PR #25897 opened by zuxy
URL: https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/25897
Patch URL: https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/25897.patch

Following the same pattern in 8-bit vp9, this series of patches removes all use 
of MMX registers in said area and replace them with bit exact implementations 
with XMM registers; performance neutral or better.


>From 51198296ae3dcebc651c8c5b0d0a2b94feb242b2 Mon Sep 17 00:00:00 2001
From: Zuxy Meng <[email protected]>
Date: Mon, 2 Nov 2026 21:09:13 -0800
Subject: [PATCH 1/4] avcodec/x86/vp9: convert v_4x4_16 predictor MMX to SSE

Convert the 16-bit vertical 4x4 predictor from MMX to XMM.

Benchmarks (checkasm, Zen 4, pinned, median of 3
interleaved pairs, cycles): both sides sit at the ~0.1
cycle noise floor for this single-load/four-store kernel,
with no measurable difference (10bpp: 0.2/0.0/0.1 -> 0.1/
0.1/0.1; 12bpp: 0.1/0.0/0.1 -> 0.2/0.1/0.1).

checkasm --test=vp9dsp passes (1272/1272 over 4 seeds).

Signed-off-by: Zuxy Meng <[email protected]>
---
 libavcodec/x86/vp9dsp_init_16bpp.c    |  7 ++-----
 libavcodec/x86/vp9intrapred_16bpp.asm | 13 +++++++------
 2 files changed, 9 insertions(+), 11 deletions(-)

diff --git a/libavcodec/x86/vp9dsp_init_16bpp.c 
b/libavcodec/x86/vp9dsp_init_16bpp.c
index 3748c316b08..0c290e0f4b2 100644
--- a/libavcodec/x86/vp9dsp_init_16bpp.c
+++ b/libavcodec/x86/vp9dsp_init_16bpp.c
@@ -43,7 +43,7 @@ decl_fpel_func(avg,  32, _16, avx2);
 decl_fpel_func(avg,  64, _16, avx2);
 decl_fpel_func(avg, 128, _16, avx2);
 
-decl_ipred_fns(v,       16, mmx,    sse);
+decl_ipred_fns(v,       16, sse,    sse);
 decl_ipred_fns(h,       16, mmxext, sse2);
 decl_ipred_fns(dc,      16, mmxext, sse2);
 decl_ipred_fns(dc_top,  16, mmxext, sse2);
@@ -71,10 +71,6 @@ av_cold void ff_vp9dsp_init_16bpp_x86(VP9DSPContext *dsp)
 {
     int cpu_flags = av_get_cpu_flags();
 
-    if (EXTERNAL_MMX(cpu_flags)) {
-        init_ipred_func(v, VERT, 4, 16, mmx);
-    }
-
     if (EXTERNAL_MMXEXT(cpu_flags)) {
         init_ipred_func(h, HOR, 4, 16, mmxext);
         init_ipred_func(dc, DC, 4, 16, mmxext);
@@ -88,6 +84,7 @@ av_cold void ff_vp9dsp_init_16bpp_x86(VP9DSPContext *dsp)
         init_fpel_func(1, 0,  64, put, , sse);
         init_fpel_func(0, 0, 128, put, , sse);
         init_8_16_32_ipred_funcs(v, VERT, 16, sse);
+        init_ipred_func(v, VERT, 4, 16, sse);
     }
 
     if (EXTERNAL_SSE2(cpu_flags)) {
diff --git a/libavcodec/x86/vp9intrapred_16bpp.asm 
b/libavcodec/x86/vp9intrapred_16bpp.asm
index 808056a809e..9ba20a986a4 100644
--- a/libavcodec/x86/vp9intrapred_16bpp.asm
+++ b/libavcodec/x86/vp9intrapred_16bpp.asm
@@ -82,16 +82,17 @@ SECTION .text
 %endif
 %endmacro
 
-INIT_MMX mmx
+INIT_XMM sse
 cglobal vp9_ipred_v_4x4_16, 2, 4, 1, dst, stride, l, a
     movifnidn               aq, amp
-    mova                    m0, [aq]
+    xorps                   m0, m0
+    movlps                  m0, [aq]
     DEFINE_ARGS dst, stride, stride3
     lea               stride3q, [strideq*3]
-    mova      [dstq+strideq*0], m0
-    mova      [dstq+strideq*1], m0
-    mova      [dstq+strideq*2], m0
-    mova      [dstq+stride3q ], m0
+    movlps    [dstq+strideq*0], m0
+    movlps    [dstq+strideq*1], m0
+    movlps    [dstq+strideq*2], m0
+    movlps    [dstq+stride3q ], m0
     RET
 
 INIT_XMM sse
-- 
2.54.0


>From 282ff3a7f1379b516066b9ca9be67fb7c28be998 Mon Sep 17 00:00:00 2001
From: Zuxy Meng <[email protected]>
Date: Mon, 2 Nov 2026 21:09:34 -0800
Subject: [PATCH 2/4] avcodec/x86/vp9: convert h_4x4_16 predictor MMXEXT to
 SSE2

Convert the 16-bit horizontal 4x4 predictor from MMXEXT to
XMM.

Benchmarks (checkasm, Zen 4, pinned, median of 3
interleaved pairs, cycles):
 vp9_hor_4x4_10bpp: 0.8 -> 0.4
 vp9_hor_4x4_12bpp: 0.8 -> 0.4

checkasm --test=vp9dsp passes (1272/1272 over 3 seeds).

Signed-off-by: Zuxy Meng <[email protected]>
---
 libavcodec/x86/vp9dsp_init_16bpp.c    |  5 ++---
 libavcodec/x86/vp9intrapred_16bpp.asm | 20 ++++++++++----------
 2 files changed, 12 insertions(+), 13 deletions(-)

diff --git a/libavcodec/x86/vp9dsp_init_16bpp.c 
b/libavcodec/x86/vp9dsp_init_16bpp.c
index 0c290e0f4b2..18c7006eb72 100644
--- a/libavcodec/x86/vp9dsp_init_16bpp.c
+++ b/libavcodec/x86/vp9dsp_init_16bpp.c
@@ -44,7 +44,7 @@ decl_fpel_func(avg,  64, _16, avx2);
 decl_fpel_func(avg, 128, _16, avx2);
 
 decl_ipred_fns(v,       16, sse,    sse);
-decl_ipred_fns(h,       16, mmxext, sse2);
+decl_ipred_fns(h,       16, sse2,   sse2);
 decl_ipred_fns(dc,      16, mmxext, sse2);
 decl_ipred_fns(dc_top,  16, mmxext, sse2);
 decl_ipred_fns(dc_left, 16, mmxext, sse2);
@@ -72,7 +72,6 @@ av_cold void ff_vp9dsp_init_16bpp_x86(VP9DSPContext *dsp)
     int cpu_flags = av_get_cpu_flags();
 
     if (EXTERNAL_MMXEXT(cpu_flags)) {
-        init_ipred_func(h, HOR, 4, 16, mmxext);
         init_ipred_func(dc, DC, 4, 16, mmxext);
         init_ipred_func(dc_top,  TOP_DC,  4, 16, mmxext);
         init_ipred_func(dc_left, LEFT_DC, 4, 16, mmxext);
@@ -94,7 +93,7 @@ av_cold void ff_vp9dsp_init_16bpp_x86(VP9DSPContext *dsp)
         init_fpel_func(2, 1,  32, avg, _16, sse2);
         init_fpel_func(1, 1,  64, avg, _16, sse2);
         init_fpel_func(0, 1, 128, avg, _16, sse2);
-        init_8_16_32_ipred_funcs(h, HOR, 16, sse2);
+        init_ipred_funcs(h, HOR, 16, sse2);
         init_8_16_32_ipred_funcs(dc, DC, 16, sse2);
         init_8_16_32_ipred_funcs(dc_top,  TOP_DC,  16, sse2);
         init_8_16_32_ipred_funcs(dc_left, LEFT_DC, 16, sse2);
diff --git a/libavcodec/x86/vp9intrapred_16bpp.asm 
b/libavcodec/x86/vp9intrapred_16bpp.asm
index 9ba20a986a4..91eb181aa97 100644
--- a/libavcodec/x86/vp9intrapred_16bpp.asm
+++ b/libavcodec/x86/vp9intrapred_16bpp.asm
@@ -157,19 +157,19 @@ cglobal vp9_ipred_v_32x32_16, 2, 4, 4, dst, stride, l, a
     jg .loop
     RET
 
-INIT_MMX mmxext
+INIT_XMM sse2
 cglobal vp9_ipred_h_4x4_16, 3, 3, 4, dst, stride, l, a
-    mova                    m3, [lq]
+    movq                    m3, [lq]
     DEFINE_ARGS dst, stride, stride3
     lea               stride3q, [strideq*3]
-    pshufw                  m0, m3, q3333
-    pshufw                  m1, m3, q2222
-    pshufw                  m2, m3, q1111
-    pshufw                  m3, m3, q0000
-    mova      [dstq+strideq*0], m0
-    mova      [dstq+strideq*1], m1
-    mova      [dstq+strideq*2], m2
-    mova      [dstq+stride3q ], m3
+    pshuflw                 m0, m3, q3333
+    pshuflw                 m1, m3, q2222
+    pshuflw                 m2, m3, q1111
+    pshuflw                 m3, m3, q0000
+    movq      [dstq+strideq*0], m0
+    movq      [dstq+strideq*1], m1
+    movq      [dstq+strideq*2], m2
+    movq      [dstq+stride3q ], m3
     RET
 
 INIT_XMM sse2
-- 
2.54.0


>From 4046b1ea16fec660f334148ec992a0d526cfd95e Mon Sep 17 00:00:00 2001
From: Zuxy Meng <[email protected]>
Date: Mon, 2 Nov 2026 21:10:20 -0800
Subject: [PATCH 3/4] avcodec/x86/vp9: convert dc/dc_top/dc_left 4x4_16
 predictors MMXEXT to SSE2

Convert the 16-bit DC 4x4 predictors (dc, dc_top and dc_left,
the latter two sharing the DC_1D_FNS macro) from MMXEXT to XMM.
The EXTERNAL_MMXEXT block in vp9dsp_init_16bpp.c is now empty
and removed.

Benchmarks (checkasm, Zen 4, pinned, median of 3
interleaved pairs, cycles):
 vp9_dc_4x4_10bpp:      1.1 -> 0.9
 vp9_dc_4x4_12bpp:      1.1 -> 0.9
 vp9_dc_left_4x4_10bpp: 1.0 -> 0.4
 vp9_dc_left_4x4_12bpp: 0.9 -> 0.4
 vp9_dc_top_4x4_10bpp:  1.0 -> 0.5
 vp9_dc_top_4x4_12bpp:  1.0 -> 0.5

checkasm --test=vp9dsp passes (1272/1272 over 3 seeds).

Signed-off-by: Zuxy Meng <[email protected]>
---
 libavcodec/x86/vp9dsp_init_16bpp.c    | 18 +++++---------
 libavcodec/x86/vp9intrapred_16bpp.asm | 35 ++++++++++++++-------------
 2 files changed, 24 insertions(+), 29 deletions(-)

diff --git a/libavcodec/x86/vp9dsp_init_16bpp.c 
b/libavcodec/x86/vp9dsp_init_16bpp.c
index 18c7006eb72..bc475290940 100644
--- a/libavcodec/x86/vp9dsp_init_16bpp.c
+++ b/libavcodec/x86/vp9dsp_init_16bpp.c
@@ -45,9 +45,9 @@ decl_fpel_func(avg, 128, _16, avx2);
 
 decl_ipred_fns(v,       16, sse,    sse);
 decl_ipred_fns(h,       16, sse2,   sse2);
-decl_ipred_fns(dc,      16, mmxext, sse2);
-decl_ipred_fns(dc_top,  16, mmxext, sse2);
-decl_ipred_fns(dc_left, 16, mmxext, sse2);
+decl_ipred_fns(dc,      16, sse2,   sse2);
+decl_ipred_fns(dc_top,  16, sse2,   sse2);
+decl_ipred_fns(dc_left, 16, sse2,   sse2);
 decl_ipred_fn(dl,       16,     16, avx2);
 decl_ipred_fn(dl,       32,     16, avx2);
 decl_ipred_fn(dr,       16,     16, avx2);
@@ -71,12 +71,6 @@ av_cold void ff_vp9dsp_init_16bpp_x86(VP9DSPContext *dsp)
 {
     int cpu_flags = av_get_cpu_flags();
 
-    if (EXTERNAL_MMXEXT(cpu_flags)) {
-        init_ipred_func(dc, DC, 4, 16, mmxext);
-        init_ipred_func(dc_top,  TOP_DC,  4, 16, mmxext);
-        init_ipred_func(dc_left, LEFT_DC, 4, 16, mmxext);
-    }
-
     if (EXTERNAL_SSE(cpu_flags)) {
         init_fpel_func(3, 0,  16, put, , sse);
         init_fpel_func(2, 0,  32, put, , sse);
@@ -94,9 +88,9 @@ av_cold void ff_vp9dsp_init_16bpp_x86(VP9DSPContext *dsp)
         init_fpel_func(1, 1,  64, avg, _16, sse2);
         init_fpel_func(0, 1, 128, avg, _16, sse2);
         init_ipred_funcs(h, HOR, 16, sse2);
-        init_8_16_32_ipred_funcs(dc, DC, 16, sse2);
-        init_8_16_32_ipred_funcs(dc_top,  TOP_DC,  16, sse2);
-        init_8_16_32_ipred_funcs(dc_left, LEFT_DC, 16, sse2);
+        init_ipred_funcs(dc, DC, 16, sse2);
+        init_ipred_funcs(dc_top,  TOP_DC,  16, sse2);
+        init_ipred_funcs(dc_left, LEFT_DC, 16, sse2);
         init_ipred_funcs(dl, DIAG_DOWN_LEFT, 16, sse2);
         init_ipred_funcs(dr, DIAG_DOWN_RIGHT, 16, sse2);
         init_ipred_funcs(vl, VERT_LEFT, 16, sse2);
diff --git a/libavcodec/x86/vp9intrapred_16bpp.asm 
b/libavcodec/x86/vp9intrapred_16bpp.asm
index 91eb181aa97..196e53fe10b 100644
--- a/libavcodec/x86/vp9intrapred_16bpp.asm
+++ b/libavcodec/x86/vp9intrapred_16bpp.asm
@@ -254,22 +254,23 @@ cglobal vp9_ipred_h_32x32_16, 3, 5, 4, dst, stride, l, 
stride3, cnt
     jge .loop
     RET
 
-INIT_MMX mmxext
+INIT_XMM sse2
 cglobal vp9_ipred_dc_4x4_16, 4, 4, 2, dst, stride, l, a
-    mova                    m0, [lq]
-    paddw                   m0, [aq]
+    movq                    m0, [lq]
+    movq                    m1, [aq]
+    paddw                   m0, m1
     DEFINE_ARGS dst, stride, stride3
     lea               stride3q, [strideq*3]
     pmaddwd                 m0, [pw_1]
-    pshufw                  m1, m0, q3232
+    pshuflw                 m1, m0, q3232
     paddd                   m0, [pd_4]
     paddd                   m0, m1
     psrad                   m0, 3
-    pshufw                  m0, m0, q0000
-    mova      [dstq+strideq*0], m0
-    mova      [dstq+strideq*1], m0
-    mova      [dstq+strideq*2], m0
-    mova      [dstq+stride3q ], m0
+    pshuflw                 m0, m0, 0
+    movq      [dstq+strideq*0], m0
+    movq      [dstq+strideq*1], m0
+    movq      [dstq+strideq*2], m0
+    movq      [dstq+stride3q ], m0
     RET
 
 INIT_XMM sse2
@@ -367,21 +368,21 @@ cglobal vp9_ipred_dc_32x32_16, 4, 4, 2, dst, stride, l, a
     RET
 
 %macro DC_1D_FNS 2
-INIT_MMX mmxext
+INIT_XMM sse2
 cglobal vp9_ipred_dc_%1_4x4_16, 4, 4, 2, dst, stride, l, a
-    mova                    m0, [%2]
+    movq                    m0, [%2]
     DEFINE_ARGS dst, stride, stride3
     lea               stride3q, [strideq*3]
     pmaddwd                 m0, [pw_1]
-    pshufw                  m1, m0, q3232
+    pshuflw                 m1, m0, q3232
     paddd                   m0, [pd_2]
     paddd                   m0, m1
     psrad                   m0, 2
-    pshufw                  m0, m0, q0000
-    mova      [dstq+strideq*0], m0
-    mova      [dstq+strideq*1], m0
-    mova      [dstq+strideq*2], m0
-    mova      [dstq+stride3q ], m0
+    pshuflw                 m0, m0, 0
+    movq      [dstq+strideq*0], m0
+    movq      [dstq+strideq*1], m0
+    movq      [dstq+strideq*2], m0
+    movq      [dstq+stride3q ], m0
     RET
 
 INIT_XMM sse2
-- 
2.54.0


>From c12df7c8ffbe1580cdc868060e92e71830f229ee Mon Sep 17 00:00:00 2001
From: Zuxy Meng <[email protected]>
Date: Mon, 2 Nov 2026 21:10:49 -0800
Subject: [PATCH 4/4] avcodec/x86/vp9: convert tm_4x4_10/12 predictors MMXEXT
 to SSE2

Convert the 16-bit TM_VP8 4x4 predictors (10- and 12-bit,
sharing one .body) from MMXEXT to XMM.

This completes deprecation of all MMX code from vp9's intrapred, so
we replace declare_func_emms with declare_func in test code too.

Benchmarks (checkasm, Zen 4, pinned, median of 3
interleaved pairs, cycles):
 vp9_tm_4x4_10bpp: 2.9 -> 1.8
 vp9_tm_4x4_12bpp: 2.9 -> 1.9

checkasm --test=vp9dsp passes (1272/1272 over 3 seeds).

Signed-off-by: Zuxy Meng <[email protected]>
---
 libavcodec/x86/vp9dsp_init_16bpp_template.c |  5 ++---
 libavcodec/x86/vp9intrapred_16bpp.asm       | 24 ++++++++++-----------
 tests/checkasm/vp9dsp.c                     |  4 ++--
 3 files changed, 16 insertions(+), 17 deletions(-)

diff --git a/libavcodec/x86/vp9dsp_init_16bpp_template.c 
b/libavcodec/x86/vp9dsp_init_16bpp_template.c
index 969db94d3c8..2817d0ec81c 100644
--- a/libavcodec/x86/vp9dsp_init_16bpp_template.c
+++ b/libavcodec/x86/vp9dsp_init_16bpp_template.c
@@ -119,7 +119,7 @@ lpf_mix2_wrappers_set(BPC, sse2)
 lpf_mix2_wrappers_set(BPC, ssse3)
 lpf_mix2_wrappers_set(BPC, avx)
 
-decl_ipred_fns(tm, BPC, mmxext, sse2);
+decl_ipred_fns(tm, BPC, sse2, sse2);
 
 decl_itxfm_func(iwht, iwht, 4, BPC, mmxext);
 #if BPC == 10
@@ -180,7 +180,6 @@ av_cold void INIT_FUNC(VP9DSPContext *dsp, int bitexact)
     init_itx_func(idx, ADST_ADST, iadst, iadst, size, bpp, opt); \
 
     if (EXTERNAL_MMXEXT(cpu_flags)) {
-        init_ipred_func(tm, TM_VP8, 4, BPC, mmxext);
         if (!bitexact) {
             init_itx_func_one(4 /* lossless */, iwht, iwht, 4, BPC, mmxext);
         }
@@ -190,7 +189,7 @@ av_cold void INIT_FUNC(VP9DSPContext *dsp, int bitexact)
         init_subpel3(0, put, BPC, sse2);
         init_subpel3(1, avg, BPC, sse2);
         init_lpf_funcs(BPC, sse2);
-        init_8_16_32_ipred_funcs(tm, TM_VP8, BPC, sse2);
+        init_ipred_funcs(tm, TM_VP8, BPC, sse2);
 #if BPC == 10
         if (!bitexact) {
             init_itx_func(TX_4X4, ADST_DCT,  idct,  iadst, 4, 10, sse2);
diff --git a/libavcodec/x86/vp9intrapred_16bpp.asm 
b/libavcodec/x86/vp9intrapred_16bpp.asm
index 196e53fe10b..d5881516749 100644
--- a/libavcodec/x86/vp9intrapred_16bpp.asm
+++ b/libavcodec/x86/vp9intrapred_16bpp.asm
@@ -475,21 +475,21 @@ cglobal vp9_ipred_dc_%1_32x32_16, 4, 4, 2, dst, stride, 
l, a
 DC_1D_FNS top,  aq
 DC_1D_FNS left, lq
 
-INIT_MMX mmxext
+INIT_XMM sse2
 cglobal vp9_ipred_tm_4x4_10, 4, 4, 6, dst, stride, l, a
     mova                    m5, [pw_1023]
 .body:
-    mova                    m4, [aq]
-    mova                    m3, [lq]
+    movq                    m4, [aq]
+    movq                    m3, [lq]
     movd                    m0, [aq-4]
-    pshufw                  m0, m0, q1111
+    pshuflw                 m0, m0, q1111
     psubw                   m4, m0
     DEFINE_ARGS dst, stride, stride3
     lea               stride3q, [strideq*3]
-    pshufw                  m0, m3, q3333
-    pshufw                  m1, m3, q2222
-    pshufw                  m2, m3, q1111
-    pshufw                  m3, m3, q0000
+    pshuflw                 m0, m3, q3333
+    pshuflw                 m1, m3, q2222
+    pshuflw                 m2, m3, q1111
+    pshuflw                 m3, m3, q0000
     paddw                   m0, m4
     paddw                   m1, m4
     paddw                   m2, m4
@@ -503,10 +503,10 @@ cglobal vp9_ipred_tm_4x4_10, 4, 4, 6, dst, stride, l, a
     pminsw                  m1, m5
     pminsw                  m2, m5
     pminsw                  m3, m5
-    mova      [dstq+strideq*0], m0
-    mova      [dstq+strideq*1], m1
-    mova      [dstq+strideq*2], m2
-    mova      [dstq+stride3q ], m3
+    movq      [dstq+strideq*0], m0
+    movq      [dstq+strideq*1], m1
+    movq      [dstq+strideq*2], m2
+    movq      [dstq+stride3q ], m3
     RET
 
 cglobal vp9_ipred_tm_4x4_12, 4, 4, 6, dst, stride, l, a
diff --git a/tests/checkasm/vp9dsp.c b/tests/checkasm/vp9dsp.c
index 993e0715aa3..986c5802a39 100644
--- a/tests/checkasm/vp9dsp.c
+++ b/tests/checkasm/vp9dsp.c
@@ -56,8 +56,8 @@ static void check_ipred(void)
     LOCAL_ALIGNED_32(uint8_t, dst1, [32 * 32 * 2]);
     VP9DSPContext dsp;
     int tx, mode, bit_depth;
-    declare_func_emms(AV_CPU_FLAG_MMX | AV_CPU_FLAG_MMXEXT, void, uint8_t 
*dst, ptrdiff_t stride,
-                      const uint8_t *left, const uint8_t *top);
+    declare_func(void, uint8_t *dst, ptrdiff_t stride,
+                 const uint8_t *left, const uint8_t *top);
     static const char *const mode_names[N_INTRA_PRED_MODES] = {
         [VERT_PRED] = "vert",
         [HOR_PRED] = "hor",
-- 
2.54.0

_______________________________________________
ffmpeg-devel mailing list -- [email protected]
To unsubscribe send an email to [email protected]

Reply via email to