avcodec/x86: add avx512icl function for v210dec

Ice Lake (Xeon Silver 4316): 2.01x faster (1147±36.8 vs. 571±38.2 decicycles) compared with avx2
pull/388/head
James Darnley 2 years ago
parent f30b4c2f47
commit 6af453ca38
  1. 10
      libavcodec/x86/v210-init.c
  2. 60
      libavcodec/x86/v210.asm
  3. 12
      tests/checkasm/v210dec.c

@ -17,7 +17,7 @@
*/
#include "libavutil/attributes.h"
#include "libavutil/cpu.h"
#include "libavutil/x86/cpu.h"
#include "libavcodec/v210dec.h"
extern void ff_v210_planar_unpack_unaligned_ssse3(const uint32_t *src, uint16_t *y, uint16_t *u, uint16_t *v, int width);
@ -28,6 +28,8 @@ extern void ff_v210_planar_unpack_aligned_ssse3(const uint32_t *src, uint16_t *y
extern void ff_v210_planar_unpack_aligned_avx(const uint32_t *src, uint16_t *y, uint16_t *u, uint16_t *v, int width);
extern void ff_v210_planar_unpack_aligned_avx2(const uint32_t *src, uint16_t *y, uint16_t *u, uint16_t *v, int width);
extern void ff_v210_planar_unpack_avx512icl(const uint32_t *src, uint16_t *y, uint16_t *u, uint16_t *v, int width);
av_cold void ff_v210_x86_init(V210DecContext *s)
{
#if HAVE_X86ASM
@ -42,6 +44,9 @@ av_cold void ff_v210_x86_init(V210DecContext *s)
if (HAVE_AVX2_EXTERNAL && cpu_flags & AV_CPU_FLAG_AVX2)
s->unpack_frame = ff_v210_planar_unpack_aligned_avx2;
if (EXTERNAL_AVX512ICL(cpu_flags))
s->unpack_frame = ff_v210_planar_unpack_avx512icl;
}
else {
if (cpu_flags & AV_CPU_FLAG_SSSE3)
@ -52,6 +57,9 @@ av_cold void ff_v210_x86_init(V210DecContext *s)
if (HAVE_AVX2_EXTERNAL && cpu_flags & AV_CPU_FLAG_AVX2)
s->unpack_frame = ff_v210_planar_unpack_unaligned_avx2;
if (EXTERNAL_AVX512ICL(cpu_flags))
s->unpack_frame = ff_v210_planar_unpack_avx512icl;
}
#endif
}

@ -22,7 +22,21 @@
%include "libavutil/x86/x86util.asm"
SECTION_RODATA 32
SECTION_RODATA 64
perm_y:
db 0,1, 4,5, 6,7, 8,9, 12,13, 14,15, 16,17, 20,21
db 22,23, 24,25, 28,29, 30,31, 32,33, 36,37, 38,39, 40,41
db 44,45, 46,47, 48,49, 52,53, 54,55, 56,57, 60,61, 62,63
times 16 db 0xff ; align to 64
perm_uv:
db 0,1, 4,5, 10,11, 16,17, 20,21, 26,27, 32,33, 36,37
db 42,43, 48,49, 52,53, 58,59
times 8 db 0xff ; align to 32
db 2,3, 8,9, 12,13, 18,19, 24,25, 28,29, 34,35, 40,41
db 44,45, 50,51, 56,57, 60,61
times 8 db 0xff ; align to 32
; for AVX2 version only
v210_luma_permute: dd 0,1,2,4,5,6,7,7 ; 32-byte alignment required
@ -34,6 +48,9 @@ v210_mult: dw 64,4,64,4,64,4,64,4
v210_luma_shuf: db 8,9,0,1,2,3,12,13,4,5,6,7,-1,-1,-1,-1
v210_chroma_shuf: db 0,1,8,9,6,7,-1,-1,2,3,4,5,12,13,-1,-1
shift: times 4 dw 6, 2
kmask: dw 0x5555, 0xaaaa
SECTION .text
%macro v210_planar_unpack 1
@ -127,3 +144,44 @@ v210_planar_unpack aligned
INIT_YMM avx2
v210_planar_unpack aligned
%endif
%if HAVE_AVX512ICL_EXTERNAL
INIT_ZMM avx512icl
cglobal v210_planar_unpack, 5, 5, 6, src, y, u, v, w
movsxdifnidn wq, wd
lea yq, [yq+2*wq]
add uq, wq
add vq, wq
neg wq
kmovw k1, [kmask] ; odd dword mask
kmovw k2, [kmask+2] ; even dword mask
VBROADCASTI128 m0, [shift]
mova m1, [perm_y]
mova m2, [perm_uv]
.loop:
movu m3, [srcq]
vpsllvw m4, m3, m0
pslld m5, m3, 12
psrlw m4, 6
psrld m5, 22
vpblendmd m3{k1}, m4, m5
vpermb m3, m1, m3 ; could use vpcompressw
movu [yq+2*wq], m3
vpblendmd m5{k2}, m4, m5
vpermb m5, m2, m5
movu [uq+wq], ym5
vextracti32x8 [vq+wq], zm5, 1
add srcq, mmsize
add wq, (mmsize*3)/8
jl .loop
RET
%endif

@ -54,12 +54,12 @@ void checkasm_check_v210dec(void)
if (check_func(h.unpack_frame, "v210_unpack")) {
uint32_t src0[NUM_SAMPLES/3];
uint32_t src1[NUM_SAMPLES/3];
uint16_t y0[NUM_SAMPLES/2];
uint16_t y1[NUM_SAMPLES/2];
uint16_t u0[NUM_SAMPLES/4];
uint16_t u1[NUM_SAMPLES/4];
uint16_t v0[NUM_SAMPLES/4];
uint16_t v1[NUM_SAMPLES/4];
uint16_t y0[NUM_SAMPLES/2 + 15];
uint16_t y1[NUM_SAMPLES/2 + 15];
uint16_t u0[NUM_SAMPLES/4 + 7];
uint16_t u1[NUM_SAMPLES/4 + 7];
uint16_t v0[NUM_SAMPLES/4 + 7];
uint16_t v1[NUM_SAMPLES/4 + 7];
declare_func(void, const uint32_t *src, uint16_t *y, uint16_t *u, uint16_t *v, int width);
const int pixels = NUM_SAMPLES / 2 / 6 * 6;

Loading…
Cancel
Save