Skip to content

L
libyuv
Commit 73a6f100 authored by Manojkumar Bhosale's avatar Manojkumar Bhosale
Browse files
Options

Add MSA optimized rotate functions (used 16x16 transpose) 

R=fbarchard@google.com
BUG=libyuv:634

Performance Gain (vs C vectorized)
TransposeWx16_MSA        - ~6.0x
TransposeWx16_Any_MSA    - ~4.7x
TransposeUVWx16_MSA      - ~6.3x
TransposeUVWx16_Any_MSA  - ~5.4x

Performance Gain (vs C non-vectorized)
TransposeWx16_MSA        - ~6.0x
TransposeWx16_Any_MSA    - ~4.8x
TransposeUVWx16_MSA      - ~6.3x
TransposeUVWx16_Any_MSA  - ~5.4x

Review-Url: https://codereview.chromium.org/2617703002 .
parent 7c64163f
Hide whitespace changes
Inline Side-by-side
Showing with 278 additions and 186 deletions
include/libyuv/rotate_row.h
View file @ 73a6f100
...@@ -61,8 +61,8 @@ extern "C" { ...@@ -61,8 +61,8 @@ extern "C" {
#endif // defined(__mips__) #endif // defined(__mips__)
#if !defined(LIBYUV_DISABLE_MSA) && defined(__mips_msa) #if !defined(LIBYUV_DISABLE_MSA) && defined(__mips_msa)
#define HAS_TRANSPOSEWX8_MSA #define HAS_TRANSPOSEWX16_MSA
#define HAS_TRANSPOSEUVWX8_MSA #define HAS_TRANSPOSEUVWX16_MSA
#endif #endif
void TransposeWxH_C(const uint8* src, void TransposeWxH_C(const uint8* src,
...@@ -77,6 +77,11 @@ void TransposeWx8_C(const uint8* src, ...@@ -77,6 +77,11 @@ void TransposeWx8_C(const uint8* src,
uint8* dst, uint8* dst,
int dst_stride, int dst_stride,
int width); int width);
void TransposeWx16_C(const uint8* src,
int src_stride,
uint8* dst,
int dst_stride,
int width);
void TransposeWx8_NEON(const uint8* src, void TransposeWx8_NEON(const uint8* src,
int src_stride, int src_stride,
uint8* dst, uint8* dst,
...@@ -102,11 +107,11 @@ void TransposeWx8_Fast_DSPR2(const uint8* src, ...@@ -102,11 +107,11 @@ void TransposeWx8_Fast_DSPR2(const uint8* src,
uint8* dst, uint8* dst,
int dst_stride, int dst_stride,
int width); int width);
void TransposeWx8_MSA(const uint8* src, void TransposeWx16_MSA(const uint8* src,
int src_stride, int src_stride,
uint8* dst, uint8* dst,
int dst_stride, int dst_stride,
int width); int width);
void TransposeWx8_Any_NEON(const uint8* src, void TransposeWx8_Any_NEON(const uint8* src,
int src_stride, int src_stride,
...@@ -128,11 +133,11 @@ void TransposeWx8_Any_DSPR2(const uint8* src, ...@@ -128,11 +133,11 @@ void TransposeWx8_Any_DSPR2(const uint8* src,
uint8* dst, uint8* dst,
int dst_stride, int dst_stride,
int width); int width);
void TransposeWx8_Any_MSA(const uint8* src, void TransposeWx16_Any_MSA(const uint8* src,
int src_stride, int src_stride,
uint8* dst, uint8* dst,
int dst_stride, int dst_stride,
int width); int width);
void TransposeUVWxH_C(const uint8* src, void TransposeUVWxH_C(const uint8* src,
int src_stride, int src_stride,
...@@ -150,6 +155,13 @@ void TransposeUVWx8_C(const uint8* src, ...@@ -150,6 +155,13 @@ void TransposeUVWx8_C(const uint8* src,
uint8* dst_b, uint8* dst_b,
int dst_stride_b, int dst_stride_b,
int width); int width);
void TransposeUVWx16_C(const uint8* src,
int src_stride,
uint8* dst_a,
int dst_stride_a,
uint8* dst_b,
int dst_stride_b,
int width);
void TransposeUVWx8_SSE2(const uint8* src, void TransposeUVWx8_SSE2(const uint8* src,
int src_stride, int src_stride,
uint8* dst_a, uint8* dst_a,
...@@ -171,13 +183,13 @@ void TransposeUVWx8_DSPR2(const uint8* src, ...@@ -171,13 +183,13 @@ void TransposeUVWx8_DSPR2(const uint8* src,
uint8* dst_b, uint8* dst_b,
int dst_stride_b, int dst_stride_b,
int width); int width);
void TransposeUVWx8_MSA(const uint8* src, void TransposeUVWx16_MSA(const uint8* src,
int src_stride, int src_stride,
uint8* dst_a, uint8* dst_a,
int dst_stride_a, int dst_stride_a,
uint8* dst_b, uint8* dst_b,
int dst_stride_b, int dst_stride_b,
int width); int width);
void TransposeUVWx8_Any_SSE2(const uint8* src, void TransposeUVWx8_Any_SSE2(const uint8* src,
int src_stride, int src_stride,
...@@ -200,13 +212,13 @@ void TransposeUVWx8_Any_DSPR2(const uint8* src, ...@@ -200,13 +212,13 @@ void TransposeUVWx8_Any_DSPR2(const uint8* src,
uint8* dst_b, uint8* dst_b,
int dst_stride_b, int dst_stride_b,
int width); int width);
void TransposeUVWx8_Any_MSA(const uint8* src, void TransposeUVWx16_Any_MSA(const uint8* src,
int src_stride, int src_stride,
uint8* dst_a, uint8* dst_a,
int dst_stride_a, int dst_stride_a,
uint8* dst_b, uint8* dst_b,
int dst_stride_b, int dst_stride_b,
int width); int width);
#ifdef __cplusplus #ifdef __cplusplus
} // extern "C" } // extern "C"
......
include/libyuv/row.h 100755 → 100644
View file @ 73a6f100
File mode changed from 100755 to 100644
source/convert_argb.cc 100755 → 100644
View file @ 73a6f100
File mode changed from 100755 to 100644
source/rotate.cc
View file @ 73a6f100
...@@ -29,8 +29,13 @@ void TransposePlane(const uint8* src, ...@@ -29,8 +29,13 @@ void TransposePlane(const uint8* src,
int width, int width,
int height) { int height) {
int i = height; int i = height;
#if defined(HAS_TRANSPOSEWX16_MSA)
void (*TransposeWx16)(const uint8* src, int src_stride, uint8* dst,
int dst_stride, int width) = TransposeWx16_C;
#else
void (*TransposeWx8)(const uint8* src, int src_stride, uint8* dst, void (*TransposeWx8)(const uint8* src, int src_stride, uint8* dst,
int dst_stride, int width) = TransposeWx8_C; int dst_stride, int width) = TransposeWx8_C;
#endif
#if defined(HAS_TRANSPOSEWX8_NEON) #if defined(HAS_TRANSPOSEWX8_NEON)
if (TestCpuFlag(kCpuHasNEON)) { if (TestCpuFlag(kCpuHasNEON)) {
TransposeWx8 = TransposeWx8_NEON; TransposeWx8 = TransposeWx8_NEON;
...@@ -62,15 +67,24 @@ void TransposePlane(const uint8* src, ...@@ -62,15 +67,24 @@ void TransposePlane(const uint8* src,
} }
} }
#endif #endif
#if defined(HAS_TRANSPOSEWX8_MSA) #if defined(HAS_TRANSPOSEWX16_MSA)
if (TestCpuFlag(kCpuHasMSA)) { if (TestCpuFlag(kCpuHasMSA)) {
TransposeWx8 = TransposeWx8_Any_MSA; TransposeWx16 = TransposeWx16_Any_MSA;
if (IS_ALIGNED(width, 16)) { if (IS_ALIGNED(width, 16)) {
TransposeWx8 = TransposeWx8_MSA; TransposeWx16 = TransposeWx16_MSA;
} }
} }
#endif #endif
#if defined(HAS_TRANSPOSEWX16_MSA)
// Work across the source in 16x16 tiles
while (i >= 16) {
TransposeWx16(src, src_stride, dst, dst_stride, width);
src += 16 * src_stride; // Go down 16 rows.
dst += 16; // Move over 16 columns.
i -= 16;
}
#else
// Work across the source in 8x8 tiles // Work across the source in 8x8 tiles
while (i >= 8) { while (i >= 8) {
TransposeWx8(src, src_stride, dst, dst_stride, width); TransposeWx8(src, src_stride, dst, dst_stride, width);
...@@ -78,6 +92,7 @@ void TransposePlane(const uint8* src, ...@@ -78,6 +92,7 @@ void TransposePlane(const uint8* src,
dst += 8; // Move over 8 columns. dst += 8; // Move over 8 columns.
i -= 8; i -= 8;
} }
#endif
if (i > 0) { if (i > 0) {
TransposeWxH_C(src, src_stride, dst, dst_stride, width, i); TransposeWxH_C(src, src_stride, dst, dst_stride, width, i);
...@@ -218,9 +233,15 @@ void TransposeUV(const uint8* src, ...@@ -218,9 +233,15 @@ void TransposeUV(const uint8* src,
int width, int width,
int height) { int height) {
int i = height; int i = height;
#if defined(HAS_TRANSPOSEUVWX16_MSA)
void (*TransposeUVWx16)(const uint8* src, int src_stride, uint8* dst_a,
int dst_stride_a, uint8* dst_b, int dst_stride_b,
int width) = TransposeUVWx16_C;
#else
void (*TransposeUVWx8)(const uint8* src, int src_stride, uint8* dst_a, void (*TransposeUVWx8)(const uint8* src, int src_stride, uint8* dst_a,
int dst_stride_a, uint8* dst_b, int dst_stride_b, int dst_stride_a, uint8* dst_b, int dst_stride_b,
int width) = TransposeUVWx8_C; int width) = TransposeUVWx8_C;
#endif
#if defined(HAS_TRANSPOSEUVWX8_NEON) #if defined(HAS_TRANSPOSEUVWX8_NEON)
if (TestCpuFlag(kCpuHasNEON)) { if (TestCpuFlag(kCpuHasNEON)) {
TransposeUVWx8 = TransposeUVWx8_NEON; TransposeUVWx8 = TransposeUVWx8_NEON;
...@@ -240,15 +261,26 @@ void TransposeUV(const uint8* src, ...@@ -240,15 +261,26 @@ void TransposeUV(const uint8* src,
TransposeUVWx8 = TransposeUVWx8_DSPR2; TransposeUVWx8 = TransposeUVWx8_DSPR2;
} }
#endif #endif
#if defined(HAS_TRANSPOSEUVWX8_MSA) #if defined(HAS_TRANSPOSEUVWX16_MSA)
if (TestCpuFlag(kCpuHasMSA)) { if (TestCpuFlag(kCpuHasMSA)) {
TransposeUVWx8 = TransposeUVWx8_Any_MSA; TransposeUVWx16 = TransposeUVWx16_Any_MSA;
if (IS_ALIGNED(width, 8)) { if (IS_ALIGNED(width, 8)) {
TransposeUVWx8 = TransposeUVWx8_MSA; TransposeUVWx16 = TransposeUVWx16_MSA;
} }
} }
#endif #endif
#if defined(HAS_TRANSPOSEUVWX16_MSA)
// Work through the source in 8x8 tiles.
while (i >= 16) {
TransposeUVWx16(src, src_stride, dst_a, dst_stride_a, dst_b, dst_stride_b,
width);
src += 16 * src_stride; // Go down 16 rows.
dst_a += 16; // Move over 8 columns.
dst_b += 16; // Move over 8 columns.
i -= 16;
}
#else
// Work through the source in 8x8 tiles. // Work through the source in 8x8 tiles.
while (i >= 8) { while (i >= 8) {
TransposeUVWx8(src, src_stride, dst_a, dst_stride_a, dst_b, dst_stride_b, TransposeUVWx8(src, src_stride, dst_a, dst_stride_a, dst_b, dst_stride_b,
...@@ -258,6 +290,7 @@ void TransposeUV(const uint8* src, ...@@ -258,6 +290,7 @@ void TransposeUV(const uint8* src,
dst_b += 8; // Move over 8 columns. dst_b += 8; // Move over 8 columns.
i -= 8; i -= 8;
} }
#endif
if (i > 0) { if (i > 0) {
TransposeUVWxH_C(src, src_stride, dst_a, dst_stride_a, dst_b, dst_stride_b, TransposeUVWxH_C(src, src_stride, dst_a, dst_stride_a, dst_b, dst_stride_b,
......
source/rotate_any.cc
View file @ 73a6f100
...@@ -41,8 +41,8 @@ TANY(TransposeWx8_Fast_Any_SSSE3, TransposeWx8_Fast_SSSE3, 15) ...@@ -41,8 +41,8 @@ TANY(TransposeWx8_Fast_Any_SSSE3, TransposeWx8_Fast_SSSE3, 15)
#ifdef HAS_TRANSPOSEWX8_DSPR2 #ifdef HAS_TRANSPOSEWX8_DSPR2
TANY(TransposeWx8_Any_DSPR2, TransposeWx8_DSPR2, 7) TANY(TransposeWx8_Any_DSPR2, TransposeWx8_DSPR2, 7)
#endif #endif
#ifdef HAS_TRANSPOSEWX8_MSA #ifdef HAS_TRANSPOSEWX16_MSA
TANY(TransposeWx8_Any_MSA, TransposeWx8_MSA, 15) TANY(TransposeWx16_Any_MSA, TransposeWx16_MSA, 15)
#endif #endif
#undef TANY #undef TANY
...@@ -67,8 +67,8 @@ TUVANY(TransposeUVWx8_Any_SSE2, TransposeUVWx8_SSE2, 7) ...@@ -67,8 +67,8 @@ TUVANY(TransposeUVWx8_Any_SSE2, TransposeUVWx8_SSE2, 7)
#ifdef HAS_TRANSPOSEUVWX8_DSPR2 #ifdef HAS_TRANSPOSEUVWX8_DSPR2
TUVANY(TransposeUVWx8_Any_DSPR2, TransposeUVWx8_DSPR2, 7) TUVANY(TransposeUVWx8_Any_DSPR2, TransposeUVWx8_DSPR2, 7)
#endif #endif
#ifdef HAS_TRANSPOSEUVWX8_MSA #ifdef HAS_TRANSPOSEUVWX16_MSA
TUVANY(TransposeUVWx8_Any_MSA, TransposeUVWx8_MSA, 7) TUVANY(TransposeUVWx16_Any_MSA, TransposeUVWx16_MSA, 7)
#endif #endif
#undef TUVANY #undef TUVANY
......
source/rotate_msa.cc
View file @ 73a6f100
...@@ -19,179 +19,226 @@ namespace libyuv { ...@@ -19,179 +19,226 @@ namespace libyuv {
extern "C" { extern "C" {
#endif #endif
void TransposeWx8_MSA(const uint8_t* src, #define ILVRL_B(in0, in1, in2, in3, out0, out1, out2, out3) \
int src_stride, { \
uint8_t* dst, out0 = (v16u8)__msa_ilvr_b((v16i8)in1, (v16i8)in0); \
int dst_stride, out1 = (v16u8)__msa_ilvl_b((v16i8)in1, (v16i8)in0); \
int width) { out2 = (v16u8)__msa_ilvr_b((v16i8)in3, (v16i8)in2); \
out3 = (v16u8)__msa_ilvl_b((v16i8)in3, (v16i8)in2); \
}
#define ILVRL_H(in0, in1, in2, in3, out0, out1, out2, out3) \
{ \
out0 = (v16u8)__msa_ilvr_h((v8i16)in1, (v8i16)in0); \
out1 = (v16u8)__msa_ilvl_h((v8i16)in1, (v8i16)in0); \
out2 = (v16u8)__msa_ilvr_h((v8i16)in3, (v8i16)in2); \
out3 = (v16u8)__msa_ilvl_h((v8i16)in3, (v8i16)in2); \
}
#define ILVRL_W(in0, in1, in2, in3, out0, out1, out2, out3) \
{ \
out0 = (v16u8)__msa_ilvr_w((v4i32)in1, (v4i32)in0); \
out1 = (v16u8)__msa_ilvl_w((v4i32)in1, (v4i32)in0); \
out2 = (v16u8)__msa_ilvr_w((v4i32)in3, (v4i32)in2); \
out3 = (v16u8)__msa_ilvl_w((v4i32)in3, (v4i32)in2); \
}
#define ILVRL_D(in0, in1, in2, in3, out0, out1, out2, out3) \
{ \
out0 = (v16u8)__msa_ilvr_d((v2i64)in1, (v2i64)in0); \
out1 = (v16u8)__msa_ilvl_d((v2i64)in1, (v2i64)in0); \
out2 = (v16u8)__msa_ilvr_d((v2i64)in3, (v2i64)in2); \
out3 = (v16u8)__msa_ilvl_d((v2i64)in3, (v2i64)in2); \
}
void TransposeWx16_C(const uint8* src,
int src_stride,
uint8* dst,
int dst_stride,
int width) {
TransposeWx8_C(src, src_stride, dst, dst_stride, width);
TransposeWx8_C((src + 8 * src_stride), src_stride, (dst + 8), dst_stride,
width);
}
void TransposeUVWx16_C(const uint8* src,
int src_stride,
uint8* dst_a,
int dst_stride_a,
uint8* dst_b,
int dst_stride_b,
int width) {
TransposeUVWx8_C(src, src_stride, dst_a, dst_stride_a, dst_b, dst_stride_b,
width);
TransposeUVWx8_C((src + 8 * src_stride), src_stride, (dst_a + 8),
dst_stride_a, (dst_b + 8), dst_stride_b, width);
}
void TransposeWx16_MSA(const uint8* src,
int src_stride,
uint8* dst,
int dst_stride,
int width) {
int x; int x;
uint64_t val0, val1, val2, val3; const uint8* s;
v16u8 src0, src1, src2, src3, src4, src5, src6, src7; v16u8 src0, src1, src2, src3, dst0, dst1, dst2, dst3, vec0, vec1, vec2, vec3;
v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
v16u8 reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7; v16u8 reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;
v16u8 dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7; v16u8 res0, res1, res2, res3, res4, res5, res6, res7, res8, res9;
for (x = 0; x < width; x += 16) { for (x = 0; x < width; x += 16) {
src0 = (v16u8)__msa_ld_b((v16i8*)src, 0); s = src;
src1 = (v16u8)__msa_ld_b((v16i8*)(src + src_stride), 0); src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
src2 = (v16u8)__msa_ld_b((v16i8*)(src + src_stride * 2), 0); s += src_stride;
src3 = (v16u8)__msa_ld_b((v16i8*)(src + src_stride * 3), 0); src1 = (v16u8)__msa_ld_b((v16i8*)s, 0);
src4 = (v16u8)__msa_ld_b((v16i8*)(src + src_stride * 4), 0); s += src_stride;
src5 = (v16u8)__msa_ld_b((v16i8*)(src + src_stride * 5), 0); src2 = (v16u8)__msa_ld_b((v16i8*)s, 0);
src6 = (v16u8)__msa_ld_b((v16i8*)(src + src_stride * 6), 0); s += src_stride;
src7 = (v16u8)__msa_ld_b((v16i8*)(src + src_stride * 7), 0); src3 = (v16u8)__msa_ld_b((v16i8*)s, 0);
vec0 = (v16u8)__msa_ilvr_b((v16i8)src2, (v16i8)src0); s += src_stride;
vec1 = (v16u8)__msa_ilvr_b((v16i8)src3, (v16i8)src1); ILVRL_B(src0, src1, src2, src3, vec0, vec1, vec2, vec3);
vec2 = (v16u8)__msa_ilvr_b((v16i8)src6, (v16i8)src4); ILVRL_H(vec0, vec2, vec1, vec3, reg0, reg1, reg2, reg3);
vec3 = (v16u8)__msa_ilvr_b((v16i8)src7, (v16i8)src5); src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
vec4 = (v16u8)__msa_ilvl_b((v16i8)src2, (v16i8)src0); s += src_stride;
vec5 = (v16u8)__msa_ilvl_b((v16i8)src3, (v16i8)src1); src1 = (v16u8)__msa_ld_b((v16i8*)s, 0);
vec6 = (v16u8)__msa_ilvl_b((v16i8)src6, (v16i8)src4); s += src_stride;
vec7 = (v16u8)__msa_ilvl_b((v16i8)src7, (v16i8)src5); src2 = (v16u8)__msa_ld_b((v16i8*)s, 0);
reg0 = (v16u8)__msa_ilvr_b((v16i8)vec1, (v16i8)vec0); s += src_stride;
reg1 = (v16u8)__msa_ilvl_b((v16i8)vec1, (v16i8)vec0); src3 = (v16u8)__msa_ld_b((v16i8*)s, 0);
reg2 = (v16u8)__msa_ilvr_b((v16i8)vec3, (v16i8)vec2); s += src_stride;
reg3 = (v16u8)__msa_ilvl_b((v16i8)vec3, (v16i8)vec2); ILVRL_B(src0, src1, src2, src3, vec0, vec1, vec2, vec3);
reg4 = (v16u8)__msa_ilvr_b((v16i8)vec5, (v16i8)vec4); ILVRL_H(vec0, vec2, vec1, vec3, reg4, reg5, reg6, reg7);
reg5 = (v16u8)__msa_ilvl_b((v16i8)vec5, (v16i8)vec4); ILVRL_W(reg0, reg4, reg1, reg5, res0, res1, res2, res3);
reg6 = (v16u8)__msa_ilvr_b((v16i8)vec7, (v16i8)vec6); ILVRL_W(reg2, reg6, reg3, reg7, res4, res5, res6, res7);
reg7 = (v16u8)__msa_ilvl_b((v16i8)vec7, (v16i8)vec6); src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
dst0 = (v16u8)__msa_ilvr_w((v4i32)reg2, (v4i32)reg0); s += src_stride;
dst1 = (v16u8)__msa_ilvl_w((v4i32)reg2, (v4i32)reg0); src1 = (v16u8)__msa_ld_b((v16i8*)s, 0);
dst2 = (v16u8)__msa_ilvr_w((v4i32)reg3, (v4i32)reg1); s += src_stride;
dst3 = (v16u8)__msa_ilvl_w((v4i32)reg3, (v4i32)reg1); src2 = (v16u8)__msa_ld_b((v16i8*)s, 0);
dst4 = (v16u8)__msa_ilvr_w((v4i32)reg6, (v4i32)reg4); s += src_stride;
dst5 = (v16u8)__msa_ilvl_w((v4i32)reg6, (v4i32)reg4); src3 = (v16u8)__msa_ld_b((v16i8*)s, 0);
dst6 = (v16u8)__msa_ilvr_w((v4i32)reg7, (v4i32)reg5); s += src_stride;
dst7 = (v16u8)__msa_ilvl_w((v4i32)reg7, (v4i32)reg5); ILVRL_B(src0, src1, src2, src3, vec0, vec1, vec2, vec3);
val0 = __msa_copy_s_d((v2i64)dst0, 0); ILVRL_H(vec0, vec2, vec1, vec3, reg0, reg1, reg2, reg3);
val1 = __msa_copy_s_d((v2i64)dst0, 1); src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
val2 = __msa_copy_s_d((v2i64)dst1, 0); s += src_stride;
val3 = __msa_copy_s_d((v2i64)dst1, 1); src1 = (v16u8)__msa_ld_b((v16i8*)s, 0);
SD(val0, dst); s += src_stride;
SD(val1, dst + dst_stride); src2 = (v16u8)__msa_ld_b((v16i8*)s, 0);
SD(val2, dst + dst_stride * 2); s += src_stride;
SD(val3, dst + dst_stride * 3); src3 = (v16u8)__msa_ld_b((v16i8*)s, 0);
s += src_stride;
ILVRL_B(src0, src1, src2, src3, vec0, vec1, vec2, vec3);
ILVRL_H(vec0, vec2, vec1, vec3, reg4, reg5, reg6, reg7);
res8 = (v16u8)__msa_ilvr_w((v4i32)reg4, (v4i32)reg0);
res9 = (v16u8)__msa_ilvl_w((v4i32)reg4, (v4i32)reg0);
ILVRL_D(res0, res8, res1, res9, dst0, dst1, dst2, dst3);
ST_UB4(dst0, dst1, dst2, dst3, dst, dst_stride);
dst += dst_stride * 4; dst += dst_stride * 4;
val0 = __msa_copy_s_d((v2i64)dst2, 0); res8 = (v16u8)__msa_ilvr_w((v4i32)reg5, (v4i32)reg1);
val1 = __msa_copy_s_d((v2i64)dst2, 1); res9 = (v16u8)__msa_ilvl_w((v4i32)reg5, (v4i32)reg1);
val2 = __msa_copy_s_d((v2i64)dst3, 0); ILVRL_D(res2, res8, res3, res9, dst0, dst1, dst2, dst3);
val3 = __msa_copy_s_d((v2i64)dst3, 1); ST_UB4(dst0, dst1, dst2, dst3, dst, dst_stride);
SD(val0, dst);
SD(val1, dst + dst_stride);
SD(val2, dst + dst_stride * 2);
SD(val3, dst + dst_stride * 3);
dst += dst_stride * 4; dst += dst_stride * 4;
val0 = __msa_copy_s_d((v2i64)dst4, 0); res8 = (v16u8)__msa_ilvr_w((v4i32)reg6, (v4i32)reg2);
val1 = __msa_copy_s_d((v2i64)dst4, 1); res9 = (v16u8)__msa_ilvl_w((v4i32)reg6, (v4i32)reg2);
val2 = __msa_copy_s_d((v2i64)dst5, 0); ILVRL_D(res4, res8, res5, res9, dst0, dst1, dst2, dst3);
val3 = __msa_copy_s_d((v2i64)dst5, 1); ST_UB4(dst0, dst1, dst2, dst3, dst, dst_stride);
SD(val0, dst);
SD(val1, dst + dst_stride);
SD(val2, dst + dst_stride * 2);
SD(val3, dst + dst_stride * 3);
dst += dst_stride * 4;
val0 = __msa_copy_s_d((v2i64)dst6, 0);
val1 = __msa_copy_s_d((v2i64)dst6, 1);
val2 = __msa_copy_s_d((v2i64)dst7, 0);
val3 = __msa_copy_s_d((v2i64)dst7, 1);
SD(val0, dst);
SD(val1, dst + dst_stride);
SD(val2, dst + dst_stride * 2);
SD(val3, dst + dst_stride * 3);
dst += dst_stride * 4; dst += dst_stride * 4;
res8 = (v16u8)__msa_ilvr_w((v4i32)reg7, (v4i32)reg3);
res9 = (v16u8)__msa_ilvl_w((v4i32)reg7, (v4i32)reg3);
ILVRL_D(res6, res8, res7, res9, dst0, dst1, dst2, dst3);
ST_UB4(dst0, dst1, dst2, dst3, dst, dst_stride);
src += 16; src += 16;
dst += dst_stride * 4;
} }
} }
void TransposeUVWx8_MSA(const uint8_t* src, void TransposeUVWx16_MSA(const uint8* src,
int src_stride, int src_stride,
uint8_t* dst_a, uint8* dst_a,
int dst_stride_a, int dst_stride_a,
uint8_t* dst_b, uint8* dst_b,
int dst_stride_b, int dst_stride_b,
int width) { int width) {
int x; int x;
uint64_t val0, val1, val2, val3; const uint8* s;
v16u8 src0, src1, src2, src3, src4, src5, src6, src7; v16u8 src0, src1, src2, src3, dst0, dst1, dst2, dst3, vec0, vec1, vec2, vec3;
v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
v16u8 reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7; v16u8 reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;
v16u8 dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7; v16u8 res0, res1, res2, res3, res4, res5, res6, res7, res8, res9;
for (x = 0; x < width; x += 8) { for (x = 0; x < width; x += 8) {
src0 = (v16u8)__msa_ld_b((v16i8*)src, 0); s = src;
src1 = (v16u8)__msa_ld_b((v16i8*)(src + src_stride), 0); src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
src2 = (v16u8)__msa_ld_b((v16i8*)(src + src_stride * 2), 0); s += src_stride;
src3 = (v16u8)__msa_ld_b((v16i8*)(src + src_stride * 3), 0); src1 = (v16u8)__msa_ld_b((v16i8*)s, 0);
src4 = (v16u8)__msa_ld_b((v16i8*)(src + src_stride * 4), 0); s += src_stride;
src5 = (v16u8)__msa_ld_b((v16i8*)(src + src_stride * 5), 0); src2 = (v16u8)__msa_ld_b((v16i8*)s, 0);
src6 = (v16u8)__msa_ld_b((v16i8*)(src + src_stride * 6), 0); s += src_stride;
src7 = (v16u8)__msa_ld_b((v16i8*)(src + src_stride * 7), 0); src3 = (v16u8)__msa_ld_b((v16i8*)s, 0);
vec0 = (v16u8)__msa_ilvr_b((v16i8)src1, (v16i8)src0); s += src_stride;
vec1 = (v16u8)__msa_ilvr_b((v16i8)src3, (v16i8)src2); ILVRL_B(src0, src1, src2, src3, vec0, vec1, vec2, vec3);
vec2 = (v16u8)__msa_ilvr_b((v16i8)src5, (v16i8)src4); ILVRL_H(vec0, vec2, vec1, vec3, reg0, reg1, reg2, reg3);
vec3 = (v16u8)__msa_ilvr_b((v16i8)src7, (v16i8)src6); src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
vec4 = (v16u8)__msa_ilvl_b((v16i8)src1, (v16i8)src0); s += src_stride;
vec5 = (v16u8)__msa_ilvl_b((v16i8)src3, (v16i8)src2); src1 = (v16u8)__msa_ld_b((v16i8*)s, 0);
vec6 = (v16u8)__msa_ilvl_b((v16i8)src5, (v16i8)src4); s += src_stride;
vec7 = (v16u8)__msa_ilvl_b((v16i8)src7, (v16i8)src6); src2 = (v16u8)__msa_ld_b((v16i8*)s, 0);
reg0 = (v16u8)__msa_ilvr_h((v8i16)vec1, (v8i16)vec0); s += src_stride;
reg1 = (v16u8)__msa_ilvr_h((v8i16)vec3, (v8i16)vec2); src3 = (v16u8)__msa_ld_b((v16i8*)s, 0);
reg2 = (v16u8)__msa_ilvl_h((v8i16)vec1, (v8i16)vec0); s += src_stride;
reg3 = (v16u8)__msa_ilvl_h((v8i16)vec3, (v8i16)vec2); ILVRL_B(src0, src1, src2, src3, vec0, vec1, vec2, vec3);
reg4 = (v16u8)__msa_ilvr_h((v8i16)vec5, (v8i16)vec4); ILVRL_H(vec0, vec2, vec1, vec3, reg4, reg5, reg6, reg7);
reg5 = (v16u8)__msa_ilvr_h((v8i16)vec7, (v8i16)vec6); ILVRL_W(reg0, reg4, reg1, reg5, res0, res1, res2, res3);
reg6 = (v16u8)__msa_ilvl_h((v8i16)vec5, (v8i16)vec5); ILVRL_W(reg2, reg6, reg3, reg7, res4, res5, res6, res7);
reg7 = (v16u8)__msa_ilvl_h((v8i16)vec7, (v8i16)vec6); src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
dst0 = (v16u8)__msa_ilvr_w((v4i32)reg1, (v4i32)reg0); s += src_stride;
dst1 = (v16u8)__msa_ilvl_w((v4i32)reg1, (v4i32)reg0); src1 = (v16u8)__msa_ld_b((v16i8*)s, 0);
dst2 = (v16u8)__msa_ilvr_w((v4i32)reg3, (v4i32)reg2); s += src_stride;
dst3 = (v16u8)__msa_ilvl_w((v4i32)reg3, (v4i32)reg2); src2 = (v16u8)__msa_ld_b((v16i8*)s, 0);
dst4 = (v16u8)__msa_ilvr_w((v4i32)reg5, (v4i32)reg4); s += src_stride;
dst5 = (v16u8)__msa_ilvl_w((v4i32)reg5, (v4i32)reg4); src3 = (v16u8)__msa_ld_b((v16i8*)s, 0);
dst6 = (v16u8)__msa_ilvr_w((v4i32)reg7, (v4i32)reg6); s += src_stride;
dst7 = (v16u8)__msa_ilvl_w((v4i32)reg7, (v4i32)reg6); ILVRL_B(src0, src1, src2, src3, vec0, vec1, vec2, vec3);
val0 = __msa_copy_s_d((v2i64)dst0, 0); ILVRL_H(vec0, vec2, vec1, vec3, reg0, reg1, reg2, reg3);
val1 = __msa_copy_s_d((v2i64)dst0, 1); src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
val2 = __msa_copy_s_d((v2i64)dst1, 0); s += src_stride;
val3 = __msa_copy_s_d((v2i64)dst1, 1); src1 = (v16u8)__msa_ld_b((v16i8*)s, 0);
SD(val0, dst_a); s += src_stride;
SD(val2, dst_a + dst_stride_a); src2 = (v16u8)__msa_ld_b((v16i8*)s, 0);
SD(val1, dst_b); s += src_stride;
SD(val3, dst_b + dst_stride_b); src3 = (v16u8)__msa_ld_b((v16i8*)s, 0);
s += src_stride;
ILVRL_B(src0, src1, src2, src3, vec0, vec1, vec2, vec3);
ILVRL_H(vec0, vec2, vec1, vec3, reg4, reg5, reg6, reg7);
res8 = (v16u8)__msa_ilvr_w((v4i32)reg4, (v4i32)reg0);
res9 = (v16u8)__msa_ilvl_w((v4i32)reg4, (v4i32)reg0);
ILVRL_D(res0, res8, res1, res9, dst0, dst1, dst2, dst3);
ST_UB2(dst0, dst2, dst_a, dst_stride_a);
ST_UB2(dst1, dst3, dst_b, dst_stride_b);
dst_a += dst_stride_a * 2; dst_a += dst_stride_a * 2;
dst_b += dst_stride_b * 2; dst_b += dst_stride_b * 2;
val0 = __msa_copy_s_d((v2i64)dst2, 0); res8 = (v16u8)__msa_ilvr_w((v4i32)reg5, (v4i32)reg1);
val1 = __msa_copy_s_d((v2i64)dst2, 1); res9 = (v16u8)__msa_ilvl_w((v4i32)reg5, (v4i32)reg1);
val2 = __msa_copy_s_d((v2i64)dst3, 0); ILVRL_D(res2, res8, res3, res9, dst0, dst1, dst2, dst3);
val3 = __msa_copy_s_d((v2i64)dst3, 1); ST_UB2(dst0, dst2, dst_a, dst_stride_a);
SD(val0, dst_a); ST_UB2(dst1, dst3, dst_b, dst_stride_b);
SD(val2, dst_a + dst_stride_a);
SD(val1, dst_b);
SD(val3, dst_b + dst_stride_b);
dst_a += dst_stride_a * 2; dst_a += dst_stride_a * 2;
dst_b += dst_stride_b * 2; dst_b += dst_stride_b * 2;
val0 = __msa_copy_s_d((v2i64)dst4, 0); res8 = (v16u8)__msa_ilvr_w((v4i32)reg6, (v4i32)reg2);
val1 = __msa_copy_s_d((v2i64)dst4, 1); res9 = (v16u8)__msa_ilvl_w((v4i32)reg6, (v4i32)reg2);
val2 = __msa_copy_s_d((v2i64)dst5, 0); ILVRL_D(res4, res8, res5, res9, dst0, dst1, dst2, dst3);
val3 = __msa_copy_s_d((v2i64)dst5, 1); ST_UB2(dst0, dst2, dst_a, dst_stride_a);
SD(val0, dst_a); ST_UB2(dst1, dst3, dst_b, dst_stride_b);
SD(val2, dst_a + dst_stride_a);
SD(val1, dst_b);
SD(val3, dst_b + dst_stride_b);
dst_a += dst_stride_a * 2; dst_a += dst_stride_a * 2;
dst_b += dst_stride_b * 2; dst_b += dst_stride_b * 2;
val0 = __msa_copy_s_d((v2i64)dst6, 0); res8 = (v16u8)__msa_ilvr_w((v4i32)reg7, (v4i32)reg3);
val1 = __msa_copy_s_d((v2i64)dst6, 1); res9 = (v16u8)__msa_ilvl_w((v4i32)reg7, (v4i32)reg3);
val2 = __msa_copy_s_d((v2i64)dst7, 0); ILVRL_D(res6, res8, res7, res9, dst0, dst1, dst2, dst3);
val3 = __msa_copy_s_d((v2i64)dst7, 1); ST_UB2(dst0, dst2, dst_a, dst_stride_a);
SD(val0, dst_a); ST_UB2(dst1, dst3, dst_b, dst_stride_b);
SD(val2, dst_a + dst_stride_a); src += 16;
SD(val1, dst_b);
SD(val3, dst_b + dst_stride_b);
dst_a += dst_stride_a * 2; dst_a += dst_stride_a * 2;
dst_b += dst_stride_b * 2; dst_b += dst_stride_b * 2;
src += 16;
} }
} }
......
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