Browse Source

avcodec/utvideodsp : add avx2 version for the dsp

tags/n4.0
Martin Vignali 8 years ago
parent
commit
b5ebe38443
2 changed files with 31 additions and 0 deletions
  1. +20
    -0
      libavcodec/x86/utvideodsp.asm
  2. +11
    -0
      libavcodec/x86/utvideodsp_init.c

+ 20
- 0
libavcodec/x86/utvideodsp.asm View File

@@ -1,6 +1,7 @@
;******************************************************************************
;* SIMD-optimized UTVideo functions
;* Copyright (c) 2017 Paul B Mahol
;* Copyright (c) 2017 Jokyo Images
;*
;* This file is part of FFmpeg.
;*
@@ -45,7 +46,11 @@ DEFINE_ARGS src_r, src_g, src_b, linesize_r, linesize_g, linesize_b, x
%define wq r6m
%define hd r7mp
%endif
%if mmsize == 32
vbroadcasti128 m3, [pb_128]
%else
mova m3, [pb_128]
%endif
.nextrow:
mov xq, wq

@@ -72,6 +77,11 @@ DEFINE_ARGS src_r, src_g, src_b, linesize_r, linesize_g, linesize_b, x
INIT_XMM sse2
RESTORE_RGB_PLANES

%if HAVE_AVX2_EXTERNAL
INIT_YMM avx2
RESTORE_RGB_PLANES
%endif

%macro RESTORE_RGB_PLANES10 0
cglobal restore_rgb_planes10, 7 + ARCH_X86_64, 7 + ARCH_X86_64 * 2, 5, src_r, src_g, src_b, linesize_r, linesize_g, linesize_b, w, h, x
shl wd, 1
@@ -81,8 +91,13 @@ cglobal restore_rgb_planes10, 7 + ARCH_X86_64, 7 + ARCH_X86_64 * 2, 5, src_r, sr
add src_rq, wq
add src_gq, wq
add src_bq, wq
%if mmsize == 32
vbroadcasti128 m3, [pw_512]
vbroadcasti128 m4, [pw_1023]
%else
mova m3, [pw_512]
mova m4, [pw_1023]
%endif
neg wq
%if ARCH_X86_64 == 0
mov wm, wq
@@ -117,3 +132,8 @@ DEFINE_ARGS src_r, src_g, src_b, linesize_r, linesize_g, linesize_b, x

INIT_XMM sse2
RESTORE_RGB_PLANES10

%if HAVE_AVX2_EXTERNAL
INIT_YMM avx2
RESTORE_RGB_PLANES10
%endif

+ 11
- 0
libavcodec/x86/utvideodsp_init.c View File

@@ -28,9 +28,16 @@
void ff_restore_rgb_planes_sse2(uint8_t *src_r, uint8_t *src_g, uint8_t *src_b,
ptrdiff_t linesize_r, ptrdiff_t linesize_g,
ptrdiff_t linesize_b, int width, int height);
void ff_restore_rgb_planes_avx2(uint8_t *src_r, uint8_t *src_g, uint8_t *src_b,
ptrdiff_t linesize_r, ptrdiff_t linesize_g,
ptrdiff_t linesize_b, int width, int height);

void ff_restore_rgb_planes10_sse2(uint16_t *src_r, uint16_t *src_g, uint16_t *src_b,
ptrdiff_t linesize_r, ptrdiff_t linesize_g,
ptrdiff_t linesize_b, int width, int height);
void ff_restore_rgb_planes10_avx2(uint16_t *src_r, uint16_t *src_g, uint16_t *src_b,
ptrdiff_t linesize_r, ptrdiff_t linesize_g,
ptrdiff_t linesize_b, int width, int height);

av_cold void ff_utvideodsp_init_x86(UTVideoDSPContext *c)
{
@@ -40,4 +47,8 @@ av_cold void ff_utvideodsp_init_x86(UTVideoDSPContext *c)
c->restore_rgb_planes = ff_restore_rgb_planes_sse2;
c->restore_rgb_planes10 = ff_restore_rgb_planes10_sse2;
}
if (EXTERNAL_AVX2_FAST(cpu_flags)) {
c->restore_rgb_planes = ff_restore_rgb_planes_avx2;
c->restore_rgb_planes10 = ff_restore_rgb_planes10_avx2;
}
}

Loading…
Cancel
Save