har0ke/FFmpeg
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

swscale/la: Add yuv2rgb_lasx.c and rgb2rgb_lasx.c files

Browse Source 
ffmpeg -i 1_h264_1080p_30fps_3Mbps.mp4 -f rawvideo -pix_fmt rgb24 -y /dev/null -an
before: 178fps
after:  210fps

Signed-off-by: Hao Chen <chenhao@loongson.cn>
Reviewed-by: yinshiyou-hf@loongson.cn
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
This commit is contained in:
Hao Chen 2022-09-09 17:00:25 +08:00 committed by Michael Niedermayer
parent 38cacce22a
commit 74d09b068d
No known key found for this signature in database
GPG Key ID: B18E8928B3948D64
8 changed files with 444 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
libswscale/loongarch/Makefile
View File

@ -1,3 +1,5 @@
OBJS-$(CONFIG_SWSCALE) += loongarch/swscale_init_loongarch.o
LASX-OBJS-$(CONFIG_SWSCALE) += loongarch/swscale_lasx.o \
loongarch/input_lasx.o \
loongarch/yuv2rgb_lasx.o \
loongarch/rgb2rgb_lasx.o

52
libswscale/loongarch/rgb2rgb_lasx.c Normal file
View File

@ -0,0 +1,52 @@
/*
* Copyright (c) 2022 Loongson Technology Corporation Limited
* Contributed by Hao Chen(chenhao@loongson.cn)
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "swscale_loongarch.h"
#include "libavutil/loongarch/loongson_intrinsics.h"
void ff_interleave_bytes_lasx(const uint8_t *src1, const uint8_t *src2,
uint8_t *dest, int width, int height,
int src1Stride, int src2Stride, int dstStride)
{
int h;
int len = width & (0xFFFFFFF0);
for (h = 0; h < height; h++) {
int w, index = 0;
__m256i src_1, src_2, dst;
for (w = 0; w < len; w += 16) {
DUP2_ARG2(__lasx_xvld, src1 + w, 0, src2 + w, 0, src_1, src_2);
src_1 = __lasx_xvpermi_d(src_1, 0xD8);
src_2 = __lasx_xvpermi_d(src_2, 0xD8);
dst = __lasx_xvilvl_b(src_2, src_1);
__lasx_xvst(dst, dest + index, 0);
index += 32;
}
for (; w < width; w++) {
dest[(w << 1) + 0] = src1[w];
dest[(w << 1) + 1] = src2[w];
}
dest += dstStride;
src1 += src1Stride;
src2 += src2Stride;
}
}

42
libswscale/loongarch/swscale_init_loongarch.c
View File

@ -21,6 +21,7 @@
#include "swscale_loongarch.h"
#include "libswscale/swscale_internal.h"
#include "libswscale/rgb2rgb.h"
#include "libavutil/loongarch/cpu.h"
av_cold void ff_sws_init_swscale_loongarch(SwsContext *c)
@ -48,3 +49,44 @@ av_cold void ff_sws_init_swscale_loongarch(SwsContext *c)
}
}
}
av_cold void rgb2rgb_init_loongarch(void)
{
int cpu_flags = av_get_cpu_flags();
if (have_lasx(cpu_flags))
interleaveBytes = ff_interleave_bytes_lasx;
}
av_cold SwsFunc ff_yuv2rgb_init_loongarch(SwsContext *c)
{
int cpu_flags = av_get_cpu_flags();
if (have_lasx(cpu_flags)) {
switch (c->dstFormat) {
case AV_PIX_FMT_RGB24:
return yuv420_rgb24_lasx;
case AV_PIX_FMT_BGR24:
return yuv420_bgr24_lasx;
case AV_PIX_FMT_RGBA:
if (CONFIG_SWSCALE_ALPHA && isALPHA(c->srcFormat)) {
break;
} else
return yuv420_rgba32_lasx;
case AV_PIX_FMT_ARGB:
if (CONFIG_SWSCALE_ALPHA && isALPHA(c->srcFormat)) {
break;
} else
return yuv420_argb32_lasx;
case AV_PIX_FMT_BGRA:
if (CONFIG_SWSCALE_ALPHA && isALPHA(c->srcFormat)) {
break;
} else
return yuv420_bgra32_lasx;
case AV_PIX_FMT_ABGR:
if (CONFIG_SWSCALE_ALPHA && isALPHA(c->srcFormat)) {
break;
} else
return yuv420_abgr32_lasx;
}
}
return NULL;
}

22
libswscale/loongarch/swscale_loongarch.h
View File

@ -47,4 +47,26 @@ void planar_rgb_to_uv_lasx(uint8_t *_dstU, uint8_t *_dstV, const uint8_t *src[4]
void planar_rgb_to_y_lasx(uint8_t *_dst, const uint8_t *src[4], int width,
int32_t *rgb2yuv, void *opq);
int yuv420_rgb24_lasx(SwsContext *c, const uint8_t *src[], int srcStride[],
int srcSliceY, int srcSliceH, uint8_t *dst[], int dstStride[]);
int yuv420_bgr24_lasx(SwsContext *c, const uint8_t *src[], int srcStride[],
int srcSliceY, int srcSliceH, uint8_t *dst[], int dstStride[]);
int yuv420_rgba32_lasx(SwsContext *c, const uint8_t *src[], int srcStride[],
int srcSliceY, int srcSliceH, uint8_t *dst[], int dstStride[]);
int yuv420_bgra32_lasx(SwsContext *c, const uint8_t *src[], int srcStride[],
int srcSliceY, int srcSliceH, uint8_t *dst[], int dstStride[]);
int yuv420_argb32_lasx(SwsContext *c, const uint8_t *src[], int srcStride[],
int srcSliceY, int srcSliceH, uint8_t *dst[], int dstStride[]);
int yuv420_abgr32_lasx(SwsContext *c, const uint8_t *src[], int srcStride[],
int srcSliceY, int srcSliceH, uint8_t *dst[], int dstStride[]);
void ff_interleave_bytes_lasx(const uint8_t *src1, const uint8_t *src2,
uint8_t *dest, int width, int height,
int src1Stride, int src2Stride, int dstStride);
#endif /* SWSCALE_LOONGARCH_SWSCALE_LOONGARCH_H */

321
libswscale/loongarch/yuv2rgb_lasx.c Normal file
View File

@ -0,0 +1,321 @@
/*
* Copyright (C) 2022 Loongson Technology Corporation Limited
* Contributed by Hao Chen(chenhao@loongson.cn)
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "swscale_loongarch.h"
#include "libavutil/loongarch/loongson_intrinsics.h"
#define YUV2RGB_LOAD_COE \
/* Load x_offset */ \
__m256i y_offset = __lasx_xvreplgr2vr_d(c->yOffset); \
__m256i u_offset = __lasx_xvreplgr2vr_d(c->uOffset); \
__m256i v_offset = __lasx_xvreplgr2vr_d(c->vOffset); \
/* Load x_coeff */ \
__m256i ug_coeff = __lasx_xvreplgr2vr_d(c->ugCoeff); \
__m256i vg_coeff = __lasx_xvreplgr2vr_d(c->vgCoeff); \
__m256i y_coeff = __lasx_xvreplgr2vr_d(c->yCoeff); \
__m256i ub_coeff = __lasx_xvreplgr2vr_d(c->ubCoeff); \
__m256i vr_coeff = __lasx_xvreplgr2vr_d(c->vrCoeff); \
#define LOAD_YUV_16 \
m_y1 = __lasx_xvld(py_1, 0); \
m_y2 = __lasx_xvld(py_2, 0); \
m_u = __lasx_xvldrepl_d(pu, 0); \
m_v = __lasx_xvldrepl_d(pv, 0); \
m_u = __lasx_xvilvl_b(m_u, m_u); \
m_v = __lasx_xvilvl_b(m_v, m_v); \
DUP4_ARG1(__lasx_vext2xv_hu_bu, m_y1, m_y2, m_u, m_v, \
m_y1, m_y2, m_u, m_v); \
/* YUV2RGB method
* The conversion method is as follows:
* R = Y' * y_coeff + V' * vr_coeff
* G = Y' * y_coeff + V' * vg_coeff + U' * ug_coeff
* B = Y' * y_coeff + U' * ub_coeff
*
* where X' = X * 8 - x_offset
*
*/
#define YUV2RGB \
m_y1 = __lasx_xvslli_h(m_y1, 3); \
m_y2 = __lasx_xvslli_h(m_y2, 3); \
m_u = __lasx_xvslli_h(m_u, 3); \
m_v = __lasx_xvslli_h(m_v, 3); \
m_y1 = __lasx_xvsub_h(m_y1, y_offset); \
m_y2 = __lasx_xvsub_h(m_y2, y_offset); \
m_u = __lasx_xvsub_h(m_u, u_offset); \
m_v = __lasx_xvsub_h(m_v, v_offset); \
y_1 = __lasx_xvmuh_h(m_y1, y_coeff); \
y_2 = __lasx_xvmuh_h(m_y2, y_coeff); \
u2g = __lasx_xvmuh_h(m_u, ug_coeff); \
u2b = __lasx_xvmuh_h(m_u, ub_coeff); \
v2r = __lasx_xvmuh_h(m_v, vr_coeff); \
v2g = __lasx_xvmuh_h(m_v, vg_coeff); \
r1 = __lasx_xvsadd_h(y_1, v2r); \
v2g = __lasx_xvsadd_h(v2g, u2g); \
g1 = __lasx_xvsadd_h(y_1, v2g); \
b1 = __lasx_xvsadd_h(y_1, u2b); \
r2 = __lasx_xvsadd_h(y_2, v2r); \
g2 = __lasx_xvsadd_h(y_2, v2g); \
b2 = __lasx_xvsadd_h(y_2, u2b); \
DUP4_ARG1(__lasx_xvclip255_h, r1, g1, b1, r2, r1, g1, b1, r2); \
DUP2_ARG1(__lasx_xvclip255_h, g2, b2, g2, b2); \
#define YUV2RGB_RES \
m_y1 = __lasx_xvldrepl_d(py_1, 0); \
m_y2 = __lasx_xvldrepl_d(py_2, 0); \
m_u = __lasx_xvldrepl_d(pu, 0); \
m_v = __lasx_xvldrepl_d(pv, 0); \
m_y1 = __lasx_xvilvl_d(m_y2, m_y1); \
m_u = __lasx_xvilvl_b(m_u, m_u); \
m_v = __lasx_xvilvl_b(m_v, m_v); \
m_y1 = __lasx_vext2xv_hu_bu(m_y1); \
m_u = __lasx_vext2xv_hu_bu(m_u); \
m_v = __lasx_vext2xv_hu_bu(m_v); \
m_y1 = __lasx_xvslli_h(m_y1, 3); \
m_u = __lasx_xvslli_h(m_u, 3); \
m_v = __lasx_xvslli_h(m_v, 3); \
m_y1 = __lasx_xvsub_h(m_y1, y_offset); \
m_u = __lasx_xvsub_h(m_u, u_offset); \
m_v = __lasx_xvsub_h(m_v, v_offset); \
y_1 = __lasx_xvmuh_h(m_y1, y_coeff); \
u2g = __lasx_xvmuh_h(m_u, ug_coeff); \
u2b = __lasx_xvmuh_h(m_u, ub_coeff); \
v2r = __lasx_xvmuh_h(m_v, vr_coeff); \
v2g = __lasx_xvmuh_h(m_v, vg_coeff); \
r1 = __lasx_xvsadd_h(y_1, v2r); \
v2g = __lasx_xvsadd_h(v2g, u2g); \
g1 = __lasx_xvsadd_h(y_1, v2g); \
b1 = __lasx_xvsadd_h(y_1, u2b); \
r1 = __lasx_xvclip255_h(r1); \
g1 = __lasx_xvclip255_h(g1); \
b1 = __lasx_xvclip255_h(b1); \
#define RGB_PACK(r, g, b, rgb_l, rgb_h) \
{ \
__m256i rg; \
rg = __lasx_xvpackev_b(g, r); \
DUP2_ARG3(__lasx_xvshuf_b, b, rg, shuf2, b, rg, shuf3, rgb_l, rgb_h); \
}
#define RGB32_PACK(a, r, g, b, rgb_l, rgb_h) \
{ \
__m256i ra, bg, tmp0, tmp1; \
ra = __lasx_xvpackev_b(r, a); \
bg = __lasx_xvpackev_b(b, g); \
tmp0 = __lasx_xvilvl_h(bg, ra); \
tmp1 = __lasx_xvilvh_h(bg, ra); \
rgb_l = __lasx_xvpermi_q(tmp1, tmp0, 0x20); \
rgb_h = __lasx_xvpermi_q(tmp1, tmp0, 0x31); \
}
#define RGB_STORE_RES(rgb_l, rgb_h, image_1, image_2) \
{ \
__lasx_xvstelm_d(rgb_l, image_1, 0, 0); \
__lasx_xvstelm_d(rgb_l, image_1, 8, 1); \
__lasx_xvstelm_d(rgb_h, image_1, 16, 0); \
__lasx_xvstelm_d(rgb_l, image_2, 0, 2); \
__lasx_xvstelm_d(rgb_l, image_2, 8, 3); \
__lasx_xvstelm_d(rgb_h, image_2, 16, 2); \
}
#define RGB_STORE(rgb_l, rgb_h, image) \
{ \
__lasx_xvstelm_d(rgb_l, image, 0, 0); \
__lasx_xvstelm_d(rgb_l, image, 8, 1); \
__lasx_xvstelm_d(rgb_h, image, 16, 0); \
__lasx_xvstelm_d(rgb_l, image, 24, 2); \
__lasx_xvstelm_d(rgb_l, image, 32, 3); \
__lasx_xvstelm_d(rgb_h, image, 40, 2); \
}
#define RGB32_STORE(rgb_l, rgb_h, image) \
{ \
__lasx_xvst(rgb_l, image, 0); \
__lasx_xvst(rgb_h, image, 32); \
}
#define RGB32_STORE_RES(rgb_l, rgb_h, image_1, image_2) \
{ \
__lasx_xvst(rgb_l, image_1, 0); \
__lasx_xvst(rgb_h, image_2, 0); \
}
#define YUV2RGBFUNC(func_name, dst_type, alpha) \
int func_name(SwsContext *c, const uint8_t *src[], \
int srcStride[], int srcSliceY, int srcSliceH, \
uint8_t *dst[], int dstStride[]) \
{ \
int x, y, h_size, vshift, res; \
__m256i m_y1, m_y2, m_u, m_v; \
__m256i y_1, y_2, u2g, v2g, u2b, v2r, rgb1_l, rgb1_h; \
__m256i rgb2_l, rgb2_h, r1, g1, b1, r2, g2, b2; \
__m256i shuf2 = {0x0504120302100100, 0x0A18090816070614, \
0x0504120302100100, 0x0A18090816070614}; \
__m256i shuf3 = {0x1E0F0E1C0D0C1A0B, 0x0101010101010101, \
0x1E0F0E1C0D0C1A0B, 0x0101010101010101}; \
YUV2RGB_LOAD_COE \
y = (c->dstW + 7) & ~7; \
h_size = y >> 4; \
res = y & 15; \
\
vshift = c->srcFormat != AV_PIX_FMT_YUV422P; \
for (y = 0; y < srcSliceH; y += 2) { \
dst_type *image1 = (dst_type *)(dst[0] + (y + srcSliceY) * dstStride[0]);\
dst_type *image2 = (dst_type *)(image1 + dstStride[0]);\
const uint8_t *py_1 = src[0] + y * srcStride[0]; \
const uint8_t *py_2 = py_1 + srcStride[0]; \
const uint8_t *pu = src[1] + (y >> vshift) * srcStride[1]; \
const uint8_t *pv = src[2] + (y >> vshift) * srcStride[2]; \
for(x = 0; x < h_size; x++) { \
#define YUV2RGBFUNC32(func_name, dst_type, alpha) \
int func_name(SwsContext *c, const uint8_t *src[], \
int srcStride[], int srcSliceY, int srcSliceH, \
uint8_t *dst[], int dstStride[]) \
{ \
int x, y, h_size, vshift, res; \
__m256i m_y1, m_y2, m_u, m_v; \
__m256i y_1, y_2, u2g, v2g, u2b, v2r, rgb1_l, rgb1_h; \
__m256i rgb2_l, rgb2_h, r1, g1, b1, r2, g2, b2; \
__m256i a = __lasx_xvldi(0xFF); \
\
YUV2RGB_LOAD_COE \
y = (c->dstW + 7) & ~7; \
h_size = y >> 4; \
res = y & 15; \
\
vshift = c->srcFormat != AV_PIX_FMT_YUV422P; \
for (y = 0; y < srcSliceH; y += 2) { \
int yd = y + srcSliceY; \
dst_type av_unused *r, *g, *b; \
dst_type *image1 = (dst_type *)(dst[0] + (yd) * dstStride[0]); \
dst_type *image2 = (dst_type *)(dst[0] + (yd + 1) * dstStride[0]); \
const uint8_t *py_1 = src[0] + y * srcStride[0]; \
const uint8_t *py_2 = py_1 + srcStride[0]; \
const uint8_t *pu = src[1] + (y >> vshift) * srcStride[1]; \
const uint8_t *pv = src[2] + (y >> vshift) * srcStride[2]; \
for(x = 0; x < h_size; x++) { \
#define DEALYUV2RGBREMAIN \
py_1 += 16; \
py_2 += 16; \
pu += 8; \
pv += 8; \
image1 += 48; \
image2 += 48; \
} \
if (res) { \
#define DEALYUV2RGBREMAIN32 \
py_1 += 16; \
py_2 += 16; \
pu += 8; \
pv += 8; \
image1 += 16; \
image2 += 16; \
} \
if (res) { \
#define END_FUNC() \
} \
} \
return srcSliceH; \
}
YUV2RGBFUNC(yuv420_rgb24_lasx, uint8_t, 0)
LOAD_YUV_16
YUV2RGB
RGB_PACK(r1, g1, b1, rgb1_l, rgb1_h);
RGB_PACK(r2, g2, b2, rgb2_l, rgb2_h);
RGB_STORE(rgb1_l, rgb1_h, image1);
RGB_STORE(rgb2_l, rgb2_h, image2);
DEALYUV2RGBREMAIN
YUV2RGB_RES
RGB_PACK(r1, g1, b1, rgb1_l, rgb1_h);
RGB_STORE_RES(rgb1_l, rgb1_h, image1, image2);
END_FUNC()
YUV2RGBFUNC(yuv420_bgr24_lasx, uint8_t, 0)
LOAD_YUV_16
YUV2RGB
RGB_PACK(b1, g1, r1, rgb1_l, rgb1_h);
RGB_PACK(b2, g2, r2, rgb2_l, rgb2_h);
RGB_STORE(rgb1_l, rgb1_h, image1);
RGB_STORE(rgb2_l, rgb2_h, image2);
DEALYUV2RGBREMAIN
YUV2RGB_RES
RGB_PACK(b1, g1, r1, rgb1_l, rgb1_h);
RGB_STORE_RES(rgb1_l, rgb1_h, image1, image2);
END_FUNC()
YUV2RGBFUNC32(yuv420_rgba32_lasx, uint32_t, 0)
LOAD_YUV_16
YUV2RGB
RGB32_PACK(r1, g1, b1, a, rgb1_l, rgb1_h);
RGB32_PACK(r2, g2, b2, a, rgb2_l, rgb2_h);
RGB32_STORE(rgb1_l, rgb1_h, image1);
RGB32_STORE(rgb2_l, rgb2_h, image2);
DEALYUV2RGBREMAIN32
YUV2RGB_RES
RGB32_PACK(r1, g1, b1, a, rgb1_l, rgb1_h);
RGB32_STORE_RES(rgb1_l, rgb1_h, image1, image2);
END_FUNC()
YUV2RGBFUNC32(yuv420_bgra32_lasx, uint32_t, 0)
LOAD_YUV_16
YUV2RGB
RGB32_PACK(b1, g1, r1, a, rgb1_l, rgb1_h);
RGB32_PACK(b2, g2, r2, a, rgb2_l, rgb2_h);
RGB32_STORE(rgb1_l, rgb1_h, image1);
RGB32_STORE(rgb2_l, rgb2_h, image2);
DEALYUV2RGBREMAIN32
YUV2RGB_RES
RGB32_PACK(b1, g1, r1, a, rgb1_l, rgb1_h);
RGB32_STORE_RES(rgb1_l, rgb1_h, image1, image2);
END_FUNC()
YUV2RGBFUNC32(yuv420_argb32_lasx, uint32_t, 0)
LOAD_YUV_16
YUV2RGB
RGB32_PACK(a, r1, g1, b1, rgb1_l, rgb1_h);
RGB32_PACK(a, r2, g2, b2, rgb2_l, rgb2_h);
RGB32_STORE(rgb1_l, rgb1_h, image1);
RGB32_STORE(rgb2_l, rgb2_h, image2);
DEALYUV2RGBREMAIN32
YUV2RGB_RES
RGB32_PACK(a, r1, g1, b1, rgb1_l, rgb1_h);
RGB32_STORE_RES(rgb1_l, rgb1_h, image1, image2);
END_FUNC()
YUV2RGBFUNC32(yuv420_abgr32_lasx, uint32_t, 0)
LOAD_YUV_16
YUV2RGB
RGB32_PACK(a, b1, g1, r1, rgb1_l, rgb1_h);
RGB32_PACK(a, b2, g2, r2, rgb2_l, rgb2_h);
RGB32_STORE(rgb1_l, rgb1_h, image1);
RGB32_STORE(rgb2_l, rgb2_h, image2);
DEALYUV2RGBREMAIN32
YUV2RGB_RES
RGB32_PACK(a, b1, g1, r1, rgb1_l, rgb1_h);
RGB32_STORE_RES(rgb1_l, rgb1_h, image1, image2);
END_FUNC()

2
libswscale/rgb2rgb.c
View File

@ -141,6 +141,8 @@ av_cold void ff_sws_rgb2rgb_init(void)
rgb2rgb_init_aarch64();
#elif ARCH_X86
rgb2rgb_init_x86();
#elif ARCH_LOONGARCH64
rgb2rgb_init_loongarch();
#endif
}

1
libswscale/rgb2rgb.h
View File

@ -168,5 +168,6 @@ void ff_sws_rgb2rgb_init(void);
void rgb2rgb_init_aarch64(void);
void rgb2rgb_init_x86(void);
void rgb2rgb_init_loongarch(void);
#endif /* SWSCALE_RGB2RGB_H */

2
libswscale/yuv2rgb.c
View File

@ -683,6 +683,8 @@ SwsFunc ff_yuv2rgb_get_func_ptr(SwsContext *c)
t = ff_yuv2rgb_init_ppc(c);
#elif ARCH_X86
t = ff_yuv2rgb_init_x86(c);
#elif ARCH_LOONGARCH64
t = ff_yuv2rgb_init_loongarch(c);
#endif
if (t)