[FFmpeg-devel] [PATCH] avfilter/x86/vf_360: add most of >8 depth asm
Paul B Mahol
onemda at gmail.com
Sun Sep 15 19:15:41 EEST 2019
Signed-off-by: Paul B Mahol <onemda at gmail.com>
---
libavfilter/x86/vf_v360.asm | 67 ++++++++++++++++++++++++++++++++++
libavfilter/x86/vf_v360_init.c | 12 ++++++
2 files changed, 79 insertions(+)
diff --git a/libavfilter/x86/vf_v360.asm b/libavfilter/x86/vf_v360.asm
index a0936eb6dc..5b241220d8 100644
--- a/libavfilter/x86/vf_v360.asm
+++ b/libavfilter/x86/vf_v360.asm
@@ -26,7 +26,9 @@
SECTION_RODATA
pb_mask: db 0,4,8,12,-1,-1,-1,-1,-1,-1,-1,-1,-1,-1,-1,-1
+pw_mask: db 0,1,4, 5, 8, 9,12,13,-1,-1,-1,-1,-1,-1,-1,-1
pd_255: times 4 dd 255
+pd_65535: times 4 dd 65535
SECTION .text
@@ -60,6 +62,34 @@ cglobal remap1_8bit_line, 6, 7, 6, dst, width, src, in_linesize, u, v, x
jl .loop
RET
+INIT_YMM avx2
+cglobal remap1_16bit_line, 6, 7, 6, dst, width, src, in_linesize, u, v, x
+ movsxdifnidn widthq, widthd
+ xor xq, xq
+ movd xm0, in_linesized
+ pcmpeqw m4, m4
+ VBROADCASTI128 m3, [pw_mask]
+ vpbroadcastd m0, xm0
+
+ .loop:
+ pmovsxwd m1, [vq + xq * 2]
+ pmovsxwd m2, [uq + xq * 2]
+
+ pslld m2, 0x1
+ pmulld m1, m0
+ paddd m1, m2
+ mova m2, m4
+ vpgatherdd m5, [srcq + m1], m2
+ pshufb m1, m5, m3
+ vextracti128 xm2, m1, 1
+ movq [dstq+xq*2], xm1
+ movq [dstq+xq*2+8], xm2
+
+ add xq, mmsize / 4
+ cmp xq, widthq
+ jl .loop
+ RET
+
INIT_YMM avx2
cglobal remap2_8bit_line, 7, 8, 8, dst, width, src, in_linesize, u, v, ker, x
movsxdifnidn widthq, widthd
@@ -96,6 +126,43 @@ DEFINE_ARGS dst, width, src, x, u, v, ker
jl .loop
RET
+INIT_YMM avx2
+cglobal remap2_16bit_line, 7, 8, 8, dst, width, src, in_linesize, u, v, ker, x
+ movsxdifnidn widthq, widthd
+ movd xm0, in_linesized
+%if ARCH_X86_32
+DEFINE_ARGS dst, width, src, x, u, v, ker
+%endif
+ xor xq, xq
+ pcmpeqw m7, m7
+ vpbroadcastd m0, xm0
+ vpbroadcastd m6, [pd_65535]
+
+ .loop:
+ pmovsxwd m1, [kerq + xq * 8]
+ pmovsxwd m2, [vq + xq * 8]
+ pmovsxwd m3, [uq + xq * 8]
+
+ pslld m3, 0x1
+ pmulld m4, m2, m0
+ paddd m4, m3
+ mova m3, m7
+ vpgatherdd m2, [srcq + m4], m3
+ pand m2, m6
+ pmulld m2, m1
+ phaddd m2, m2
+ phaddd m1, m2, m2
+ psrld m1, m1, 0xe
+ vextracti128 xm2, m1, 1
+
+ pextrw [dstq+xq*2], xm1, 0
+ pextrw [dstq+xq*2+2], xm2, 0
+
+ add xq, mmsize / 16
+ cmp xq, widthq
+ jl .loop
+ RET
+
%if ARCH_X86_64
INIT_YMM avx2
diff --git a/libavfilter/x86/vf_v360_init.c b/libavfilter/x86/vf_v360_init.c
index 8c1a10c705..c7f4a3dd6d 100644
--- a/libavfilter/x86/vf_v360_init.c
+++ b/libavfilter/x86/vf_v360_init.c
@@ -32,6 +32,12 @@ void ff_remap2_8bit_line_avx2(uint8_t *dst, int width, const uint8_t *src, ptrdi
void ff_remap4_8bit_line_avx2(uint8_t *dst, int width, const uint8_t *src, ptrdiff_t in_linesize,
const uint16_t *u, const uint16_t *v, const int16_t *ker);
+void ff_remap1_16bit_line_avx2(uint8_t *dst, int width, const uint8_t *src, ptrdiff_t in_linesize,
+ const uint16_t *u, const uint16_t *v, const int16_t *ker);
+
+void ff_remap2_16bit_line_avx2(uint8_t *dst, int width, const uint8_t *src, ptrdiff_t in_linesize,
+ const uint16_t *u, const uint16_t *v, const int16_t *ker);
+
av_cold void ff_v360_init_x86(V360Context *s, int depth)
{
int cpu_flags = av_get_cpu_flags();
@@ -42,6 +48,12 @@ av_cold void ff_v360_init_x86(V360Context *s, int depth)
if (EXTERNAL_AVX2_FAST(cpu_flags) && s->interp == BILINEAR && depth <= 8)
s->remap_line = ff_remap2_8bit_line_avx2;
+ if (EXTERNAL_AVX2_FAST(cpu_flags) && s->interp == NEAREST && depth > 8)
+ s->remap_line = ff_remap1_16bit_line_avx2;
+
+ if (EXTERNAL_AVX2_FAST(cpu_flags) && s->interp == BILINEAR && depth > 8)
+ s->remap_line = ff_remap2_16bit_line_avx2;
+
#if ARCH_X86_64
if (EXTERNAL_AVX2_FAST(cpu_flags) && (s->interp == BICUBIC ||
s->interp == LANCZOS) && depth <= 8)
--
2.17.1
More information about the ffmpeg-devel
mailing list