From 3d190ee9f1f9e9e1be1c67cfd4cd24b039d7bc6f Mon Sep 17 00:00:00 2001 From: Frank Barchard Date: Tue, 14 Jul 2015 10:23:10 -0700 Subject: [PATCH] break rotate into files by cpu in preparation for optimization. R=bcornell@google.com BUG=libyuv:464 Review URL: https://webrtc-codereview.appspot.com/51289004. --- Android.mk | 2 + BUILD.gn | 4 + CMakeLists.txt | 4 + README.chromium | 2 +- include/libyuv/version.h | 2 +- libyuv.gypi | 3 + linux.mk | 2 + source/rotate_common.cc | 97 ++++++++ source/rotate_gcc.cc | 494 +++++++++++++++++++++++++++++++++++++++ source/rotate_win.cc | 248 ++++++++++++++++++++ winarm.mk | 1 + 11 files changed, 857 insertions(+), 2 deletions(-) create mode 100644 source/rotate_common.cc create mode 100644 source/rotate_gcc.cc create mode 100644 source/rotate_win.cc diff --git a/Android.mk b/Android.mk index bb8802373..2696f5821 100644 --- a/Android.mk +++ b/Android.mk @@ -20,8 +20,10 @@ LOCAL_SRC_FILES := \ source/planar_functions.cc \ source/rotate.cc \ source/rotate_argb.cc \ + source/rotate_common.cc \ source/rotate_mips.cc \ source/rotate_neon64.cc \ + source/rotate_gcc.cc \ source/row_any.cc \ source/row_common.cc \ source/row_mips.cc \ diff --git a/BUILD.gn b/BUILD.gn index 43a25e018..455274f8e 100644 --- a/BUILD.gn +++ b/BUILD.gn @@ -31,6 +31,7 @@ source_set("libyuv") { "include/libyuv/planar_functions.h", "include/libyuv/rotate.h", "include/libyuv/rotate_argb.h", + "include/libyuv/rotate_row.h", "include/libyuv/row.h", "include/libyuv/scale.h", "include/libyuv/scale_argb.h", @@ -56,7 +57,10 @@ source_set("libyuv") { "source/planar_functions.cc", "source/rotate.cc", "source/rotate_argb.cc", + "source/rotate_common.cc", "source/rotate_mips.cc", + "source/rotate_gcc.cc", + "source/rotate_win.cc", "source/row_any.cc", "source/row_common.cc", "source/row_mips.cc", diff --git a/CMakeLists.txt b/CMakeLists.txt index 1e0acc56f..6c7082503 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -29,8 +29,11 @@ set(ly_source_files ${ly_src_dir}/planar_functions.cc ${ly_src_dir}/rotate.cc ${ly_src_dir}/rotate_argb.cc + ${ly_src_dir}/rotate_common.cc ${ly_src_dir}/rotate_mips.cc ${ly_src_dir}/rotate_neon.cc + ${ly_src_dir}/rotate_gcc.cc + ${ly_src_dir}/rotate_win.cc ${ly_src_dir}/row_any.cc ${ly_src_dir}/row_common.cc ${ly_src_dir}/row_mips.cc @@ -77,6 +80,7 @@ set(ly_header_files ${ly_inc_dir}/libyuv/planar_functions.h ${ly_inc_dir}/libyuv/rotate.h ${ly_inc_dir}/libyuv/rotate_argb.h + ${ly_inc_dir}/libyuv/rotate_row.h ${ly_inc_dir}/libyuv/row.h ${ly_inc_dir}/libyuv/scale.h ${ly_inc_dir}/libyuv/scale_argb.h diff --git a/README.chromium b/README.chromium index 505b420a2..a3095c49f 100644 --- a/README.chromium +++ b/README.chromium @@ -1,6 +1,6 @@ Name: libyuv URL: http://code.google.com/p/libyuv/ -Version: 1444 +Version: 1445 License: BSD License File: LICENSE diff --git a/include/libyuv/version.h b/include/libyuv/version.h index 0c2e7be82..ba35193f5 100644 --- a/include/libyuv/version.h +++ b/include/libyuv/version.h @@ -11,6 +11,6 @@ #ifndef INCLUDE_LIBYUV_VERSION_H_ // NOLINT #define INCLUDE_LIBYUV_VERSION_H_ -#define LIBYUV_VERSION 1444 +#define LIBYUV_VERSION 1445 #endif // INCLUDE_LIBYUV_VERSION_H_ NOLINT diff --git a/libyuv.gypi b/libyuv.gypi index 47209c9cb..eccde39c8 100644 --- a/libyuv.gypi +++ b/libyuv.gypi @@ -47,7 +47,10 @@ 'source/planar_functions.cc', 'source/rotate.cc', 'source/rotate_argb.cc', + 'source/rotate_common.cc', 'source/rotate_mips.cc', + 'source/rotate_gcc.cc', + 'source/rotate_win.cc', 'source/row_any.cc', 'source/row_common.cc', 'source/row_mips.cc', diff --git a/linux.mk b/linux.mk index aca8bed3b..d0b47a690 100644 --- a/linux.mk +++ b/linux.mk @@ -19,6 +19,8 @@ LOCAL_OBJ_FILES := \ source/planar_functions.o \ source/rotate.o \ source/rotate_argb.o \ + source/rotate_common.o \ + source/rotate_gcc.o \ source/rotate_mips.o \ source/row_any.o \ source/row_common.o \ diff --git a/source/rotate_common.cc b/source/rotate_common.cc new file mode 100644 index 000000000..b53326ee8 --- /dev/null +++ b/source/rotate_common.cc @@ -0,0 +1,97 @@ +/* + * Copyright 2011 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.h" + +#include "libyuv/cpu_id.h" +#include "libyuv/convert.h" +#include "libyuv/planar_functions.h" +#include "libyuv/rotate_row.h" +#include "libyuv/row.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +void TransposeWx8_C(const uint8* src, int src_stride, + uint8* dst, int dst_stride, int width) { + int i; + for (i = 0; i < width; ++i) { + dst[0] = src[0 * src_stride]; + dst[1] = src[1 * src_stride]; + dst[2] = src[2 * src_stride]; + dst[3] = src[3 * src_stride]; + dst[4] = src[4 * src_stride]; + dst[5] = src[5 * src_stride]; + dst[6] = src[6 * src_stride]; + dst[7] = src[7 * src_stride]; + ++src; + dst += dst_stride; + } +} + +void TransposeUVWx8_C(const uint8* src, int src_stride, + uint8* dst_a, int dst_stride_a, + uint8* dst_b, int dst_stride_b, int width) { + int i; + for (i = 0; i < width; ++i) { + dst_a[0] = src[0 * src_stride + 0]; + dst_b[0] = src[0 * src_stride + 1]; + dst_a[1] = src[1 * src_stride + 0]; + dst_b[1] = src[1 * src_stride + 1]; + dst_a[2] = src[2 * src_stride + 0]; + dst_b[2] = src[2 * src_stride + 1]; + dst_a[3] = src[3 * src_stride + 0]; + dst_b[3] = src[3 * src_stride + 1]; + dst_a[4] = src[4 * src_stride + 0]; + dst_b[4] = src[4 * src_stride + 1]; + dst_a[5] = src[5 * src_stride + 0]; + dst_b[5] = src[5 * src_stride + 1]; + dst_a[6] = src[6 * src_stride + 0]; + dst_b[6] = src[6 * src_stride + 1]; + dst_a[7] = src[7 * src_stride + 0]; + dst_b[7] = src[7 * src_stride + 1]; + src += 2; + dst_a += dst_stride_a; + dst_b += dst_stride_b; + } +} + +void TransposeWxH_C(const uint8* src, int src_stride, + uint8* dst, int dst_stride, + int width, int height) { + int i; + for (i = 0; i < width; ++i) { + int j; + for (j = 0; j < height; ++j) { + dst[i * dst_stride + j] = src[j * src_stride + i]; + } + } +} + +void TransposeUVWxH_C(const uint8* src, int src_stride, + uint8* dst_a, int dst_stride_a, + uint8* dst_b, int dst_stride_b, + int width, int height) { + int i; + for (i = 0; i < width * 2; i += 2) { + int j; + for (j = 0; j < height; ++j) { + dst_a[j + ((i >> 1) * dst_stride_a)] = src[i + (j * src_stride)]; + dst_b[j + ((i >> 1) * dst_stride_b)] = src[i + (j * src_stride) + 1]; + } + } +} + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/source/rotate_gcc.cc b/source/rotate_gcc.cc new file mode 100644 index 000000000..609f92306 --- /dev/null +++ b/source/rotate_gcc.cc @@ -0,0 +1,494 @@ +/* + * Copyright 2015 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/row.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +// This module is for GCC x86 and x64. +#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__)) + +#if !defined(LIBYUV_DISABLE_X86) && \ + (defined(__i386__) || (defined(__x86_64__) && !defined(__native_client__))) +void TransposeWx8_SSSE3(const uint8* src, int src_stride, + uint8* dst, int dst_stride, int width) { + asm volatile ( + // Read in the data from the source pointer. + // First round of bit swap. + ".p2align 2 \n" + "1: \n" + "movq (%0),%%xmm0 \n" + "movq (%0,%3),%%xmm1 \n" + "lea (%0,%3,2),%0 \n" + "punpcklbw %%xmm1,%%xmm0 \n" + "movq (%0),%%xmm2 \n" + "movdqa %%xmm0,%%xmm1 \n" + "palignr $0x8,%%xmm1,%%xmm1 \n" + "movq (%0,%3),%%xmm3 \n" + "lea (%0,%3,2),%0 \n" + "punpcklbw %%xmm3,%%xmm2 \n" + "movdqa %%xmm2,%%xmm3 \n" + "movq (%0),%%xmm4 \n" + "palignr $0x8,%%xmm3,%%xmm3 \n" + "movq (%0,%3),%%xmm5 \n" + "lea (%0,%3,2),%0 \n" + "punpcklbw %%xmm5,%%xmm4 \n" + "movdqa %%xmm4,%%xmm5 \n" + "movq (%0),%%xmm6 \n" + "palignr $0x8,%%xmm5,%%xmm5 \n" + "movq (%0,%3),%%xmm7 \n" + "lea (%0,%3,2),%0 \n" + "punpcklbw %%xmm7,%%xmm6 \n" + "neg %3 \n" + "movdqa %%xmm6,%%xmm7 \n" + "lea 0x8(%0,%3,8),%0 \n" + "palignr $0x8,%%xmm7,%%xmm7 \n" + "neg %3 \n" + // Second round of bit swap. + "punpcklwd %%xmm2,%%xmm0 \n" + "punpcklwd %%xmm3,%%xmm1 \n" + "movdqa %%xmm0,%%xmm2 \n" + "movdqa %%xmm1,%%xmm3 \n" + "palignr $0x8,%%xmm2,%%xmm2 \n" + "palignr $0x8,%%xmm3,%%xmm3 \n" + "punpcklwd %%xmm6,%%xmm4 \n" + "punpcklwd %%xmm7,%%xmm5 \n" + "movdqa %%xmm4,%%xmm6 \n" + "movdqa %%xmm5,%%xmm7 \n" + "palignr $0x8,%%xmm6,%%xmm6 \n" + "palignr $0x8,%%xmm7,%%xmm7 \n" + // Third round of bit swap. + // Write to the destination pointer. + "punpckldq %%xmm4,%%xmm0 \n" + "movq %%xmm0,(%1) \n" + "movdqa %%xmm0,%%xmm4 \n" + "palignr $0x8,%%xmm4,%%xmm4 \n" + "movq %%xmm4,(%1,%4) \n" + "lea (%1,%4,2),%1 \n" + "punpckldq %%xmm6,%%xmm2 \n" + "movdqa %%xmm2,%%xmm6 \n" + "movq %%xmm2,(%1) \n" + "palignr $0x8,%%xmm6,%%xmm6 \n" + "punpckldq %%xmm5,%%xmm1 \n" + "movq %%xmm6,(%1,%4) \n" + "lea (%1,%4,2),%1 \n" + "movdqa %%xmm1,%%xmm5 \n" + "movq %%xmm1,(%1) \n" + "palignr $0x8,%%xmm5,%%xmm5 \n" + "movq %%xmm5,(%1,%4) \n" + "lea (%1,%4,2),%1 \n" + "punpckldq %%xmm7,%%xmm3 \n" + "movq %%xmm3,(%1) \n" + "movdqa %%xmm3,%%xmm7 \n" + "palignr $0x8,%%xmm7,%%xmm7 \n" + "sub $0x8,%2 \n" + "movq %%xmm7,(%1,%4) \n" + "lea (%1,%4,2),%1 \n" + "jg 1b \n" + : "+r"(src), // %0 + "+r"(dst), // %1 + "+r"(width) // %2 + : "r"((intptr_t)(src_stride)), // %3 + "r"((intptr_t)(dst_stride)) // %4 + : "memory", "cc", + "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" + ); +} + +#if !defined(LIBYUV_DISABLE_X86) && defined(__i386__) && !defined(__clang__) +void TransposeUVWx8_SSE2(const uint8* src, int src_stride, + uint8* dst_a, int dst_stride_a, + uint8* dst_b, int dst_stride_b, + int w); + asm ( + DECLARE_FUNCTION(TransposeUVWx8_SSE2) + "push %ebx \n" + "push %esi \n" + "push %edi \n" + "push %ebp \n" + "mov 0x14(%esp),%eax \n" + "mov 0x18(%esp),%edi \n" + "mov 0x1c(%esp),%edx \n" + "mov 0x20(%esp),%esi \n" + "mov 0x24(%esp),%ebx \n" + "mov 0x28(%esp),%ebp \n" + "mov %esp,%ecx \n" + "sub $0x14,%esp \n" + "and $0xfffffff0,%esp \n" + "mov %ecx,0x10(%esp) \n" + "mov 0x2c(%ecx),%ecx \n" + +"1: \n" + "movdqu (%eax),%xmm0 \n" + "movdqu (%eax,%edi,1),%xmm1 \n" + "lea (%eax,%edi,2),%eax \n" + "movdqa %xmm0,%xmm7 \n" + "punpcklbw %xmm1,%xmm0 \n" + "punpckhbw %xmm1,%xmm7 \n" + "movdqa %xmm7,%xmm1 \n" + "movdqu (%eax),%xmm2 \n" + "movdqu (%eax,%edi,1),%xmm3 \n" + "lea (%eax,%edi,2),%eax \n" + "movdqa %xmm2,%xmm7 \n" + "punpcklbw %xmm3,%xmm2 \n" + "punpckhbw %xmm3,%xmm7 \n" + "movdqa %xmm7,%xmm3 \n" + "movdqu (%eax),%xmm4 \n" + "movdqu (%eax,%edi,1),%xmm5 \n" + "lea (%eax,%edi,2),%eax \n" + "movdqa %xmm4,%xmm7 \n" + "punpcklbw %xmm5,%xmm4 \n" + "punpckhbw %xmm5,%xmm7 \n" + "movdqa %xmm7,%xmm5 \n" + "movdqu (%eax),%xmm6 \n" + "movdqu (%eax,%edi,1),%xmm7 \n" + "lea (%eax,%edi,2),%eax \n" + "movdqu %xmm5,(%esp) \n" + "neg %edi \n" + "movdqa %xmm6,%xmm5 \n" + "punpcklbw %xmm7,%xmm6 \n" + "punpckhbw %xmm7,%xmm5 \n" + "movdqa %xmm5,%xmm7 \n" + "lea 0x10(%eax,%edi,8),%eax \n" + "neg %edi \n" + "movdqa %xmm0,%xmm5 \n" + "punpcklwd %xmm2,%xmm0 \n" + "punpckhwd %xmm2,%xmm5 \n" + "movdqa %xmm5,%xmm2 \n" + "movdqa %xmm1,%xmm5 \n" + "punpcklwd %xmm3,%xmm1 \n" + "punpckhwd %xmm3,%xmm5 \n" + "movdqa %xmm5,%xmm3 \n" + "movdqa %xmm4,%xmm5 \n" + "punpcklwd %xmm6,%xmm4 \n" + "punpckhwd %xmm6,%xmm5 \n" + "movdqa %xmm5,%xmm6 \n" + "movdqu (%esp),%xmm5 \n" + "movdqu %xmm6,(%esp) \n" + "movdqa %xmm5,%xmm6 \n" + "punpcklwd %xmm7,%xmm5 \n" + "punpckhwd %xmm7,%xmm6 \n" + "movdqa %xmm6,%xmm7 \n" + "movdqa %xmm0,%xmm6 \n" + "punpckldq %xmm4,%xmm0 \n" + "punpckhdq %xmm4,%xmm6 \n" + "movdqa %xmm6,%xmm4 \n" + "movdqu (%esp),%xmm6 \n" + "movlpd %xmm0,(%edx) \n" + "movhpd %xmm0,(%ebx) \n" + "movlpd %xmm4,(%edx,%esi,1) \n" + "lea (%edx,%esi,2),%edx \n" + "movhpd %xmm4,(%ebx,%ebp,1) \n" + "lea (%ebx,%ebp,2),%ebx \n" + "movdqa %xmm2,%xmm0 \n" + "punpckldq %xmm6,%xmm2 \n" + "movlpd %xmm2,(%edx) \n" + "movhpd %xmm2,(%ebx) \n" + "punpckhdq %xmm6,%xmm0 \n" + "movlpd %xmm0,(%edx,%esi,1) \n" + "lea (%edx,%esi,2),%edx \n" + "movhpd %xmm0,(%ebx,%ebp,1) \n" + "lea (%ebx,%ebp,2),%ebx \n" + "movdqa %xmm1,%xmm0 \n" + "punpckldq %xmm5,%xmm1 \n" + "movlpd %xmm1,(%edx) \n" + "movhpd %xmm1,(%ebx) \n" + "punpckhdq %xmm5,%xmm0 \n" + "movlpd %xmm0,(%edx,%esi,1) \n" + "lea (%edx,%esi,2),%edx \n" + "movhpd %xmm0,(%ebx,%ebp,1) \n" + "lea (%ebx,%ebp,2),%ebx \n" + "movdqa %xmm3,%xmm0 \n" + "punpckldq %xmm7,%xmm3 \n" + "movlpd %xmm3,(%edx) \n" + "movhpd %xmm3,(%ebx) \n" + "punpckhdq %xmm7,%xmm0 \n" + "sub $0x8,%ecx \n" + "movlpd %xmm0,(%edx,%esi,1) \n" + "lea (%edx,%esi,2),%edx \n" + "movhpd %xmm0,(%ebx,%ebp,1) \n" + "lea (%ebx,%ebp,2),%ebx \n" + "jg 1b \n" + "mov 0x10(%esp),%esp \n" + "pop %ebp \n" + "pop %edi \n" + "pop %esi \n" + "pop %ebx \n" +#if defined(__native_client__) + "pop %ecx \n" + "and $0xffffffe0,%ecx \n" + "jmp *%ecx \n" +#else + "ret \n" +#endif +); +#endif +#if !defined(LIBYUV_DISABLE_X86) && !defined(__native_client__) && \ + defined(__x86_64__) +// 64 bit version has enough registers to do 16x8 to 8x16 at a time. +void TransposeWx8_FAST_SSSE3(const uint8* src, int src_stride, + uint8* dst, int dst_stride, int width) { + asm volatile ( + // Read in the data from the source pointer. + // First round of bit swap. + ".p2align 2 \n" +"1: \n" + "movdqu (%0),%%xmm0 \n" + "movdqu (%0,%3),%%xmm1 \n" + "lea (%0,%3,2),%0 \n" + "movdqa %%xmm0,%%xmm8 \n" + "punpcklbw %%xmm1,%%xmm0 \n" + "punpckhbw %%xmm1,%%xmm8 \n" + "movdqu (%0),%%xmm2 \n" + "movdqa %%xmm0,%%xmm1 \n" + "movdqa %%xmm8,%%xmm9 \n" + "palignr $0x8,%%xmm1,%%xmm1 \n" + "palignr $0x8,%%xmm9,%%xmm9 \n" + "movdqu (%0,%3),%%xmm3 \n" + "lea (%0,%3,2),%0 \n" + "movdqa %%xmm2,%%xmm10 \n" + "punpcklbw %%xmm3,%%xmm2 \n" + "punpckhbw %%xmm3,%%xmm10 \n" + "movdqa %%xmm2,%%xmm3 \n" + "movdqa %%xmm10,%%xmm11 \n" + "movdqu (%0),%%xmm4 \n" + "palignr $0x8,%%xmm3,%%xmm3 \n" + "palignr $0x8,%%xmm11,%%xmm11 \n" + "movdqu (%0,%3),%%xmm5 \n" + "lea (%0,%3,2),%0 \n" + "movdqa %%xmm4,%%xmm12 \n" + "punpcklbw %%xmm5,%%xmm4 \n" + "punpckhbw %%xmm5,%%xmm12 \n" + "movdqa %%xmm4,%%xmm5 \n" + "movdqa %%xmm12,%%xmm13 \n" + "movdqu (%0),%%xmm6 \n" + "palignr $0x8,%%xmm5,%%xmm5 \n" + "palignr $0x8,%%xmm13,%%xmm13 \n" + "movdqu (%0,%3),%%xmm7 \n" + "lea (%0,%3,2),%0 \n" + "movdqa %%xmm6,%%xmm14 \n" + "punpcklbw %%xmm7,%%xmm6 \n" + "punpckhbw %%xmm7,%%xmm14 \n" + "neg %3 \n" + "movdqa %%xmm6,%%xmm7 \n" + "movdqa %%xmm14,%%xmm15 \n" + "lea 0x10(%0,%3,8),%0 \n" + "palignr $0x8,%%xmm7,%%xmm7 \n" + "palignr $0x8,%%xmm15,%%xmm15 \n" + "neg %3 \n" + // Second round of bit swap. + "punpcklwd %%xmm2,%%xmm0 \n" + "punpcklwd %%xmm3,%%xmm1 \n" + "movdqa %%xmm0,%%xmm2 \n" + "movdqa %%xmm1,%%xmm3 \n" + "palignr $0x8,%%xmm2,%%xmm2 \n" + "palignr $0x8,%%xmm3,%%xmm3 \n" + "punpcklwd %%xmm6,%%xmm4 \n" + "punpcklwd %%xmm7,%%xmm5 \n" + "movdqa %%xmm4,%%xmm6 \n" + "movdqa %%xmm5,%%xmm7 \n" + "palignr $0x8,%%xmm6,%%xmm6 \n" + "palignr $0x8,%%xmm7,%%xmm7 \n" + "punpcklwd %%xmm10,%%xmm8 \n" + "punpcklwd %%xmm11,%%xmm9 \n" + "movdqa %%xmm8,%%xmm10 \n" + "movdqa %%xmm9,%%xmm11 \n" + "palignr $0x8,%%xmm10,%%xmm10 \n" + "palignr $0x8,%%xmm11,%%xmm11 \n" + "punpcklwd %%xmm14,%%xmm12 \n" + "punpcklwd %%xmm15,%%xmm13 \n" + "movdqa %%xmm12,%%xmm14 \n" + "movdqa %%xmm13,%%xmm15 \n" + "palignr $0x8,%%xmm14,%%xmm14 \n" + "palignr $0x8,%%xmm15,%%xmm15 \n" + // Third round of bit swap. + // Write to the destination pointer. + "punpckldq %%xmm4,%%xmm0 \n" + "movq %%xmm0,(%1) \n" + "movdqa %%xmm0,%%xmm4 \n" + "palignr $0x8,%%xmm4,%%xmm4 \n" + "movq %%xmm4,(%1,%4) \n" + "lea (%1,%4,2),%1 \n" + "punpckldq %%xmm6,%%xmm2 \n" + "movdqa %%xmm2,%%xmm6 \n" + "movq %%xmm2,(%1) \n" + "palignr $0x8,%%xmm6,%%xmm6 \n" + "punpckldq %%xmm5,%%xmm1 \n" + "movq %%xmm6,(%1,%4) \n" + "lea (%1,%4,2),%1 \n" + "movdqa %%xmm1,%%xmm5 \n" + "movq %%xmm1,(%1) \n" + "palignr $0x8,%%xmm5,%%xmm5 \n" + "movq %%xmm5,(%1,%4) \n" + "lea (%1,%4,2),%1 \n" + "punpckldq %%xmm7,%%xmm3 \n" + "movq %%xmm3,(%1) \n" + "movdqa %%xmm3,%%xmm7 \n" + "palignr $0x8,%%xmm7,%%xmm7 \n" + "movq %%xmm7,(%1,%4) \n" + "lea (%1,%4,2),%1 \n" + "punpckldq %%xmm12,%%xmm8 \n" + "movq %%xmm8,(%1) \n" + "movdqa %%xmm8,%%xmm12 \n" + "palignr $0x8,%%xmm12,%%xmm12 \n" + "movq %%xmm12,(%1,%4) \n" + "lea (%1,%4,2),%1 \n" + "punpckldq %%xmm14,%%xmm10 \n" + "movdqa %%xmm10,%%xmm14 \n" + "movq %%xmm10,(%1) \n" + "palignr $0x8,%%xmm14,%%xmm14 \n" + "punpckldq %%xmm13,%%xmm9 \n" + "movq %%xmm14,(%1,%4) \n" + "lea (%1,%4,2),%1 \n" + "movdqa %%xmm9,%%xmm13 \n" + "movq %%xmm9,(%1) \n" + "palignr $0x8,%%xmm13,%%xmm13 \n" + "movq %%xmm13,(%1,%4) \n" + "lea (%1,%4,2),%1 \n" + "punpckldq %%xmm15,%%xmm11 \n" + "movq %%xmm11,(%1) \n" + "movdqa %%xmm11,%%xmm15 \n" + "palignr $0x8,%%xmm15,%%xmm15 \n" + "sub $0x10,%2 \n" + "movq %%xmm15,(%1,%4) \n" + "lea (%1,%4,2),%1 \n" + "jg 1b \n" + : "+r"(src), // %0 + "+r"(dst), // %1 + "+r"(width) // %2 + : "r"((intptr_t)(src_stride)), // %3 + "r"((intptr_t)(dst_stride)) // %4 + : "memory", "cc", + "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7", + "xmm8", "xmm9", "xmm10", "xmm11", "xmm12", "xmm13", "xmm14", "xmm15" +); +} + +void TransposeUVWx8_SSE2(const uint8* src, int src_stride, + uint8* dst_a, int dst_stride_a, + uint8* dst_b, int dst_stride_b, + int w) { + asm volatile ( + // Read in the data from the source pointer. + // First round of bit swap. + ".p2align 2 \n" +"1: \n" + "movdqu (%0),%%xmm0 \n" + "movdqu (%0,%4),%%xmm1 \n" + "lea (%0,%4,2),%0 \n" + "movdqa %%xmm0,%%xmm8 \n" + "punpcklbw %%xmm1,%%xmm0 \n" + "punpckhbw %%xmm1,%%xmm8 \n" + "movdqa %%xmm8,%%xmm1 \n" + "movdqu (%0),%%xmm2 \n" + "movdqu (%0,%4),%%xmm3 \n" + "lea (%0,%4,2),%0 \n" + "movdqa %%xmm2,%%xmm8 \n" + "punpcklbw %%xmm3,%%xmm2 \n" + "punpckhbw %%xmm3,%%xmm8 \n" + "movdqa %%xmm8,%%xmm3 \n" + "movdqu (%0),%%xmm4 \n" + "movdqu (%0,%4),%%xmm5 \n" + "lea (%0,%4,2),%0 \n" + "movdqa %%xmm4,%%xmm8 \n" + "punpcklbw %%xmm5,%%xmm4 \n" + "punpckhbw %%xmm5,%%xmm8 \n" + "movdqa %%xmm8,%%xmm5 \n" + "movdqu (%0),%%xmm6 \n" + "movdqu (%0,%4),%%xmm7 \n" + "lea (%0,%4,2),%0 \n" + "movdqa %%xmm6,%%xmm8 \n" + "punpcklbw %%xmm7,%%xmm6 \n" + "neg %4 \n" + "lea 0x10(%0,%4,8),%0 \n" + "punpckhbw %%xmm7,%%xmm8 \n" + "movdqa %%xmm8,%%xmm7 \n" + "neg %4 \n" + // Second round of bit swap. + "movdqa %%xmm0,%%xmm8 \n" + "movdqa %%xmm1,%%xmm9 \n" + "punpckhwd %%xmm2,%%xmm8 \n" + "punpckhwd %%xmm3,%%xmm9 \n" + "punpcklwd %%xmm2,%%xmm0 \n" + "punpcklwd %%xmm3,%%xmm1 \n" + "movdqa %%xmm8,%%xmm2 \n" + "movdqa %%xmm9,%%xmm3 \n" + "movdqa %%xmm4,%%xmm8 \n" + "movdqa %%xmm5,%%xmm9 \n" + "punpckhwd %%xmm6,%%xmm8 \n" + "punpckhwd %%xmm7,%%xmm9 \n" + "punpcklwd %%xmm6,%%xmm4 \n" + "punpcklwd %%xmm7,%%xmm5 \n" + "movdqa %%xmm8,%%xmm6 \n" + "movdqa %%xmm9,%%xmm7 \n" + // Third round of bit swap. + // Write to the destination pointer. + "movdqa %%xmm0,%%xmm8 \n" + "punpckldq %%xmm4,%%xmm0 \n" + "movlpd %%xmm0,(%1) \n" // Write back U channel + "movhpd %%xmm0,(%2) \n" // Write back V channel + "punpckhdq %%xmm4,%%xmm8 \n" + "movlpd %%xmm8,(%1,%5) \n" + "lea (%1,%5,2),%1 \n" + "movhpd %%xmm8,(%2,%6) \n" + "lea (%2,%6,2),%2 \n" + "movdqa %%xmm2,%%xmm8 \n" + "punpckldq %%xmm6,%%xmm2 \n" + "movlpd %%xmm2,(%1) \n" + "movhpd %%xmm2,(%2) \n" + "punpckhdq %%xmm6,%%xmm8 \n" + "movlpd %%xmm8,(%1,%5) \n" + "lea (%1,%5,2),%1 \n" + "movhpd %%xmm8,(%2,%6) \n" + "lea (%2,%6,2),%2 \n" + "movdqa %%xmm1,%%xmm8 \n" + "punpckldq %%xmm5,%%xmm1 \n" + "movlpd %%xmm1,(%1) \n" + "movhpd %%xmm1,(%2) \n" + "punpckhdq %%xmm5,%%xmm8 \n" + "movlpd %%xmm8,(%1,%5) \n" + "lea (%1,%5,2),%1 \n" + "movhpd %%xmm8,(%2,%6) \n" + "lea (%2,%6,2),%2 \n" + "movdqa %%xmm3,%%xmm8 \n" + "punpckldq %%xmm7,%%xmm3 \n" + "movlpd %%xmm3,(%1) \n" + "movhpd %%xmm3,(%2) \n" + "punpckhdq %%xmm7,%%xmm8 \n" + "sub $0x8,%3 \n" + "movlpd %%xmm8,(%1,%5) \n" + "lea (%1,%5,2),%1 \n" + "movhpd %%xmm8,(%2,%6) \n" + "lea (%2,%6,2),%2 \n" + "jg 1b \n" + : "+r"(src), // %0 + "+r"(dst_a), // %1 + "+r"(dst_b), // %2 + "+r"(w) // %3 + : "r"((intptr_t)(src_stride)), // %4 + "r"((intptr_t)(dst_stride_a)), // %5 + "r"((intptr_t)(dst_stride_b)) // %6 + : "memory", "cc", + "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7", + "xmm8", "xmm9" +); +} +#endif +#endif + +#endif // defined(__x86_64__) || defined(__i386__) + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/source/rotate_win.cc b/source/rotate_win.cc new file mode 100644 index 000000000..2760066df --- /dev/null +++ b/source/rotate_win.cc @@ -0,0 +1,248 @@ +/* + * Copyright 2013 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/row.h" +#include "libyuv/rotate_row.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +// This module is for Visual C x86. +#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && \ + defined(_MSC_VER) && !defined(__clang__) + +__declspec(naked) +void TransposeWx8_SSSE3(const uint8* src, int src_stride, + uint8* dst, int dst_stride, int width) { + __asm { + push edi + push esi + push ebp + mov eax, [esp + 12 + 4] // src + mov edi, [esp + 12 + 8] // src_stride + mov edx, [esp + 12 + 12] // dst + mov esi, [esp + 12 + 16] // dst_stride + mov ecx, [esp + 12 + 20] // width + + // Read in the data from the source pointer. + // First round of bit swap. + align 4 + convertloop: + movq xmm0, qword ptr [eax] + lea ebp, [eax + 8] + movq xmm1, qword ptr [eax + edi] + lea eax, [eax + 2 * edi] + punpcklbw xmm0, xmm1 + movq xmm2, qword ptr [eax] + movdqa xmm1, xmm0 + palignr xmm1, xmm1, 8 + movq xmm3, qword ptr [eax + edi] + lea eax, [eax + 2 * edi] + punpcklbw xmm2, xmm3 + movdqa xmm3, xmm2 + movq xmm4, qword ptr [eax] + palignr xmm3, xmm3, 8 + movq xmm5, qword ptr [eax + edi] + punpcklbw xmm4, xmm5 + lea eax, [eax + 2 * edi] + movdqa xmm5, xmm4 + movq xmm6, qword ptr [eax] + palignr xmm5, xmm5, 8 + movq xmm7, qword ptr [eax + edi] + punpcklbw xmm6, xmm7 + mov eax, ebp + movdqa xmm7, xmm6 + palignr xmm7, xmm7, 8 + // Second round of bit swap. + punpcklwd xmm0, xmm2 + punpcklwd xmm1, xmm3 + movdqa xmm2, xmm0 + movdqa xmm3, xmm1 + palignr xmm2, xmm2, 8 + palignr xmm3, xmm3, 8 + punpcklwd xmm4, xmm6 + punpcklwd xmm5, xmm7 + movdqa xmm6, xmm4 + movdqa xmm7, xmm5 + palignr xmm6, xmm6, 8 + palignr xmm7, xmm7, 8 + // Third round of bit swap. + // Write to the destination pointer. + punpckldq xmm0, xmm4 + movq qword ptr [edx], xmm0 + movdqa xmm4, xmm0 + palignr xmm4, xmm4, 8 + movq qword ptr [edx + esi], xmm4 + lea edx, [edx + 2 * esi] + punpckldq xmm2, xmm6 + movdqa xmm6, xmm2 + palignr xmm6, xmm6, 8 + movq qword ptr [edx], xmm2 + punpckldq xmm1, xmm5 + movq qword ptr [edx + esi], xmm6 + lea edx, [edx + 2 * esi] + movdqa xmm5, xmm1 + movq qword ptr [edx], xmm1 + palignr xmm5, xmm5, 8 + punpckldq xmm3, xmm7 + movq qword ptr [edx + esi], xmm5 + lea edx, [edx + 2 * esi] + movq qword ptr [edx], xmm3 + movdqa xmm7, xmm3 + palignr xmm7, xmm7, 8 + sub ecx, 8 + movq qword ptr [edx + esi], xmm7 + lea edx, [edx + 2 * esi] + jg convertloop + + pop ebp + pop esi + pop edi + ret + } +} + +__declspec(naked) +void TransposeUVWx8_SSE2(const uint8* src, int src_stride, + uint8* dst_a, int dst_stride_a, + uint8* dst_b, int dst_stride_b, + int w) { + __asm { + push ebx + push esi + push edi + push ebp + mov eax, [esp + 16 + 4] // src + mov edi, [esp + 16 + 8] // src_stride + mov edx, [esp + 16 + 12] // dst_a + mov esi, [esp + 16 + 16] // dst_stride_a + mov ebx, [esp + 16 + 20] // dst_b + mov ebp, [esp + 16 + 24] // dst_stride_b + mov ecx, esp + sub esp, 4 + 16 + and esp, ~15 + mov [esp + 16], ecx + mov ecx, [ecx + 16 + 28] // w + + align 4 + convertloop: + // Read in the data from the source pointer. + // First round of bit swap. + movdqu xmm0, [eax] + movdqu xmm1, [eax + edi] + lea eax, [eax + 2 * edi] + movdqa xmm7, xmm0 // use xmm7 as temp register. + punpcklbw xmm0, xmm1 + punpckhbw xmm7, xmm1 + movdqa xmm1, xmm7 + movdqu xmm2, [eax] + movdqu xmm3, [eax + edi] + lea eax, [eax + 2 * edi] + movdqa xmm7, xmm2 + punpcklbw xmm2, xmm3 + punpckhbw xmm7, xmm3 + movdqa xmm3, xmm7 + movdqu xmm4, [eax] + movdqu xmm5, [eax + edi] + lea eax, [eax + 2 * edi] + movdqa xmm7, xmm4 + punpcklbw xmm4, xmm5 + punpckhbw xmm7, xmm5 + movdqa xmm5, xmm7 + movdqu xmm6, [eax] + movdqu xmm7, [eax + edi] + lea eax, [eax + 2 * edi] + movdqu [esp], xmm5 // backup xmm5 + neg edi + movdqa xmm5, xmm6 // use xmm5 as temp register. + punpcklbw xmm6, xmm7 + punpckhbw xmm5, xmm7 + movdqa xmm7, xmm5 + lea eax, [eax + 8 * edi + 16] + neg edi + // Second round of bit swap. + movdqa xmm5, xmm0 + punpcklwd xmm0, xmm2 + punpckhwd xmm5, xmm2 + movdqa xmm2, xmm5 + movdqa xmm5, xmm1 + punpcklwd xmm1, xmm3 + punpckhwd xmm5, xmm3 + movdqa xmm3, xmm5 + movdqa xmm5, xmm4 + punpcklwd xmm4, xmm6 + punpckhwd xmm5, xmm6 + movdqa xmm6, xmm5 + movdqu xmm5, [esp] // restore xmm5 + movdqu [esp], xmm6 // backup xmm6 + movdqa xmm6, xmm5 // use xmm6 as temp register. + punpcklwd xmm5, xmm7 + punpckhwd xmm6, xmm7 + movdqa xmm7, xmm6 + // Third round of bit swap. + // Write to the destination pointer. + movdqa xmm6, xmm0 + punpckldq xmm0, xmm4 + punpckhdq xmm6, xmm4 + movdqa xmm4, xmm6 + movdqu xmm6, [esp] // restore xmm6 + movlpd qword ptr [edx], xmm0 + movhpd qword ptr [ebx], xmm0 + movlpd qword ptr [edx + esi], xmm4 + lea edx, [edx + 2 * esi] + movhpd qword ptr [ebx + ebp], xmm4 + lea ebx, [ebx + 2 * ebp] + movdqa xmm0, xmm2 // use xmm0 as the temp register. + punpckldq xmm2, xmm6 + movlpd qword ptr [edx], xmm2 + movhpd qword ptr [ebx], xmm2 + punpckhdq xmm0, xmm6 + movlpd qword ptr [edx + esi], xmm0 + lea edx, [edx + 2 * esi] + movhpd qword ptr [ebx + ebp], xmm0 + lea ebx, [ebx + 2 * ebp] + movdqa xmm0, xmm1 // use xmm0 as the temp register. + punpckldq xmm1, xmm5 + movlpd qword ptr [edx], xmm1 + movhpd qword ptr [ebx], xmm1 + punpckhdq xmm0, xmm5 + movlpd qword ptr [edx + esi], xmm0 + lea edx, [edx + 2 * esi] + movhpd qword ptr [ebx + ebp], xmm0 + lea ebx, [ebx + 2 * ebp] + movdqa xmm0, xmm3 // use xmm0 as the temp register. + punpckldq xmm3, xmm7 + movlpd qword ptr [edx], xmm3 + movhpd qword ptr [ebx], xmm3 + punpckhdq xmm0, xmm7 + sub ecx, 8 + movlpd qword ptr [edx + esi], xmm0 + lea edx, [edx + 2 * esi] + movhpd qword ptr [ebx + ebp], xmm0 + lea ebx, [ebx + 2 * ebp] + jg convertloop + + mov esp, [esp + 16] + pop ebp + pop edi + pop esi + pop ebx + ret + } +} + +#endif // !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/winarm.mk b/winarm.mk index 6f559c4b4..442fe75f9 100644 --- a/winarm.mk +++ b/winarm.mk @@ -22,6 +22,7 @@ LOCAL_OBJ_FILES = \ source/planar_functions.o\ source/rotate.o\ source/rotate_argb.o\ + source/rotate_common.o\ source/row_any.o\ source/row_common.o\ source/scale.o\