x86/aacpsdsp: add ps_hybrid_analysis_fma3

This replace the sse3 version, which was not really faster than the sse one.

Signed-off-by: James Almer <jamrial@gmail.com>

2 files changed, 25 insertions, 23 deletions

diff --git a/libavcodec/x86/aacpsdsp.asm b/libavcodec/x86/aacpsdsp.asm
index 543d33e68d..105e1af5c5 100644
--- a/libavcodec/x86/aacpsdsp.asm
+++ b/libavcodec/x86/aacpsdsp.asm
@@ -403,10 +403,8 @@ HYBRID_SYNTHESIS_DEINT
 %macro PS_HYBRID_ANALYSIS_IN 1
     movu     m0, [inq+mmsize*%1]
     movu     m1, [inq+mmsize*(5-%1)+8]
-    mova     m3, m0
-    mova     m4, m1
-    shufps   m3, m3, q2301
-    shufps   m4, m4, q0123
+    shufps   m3, m0, m0, q2301
+    shufps   m4, m1, m1, q0123
     shufps   m1, m1, q1032
 %if cpuflag(sse3)
     addsubps m3, m4
@@ -424,6 +422,15 @@ HYBRID_SYNTHESIS_DEINT
 %macro PS_HYBRID_ANALYSIS_LOOP 3
     mova     m2, [filterq+nq+mmsize*%3]
     shufps   m2, m2, q2301
+%if cpuflag(fma3)
+%if %3
+    fmaddps  m3, m2, [rsp+mmsize*%3*2], m3
+    fmaddps  m0, m2, [rsp+mmsize+mmsize*%3*2], m0
+%else
+    mulps    m3, m2, [rsp]
+    mulps    m0, m2, [rsp+mmsize]
+%endif
+%else ; cpuflag(sse)
     mova     %2, [rsp+mmsize*%3*2]
     mova     %1, [rsp+mmsize+mmsize*%3*2]
     mulps    %2, m2
@@ -432,20 +439,21 @@ HYBRID_SYNTHESIS_DEINT
     addps    m3, %2
     addps    m0, %1
 %endif
+%endif
 %endmacro
 
 %macro PS_HYBRID_ANALYSIS 0
-cglobal ps_hybrid_analysis, 5, 5, 8, 24 * 4, out, in, filter, stride, n
+cglobal ps_hybrid_analysis, 5, 5, 5 + notcpuflag(fma3) * 3, 24 * 4, out, in, filter, stride, n
 %if cpuflag(sse3)
 %define MOVH movsd
 %else
 %define MOVH movlps
+    mova m7, [ps_p1m1p1m1]
 %endif
     shl strideq, 3
     shl nd, 6
     add filterq, nq
     neg nq
-    mova m7, [ps_p1m1p1m1]
     PS_HYBRID_ANALYSIS_IN 0
     PS_HYBRID_ANALYSIS_IN 1
     PS_HYBRID_ANALYSIS_IN 2
@@ -456,30 +464,22 @@ align 16
     PS_HYBRID_ANALYSIS_LOOP m5, m6, 1
     PS_HYBRID_ANALYSIS_LOOP m5, m6, 2
 
-%if cpuflag(sse3)
-    pshufd   m3, m3, q2301
-    xorps    m0, m7
-    hsubps   m3, m0
-    pshufd   m1, m3, q0020
-    pshufd   m3, m3, q0031
-    addps    m1, m3
-    movsd    m2, [inq+6*8]
-%else
-    mova     m1, m3
-    mova     m2, m0
-    shufps   m1, m1, q2301
-    shufps   m2, m2, q2301
+    shufps   m1, m3, m3, q2301
+    shufps   m2, m0, m0, q2301
     subps    m1, m3
     addps    m2, m0
     unpcklps m3, m1, m2
     unpckhps m1, m2
     addps    m1, m3
     movu     m2, [inq+6*8] ; faster than movlps and no risk of overread
-%endif
     movss    m3, [filterq+nq+8*6]
     SPLATD   m3
+%if cpuflag(fma3)
+    fmaddps  m1, m2, m3, m1
+%else
     mulps    m2, m3
     addps    m1, m2
+%endif
     MOVH [outq], m1
     add    outq, strideq
     add      nq, 64
@@ -489,5 +489,5 @@ align 16
 
 INIT_XMM sse
 PS_HYBRID_ANALYSIS
-INIT_XMM sse3
+INIT_XMM fma3
 PS_HYBRID_ANALYSIS
diff --git a/libavcodec/x86/aacpsdsp_init.c b/libavcodec/x86/aacpsdsp_init.c
index 21f00efa24..0b0ee07db4 100644
--- a/libavcodec/x86/aacpsdsp_init.c
+++ b/libavcodec/x86/aacpsdsp_init.c
@@ -33,7 +33,7 @@ void ff_ps_mul_pair_single_sse (float (*dst)[2], float (*src0)[2],
 void ff_ps_hybrid_analysis_sse (float (*out)[2], float (*in)[2],
                                 const float (*filter)[8][2],
                                 ptrdiff_t stride, int n);
-void ff_ps_hybrid_analysis_sse3(float (*out)[2], float (*in)[2],
+void ff_ps_hybrid_analysis_fma3(float (*out)[2], float (*in)[2],
                                 const float (*filter)[8][2],
                                 ptrdiff_t stride, int n);
 void ff_ps_stereo_interpolate_sse3(float (*l)[2], float (*r)[2],
@@ -64,9 +64,11 @@ av_cold void ff_psdsp_init_x86(PSDSPContext *s)
         s->add_squares            = ff_ps_add_squares_sse3;
         s->stereo_interpolate[0]  = ff_ps_stereo_interpolate_sse3;
         s->stereo_interpolate[1]  = ff_ps_stereo_interpolate_ipdopd_sse3;
-        s->hybrid_analysis        = ff_ps_hybrid_analysis_sse3;
     }
     if (EXTERNAL_SSE4(cpu_flags)) {
         s->hybrid_synthesis_deint = ff_ps_hybrid_synthesis_deint_sse4;
     }
+    if (EXTERNAL_FMA3(cpu_flags)) {
+        s->hybrid_analysis        = ff_ps_hybrid_analysis_fma3;
+    }
 }