, reg, offset, base, index, scale) \
BUNDLELOCK \
"lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
#opcode " %%" #reg ",(%%r15,%%r14)\n" \
BUNDLEUNLOCK
#define MEMOPARG(opcode, offset, base, index, scale, arg) \
BUNDLELOCK \
"lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
#opcode " (%%r15,%%r14),%" #arg "\n" \
BUNDLEUNLOCK
#else
#define BUNDLELOCK "\n"
#define BUNDLEUNLOCK "\n"
#define BUNDLEALIGN "\n"
#define LABELALIGN ".p2align 2\n"
#define MEMACCESS(base) "(%" #base ")"
#define MEMACCESS2(offset, base) #offset "(%" #base ")"
#define MEMLEA(offset, base) #offset "(%" #base ")"
#define MEMLEA3(offset, index, scale) \
#offset "(,%" #index "," #scale ")"
#define MEMLEA4(offset, base, index, scale) \
#offset "(%" #base ",%" #index "," #scale ")"
#define MEMMOVESTRING(s, d)
#define MEMSTORESTRING(reg, d)
#define MEMOPREG(opcode, offset, base, index, scale, reg) \
#opcode " " #offset "(%" #base ",%" #index "," #scale "),%%" #reg "\n"
#define MEMOPMEM(opcode, reg, offset, base, index, scale) \
#opcode " %%" #reg ","#offset "(%" #base ",%" #index "," #scale ")\n"
#define MEMOPARG(opcode, offset, base, index, scale, arg) \
#opcode " " #offset "(%" #base ",%" #index "," #scale "),%" #arg "\n"
#endif
// For functions that use rowbuffer and have runtime checks for overflow, // For functions that use rowbuffer and have runtime checks for overflow,
// use SAFEBUFFERS to avoid additional check. // use SAFEBUFFERS to avoid additional check.
#if defined(_MSC_VER) && (_MSC_FULL_VER >= 160040219) #if defined(_MSC_VER) && (_MSC_FULL_VER >= 160040219)
......
source/compare_posix.cc
View file @ b14f46fa
...@@ -18,21 +18,13 @@ extern "C" { ...@@ -18,21 +18,13 @@ extern "C" {
#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__)) #if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
#if defined(__native_client__) && defined(__x86_64__)
#define MEMACCESS(base) "%%nacl:(%%r15,%q" #base ")"
#define MEMLEA(offset, base) #offset "(%q" #base ")"
#else
#define MEMACCESS(base) "(%" #base ")"
#define MEMLEA(offset, base) #offset "(%" #base ")"
#endif
uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count) { uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count) {
uint32 sse; uint32 sse;
asm volatile ( // NOLINT asm volatile ( // NOLINT
"pxor %%xmm0,%%xmm0 \n" "pxor %%xmm0,%%xmm0 \n"
"pxor %%xmm5,%%xmm5 \n" "pxor %%xmm5,%%xmm5 \n"
".p2align 2 \n" LABELALIGN
"1: \n" "1: \n"
"movdqa " MEMACCESS(0) ",%%xmm1 \n" "movdqa " MEMACCESS(0) ",%%xmm1 \n"
"lea " MEMLEA(0x10, 0) ",%0 \n" "lea " MEMLEA(0x10, 0) ",%0 \n"
"movdqa " MEMACCESS(1) ",%%xmm2 \n" "movdqa " MEMACCESS(1) ",%%xmm2 \n"
...@@ -107,7 +99,7 @@ uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed) { ...@@ -107,7 +99,7 @@ uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed) {
"movd %2,%%xmm0 \n" "movd %2,%%xmm0 \n"
"pxor %%xmm7,%%xmm7 \n" "pxor %%xmm7,%%xmm7 \n"
"movdqa %4,%%xmm6 \n" "movdqa %4,%%xmm6 \n"
".p2align 2 \n" LABELALIGN
"1: \n" "1: \n"
"movdqu " MEMACCESS(0) ",%%xmm1 \n" "movdqu " MEMACCESS(0) ",%%xmm1 \n"
"lea " MEMLEA(0x10, 0) ",%0 \n" "lea " MEMLEA(0x10, 0) ",%0 \n"
......
source/row_posix.cc
View file @ b14f46fa
...@@ -18,61 +18,6 @@ extern "C" { ...@@ -18,61 +18,6 @@ extern "C" {
// This module is for GCC x86 and x64. // This module is for GCC x86 and x64.
#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__)) #if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
#if defined(__native_client__) && defined(__x86_64__)
// TODO(nfullagar): When pepper_33 toolchain is distributed, default to
// NEW_BINUTILS and remove all BUNDLEALIGN occurances.
#if defined(NEW_BINUTILS)
#define BUNDLELOCK ".bundle_lock\n"
#define BUNDLEUNLOCK ".bundle_unlock\n"
#define BUNDLEALIGN "\n"
#else
#define BUNDLELOCK "\n"
#define BUNDLEUNLOCK "\n"
#define BUNDLEALIGN ".p2align 5\n"
#endif
#define MEMACCESS(base) "%%nacl:(%%r15,%q" #base ")"
#define MEMACCESS2(offset, base) "%%nacl:" #offset "(%%r15,%q" #base ")"
#define MEMLEA(offset, base) #offset "(%q" #base ")"
#define MEMLEA4(offset, base, index, scale) \
#offset "(%q" #base ",%q" #index "," #scale ")"
#define MEMMOVESTRING(s, d) "%%nacl:(%q" #s "),%%nacl:(%q" #d "), %%r15"
#define MEMSTORESTRING(reg, d) "%%" #reg ",%%nacl:(%q" #d "), %%r15"
#define MEMOPREG(opcode, offset, base, index, scale, reg) \
BUNDLELOCK \
"lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
#opcode " (%%r15,%%r14),%%" #reg "\n" \
BUNDLEUNLOCK
#define MEMOPMEM(opcode, reg, offset, base, index, scale) \
BUNDLELOCK \
"lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
#opcode " %%" #reg ",(%%r15,%%r14)\n" \
BUNDLEUNLOCK
#define MEMOPARG(opcode, offset, base, index, scale, arg) \
BUNDLELOCK \
"lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
#opcode " (%%r15,%%r14),%" #arg "\n" \
BUNDLEUNLOCK
#define LABELALIGN ".p2align 5\n"
#else
#define BUNDLELOCK "\n"
#define BUNDLEUNLOCK "\n"
#define BUNDLEALIGN "\n"
#define MEMACCESS(base) "(%" #base ")"
#define MEMACCESS2(offset, base) #offset "(%" #base ")"
#define MEMLEA(offset, base) #offset "(%" #base ")"
#define MEMLEA4(offset, base, index, scale) \
#offset "(%" #base ",%" #index "," #scale ")"
#define MEMMOVESTRING(s, d)
#define MEMSTORESTRING(reg, d)
#define MEMOPREG(opcode, offset, base, index, scale, reg) \
#opcode " " #offset "(%" #base ",%" #index "," #scale "),%%" #reg "\n"
#define MEMOPMEM(opcode, reg, offset, base, index, scale) \
#opcode " %%" #reg ","#offset "(%" #base ",%" #index "," #scale ")\n"
#define MEMOPARG(opcode, offset, base, index, scale, arg) \
#opcode " " #offset "(%" #base ",%" #index "," #scale "),%" #arg "\n"
#define LABELALIGN ".p2align 2\n"
#endif
#if defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_ARGBGRAYROW_SSSE3) #if defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_ARGBGRAYROW_SSSE3)
// Constants for ARGB // Constants for ARGB
......
source/scale_posix.cc
View file @ b14f46fa
...@@ -92,46 +92,6 @@ static uvec8 kShufAb2 = ...@@ -92,46 +92,6 @@ static uvec8 kShufAb2 =
static uvec16 kScaleAb2 = static uvec16 kScaleAb2 =
{ 65536 / 3, 65536 / 3, 65536 / 2, 65536 / 3, 65536 / 3, 65536 / 2, 0, 0 }; { 65536 / 3, 65536 / 3, 65536 / 2, 65536 / 3, 65536 / 3, 65536 / 2, 0, 0 };
// TODO(nfullagar): For Native Client: When new toolchain becomes available,
// take advantage of bundle lock / unlock feature. This will reduce the amount
// of manual bundle alignment done below, and bundle alignment could even be
// moved into each macro that doesn't use %%nacl: such as MEMOPREG.
#if defined(__native_client__) && defined(__x86_64__)
#define MEMACCESS(base) "%%nacl:(%%r15,%q" #base ")"
#define MEMACCESS2(offset, base) "%%nacl:" #offset "(%%r15,%q" #base ")"
#define MEMLEA(offset, base) #offset "(%q" #base ")"
#define MEMLEA3(offset, index, scale) \
#offset "(,%q" #index "," #scale ")"
#define MEMLEA4(offset, base, index, scale) \
#offset "(%q" #base ",%q" #index "," #scale ")"
#define MEMOPREG(opcode, offset, base, index, scale, reg) \
"lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
#opcode " (%%r15,%%r14),%%" #reg "\n"
#define MEMOPMEM(opcode, reg, offset, base, index, scale) \
"lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
#opcode " %%" #reg ",(%%r15,%%r14)\n"
#define MEMOP(opcode, offset, base, index, scale) \
"lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
#opcode " (%%r15,%%r14)"
#define BUNDLEALIGN ".p2align 5\n"
#else
#define MEMACCESS(base) "(%" #base ")"
#define MEMACCESS2(offset, base) #offset "(%" #base ")"
#define MEMLEA(offset, base) #offset "(%" #base ")"
#define MEMLEA3(offset, index, scale) \
#offset "(,%" #index "," #scale ")"
#define MEMLEA4(offset, base, index, scale) \
#offset "(%" #base ",%" #index "," #scale ")"
#define MEMOPREG(opcode, offset, base, index, scale, reg) \
#opcode " " #offset "(%" #base ",%" #index "," #scale "),%%" #reg "\n"
#define MEMOPMEM(opcode, reg, offset, base, index, scale) \
#opcode " %%" #reg ","#offset "(%" #base ",%" #index "," #scale ")\n"
#define MEMOP(opcode, offset, base, index, scale) \
#opcode " " #offset "(%" #base ",%" #index "," #scale ")"
#define BUNDLEALIGN
#endif
// GCC versions of row functions are verbatim conversions from Visual C. // GCC versions of row functions are verbatim conversions from Visual C.
// Generated using gcc disassembly on Visual C object file: // Generated using gcc disassembly on Visual C object file:
// objdump -D yuvscaler.obj >yuvscaler.txt // objdump -D yuvscaler.obj >yuvscaler.txt
...@@ -139,8 +99,7 @@ static uvec16 kScaleAb2 = ...@@ -139,8 +99,7 @@ static uvec16 kScaleAb2 =
void ScaleRowDown2_SSE2(const uint8* src_ptr, ptrdiff_t src_stride, void ScaleRowDown2_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
uint8* dst_ptr, int dst_width) { uint8* dst_ptr, int dst_width) {
asm volatile ( asm volatile (
".p2align 2 \n" LABELALIGN
BUNDLEALIGN
"1: \n" "1: \n"
"movdqa " MEMACCESS(0) ",%%xmm0 \n" "movdqa " MEMACCESS(0) ",%%xmm0 \n"
"movdqa " MEMACCESS2(0x10,0) ",%%xmm1 \n" "movdqa " MEMACCESS2(0x10,0) ",%%xmm1 \n"
...@@ -168,8 +127,8 @@ void ScaleRowDown2Linear_SSE2(const uint8* src_ptr, ptrdiff_t, ...@@ -168,8 +127,8 @@ void ScaleRowDown2Linear_SSE2(const uint8* src_ptr, ptrdiff_t,
asm volatile ( asm volatile (
"pcmpeqb %%xmm5,%%xmm5 \n" "pcmpeqb %%xmm5,%%xmm5 \n"
"psrlw $0x8,%%xmm5 \n" "psrlw $0x8,%%xmm5 \n"
".p2align 2 \n"
BUNDLEALIGN LABELALIGN
"1: \n" "1: \n"
"movdqa " MEMACCESS(0) ",%%xmm0 \n" "movdqa " MEMACCESS(0) ",%%xmm0 \n"
"movdqa " MEMACCESS2(0x10, 0) ",%%xmm1 \n" "movdqa " MEMACCESS2(0x10, 0) ",%%xmm1 \n"
...@@ -203,8 +162,8 @@ void ScaleRowDown2Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride, ...@@ -203,8 +162,8 @@ void ScaleRowDown2Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
asm volatile ( asm volatile (
"pcmpeqb %%xmm5,%%xmm5 \n" "pcmpeqb %%xmm5,%%xmm5 \n"
"psrlw $0x8,%%xmm5 \n" "psrlw $0x8,%%xmm5 \n"
".p2align 2 \n"
BUNDLEALIGN LABELALIGN
"1: \n" "1: \n"
"movdqa " MEMACCESS(0) ",%%xmm0 \n" "movdqa " MEMACCESS(0) ",%%xmm0 \n"
"movdqa " MEMACCESS2(0x10,0) ",%%xmm1 \n" "movdqa " MEMACCESS2(0x10,0) ",%%xmm1 \n"
...@@ -245,8 +204,7 @@ void ScaleRowDown2_Unaligned_SSE2(const uint8* src_ptr, ...@@ -245,8 +204,7 @@ void ScaleRowDown2_Unaligned_SSE2(const uint8* src_ptr,
ptrdiff_t src_stride, ptrdiff_t src_stride,
uint8* dst_ptr, int dst_width) { uint8* dst_ptr, int dst_width) {
asm volatile ( asm volatile (
".p2align 2 \n" LABELALIGN
BUNDLEALIGN
"1: \n" "1: \n"
"movdqu " MEMACCESS(0) ",%%xmm0 \n" "movdqu " MEMACCESS(0) ",%%xmm0 \n"
"movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n"
...@@ -274,8 +232,8 @@ void ScaleRowDown2Linear_Unaligned_SSE2(const uint8* src_ptr, ptrdiff_t, ...@@ -274,8 +232,8 @@ void ScaleRowDown2Linear_Unaligned_SSE2(const uint8* src_ptr, ptrdiff_t,
asm volatile ( asm volatile (
"pcmpeqb %%xmm5,%%xmm5 \n" "pcmpeqb %%xmm5,%%xmm5 \n"
"psrlw $0x8,%%xmm5 \n" "psrlw $0x8,%%xmm5 \n"
".p2align 2 \n"
BUNDLEALIGN LABELALIGN
"1: \n" "1: \n"
"movdqu " MEMACCESS(0) ",%%xmm0 \n" "movdqu " MEMACCESS(0) ",%%xmm0 \n"
"movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n"
...@@ -310,8 +268,8 @@ void ScaleRowDown2Box_Unaligned_SSE2(const uint8* src_ptr, ...@@ -310,8 +268,8 @@ void ScaleRowDown2Box_Unaligned_SSE2(const uint8* src_ptr,
asm volatile ( asm volatile (
"pcmpeqb %%xmm5,%%xmm5 \n" "pcmpeqb %%xmm5,%%xmm5 \n"
"psrlw $0x8,%%xmm5 \n" "psrlw $0x8,%%xmm5 \n"
".p2align 2 \n"
BUNDLEALIGN LABELALIGN
"1: \n" "1: \n"
"movdqu " MEMACCESS(0) ",%%xmm0 \n" "movdqu " MEMACCESS(0) ",%%xmm0 \n"
"movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n"
...@@ -354,8 +312,8 @@ void ScaleRowDown4_SSE2(const uint8* src_ptr, ptrdiff_t src_stride, ...@@ -354,8 +312,8 @@ void ScaleRowDown4_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
"pcmpeqb %%xmm5,%%xmm5 \n" "pcmpeqb %%xmm5,%%xmm5 \n"
"psrld $0x18,%%xmm5 \n" "psrld $0x18,%%xmm5 \n"
"pslld $0x10,%%xmm5 \n" "pslld $0x10,%%xmm5 \n"
".p2align 2 \n"
BUNDLEALIGN LABELALIGN
"1: \n" "1: \n"
"movdqa " MEMACCESS(0) ",%%xmm0 \n" "movdqa " MEMACCESS(0) ",%%xmm0 \n"
"movdqa " MEMACCESS2(0x10,0) ",%%xmm1 \n" "movdqa " MEMACCESS2(0x10,0) ",%%xmm1 \n"
...@@ -387,8 +345,8 @@ void ScaleRowDown4Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride, ...@@ -387,8 +345,8 @@ void ScaleRowDown4Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
"pcmpeqb %%xmm7,%%xmm7 \n" "pcmpeqb %%xmm7,%%xmm7 \n"
"psrlw $0x8,%%xmm7 \n" "psrlw $0x8,%%xmm7 \n"
"lea " MEMLEA4(0x00,4,4,2) ",%3 \n" "lea " MEMLEA4(0x00,4,4,2) ",%3 \n"
".p2align 2 \n"
BUNDLEALIGN LABELALIGN
"1: \n" "1: \n"
"movdqa " MEMACCESS(0) ",%%xmm0 \n" "movdqa " MEMACCESS(0) ",%%xmm0 \n"
"movdqa " MEMACCESS2(0x10,0) ",%%xmm1 \n" "movdqa " MEMACCESS2(0x10,0) ",%%xmm1 \n"
...@@ -452,8 +410,7 @@ void ScaleRowDown34_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride, ...@@ -452,8 +410,7 @@ void ScaleRowDown34_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
"m"(kShuf2) // %2 "m"(kShuf2) // %2
); );
asm volatile ( asm volatile (
".p2align 2 \n" LABELALIGN
BUNDLEALIGN
"1: \n" "1: \n"
"movdqa " MEMACCESS(0) ",%%xmm0 \n" "movdqa " MEMACCESS(0) ",%%xmm0 \n"
"movdqa " MEMACCESS2(0x10,0) ",%%xmm2 \n" "movdqa " MEMACCESS2(0x10,0) ",%%xmm2 \n"
...@@ -502,8 +459,7 @@ void ScaleRowDown34_1_Box_SSSE3(const uint8* src_ptr, ...@@ -502,8 +459,7 @@ void ScaleRowDown34_1_Box_SSSE3(const uint8* src_ptr,
"m"(kRound34) // %2 "m"(kRound34) // %2
); );
asm volatile ( asm volatile (
".p2align 2 \n" LABELALIGN
BUNDLEALIGN
"1: \n" "1: \n"
"movdqa " MEMACCESS(0) ",%%xmm6 \n" "movdqa " MEMACCESS(0) ",%%xmm6 \n"
MEMOPREG(movdqa,0x00,0,3,1,xmm7) // movdqa (%0,%3),%%xmm7 MEMOPREG(movdqa,0x00,0,3,1,xmm7) // movdqa (%0,%3),%%xmm7
...@@ -575,8 +531,7 @@ void ScaleRowDown34_0_Box_SSSE3(const uint8* src_ptr, ...@@ -575,8 +531,7 @@ void ScaleRowDown34_0_Box_SSSE3(const uint8* src_ptr,
); );
asm volatile ( asm volatile (
".p2align 2 \n" LABELALIGN
BUNDLEALIGN
"1: \n" "1: \n"
"movdqa " MEMACCESS(0) ",%%xmm6 \n" "movdqa " MEMACCESS(0) ",%%xmm6 \n"
MEMOPREG(movdqa,0x00,0,3,1,xmm7) // movdqa (%0,%3,1),%%xmm7 MEMOPREG(movdqa,0x00,0,3,1,xmm7) // movdqa (%0,%3,1),%%xmm7
...@@ -632,8 +587,8 @@ void ScaleRowDown38_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride, ...@@ -632,8 +587,8 @@ void ScaleRowDown38_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
asm volatile ( asm volatile (
"movdqa %3,%%xmm4 \n" "movdqa %3,%%xmm4 \n"
"movdqa %4,%%xmm5 \n" "movdqa %4,%%xmm5 \n"
".p2align 2 \n"
BUNDLEALIGN LABELALIGN
"1: \n" "1: \n"
"movdqa " MEMACCESS(0) ",%%xmm0 \n" "movdqa " MEMACCESS(0) ",%%xmm0 \n"
"movdqa " MEMACCESS2(0x10,0) ",%%xmm1 \n" "movdqa " MEMACCESS2(0x10,0) ",%%xmm1 \n"
...@@ -674,8 +629,7 @@ void ScaleRowDown38_2_Box_SSSE3(const uint8* src_ptr, ...@@ -674,8 +629,7 @@ void ScaleRowDown38_2_Box_SSSE3(const uint8* src_ptr,
"m"(kScaleAb2) // %3 "m"(kScaleAb2) // %3
); );
asm volatile ( asm volatile (
".p2align 2 \n" LABELALIGN
BUNDLEALIGN
"1: \n" "1: \n"
"movdqa " MEMACCESS(0) ",%%xmm0 \n" "movdqa " MEMACCESS(0) ",%%xmm0 \n"
MEMOPREG(pavgb,0x00,0,3,1,xmm0) // pavgb (%0,%3,1),%%xmm0 MEMOPREG(pavgb,0x00,0,3,1,xmm0) // pavgb (%0,%3,1),%%xmm0
...@@ -723,8 +677,7 @@ void ScaleRowDown38_3_Box_SSSE3(const uint8* src_ptr, ...@@ -723,8 +677,7 @@ void ScaleRowDown38_3_Box_SSSE3(const uint8* src_ptr,
"m"(kScaleAc33) // %2 "m"(kScaleAc33) // %2
); );
asm volatile ( asm volatile (
".p2align 2 \n" LABELALIGN
BUNDLEALIGN
"1: \n" "1: \n"
"movdqa " MEMACCESS(0) ",%%xmm0 \n" "movdqa " MEMACCESS(0) ",%%xmm0 \n"
MEMOPREG(movdqa,0x00,0,3,1,xmm6) // movdqa (%0,%3,1),%%xmm6 MEMOPREG(movdqa,0x00,0,3,1,xmm6) // movdqa (%0,%3,1),%%xmm6
...@@ -785,8 +738,8 @@ void ScaleAddRows_SSE2(const uint8* src_ptr, ptrdiff_t src_stride, ...@@ -785,8 +738,8 @@ void ScaleAddRows_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
asm volatile ( asm volatile (
"pxor %%xmm4,%%xmm4 \n" "pxor %%xmm4,%%xmm4 \n"
"sub $0x1,%5 \n" "sub $0x1,%5 \n"
".p2align 2 \n"
BUNDLEALIGN LABELALIGN
"1: \n" "1: \n"
"movdqa " MEMACCESS(0) ",%%xmm0 \n" "movdqa " MEMACCESS(0) ",%%xmm0 \n"
"mov %0,%3 \n" "mov %0,%3 \n"
...@@ -797,8 +750,8 @@ void ScaleAddRows_SSE2(const uint8* src_ptr, ptrdiff_t src_stride, ...@@ -797,8 +750,8 @@ void ScaleAddRows_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
"mov %5,%2 \n" "mov %5,%2 \n"
"test %2,%2 \n" "test %2,%2 \n"
"je 3f \n" "je 3f \n"
".p2align 2 \n"
BUNDLEALIGN LABELALIGN
"2: \n" "2: \n"
"movdqa " MEMACCESS(0) ",%%xmm2 \n" "movdqa " MEMACCESS(0) ",%%xmm2 \n"
"add %6,%0 \n" "add %6,%0 \n"
...@@ -809,9 +762,9 @@ void ScaleAddRows_SSE2(const uint8* src_ptr, ptrdiff_t src_stride, ...@@ -809,9 +762,9 @@ void ScaleAddRows_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
"paddusw %%xmm3,%%xmm1 \n" "paddusw %%xmm3,%%xmm1 \n"
"sub $0x1,%2 \n" "sub $0x1,%2 \n"
"jg 2b \n" "jg 2b \n"
".p2align 2 \n"
LABELALIGN
"3: \n" "3: \n"
BUNDLEALIGN
"movdqa %%xmm0," MEMACCESS(1) " \n" "movdqa %%xmm0," MEMACCESS(1) " \n"
"movdqa %%xmm1," MEMACCESS2(0x10,1) " \n" "movdqa %%xmm1," MEMACCESS2(0x10,1) " \n"
"lea " MEMLEA(0x10,3) ",%0 \n" "lea " MEMLEA(0x10,3) ",%0 \n"
...@@ -852,16 +805,16 @@ void ScaleFilterCols_SSSE3(uint8* dst_ptr, const uint8* src_ptr, ...@@ -852,16 +805,16 @@ void ScaleFilterCols_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
"punpckldq %%xmm3,%%xmm3 \n" "punpckldq %%xmm3,%%xmm3 \n"
"paddd %%xmm3,%%xmm3 \n" "paddd %%xmm3,%%xmm3 \n"
"pextrw $0x3,%%xmm2,%k4 \n" "pextrw $0x3,%%xmm2,%k4 \n"
".p2align 2 \n"
BUNDLEALIGN LABELALIGN
"2: \n" "2: \n"
"movdqa %%xmm2,%%xmm1 \n" "movdqa %%xmm2,%%xmm1 \n"
"paddd %%xmm3,%%xmm2 \n" "paddd %%xmm3,%%xmm2 \n"
MEMOP(movzwl,0x00,1,3,1) ",%k2 \n" // movzwl (%1,%3,1),%k2 MEMOPARG(movzwl,0x00,1,3,1,k2) // movzwl (%1,%3,1),%k2
"movd %k2,%%xmm0 \n" "movd %k2,%%xmm0 \n"
"psrlw $0x9,%%xmm1 \n" "psrlw $0x9,%%xmm1 \n"
BUNDLEALIGN BUNDLEALIGN
MEMOP(movzwl,0x00,1,4,1) ",%k2 \n" // movzwl (%1,%4,1),%k2 MEMOPARG(movzwl,0x00,1,4,1,k2) // movzwl (%1,%4,1),%k2
"movd %k2,%%xmm4 \n" "movd %k2,%%xmm4 \n"
"pshufb %%xmm5,%%xmm1 \n" "pshufb %%xmm5,%%xmm1 \n"
"punpcklwd %%xmm4,%%xmm0 \n" "punpcklwd %%xmm4,%%xmm0 \n"
...@@ -876,12 +829,12 @@ void ScaleFilterCols_SSSE3(uint8* dst_ptr, const uint8* src_ptr, ...@@ -876,12 +829,12 @@ void ScaleFilterCols_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
"lea " MEMLEA(0x2,0) ",%0 \n" "lea " MEMLEA(0x2,0) ",%0 \n"
"sub $0x2,%5 \n" "sub $0x2,%5 \n"
"jge 2b \n" "jge 2b \n"
".p2align 2 \n"
BUNDLEALIGN LABELALIGN
"29: \n" "29: \n"
"addl $0x1,%5 \n" "addl $0x1,%5 \n"
"jl 99f \n" "jl 99f \n"
MEMOP(movzwl,0x00,1,3,1) ",%k2 \n" // movzwl (%1,%3,1),%k2 MEMOPARG(movzwl,0x00,1,3,1,k2) // movzwl (%1,%3,1),%k2
"movd %k2,%%xmm0 \n" "movd %k2,%%xmm0 \n"
"psrlw $0x9,%%xmm2 \n" "psrlw $0x9,%%xmm2 \n"
"pshufb %%xmm5,%%xmm2 \n" "pshufb %%xmm5,%%xmm2 \n"
...@@ -915,8 +868,7 @@ void ScaleFilterCols_SSSE3(uint8* dst_ptr, const uint8* src_ptr, ...@@ -915,8 +868,7 @@ void ScaleFilterCols_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
void ScaleColsUp2_SSE2(uint8* dst_ptr, const uint8* src_ptr, void ScaleColsUp2_SSE2(uint8* dst_ptr, const uint8* src_ptr,
int dst_width, int /* x */, int /* dx */) { int dst_width, int /* x */, int /* dx */) {
asm volatile ( asm volatile (
".p2align 2 \n" LABELALIGN
BUNDLEALIGN
"1: \n" "1: \n"
"movdqa " MEMACCESS(1) ",%%xmm0 \n" "movdqa " MEMACCESS(1) ",%%xmm0 \n"
"lea " MEMLEA(0x10,1) ",%1 \n" "lea " MEMLEA(0x10,1) ",%1 \n"
...@@ -944,8 +896,7 @@ void ScaleARGBRowDown2_SSE2(const uint8* src_argb, ...@@ -944,8 +896,7 @@ void ScaleARGBRowDown2_SSE2(const uint8* src_argb,
ptrdiff_t /* src_stride */, ptrdiff_t /* src_stride */,
uint8* dst_argb, int dst_width) { uint8* dst_argb, int dst_width) {
asm volatile ( asm volatile (
".p2align 2 \n" LABELALIGN
BUNDLEALIGN
"1: \n" "1: \n"
"movdqa " MEMACCESS(0) ",%%xmm0 \n" "movdqa " MEMACCESS(0) ",%%xmm0 \n"
"movdqa " MEMACCESS2(0x10,0) ",%%xmm1 \n" "movdqa " MEMACCESS2(0x10,0) ",%%xmm1 \n"
...@@ -970,8 +921,7 @@ void ScaleARGBRowDown2Linear_SSE2(const uint8* src_argb, ...@@ -970,8 +921,7 @@ void ScaleARGBRowDown2Linear_SSE2(const uint8* src_argb,
ptrdiff_t /* src_stride */, ptrdiff_t /* src_stride */,
uint8* dst_argb, int dst_width) { uint8* dst_argb, int dst_width) {
asm volatile ( asm volatile (
".p2align 2 \n" LABELALIGN
BUNDLEALIGN
"1: \n" "1: \n"
"movdqa " MEMACCESS(0) ",%%xmm0 \n" "movdqa " MEMACCESS(0) ",%%xmm0 \n"
"movdqa " MEMACCESS2(0x10,0) ",%%xmm1 \n" "movdqa " MEMACCESS2(0x10,0) ",%%xmm1 \n"
...@@ -999,8 +949,7 @@ void ScaleARGBRowDown2Box_SSE2(const uint8* src_argb, ...@@ -999,8 +949,7 @@ void ScaleARGBRowDown2Box_SSE2(const uint8* src_argb,
ptrdiff_t src_stride, ptrdiff_t src_stride,
uint8* dst_argb, int dst_width) { uint8* dst_argb, int dst_width) {
asm volatile ( asm volatile (
".p2align 2 \n" LABELALIGN
BUNDLEALIGN
"1: \n" "1: \n"
"movdqa " MEMACCESS(0) ",%%xmm0 \n" "movdqa " MEMACCESS(0) ",%%xmm0 \n"
"movdqa " MEMACCESS2(0x10,0) ",%%xmm1 \n" "movdqa " MEMACCESS2(0x10,0) ",%%xmm1 \n"
...@@ -1042,8 +991,7 @@ void ScaleARGBRowDownEven_SSE2(const uint8* src_argb, ptrdiff_t src_stride, ...@@ -1042,8 +991,7 @@ void ScaleARGBRowDownEven_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
asm volatile ( asm volatile (
"lea " MEMLEA3(0x00,1,4) ",%1 \n" "lea " MEMLEA3(0x00,1,4) ",%1 \n"
"lea " MEMLEA4(0x00,1,1,2) ",%4 \n" "lea " MEMLEA4(0x00,1,1,2) ",%4 \n"
".p2align 2 \n" LABELALIGN
BUNDLEALIGN
"1: \n" "1: \n"
"movd " MEMACCESS(0) ",%%xmm0 \n" "movd " MEMACCESS(0) ",%%xmm0 \n"
MEMOPREG(movd,0x00,0,1,1,xmm1) // movd (%0,%1,1),%%xmm1 MEMOPREG(movd,0x00,0,1,1,xmm1) // movd (%0,%1,1),%%xmm1
...@@ -1086,8 +1034,8 @@ void ScaleARGBRowDownEvenBox_SSE2(const uint8* src_argb, ...@@ -1086,8 +1034,8 @@ void ScaleARGBRowDownEvenBox_SSE2(const uint8* src_argb,
"lea " MEMLEA3(0x00,1,4) ",%1 \n" "lea " MEMLEA3(0x00,1,4) ",%1 \n"
"lea " MEMLEA4(0x00,1,1,2) ",%4 \n" "lea " MEMLEA4(0x00,1,1,2) ",%4 \n"
"lea " MEMLEA4(0x00,0,5,1) ",%5 \n" "lea " MEMLEA4(0x00,0,5,1) ",%5 \n"
".p2align 2 \n"
BUNDLEALIGN LABELALIGN
"1: \n" "1: \n"
"movq " MEMACCESS(0) ",%%xmm0 \n" "movq " MEMACCESS(0) ",%%xmm0 \n"
MEMOPREG(movhps,0x00,0,1,1,xmm0) // movhps (%0,%1,1),%%xmm0 MEMOPREG(movhps,0x00,0,1,1,xmm0) // movhps (%0,%1,1),%%xmm0
...@@ -1148,8 +1096,8 @@ void ScaleARGBCols_SSE2(uint8* dst_argb, const uint8* src_argb, ...@@ -1148,8 +1096,8 @@ void ScaleARGBCols_SSE2(uint8* dst_argb, const uint8* src_argb,
"jl 99f \n" "jl 99f \n"
"sub $0x4,%4 \n" "sub $0x4,%4 \n"
"jl 49f \n" "jl 49f \n"
".p2align 2 \n"
BUNDLEALIGN LABELALIGN
"40: \n" "40: \n"
MEMOPREG(movd,0x00,3,0,4,xmm0) // movd (%3,%0,4),%%xmm0 MEMOPREG(movd,0x00,3,0,4,xmm0) // movd (%3,%0,4),%%xmm0
MEMOPREG(movd,0x00,3,1,4,xmm1) // movd (%3,%1,4),%%xmm1 MEMOPREG(movd,0x00,3,1,4,xmm1) // movd (%3,%1,4),%%xmm1
...@@ -1206,8 +1154,7 @@ void ScaleARGBCols_SSE2(uint8* dst_argb, const uint8* src_argb, ...@@ -1206,8 +1154,7 @@ void ScaleARGBCols_SSE2(uint8* dst_argb, const uint8* src_argb,
void ScaleARGBColsUp2_SSE2(uint8* dst_argb, const uint8* src_argb, void ScaleARGBColsUp2_SSE2(uint8* dst_argb, const uint8* src_argb,
int dst_width, int /* x */, int /* dx */) { int dst_width, int /* x */, int /* dx */) {
asm volatile ( asm volatile (
".p2align 2 \n" LABELALIGN
BUNDLEALIGN
"1: \n" "1: \n"
"movdqa " MEMACCESS(1) ",%%xmm0 \n" "movdqa " MEMACCESS(1) ",%%xmm0 \n"
"lea " MEMLEA(0x10,1) ",%1 \n" "lea " MEMLEA(0x10,1) ",%1 \n"
...@@ -1272,8 +1219,7 @@ void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb, ...@@ -1272,8 +1219,7 @@ void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb,
"paddd %%xmm3,%%xmm3 \n" "paddd %%xmm3,%%xmm3 \n"
"pextrw $0x3,%%xmm2,%k4 \n" "pextrw $0x3,%%xmm2,%k4 \n"
".p2align 2 \n" LABELALIGN
BUNDLEALIGN
"2: \n" "2: \n"
"movdqa %%xmm2,%%xmm1 \n" "movdqa %%xmm2,%%xmm1 \n"
"paddd %%xmm3,%%xmm2 \n" "paddd %%xmm3,%%xmm2 \n"
...@@ -1294,8 +1240,7 @@ void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb, ...@@ -1294,8 +1240,7 @@ void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb,
"sub $0x2,%2 \n" "sub $0x2,%2 \n"
"jge 2b \n" "jge 2b \n"
".p2align 2 \n" LABELALIGN
BUNDLEALIGN
"29: \n" "29: \n"
"add $0x1,%2 \n" "add $0x1,%2 \n"
"jl 99f \n" "jl 99f \n"
...@@ -1310,7 +1255,7 @@ void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb, ...@@ -1310,7 +1255,7 @@ void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb,
"packuswb %%xmm0,%%xmm0 \n" "packuswb %%xmm0,%%xmm0 \n"
"movd %%xmm0," MEMACCESS(0) " \n" "movd %%xmm0," MEMACCESS(0) " \n"
".p2align 2 \n" LABELALIGN
"99: \n" "99: \n"
: "+r"(dst_argb), // %0 : "+r"(dst_argb), // %0
"+r"(src_argb), // %1 "+r"(src_argb), // %1
......
unit_test/scale_argb_test.cc
View file @ b14f46fa
...@@ -218,16 +218,10 @@ static int ARGBClipTestFilter(int src_width, int src_height, ...@@ -218,16 +218,10 @@ static int ARGBClipTestFilter(int src_width, int src_height,
TEST_FACTOR1(name, Bilinear, hfactor, vfactor, 2) \ TEST_FACTOR1(name, Bilinear, hfactor, vfactor, 2) \
TEST_FACTOR1(name, Box, hfactor, vfactor, 2) TEST_FACTOR1(name, Box, hfactor, vfactor, 2)
// TODO(fbarchard): ScaleDownBy1 should be lossless, but Box has error of 2.
TEST_FACTOR(1, 1 / 1, 1 / 1)
TEST_FACTOR(2, 1 / 2, 1 / 2) TEST_FACTOR(2, 1 / 2, 1 / 2)
TEST_FACTOR(4, 1 / 4, 1 / 4) TEST_FACTOR(4, 1 / 4, 1 / 4)
TEST_FACTOR(8, 1 / 8, 1 / 8) TEST_FACTOR(8, 1 / 8, 1 / 8)
TEST_FACTOR(16, 1 / 16, 1 / 16)
TEST_FACTOR(2by3, 2 / 3, 2 / 3)
TEST_FACTOR(3by4, 3 / 4, 3 / 4) TEST_FACTOR(3by4, 3 / 4, 3 / 4)
TEST_FACTOR(3by8, 3 / 8, 3 / 8)
TEST_FACTOR(Vertical2by3, 1, 2 / 3)
#undef TEST_FACTOR1 #undef TEST_FACTOR1
#undef TEST_FACTOR #undef TEST_FACTOR
...@@ -268,9 +262,7 @@ TEST_SCALETO(ARGBScale, 1, 1) ...@@ -268,9 +262,7 @@ TEST_SCALETO(ARGBScale, 1, 1)
TEST_SCALETO(ARGBScale, 320, 240) TEST_SCALETO(ARGBScale, 320, 240)
TEST_SCALETO(ARGBScale, 352, 288) TEST_SCALETO(ARGBScale, 352, 288)
TEST_SCALETO(ARGBScale, 640, 360) TEST_SCALETO(ARGBScale, 640, 360)
TEST_SCALETO(ARGBScale, 853, 480)
TEST_SCALETO(ARGBScale, 1280, 720) TEST_SCALETO(ARGBScale, 1280, 720)
TEST_SCALETO(ARGBScale, 1920, 1080)
#undef TEST_SCALETO1 #undef TEST_SCALETO1
#undef TEST_SCALETO #undef TEST_SCALETO
......
unit_test/scale_test.cc
View file @ b14f46fa
...@@ -149,16 +149,10 @@ static int TestFilter(int src_width, int src_height, ...@@ -149,16 +149,10 @@ static int TestFilter(int src_width, int src_height,
TEST_FACTOR1(name, Bilinear, hfactor, vfactor, 3) \ TEST_FACTOR1(name, Bilinear, hfactor, vfactor, 3) \
TEST_FACTOR1(name, Box, hfactor, vfactor, 3) \ TEST_FACTOR1(name, Box, hfactor, vfactor, 3) \
// TODO(fbarchard): ScaleDownBy1 should be lossless, but Box has error of 2.
TEST_FACTOR(1, 1 / 1, 1 / 1)
TEST_FACTOR(2, 1 / 2, 1 / 2) TEST_FACTOR(2, 1 / 2, 1 / 2)
TEST_FACTOR(4, 1 / 4, 1 / 4) TEST_FACTOR(4, 1 / 4, 1 / 4)
TEST_FACTOR(8, 1 / 8, 1 / 8) TEST_FACTOR(8, 1 / 8, 1 / 8)
TEST_FACTOR(16, 1 / 16, 1 / 16)
TEST_FACTOR(2by3, 2 / 3, 2 / 3)
TEST_FACTOR(3by4, 3 / 4, 3 / 4) TEST_FACTOR(3by4, 3 / 4, 3 / 4)
TEST_FACTOR(3by8, 3 / 8, 3 / 8)
TEST_FACTOR(Vertical2by3, 1, 2 / 3)
#undef TEST_FACTOR1 #undef TEST_FACTOR1
#undef TEST_FACTOR #undef TEST_FACTOR
...@@ -187,9 +181,7 @@ TEST_SCALETO(Scale, 1, 1) ...@@ -187,9 +181,7 @@ TEST_SCALETO(Scale, 1, 1)
TEST_SCALETO(Scale, 320, 240) TEST_SCALETO(Scale, 320, 240)
TEST_SCALETO(Scale, 352, 288) TEST_SCALETO(Scale, 352, 288)
TEST_SCALETO(Scale, 640, 360) TEST_SCALETO(Scale, 640, 360)
TEST_SCALETO(Scale, 853, 480)
TEST_SCALETO(Scale, 1280, 720) TEST_SCALETO(Scale, 1280, 720)
TEST_SCALETO(Scale, 1920, 1080)
#undef TEST_SCALETO1 #undef TEST_SCALETO1
#undef TEST_SCALETO #undef TEST_SCALETO
......
unit_test/unit_test.cc
View file @ b14f46fa
...@@ -19,8 +19,8 @@ ...@@ -19,8 +19,8 @@
#define BENCHMARK_ITERATIONS 1 #define BENCHMARK_ITERATIONS 1
libyuvTest::libyuvTest() : rotate_max_w_(128), rotate_max_h_(128), libyuvTest::libyuvTest() : rotate_max_w_(128), rotate_max_h_(128),
benchmark_iterations_(BENCHMARK_ITERATIONS), benchmark_width_(22), benchmark_iterations_(BENCHMARK_ITERATIONS), benchmark_width_(128),
benchmark_height_(14) { benchmark_height_(72) {
const char* repeat = getenv("LIBYUV_REPEAT"); const char* repeat = getenv("LIBYUV_REPEAT");
if (repeat) { if (repeat) {
benchmark_iterations_ = atoi(repeat); // NOLINT benchmark_iterations_ = atoi(repeat); // NOLINT
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment