[FFmpeg-devel] [PATCH] vp9: add mxext versions of the single-block (w=8, npx=8) h/v loopfilters.

[Ronald S. Bultje]

Each takes about 0.1% of runtime in my profiles, and they didn't have
any SIMD yet so far (we only had simd for npx=16 double-block versions).
---
 libavcodec/x86/vp9dsp_init.c |  3 +++
 libavcodec/x86/vp9lpf.asm    | 40 ++++++++++++++++++++++++++++------------
 2 files changed, 31 insertions(+), 12 deletions(-)

diff --git a/libavcodec/x86/vp9dsp_init.c b/libavcodec/x86/vp9dsp_init.c
index 359d38e..cc781a0 100644
--- a/libavcodec/x86/vp9dsp_init.c
+++ b/libavcodec/x86/vp9dsp_init.c
@@ -127,6 +127,7 @@ void ff_vp9_loop_filter_h_##size1##_##size2##_##opt(uint8_t *dst, ptrdiff_t stri
                                                     int E, int I, int H)
 
 lpf_funcs(4, 8, mmxext);
+lpf_funcs(8, 8, mmxext);
 lpf_funcs(16, 16, sse2);
 lpf_funcs(16, 16, ssse3);
 lpf_funcs(16, 16, avx);
@@ -284,6 +285,8 @@ av_cold void ff_vp9dsp_init_x86(VP9DSPContext *dsp, int bpp, int bitexact)
     if (EXTERNAL_MMXEXT(cpu_flags)) {
         dsp->loop_filter_8[0][0] = ff_vp9_loop_filter_h_4_8_mmxext;
         dsp->loop_filter_8[0][1] = ff_vp9_loop_filter_v_4_8_mmxext;
+        dsp->loop_filter_8[1][0] = ff_vp9_loop_filter_h_8_8_mmxext;
+        dsp->loop_filter_8[1][1] = ff_vp9_loop_filter_v_8_8_mmxext;
         init_subpel2(4, 0, 4, put, 8, mmxext);
         init_subpel2(4, 1, 4, avg, 8, mmxext);
         init_fpel_func(4, 1,  4, avg, _8, mmxext);
diff --git a/libavcodec/x86/vp9lpf.asm b/libavcodec/x86/vp9lpf.asm
index 56bba4d..b25fb1c 100644
--- a/libavcodec/x86/vp9lpf.asm
+++ b/libavcodec/x86/vp9lpf.asm
@@ -650,7 +650,7 @@ cglobal vp9_loop_filter_%1_%2_ %+ mmsize, 2, 6, 16, %3 + %4 + %5, dst, stride, m
     ABSSUB_GT           m1, rp2, rp0, m6, m5, rb80      ; abs(p2 - p0) <= 1
     por                 m2, m1
     ABSSUB              m4, rp1, rp0, m5                ; abs(p1 - p0)
-%if %2 == 16
+%if %2 <= 16
 %if cpuflag(ssse3)
     pxor                m0, m0
 %endif
@@ -855,7 +855,7 @@ cglobal vp9_loop_filter_%1_%2_ %+ mmsize, 2, 6, 16, %3 + %4 + %5, dst, stride, m
     ; filter6()
 %if %2 != 44 && %2 != 4
     pxor                m0, m0
-%if %2 > 16
+%if %2 != 16
     pand                m3, m2
 %else
     pand                m2, m3                          ;               mask(fm) & mask(in)
@@ -1102,12 +1102,12 @@ cglobal vp9_loop_filter_%1_%2_ %+ mmsize, 2, 6, 16, %3 + %4 + %5, dst, stride, m
     mova                    m3, [P0]
     mova                    m4, [Q0]
     mova                    m5, [Q1]
-%if ARCH_X86_64
+%ifdef m8
     mova                    m6, [Q2]
 %endif
     mova                    m7, [Q3]
     DEFINE_REAL_P7_TO_Q7
-%if ARCH_X86_64
+%ifdef m8
     SBUTTERFLY  bw,  0,  1, 8
     SBUTTERFLY  bw,  2,  3, 8
     SBUTTERFLY  bw,  4,  5, 8
@@ -1122,27 +1122,32 @@ cglobal vp9_loop_filter_%1_%2_ %+ mmsize, 2, 6, 16, %3 + %4 + %5, dst, stride, m
     SBUTTERFLY  dq,  3,  7, 8
 %else
     SBUTTERFLY  bw,  0,  1, 6
-    mova  [rsp+64], m1
-    mova        m6, [rsp+96]
+    mova [rsp+mmsize*4], m1
+    mova        m6, [rsp+mmsize*6]
     SBUTTERFLY  bw,  2,  3, 1
     SBUTTERFLY  bw,  4,  5, 1
     SBUTTERFLY  bw,  6,  7, 1
     SBUTTERFLY  wd,  0,  2, 1
-    mova  [rsp+96], m2
-    mova        m1, [rsp+64]
+    mova [rsp+mmsize*6], m2
+    mova        m1, [rsp+mmsize*4]
     SBUTTERFLY  wd,  1,  3, 2
     SBUTTERFLY  wd,  4,  6, 2
     SBUTTERFLY  wd,  5,  7, 2
     SBUTTERFLY  dq,  0,  4, 2
     SBUTTERFLY  dq,  1,  5, 2
+%if mmsize == 16
     movh      [Q0], m1
     movhps    [Q1], m1
-    mova        m2, [rsp+96]
+%else
+    mova      [P3], m1
+%endif
+    mova        m2, [rsp+mmsize*6]
     SBUTTERFLY  dq,  2,  6, 1
     SBUTTERFLY  dq,  3,  7, 1
 %endif
     SWAP         3, 6
     SWAP         1, 4
+%if mmsize == 16
     movh      [P7], m0
     movhps    [P6], m0
     movh      [P5], m1
@@ -1151,7 +1156,7 @@ cglobal vp9_loop_filter_%1_%2_ %+ mmsize, 2, 6, 16, %3 + %4 + %5, dst, stride, m
     movhps    [P2], m2
     movh      [P1], m3
     movhps    [P0], m3
-%if ARCH_X86_64
+%ifdef m8
     movh      [Q0], m4
     movhps    [Q1], m4
 %endif
@@ -1161,6 +1166,15 @@ cglobal vp9_loop_filter_%1_%2_ %+ mmsize, 2, 6, 16, %3 + %4 + %5, dst, stride, m
     movhps    [Q5], m6
     movh      [Q6], m7
     movhps    [Q7], m7
+%else
+    mova      [P7], m0
+    mova      [P6], m1
+    mova      [P5], m2
+    mova      [P4], m3
+    mova      [P2], m5
+    mova      [P1], m6
+    mova      [P0], m7
+%endif
 %endif
 %endif
 
@@ -1186,5 +1200,7 @@ LPF_16_VH_ALL_OPTS 84, 256, 128, 16
 LPF_16_VH_ALL_OPTS 88, 256, 128, 16
 
 INIT_MMX mmxext
-LOOPFILTER v, 4, 0,  0, 0
-LOOPFILTER h, 4, 0, 64, 0
+LOOPFILTER v, 4,   0,  0, 0
+LOOPFILTER h, 4,   0, 64, 0
+LOOPFILTER v, 8, 128,  0, 8
+LOOPFILTER h, 8, 128, 64, 8
-- 
2.8.1