x86/synth_filter: add synth_filter_avx

Sandy Bridge Win64:
180 cycles in ff_synth_filter_inner_sse2
150 cycles in ff_synth_filter_inner_avx

Also switch some instructions to a three operand format to avoid
assembly errors with Yasm 1.1.0 or older.

Signed-off-by: James Almer <jamrial@gmail.com>
Signed-off-by: Anton Khirnov <anton@khirnov.net>

2 files changed, 57 insertions, 32 deletions

diff --git a/libavcodec/x86/dcadsp.asm b/libavcodec/x86/dcadsp.asm
index 970ec3dacb..662cb96f66 100644
--- a/libavcodec/x86/dcadsp.asm
+++ b/libavcodec/x86/dcadsp.asm
@@ -200,18 +200,22 @@ DCA_LFE_FIR 0
 DCA_LFE_FIR 1
 
 %macro SETZERO 1
-%if cpuflag(sse2)
+%if cpuflag(sse2) && notcpuflag(avx)
     pxor          %1, %1
 %else
     xorps         %1, %1, %1
 %endif
 %endmacro
 
-%macro SHUF 2
-%if cpuflag(sse2)
-    pshufd        %1, %2, q0123
+%macro SHUF 3
+%if cpuflag(avx)
+    mova          %3, [%2 - 16]
+    vperm2f128    %1, %3, %3, 1
+    vshufps       %1, %1, %1, q0123
+%elif cpuflag(sse2)
+    pshufd        %1, [%2], q0123
 %else
-    mova          %1, %2
+    mova          %1, [%2]
     shufps        %1, %1, q0123
 %endif
 %endmacro
@@ -220,43 +224,43 @@ DCA_LFE_FIR 1
     ; reading backwards:  ptr1 = synth_buf + j + i; ptr2 = synth_buf + j - i
     ;~ a += window[i + j]      * (-synth_buf[15 - i + j])
     ;~ b += window[i + j + 16] * (synth_buf[i + j])
-    SHUF          m5, [ptr2 + j + (15 - 3) * 4]
+    SHUF          m5,  ptr2 + j + (15 - 3) * 4, m6
     mova          m6, [ptr1 + j]
 %if ARCH_X86_64
-    SHUF         m11, [ptr2 + j + (15 - 3) * 4 - mmsize]
+    SHUF         m11,  ptr2 + j + (15 - 3) * 4 - mmsize, m12
     mova         m12, [ptr1 + j + mmsize]
 %endif
-    mulps         m6, [win  + %1 + j + 16 * 4]
-    mulps         m5, [win  + %1 + j]
+    mulps         m6, m6,  [win + %1 + j + 16 * 4]
+    mulps         m5, m5,  [win + %1 + j]
 %if ARCH_X86_64
-    mulps        m12, [win  + %1 + j + mmsize + 16 * 4]
-    mulps        m11, [win  + %1 + j + mmsize]
+    mulps        m12, m12, [win + %1 + j + mmsize + 16 * 4]
+    mulps        m11, m11, [win + %1 + j + mmsize]
 %endif
-    addps         m2, m6
-    subps         m1, m5
+    addps         m2, m2, m6
+    subps         m1, m1, m5
 %if ARCH_X86_64
-    addps         m8, m12
-    subps         m7, m11
+    addps         m8, m8, m12
+    subps         m7, m7, m11
 %endif
     ;~ c += window[i + j + 32] * (synth_buf[16 + i + j])
     ;~ d += window[i + j + 48] * (synth_buf[31 - i + j])
-    SHUF          m6, [ptr2 + j + (31 - 3) * 4]
+    SHUF          m6,  ptr2 + j + (31 - 3) * 4, m5
     mova          m5, [ptr1 + j + 16 * 4]
 %if ARCH_X86_64
-    SHUF         m12, [ptr2 + j + (31 - 3) * 4 - mmsize]
+    SHUF         m12,  ptr2 + j + (31 - 3) * 4 - mmsize, m11
     mova         m11, [ptr1 + j + mmsize + 16 * 4]
 %endif
-    mulps         m5, [win  + %1 + j + 32 * 4]
-    mulps         m6, [win  + %1 + j + 48 * 4]
+    mulps         m5, m5,  [win + %1 + j + 32 * 4]
+    mulps         m6, m6,  [win + %1 + j + 48 * 4]
 %if ARCH_X86_64
-    mulps        m11, [win  + %1 + j + mmsize + 32 * 4]
-    mulps        m12, [win  + %1 + j + mmsize + 48 * 4]
+    mulps        m11, m11, [win + %1 + j + mmsize + 32 * 4]
+    mulps        m12, m12, [win + %1 + j + mmsize + 48 * 4]
 %endif
-    addps         m3, m5
-    addps         m4, m6
+    addps         m3, m3, m5
+    addps         m4, m4, m6
 %if ARCH_X86_64
-    addps         m9, m11
-    addps        m10, m12
+    addps         m9, m9, m11
+    addps        m10, m10, m12
 %endif
     sub            j, 64 * 4
 %endmacro
@@ -269,25 +273,34 @@ cglobal synth_filter_inner, 0, 6 + 4 * ARCH_X86_64, 7 + 6 * ARCH_X86_64, \
                               synth_buf, synth_buf2, window, out, off, scale
 %define scale m0
 %if ARCH_X86_32 || WIN64
-%if cpuflag(sse2)
+%if cpuflag(sse2) && notcpuflag(avx)
     movd       scale, scalem
+    SPLATD        m0
 %else
-    movss      scale, scalem
+    VBROADCASTSS  m0, scalem
 %endif
 ; Make sure offset is in a register and not on the stack
 %define OFFQ  r4q
 %else
+    SPLATD      xmm0
+%if cpuflag(avx)
+    vinsertf128   m0, m0, xmm0, 1
+%endif
 %define OFFQ  offq
 %endif
-    SPLATD        m0
     ; prepare inner counter limit 1
     mov          r5q, 480
     sub          r5q, offmp
     and          r5q, -64
     shl          r5q, 2
+%if ARCH_X86_32 || notcpuflag(avx)
     mov         OFFQ, r5q
 %define i        r5q
     mov            i, 16 * 4 - (ARCH_X86_64 + 1) * mmsize  ; main loop counter
+%else
+%define i 0
+%define OFFQ  r5q
+%endif
 
 %define buf2     synth_buf2q
 %if ARCH_X86_32
@@ -306,8 +319,10 @@ cglobal synth_filter_inner, 0, 6 + 4 * ARCH_X86_64, 7 + 6 * ARCH_X86_64, \
 %define j        r3q
     mov          win, windowm
     mov         ptr1, synth_bufm
+%if ARCH_X86_32 || notcpuflag(avx)
     add          win, i
     add         ptr1, i
+%endif
 %else ; ARCH_X86_64
 %define ptr1     r6q
 %define ptr2     r7q ; must be loaded
@@ -323,7 +338,9 @@ cglobal synth_filter_inner, 0, 6 + 4 * ARCH_X86_64, 7 + 6 * ARCH_X86_64, \
     mov         ptr2, synth_bufmp
     ; prepare the inner loop counter
     mov            j, OFFQ
+%if ARCH_X86_32 || notcpuflag(avx)
     sub         ptr2, i
+%endif
 .loop1:
     INNER_LOOP  0
     jge       .loop1
@@ -346,11 +363,11 @@ cglobal synth_filter_inner, 0, 6 + 4 * ARCH_X86_64, 7 + 6 * ARCH_X86_64, \
 %endif
     ;~ out[i]      = a * scale;
     ;~ out[i + 16] = b * scale;
-    mulps         m1, scale
-    mulps         m2, scale
+    mulps         m1, m1, scale
+    mulps         m2, m2, scale
 %if ARCH_X86_64
-    mulps         m7, scale
-    mulps         m8, scale
+    mulps         m7, m7, scale
+    mulps         m8, m8, scale
 %endif
     ;~ synth_buf2[i]      = c;
     ;~ synth_buf2[i + 16] = d;
@@ -368,8 +385,10 @@ cglobal synth_filter_inner, 0, 6 + 4 * ARCH_X86_64, 7 + 6 * ARCH_X86_64, \
     mova   [outq + i +  0 * 4 + mmsize], m7
     mova   [outq + i + 16 * 4 + mmsize], m8
 %endif
+%if ARCH_X86_32 || notcpuflag(avx)
     sub            i, (ARCH_X86_64 + 1) * mmsize
     jge    .mainloop
+%endif
     RET
 %endmacro
 
@@ -379,3 +398,5 @@ SYNTH_FILTER
 %endif
 INIT_XMM sse2
 SYNTH_FILTER
+INIT_YMM avx
+SYNTH_FILTER
diff --git a/libavcodec/x86/dcadsp_init.c b/libavcodec/x86/dcadsp_init.c
index 5b77985569..d7e0d65917 100644
--- a/libavcodec/x86/dcadsp_init.c
+++ b/libavcodec/x86/dcadsp_init.c
@@ -81,6 +81,7 @@ static void synth_filter_##opt(FFTContext *imdct,                              \
 SYNTH_FILTER_FUNC(sse)
 #endif
 SYNTH_FILTER_FUNC(sse2)
+SYNTH_FILTER_FUNC(avx)
 #endif /* HAVE_YASM */
 
 av_cold void ff_synth_filter_init_x86(SynthFilterContext *s)
@@ -96,5 +97,8 @@ av_cold void ff_synth_filter_init_x86(SynthFilterContext *s)
     if (EXTERNAL_SSE2(cpu_flags)) {
         s->synth_filter_float = synth_filter_sse2;
     }
+    if (EXTERNAL_AVX(cpu_flags)) {
+        s->synth_filter_float = synth_filter_avx;
+    }
 #endif /* HAVE_YASM */
 }