This is an automated email from the git hooks/post-receive script.
Git pushed a commit to branch master
in repository ffmpeg.
The following commit(s) were added to refs/heads/master by this push:
new ec6c89c1bb avcodec/x86/h264_intrapred: SSE2 impl. of pred8x8l_dc_8
ec6c89c1bb is described below
commit ec6c89c1bb0a6b724b5c3930b7f53f1900b2af70
Author: Zuxy Meng <[email protected]>
AuthorDate: Sun Apr 26 21:12:49 2026 -0700
Commit: Zuxy Meng <[email protected]>
CommitDate: Wed Jul 8 17:35:03 2026 -0700
avcodec/x86/h264_intrapred: SSE2 impl. of pred8x8l_dc_8
Deprecate MMX. Remove the SSSE3 impl. since we no longer use palignr.
The SSE2 impl. uses several more instructions but is slightly faster.
pred8x8l_dc_8_mmxext: 20.5 ( 1.66x)
pred8x8l_dc_8_ssse3: 17.9 ( 1.90x)
pred8x8l_dc_8_sse2: 15.9 ( 2.15x)
Signed-off-by: Zuxy Meng <[email protected]>
---
libavcodec/x86/h264_intrapred.asm | 152 ++++++++++++++++-------------------
libavcodec/x86/h264_intrapred_init.c | 6 +-
2 files changed, 73 insertions(+), 85 deletions(-)
diff --git a/libavcodec/x86/h264_intrapred.asm
b/libavcodec/x86/h264_intrapred.asm
index d3bf6a627f..6dc941eb02 100644
--- a/libavcodec/x86/h264_intrapred.asm
+++ b/libavcodec/x86/h264_intrapred.asm
@@ -864,104 +864,94 @@ cglobal pred8x8l_top_dc_8, 4,4,6
; ptrdiff_t stride)
;-----------------------------------------------------------------------------
-%macro PRED8x8L_DC 0
-cglobal pred8x8l_dc_8, 4,5
+INIT_XMM sse2
+cglobal pred8x8l_dc_8, 4,5,6
sub r0, r3
lea r4, [r0+r3*2]
- movq mm0, [r0+r3*1-8]
- punpckhbw mm0, [r0+r3*0-8]
- movq mm1, [r4+r3*1-8]
- punpckhbw mm1, [r0+r3*2-8]
+ movd m0, [r0+r3*1-4]
+ movd m4, [r0+r3*0-4]
+ punpcklbw m0, m4
+ movd m1, [r4+r3*1-4]
+ movd m4, [r0+r3*2-4]
+ punpcklbw m1, m4
mov r4, r0
- punpckhwd mm1, mm0
+ punpcklwd m1, m0
lea r0, [r0+r3*4]
- movq mm2, [r0+r3*1-8]
- punpckhbw mm2, [r0+r3*0-8]
+ movd m2, [r0+r3*1-4]
+ movd m4, [r0+r3*0-4]
+ punpcklbw m2, m4
lea r0, [r0+r3*2]
- movq mm3, [r0+r3*1-8]
- punpckhbw mm3, [r0+r3*0-8]
- punpckhwd mm3, mm2
- punpckhdq mm3, mm1
+ movd m3, [r0+r3*1-4]
+ movd m4, [r0+r3*0-4]
+ punpcklbw m3, m4
+ punpcklwd m3, m2
+ shufps m3, m1, 0xed
+ pshufd m3, m3, 0x0d
lea r0, [r0+r3*2]
- movq mm0, [r0+r3*0-8]
- movq mm1, [r4]
+ movq m0, [r0+r3*0-8]
+ movq m1, [r4]
mov r0, r4
- movq mm4, mm3
- movq mm2, mm3
- PALIGNR mm4, mm0, 7, mm0
- PALIGNR mm1, mm2, 1, mm2
+ mova m4, m3
+ mova m2, m3
+ punpcklqdq m0, m4
+ psrldq m0, 7
+ punpcklqdq m2, m1
+ psrldq m2, 1
test r1d, r1d
jnz .do_left
-.fix_lt_1:
- movq mm5, mm3
- pxor mm5, mm4
- psrlq mm5, 56
- psllq mm5, 48
- pxor mm1, mm5
- jmp .do_left
-.fix_lt_2:
- movq mm5, mm3
- pxor mm5, mm2
- psllq mm5, 56
- psrlq mm5, 56
- pxor mm2, mm5
- test r2d, r2d
- jnz .body
-.fix_tr_1:
- movq mm5, mm3
- pxor mm5, mm1
- psrlq mm5, 56
- psllq mm5, 56
- pxor mm1, mm5
- jmp .body
+ pxor m5, m3, m0
+ psrlq m5, 56
+ psllq m5, 48
+ pxor m2, m5
.do_left:
- movq mm0, mm4
- PRED4x4_LOWPASS mm2, mm1, mm4, mm3, mm5
- movq mm1, mm0
- movq mm7, mm2
- PRED4x4_LOWPASS mm1, mm3, mm0, mm1, mm5
- psllq mm1, 56
- PALIGNR mm7, mm1, 7, mm3
- movq mm0, [r0-8]
- movq mm3, [r0]
- movq mm1, [r0+8]
- movq mm2, mm3
- movq mm4, mm3
- PALIGNR mm2, mm0, 7, mm0
- PALIGNR mm1, mm4, 1, mm4
+ mova m4, m0
+ PRED4x4_LOWPASS m1, m2, m4, m3, m5
+ mova m4, m0
+ PRED4x4_LOWPASS m2, m3, m0, m4, m5
+ psllq m2, 56
+ punpcklqdq m2, m1
+ psrldq m2, 7
+ movu m0, [r0-8]
+ movu m3, [r0]
+ mova m4, m3
+ psrldq m0, 7
+ psrldq m4, 1
test r1d, r1d
- jz .fix_lt_2
+ jnz .skip_fix_lt_2
+ pxor m1, m3, m0
+ psllq m1, 56
+ psrlq m1, 56
+ pxor m0, m1
+.skip_fix_lt_2:
test r2d, r2d
- jz .fix_tr_1
+ jnz .body
+ pxor m1, m3, m4
+ psrlq m1, 56
+ psllq m1, 56
+ pxor m4, m1
.body:
lea r1, [r0+r3*2]
- PRED4x4_LOWPASS mm3, mm2, mm1, mm3, mm5
- pxor mm0, mm0
- pxor mm1, mm1
+ PRED4x4_LOWPASS m3, m0, m4, m3, m1
+ pxor m1, m1
lea r2, [r1+r3*2]
- psadbw mm0, mm7
- psadbw mm1, mm3
- paddw mm0, [pw_8]
- paddw mm0, mm1
+ psadbw m2, m1
+ psadbw m3, m1
+ paddw m3, [pw_8]
+ paddw m3, m2
lea r4, [r2+r3*2]
- psrlw mm0, 4
- pshufw mm0, mm0, 0
- packuswb mm0, mm0
- movq [r0+r3*1], mm0
- movq [r0+r3*2], mm0
- movq [r1+r3*1], mm0
- movq [r1+r3*2], mm0
- movq [r2+r3*1], mm0
- movq [r2+r3*2], mm0
- movq [r4+r3*1], mm0
- movq [r4+r3*2], mm0
+ psrlw m3, 4
+ pshuflw m3, m3, 0
+ punpcklqdq m3, m3
+ packuswb m3, m3
+ movq [r0+r3*1], m3
+ movq [r0+r3*2], m3
+ movq [r1+r3*1], m3
+ movq [r1+r3*2], m3
+ movq [r2+r3*1], m3
+ movq [r2+r3*2], m3
+ movq [r4+r3*1], m3
+ movq [r4+r3*2], m3
RET
-%endmacro
-
-INIT_MMX mmxext
-PRED8x8L_DC
-INIT_MMX ssse3
-PRED8x8L_DC
;-----------------------------------------------------------------------------
; void ff_pred8x8l_horizontal_8(uint8_t *src, int has_topleft,
diff --git a/libavcodec/x86/h264_intrapred_init.c
b/libavcodec/x86/h264_intrapred_init.c
index b5d82694a2..ebcbe0891d 100644
--- a/libavcodec/x86/h264_intrapred_init.c
+++ b/libavcodec/x86/h264_intrapred_init.c
@@ -127,8 +127,7 @@ PRED8x8(tm_vp8, 8, sse2)
PRED8x8(tm_vp8, 8, ssse3)
PRED8x8L(top_dc, 8, sse2)
-PRED8x8L(dc, 8, mmxext)
-PRED8x8L(dc, 8, ssse3)
+PRED8x8L(dc, 8, sse2)
PRED8x8L(horizontal, 8, mmxext)
PRED8x8L(horizontal, 8, ssse3)
PRED8x8L(vertical, 8, mmxext)
@@ -165,7 +164,6 @@ av_cold void ff_h264_pred_init_x86(H264PredContext *h, int
codec_id,
if (bit_depth == 8) {
if (EXTERNAL_MMXEXT(cpu_flags)) {
- h->pred8x8l [DC_PRED ] = ff_pred8x8l_dc_8_mmxext;
h->pred8x8l [HOR_PRED ] =
ff_pred8x8l_horizontal_8_mmxext;
h->pred8x8l [VERT_PRED ] =
ff_pred8x8l_vertical_8_mmxext;
h->pred8x8l [HOR_UP_PRED ] =
ff_pred8x8l_horizontal_up_8_mmxext;
@@ -197,6 +195,7 @@ av_cold void ff_h264_pred_init_x86(H264PredContext *h, int
codec_id,
if (EXTERNAL_SSE2(cpu_flags)) {
h->pred16x16[HOR_PRED8x8 ] =
ff_pred16x16_horizontal_8_sse2;
h->pred16x16[DC_PRED8x8 ] = ff_pred16x16_dc_8_sse2;
+ h->pred8x8l [DC_PRED ] = ff_pred8x8l_dc_8_sse2;
h->pred8x8l [TOP_DC_PRED ] = ff_pred8x8l_top_dc_8_sse2;
h->pred8x8l [DIAG_DOWN_LEFT_PRED ] = ff_pred8x8l_down_left_8_sse2;
h->pred8x8l [DIAG_DOWN_RIGHT_PRED ] =
ff_pred8x8l_down_right_8_sse2;
@@ -232,7 +231,6 @@ av_cold void ff_h264_pred_init_x86(H264PredContext *h, int
codec_id,
h->pred16x16[DC_PRED8x8 ] = ff_pred16x16_dc_8_ssse3;
if (chroma_format_idc <= 1)
h->pred8x8 [HOR_PRED8x8 ] =
ff_pred8x8_horizontal_8_ssse3;
- h->pred8x8l [DC_PRED ] = ff_pred8x8l_dc_8_ssse3;
h->pred8x8l [HOR_PRED ] =
ff_pred8x8l_horizontal_8_ssse3;
h->pred8x8l [VERT_PRED ] = ff_pred8x8l_vertical_8_ssse3;
h->pred8x8l [DIAG_DOWN_LEFT_PRED ] =
ff_pred8x8l_down_left_8_ssse3;
_______________________________________________
ffmpeg-cvslog mailing list -- [email protected]
To unsubscribe send an email to [email protected]
