From 6fa5e4eb780b67fe3275a529c6c0da9ea7b58cff Mon Sep 17 00:00:00 2001 From: Manojkumar Bhosale Date: Thu, 15 Dec 2016 10:06:01 +0530 Subject: [PATCH] Add MSA optimized TransposeWx8_MSA and TransposeUVWx8_MSA functions R=fbarchard@google.com BUG=libyuv:634 Performance Gain (vs C vectorized) TransposeWx8_MSA - ~2.7x TransposeWx8_Any_MSA - ~2.1x TransposeUVWx8_MSA - ~2.5x TransposeUVWx8_Any_MSA - ~2.7x Performance Gain (vs C non-vectorized) TransposeWx8_MSA - ~4.6x TransposeWx8_Any_MSA - ~2.9x TransposeUVWx8_MSA - ~4.4x TransposeUVWx8_Any_MSA - ~3.7x Review URL: https://codereview.chromium.org/2553403002 . --- Android.mk | 3 +- BUILD.gn | 1 + CMakeLists.txt | 1 + include/libyuv/macros_msa.h | 115 ++++++++++++++------ include/libyuv/rotate_row.h | 29 ++++++ libyuv.gypi | 1 + source/rotate.cc | 16 +++ source/rotate_any.cc | 6 ++ source/rotate_msa.cc | 203 ++++++++++++++++++++++++++++++++++++ 9 files changed, 342 insertions(+), 33 deletions(-) create mode 100644 source/rotate_msa.cc diff --git a/Android.mk b/Android.mk index 983ee1d85..5d29d900a 100644 --- a/Android.mk +++ b/Android.mk @@ -57,7 +57,8 @@ ifeq ($(TARGET_ARCH_ABI),mips) LOCAL_CFLAGS += -DLIBYUV_MSA LOCAL_SRC_FILES += \ source/row_msa.cc \ - source/scale_msa.cc + source/scale_msa.cc \ + source/rotate_msa.cc endif LOCAL_EXPORT_C_INCLUDES := $(LOCAL_PATH)/include diff --git a/BUILD.gn b/BUILD.gn index c63faf5c5..79ee567a9 100644 --- a/BUILD.gn +++ b/BUILD.gn @@ -165,6 +165,7 @@ if (libyuv_use_msa) { # MSA Source Files "source/row_msa.cc", "source/scale_msa.cc", + "source/rotate_msa.cc", ] public_configs = [ ":libyuv_config" ] diff --git a/CMakeLists.txt b/CMakeLists.txt index 2d8f5224e..e5e2c699f 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -33,6 +33,7 @@ set(ly_source_files ${ly_src_dir}/rotate_argb.cc ${ly_src_dir}/rotate_common.cc ${ly_src_dir}/rotate_mips.cc + ${ly_src_dir}/rotate_msa.cc ${ly_src_dir}/rotate_neon.cc ${ly_src_dir}/rotate_neon64.cc ${ly_src_dir}/rotate_gcc.cc diff --git a/include/libyuv/macros_msa.h b/include/libyuv/macros_msa.h index 9be804e11..a2704ba79 100644 --- a/include/libyuv/macros_msa.h +++ b/include/libyuv/macros_msa.h @@ -16,25 +16,25 @@ #include #if (__mips_isa_rev >= 6) -#define LW(psrc) \ - ({ \ - uint8* psrc_lw_m = (uint8*)(psrc); /* NOLINT */ \ - uint32 val_m; \ - asm volatile("lw %[val_m], %[psrc_lw_m] \n\t" \ - : [val_m] "=r"(val_m) \ - : [psrc_lw_m] "m"(*psrc_lw_m)); \ - val_m; \ +#define LW(psrc) \ + ({ \ + uint8* psrc_lw_m = (uint8*)(psrc); /* NOLINT */ \ + uint32 val_m; \ + asm volatile("lw %[val_m], %[psrc_lw_m] \n" \ + : [val_m] "=r"(val_m) \ + : [psrc_lw_m] "m"(*psrc_lw_m)); \ + val_m; \ }) #if (__mips == 64) -#define LD(psrc) \ - ({ \ - uint8* psrc_ld_m = (uint8*)(psrc); /* NOLINT */ \ - uint64 val_m = 0; \ - asm volatile("ld %[val_m], %[psrc_ld_m] \n\t" \ - : [val_m] "=r"(val_m) \ - : [psrc_ld_m] "m"(*psrc_ld_m)); \ - val_m; \ +#define LD(psrc) \ + ({ \ + uint8* psrc_ld_m = (uint8*)(psrc); /* NOLINT */ \ + uint64 val_m = 0; \ + asm volatile("ld %[val_m], %[psrc_ld_m] \n" \ + : [val_m] "=r"(val_m) \ + : [psrc_ld_m] "m"(*psrc_ld_m)); \ + val_m; \ }) #else // !(__mips == 64) #define LD(psrc) \ @@ -50,26 +50,58 @@ val_m; \ }) #endif // (__mips == 64) + +#define SW(val, pdst) \ + ({ \ + uint8_t* pdst_sw_m = (uint8_t*)(pdst); \ + uint32_t val_m = (val); \ + asm volatile("sw %[val_m], %[pdst_sw_m] \n" \ + \ + : [pdst_sw_m] "=m"(*pdst_sw_m) \ + : [val_m] "r"(val_m)); \ + }) + +#if (__mips == 64) +#define SD(val, pdst) \ + ({ \ + uint8_t* pdst_sd_m = (uint8_t*)(pdst); \ + uint64_t val_m = (val); \ + asm volatile("sd %[val_m], %[pdst_sd_m] \n" \ + \ + : [pdst_sd_m] "=m"(*pdst_sd_m) \ + : [val_m] "r"(val_m)); \ + }) +#else // !(__mips == 64) +#define SD(val, pdst) \ + ({ \ + uint8_t* pdst_sd_m = (uint8_t*)(pdst); \ + uint32_t val0_m, val1_m; \ + val0_m = (uint32_t)((val)&0x00000000FFFFFFFF); \ + val1_m = (uint32_t)(((val) >> 32) & 0x00000000FFFFFFFF); \ + SW(val0_m, pdst_sd_m); \ + SW(val1_m, pdst_sd_m + 4); \ + }) +#endif // !(__mips == 64) #else // !(__mips_isa_rev >= 6) -#define LW(psrc) \ - ({ \ - uint8* psrc_lw_m = (uint8*)(psrc); /* NOLINT */ \ - uint32 val_m; \ - asm volatile("ulw %[val_m], %[psrc_lw_m] \n\t" \ - : [val_m] "=r"(val_m) \ - : [psrc_lw_m] "m"(*psrc_lw_m)); \ - val_m; \ +#define LW(psrc) \ + ({ \ + uint8* psrc_lw_m = (uint8*)(psrc); /* NOLINT */ \ + uint32 val_m; \ + asm volatile("ulw %[val_m], %[psrc_lw_m] \n" \ + : [val_m] "=r"(val_m) \ + : [psrc_lw_m] "m"(*psrc_lw_m)); \ + val_m; \ }) #if (__mips == 64) -#define LD(psrc) \ - ({ \ - uint8* psrc_ld_m = (uint8*)(psrc); /* NOLINT */ \ - uint64 val_m = 0; \ - asm volatile("uld %[val_m], %[psrc_ld_m] \n\t" \ - : [val_m] "=r"(val_m) \ - : [psrc_ld_m] "m"(*psrc_ld_m)); \ - val_m; \ +#define LD(psrc) \ + ({ \ + uint8* psrc_ld_m = (uint8*)(psrc); /* NOLINT */ \ + uint64 val_m = 0; \ + asm volatile("uld %[val_m], %[psrc_ld_m] \n" \ + : [val_m] "=r"(val_m) \ + : [psrc_ld_m] "m"(*psrc_ld_m)); \ + val_m; \ }) #else // !(__mips == 64) #define LD(psrc) \ @@ -85,6 +117,25 @@ val_m; \ }) #endif // (__mips == 64) + +#define SW(val, pdst) \ + ({ \ + uint8_t* pdst_sw_m = (uint8_t*)(pdst); \ + uint32_t val_m = (val); \ + asm volatile("usw %[val_m], %[pdst_sw_m] \n" \ + : [pdst_sw_m] "=m"(*pdst_sw_m) \ + : [val_m] "r"(val_m)); \ + }) + +#define SD(val, pdst) \ + ({ \ + uint8_t* pdst_sd_m = (uint8_t*)(pdst); \ + uint32_t val0_m, val1_m; \ + val0_m = (uint32_t)((val)&0x00000000FFFFFFFF); \ + val1_m = (uint32_t)(((val) >> 32) & 0x00000000FFFFFFFF); \ + SW(val0_m, pdst_sd_m); \ + SW(val1_m, pdst_sd_m + 4); \ + }) #endif // (__mips_isa_rev >= 6) // TODO(fbarchard): Consider removing __VAR_ARGS versions. diff --git a/include/libyuv/rotate_row.h b/include/libyuv/rotate_row.h index af1d0a05f..fda6b59f3 100644 --- a/include/libyuv/rotate_row.h +++ b/include/libyuv/rotate_row.h @@ -60,6 +60,11 @@ extern "C" { #define HAS_TRANSPOSEUVWX8_DSPR2 #endif // defined(__mips__) +#if !defined(LIBYUV_DISABLE_MSA) && defined(__mips_msa) +#define HAS_TRANSPOSEWX8_MSA +#define HAS_TRANSPOSEUVWX8_MSA +#endif + void TransposeWxH_C(const uint8* src, int src_stride, uint8* dst, @@ -97,6 +102,11 @@ void TransposeWx8_Fast_DSPR2(const uint8* src, uint8* dst, int dst_stride, int width); +void TransposeWx8_MSA(const uint8* src, + int src_stride, + uint8* dst, + int dst_stride, + int width); void TransposeWx8_Any_NEON(const uint8* src, int src_stride, @@ -118,6 +128,11 @@ void TransposeWx8_Any_DSPR2(const uint8* src, uint8* dst, int dst_stride, int width); +void TransposeWx8_Any_MSA(const uint8* src, + int src_stride, + uint8* dst, + int dst_stride, + int width); void TransposeUVWxH_C(const uint8* src, int src_stride, @@ -156,6 +171,13 @@ void TransposeUVWx8_DSPR2(const uint8* src, uint8* dst_b, int dst_stride_b, int width); +void TransposeUVWx8_MSA(const uint8* src, + int src_stride, + uint8* dst_a, + int dst_stride_a, + uint8* dst_b, + int dst_stride_b, + int width); void TransposeUVWx8_Any_SSE2(const uint8* src, int src_stride, @@ -178,6 +200,13 @@ void TransposeUVWx8_Any_DSPR2(const uint8* src, uint8* dst_b, int dst_stride_b, int width); +void TransposeUVWx8_Any_MSA(const uint8* src, + int src_stride, + uint8* dst_a, + int dst_stride_a, + uint8* dst_b, + int dst_stride_b, + int width); #ifdef __cplusplus } // extern "C" diff --git a/libyuv.gypi b/libyuv.gypi index 218d25d93..165d92973 100644 --- a/libyuv.gypi +++ b/libyuv.gypi @@ -55,6 +55,7 @@ 'source/rotate_common.cc', 'source/rotate_gcc.cc', 'source/rotate_mips.cc', + 'source/rotate_msa.cc', 'source/rotate_neon.cc', 'source/rotate_neon64.cc', 'source/rotate_win.cc', diff --git a/source/rotate.cc b/source/rotate.cc index 098fcf7c7..bfd51421b 100644 --- a/source/rotate.cc +++ b/source/rotate.cc @@ -62,6 +62,14 @@ void TransposePlane(const uint8* src, } } #endif +#if defined(HAS_TRANSPOSEWX8_MSA) + if (TestCpuFlag(kCpuHasMSA)) { + TransposeWx8 = TransposeWx8_Any_MSA; + if (IS_ALIGNED(width, 16)) { + TransposeWx8 = TransposeWx8_MSA; + } + } +#endif // Work across the source in 8x8 tiles while (i >= 8) { @@ -232,6 +240,14 @@ void TransposeUV(const uint8* src, TransposeUVWx8 = TransposeUVWx8_DSPR2; } #endif +#if defined(HAS_TRANSPOSEUVWX8_MSA) + if (TestCpuFlag(kCpuHasMSA)) { + TransposeUVWx8 = TransposeUVWx8_Any_MSA; + if (IS_ALIGNED(width, 8)) { + TransposeUVWx8 = TransposeUVWx8_MSA; + } + } +#endif // Work through the source in 8x8 tiles. while (i >= 8) { diff --git a/source/rotate_any.cc b/source/rotate_any.cc index f63329ee9..67e565e4d 100644 --- a/source/rotate_any.cc +++ b/source/rotate_any.cc @@ -41,6 +41,9 @@ TANY(TransposeWx8_Fast_Any_SSSE3, TransposeWx8_Fast_SSSE3, 15) #ifdef HAS_TRANSPOSEWX8_DSPR2 TANY(TransposeWx8_Any_DSPR2, TransposeWx8_DSPR2, 7) #endif +#ifdef HAS_TRANSPOSEWX8_MSA +TANY(TransposeWx8_Any_MSA, TransposeWx8_MSA, 15) +#endif #undef TANY #define TUVANY(NAMEANY, TPOS_SIMD, MASK) \ @@ -64,6 +67,9 @@ TUVANY(TransposeUVWx8_Any_SSE2, TransposeUVWx8_SSE2, 7) #ifdef HAS_TRANSPOSEUVWX8_DSPR2 TUVANY(TransposeUVWx8_Any_DSPR2, TransposeUVWx8_DSPR2, 7) #endif +#ifdef HAS_TRANSPOSEUVWX8_MSA +TUVANY(TransposeUVWx8_Any_MSA, TransposeUVWx8_MSA, 7) +#endif #undef TUVANY #ifdef __cplusplus diff --git a/source/rotate_msa.cc b/source/rotate_msa.cc new file mode 100644 index 000000000..08fbde019 --- /dev/null +++ b/source/rotate_msa.cc @@ -0,0 +1,203 @@ +/* + * Copyright 2016 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/rotate_row.h" + +// This module is for GCC MSA +#if !defined(LIBYUV_DISABLE_MSA) && defined(__mips_msa) +#include "libyuv/macros_msa.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +void TransposeWx8_MSA(const uint8_t* src, + int src_stride, + uint8_t* dst, + int dst_stride, + int width) { + int x; + uint64_t val0, val1, val2, val3; + v16u8 src0, src1, src2, src3, src4, src5, src6, src7; + v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7; + v16u8 reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7; + v16u8 dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7; + + for (x = 0; x < width; x += 16) { + src0 = (v16u8)__msa_ld_b((v16i8*)src, 0); + src1 = (v16u8)__msa_ld_b((v16i8*)(src + src_stride), 0); + src2 = (v16u8)__msa_ld_b((v16i8*)(src + src_stride * 2), 0); + src3 = (v16u8)__msa_ld_b((v16i8*)(src + src_stride * 3), 0); + src4 = (v16u8)__msa_ld_b((v16i8*)(src + src_stride * 4), 0); + src5 = (v16u8)__msa_ld_b((v16i8*)(src + src_stride * 5), 0); + src6 = (v16u8)__msa_ld_b((v16i8*)(src + src_stride * 6), 0); + src7 = (v16u8)__msa_ld_b((v16i8*)(src + src_stride * 7), 0); + vec0 = (v16u8)__msa_ilvr_b((v16i8)src2, (v16i8)src0); + vec1 = (v16u8)__msa_ilvr_b((v16i8)src3, (v16i8)src1); + vec2 = (v16u8)__msa_ilvr_b((v16i8)src6, (v16i8)src4); + vec3 = (v16u8)__msa_ilvr_b((v16i8)src7, (v16i8)src5); + vec4 = (v16u8)__msa_ilvl_b((v16i8)src2, (v16i8)src0); + vec5 = (v16u8)__msa_ilvl_b((v16i8)src3, (v16i8)src1); + vec6 = (v16u8)__msa_ilvl_b((v16i8)src6, (v16i8)src4); + vec7 = (v16u8)__msa_ilvl_b((v16i8)src7, (v16i8)src5); + reg0 = (v16u8)__msa_ilvr_b((v16i8)vec1, (v16i8)vec0); + reg1 = (v16u8)__msa_ilvl_b((v16i8)vec1, (v16i8)vec0); + reg2 = (v16u8)__msa_ilvr_b((v16i8)vec3, (v16i8)vec2); + reg3 = (v16u8)__msa_ilvl_b((v16i8)vec3, (v16i8)vec2); + reg4 = (v16u8)__msa_ilvr_b((v16i8)vec5, (v16i8)vec4); + reg5 = (v16u8)__msa_ilvl_b((v16i8)vec5, (v16i8)vec4); + reg6 = (v16u8)__msa_ilvr_b((v16i8)vec7, (v16i8)vec6); + reg7 = (v16u8)__msa_ilvl_b((v16i8)vec7, (v16i8)vec6); + dst0 = (v16u8)__msa_ilvr_w((v4i32)reg2, (v4i32)reg0); + dst1 = (v16u8)__msa_ilvl_w((v4i32)reg2, (v4i32)reg0); + dst2 = (v16u8)__msa_ilvr_w((v4i32)reg3, (v4i32)reg1); + dst3 = (v16u8)__msa_ilvl_w((v4i32)reg3, (v4i32)reg1); + dst4 = (v16u8)__msa_ilvr_w((v4i32)reg6, (v4i32)reg4); + dst5 = (v16u8)__msa_ilvl_w((v4i32)reg6, (v4i32)reg4); + dst6 = (v16u8)__msa_ilvr_w((v4i32)reg7, (v4i32)reg5); + dst7 = (v16u8)__msa_ilvl_w((v4i32)reg7, (v4i32)reg5); + val0 = __msa_copy_s_d((v2i64)dst0, 0); + val1 = __msa_copy_s_d((v2i64)dst0, 1); + val2 = __msa_copy_s_d((v2i64)dst1, 0); + val3 = __msa_copy_s_d((v2i64)dst1, 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; + val0 = __msa_copy_s_d((v2i64)dst2, 0); + val1 = __msa_copy_s_d((v2i64)dst2, 1); + val2 = __msa_copy_s_d((v2i64)dst3, 0); + val3 = __msa_copy_s_d((v2i64)dst3, 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; + val0 = __msa_copy_s_d((v2i64)dst4, 0); + val1 = __msa_copy_s_d((v2i64)dst4, 1); + val2 = __msa_copy_s_d((v2i64)dst5, 0); + val3 = __msa_copy_s_d((v2i64)dst5, 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; + 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; + src += 16; + } +} + +void TransposeUVWx8_MSA(const uint8_t* src, + int src_stride, + uint8_t* dst_a, + int dst_stride_a, + uint8_t* dst_b, + int dst_stride_b, + int width) { + int x; + uint64_t val0, val1, val2, val3; + v16u8 src0, src1, src2, src3, src4, src5, src6, src7; + v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7; + v16u8 reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7; + v16u8 dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7; + + for (x = 0; x < width; x += 8) { + src0 = (v16u8)__msa_ld_b((v16i8*)src, 0); + src1 = (v16u8)__msa_ld_b((v16i8*)(src + src_stride), 0); + src2 = (v16u8)__msa_ld_b((v16i8*)(src + src_stride * 2), 0); + src3 = (v16u8)__msa_ld_b((v16i8*)(src + src_stride * 3), 0); + src4 = (v16u8)__msa_ld_b((v16i8*)(src + src_stride * 4), 0); + src5 = (v16u8)__msa_ld_b((v16i8*)(src + src_stride * 5), 0); + src6 = (v16u8)__msa_ld_b((v16i8*)(src + src_stride * 6), 0); + src7 = (v16u8)__msa_ld_b((v16i8*)(src + src_stride * 7), 0); + vec0 = (v16u8)__msa_ilvr_b((v16i8)src1, (v16i8)src0); + vec1 = (v16u8)__msa_ilvr_b((v16i8)src3, (v16i8)src2); + vec2 = (v16u8)__msa_ilvr_b((v16i8)src5, (v16i8)src4); + vec3 = (v16u8)__msa_ilvr_b((v16i8)src7, (v16i8)src6); + vec4 = (v16u8)__msa_ilvl_b((v16i8)src1, (v16i8)src0); + vec5 = (v16u8)__msa_ilvl_b((v16i8)src3, (v16i8)src2); + vec6 = (v16u8)__msa_ilvl_b((v16i8)src5, (v16i8)src4); + vec7 = (v16u8)__msa_ilvl_b((v16i8)src7, (v16i8)src6); + reg0 = (v16u8)__msa_ilvr_h((v8i16)vec1, (v8i16)vec0); + reg1 = (v16u8)__msa_ilvr_h((v8i16)vec3, (v8i16)vec2); + reg2 = (v16u8)__msa_ilvl_h((v8i16)vec1, (v8i16)vec0); + reg3 = (v16u8)__msa_ilvl_h((v8i16)vec3, (v8i16)vec2); + reg4 = (v16u8)__msa_ilvr_h((v8i16)vec5, (v8i16)vec4); + reg5 = (v16u8)__msa_ilvr_h((v8i16)vec7, (v8i16)vec6); + reg6 = (v16u8)__msa_ilvl_h((v8i16)vec5, (v8i16)vec5); + reg7 = (v16u8)__msa_ilvl_h((v8i16)vec7, (v8i16)vec6); + dst0 = (v16u8)__msa_ilvr_w((v4i32)reg1, (v4i32)reg0); + dst1 = (v16u8)__msa_ilvl_w((v4i32)reg1, (v4i32)reg0); + dst2 = (v16u8)__msa_ilvr_w((v4i32)reg3, (v4i32)reg2); + dst3 = (v16u8)__msa_ilvl_w((v4i32)reg3, (v4i32)reg2); + dst4 = (v16u8)__msa_ilvr_w((v4i32)reg5, (v4i32)reg4); + dst5 = (v16u8)__msa_ilvl_w((v4i32)reg5, (v4i32)reg4); + dst6 = (v16u8)__msa_ilvr_w((v4i32)reg7, (v4i32)reg6); + dst7 = (v16u8)__msa_ilvl_w((v4i32)reg7, (v4i32)reg6); + val0 = __msa_copy_s_d((v2i64)dst0, 0); + val1 = __msa_copy_s_d((v2i64)dst0, 1); + val2 = __msa_copy_s_d((v2i64)dst1, 0); + val3 = __msa_copy_s_d((v2i64)dst1, 1); + SD(val0, dst_a); + 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_b += dst_stride_b * 2; + val0 = __msa_copy_s_d((v2i64)dst2, 0); + val1 = __msa_copy_s_d((v2i64)dst2, 1); + val2 = __msa_copy_s_d((v2i64)dst3, 0); + val3 = __msa_copy_s_d((v2i64)dst3, 1); + SD(val0, dst_a); + 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_b += dst_stride_b * 2; + val0 = __msa_copy_s_d((v2i64)dst4, 0); + val1 = __msa_copy_s_d((v2i64)dst4, 1); + val2 = __msa_copy_s_d((v2i64)dst5, 0); + val3 = __msa_copy_s_d((v2i64)dst5, 1); + SD(val0, dst_a); + 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_b += dst_stride_b * 2; + 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_a); + 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_b += dst_stride_b * 2; + src += 16; + } +} + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif + +#endif // !defined(LIBYUV_DISABLE_MSA) && defined(__mips_msa)