[FFmpeg-devel] [PATCH] x86/synth_filter: improve FMA version

[James Almer]

Replace mulps+subps with fnmaddps, resulting in two less instructions inside the
inner loops.
About 1% faster FMA3 performance.

Signed-off-by: James Almer <jamrial at gmail.com>
---
 libavcodec/x86/dcadsp.asm | 42 +++++++++++++++++++++++++++++++++++-------
 1 file changed, 35 insertions(+), 7 deletions(-)

diff --git a/libavcodec/x86/dcadsp.asm b/libavcodec/x86/dcadsp.asm
index 4746784..c5a1e77 100644
--- a/libavcodec/x86/dcadsp.asm
+++ b/libavcodec/x86/dcadsp.asm
@@ -230,14 +230,27 @@ DCA_LFE_FIR 1
     SHUF         m11,  ptr2 + j + (15 - 3) * 4 - mmsize, m12
     mova         m12, [ptr1 + j + mmsize]
 %endif
-    FMULADD_PS    m2, m6,  [win + %1 + j + 16 * 4], m2, m6
+%if cpuflag(fma3)
+    fmaddps       m2, m6,  [win + %1 + j + 16 * 4], m2
+    fnmaddps      m1, m5,  [win + %1 + j], m1
+%if ARCH_X86_64
+    fmaddps       m8, m12, [win + %1 + j + mmsize + 16 * 4], m8
+    fnmaddps      m7, m11, [win + %1 + j + mmsize], m7
+%endif
+%else ; non-FMA
+    mulps         m6, m6,  [win + %1 + j + 16 * 4]
     mulps         m5, m5,  [win + %1 + j]
-    subps         m1, m1, m5
 %if ARCH_X86_64
-    FMULADD_PS    m8, m12, [win + %1 + j + mmsize + 16 * 4], m8, m12
+    mulps        m12, m12, [win + %1 + j + mmsize + 16 * 4]
     mulps        m11, m11, [win + %1 + j + mmsize]
+%endif
+    addps         m2, m2, m6
+    subps         m1, m1, m5
+%if ARCH_X86_64
+    addps         m8, m8, m12
     subps         m7, m7, m11
 %endif
+%endif ; cpuflag(fma3)
     ;~ 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, m5
@@ -246,12 +259,27 @@ DCA_LFE_FIR 1
     SHUF         m12,  ptr2 + j + (31 - 3) * 4 - mmsize, m11
     mova         m11, [ptr1 + j + mmsize + 16 * 4]
 %endif
-    FMULADD_PS    m3, m5,  [win + %1 + j + 32 * 4], m3, m5
-    FMULADD_PS    m4, m6,  [win + %1 + j + 48 * 4], m4, m6
+%if cpuflag(fma3)
+    fmaddps       m3, m5,  [win + %1 + j + 32 * 4], m3
+    fmaddps       m4, m6,  [win + %1 + j + 48 * 4], m4
+%if ARCH_X86_64
+    fmaddps       m9, m11, [win + %1 + j + mmsize + 32 * 4], m9
+    fmaddps      m10, m12, [win + %1 + j + mmsize + 48 * 4], m10
+%endif
+%else ; non-FMA
+    mulps         m5, m5,  [win + %1 + j + 32 * 4]
+    mulps         m6, m6,  [win + %1 + j + 48 * 4]
+%if ARCH_X86_64
+    mulps        m11, m11, [win + %1 + j + mmsize + 32 * 4]
+    mulps        m12, m12, [win + %1 + j + mmsize + 48 * 4]
+%endif
+    addps         m3, m3, m5
+    addps         m4, m4, m6
 %if ARCH_X86_64
-    FMULADD_PS    m9, m11, [win + %1 + j + mmsize + 32 * 4], m9, m11
-    FMULADD_PS   m10, m12, [win + %1 + j + mmsize + 48 * 4], m10, m12
+    addps         m9, m9, m11
+    addps        m10, m10, m12
 %endif
+%endif ; cpuflag(fma3)
     sub            j, 64 * 4
 %endmacro
 
-- 
1.8.3.2