[FFmpeg-devel] [PATCH] x86/swr: replace sse4 instructions in pack_6ch with sse ones

[James Almer]

There's no benefit from using blendps here except on CPUs with AVX, where
it's faster than shufps according to Intel's documentation.
As such, rename the sse4 functions to sse/sse2 and use shufps instead.

Signed-off-by: James Almer <jamrial at gmail.com>
---
 libswresample/x86/audio_convert.asm    | 31 +++++++++++++++++++++----------
 libswresample/x86/audio_convert_init.c | 23 ++++++++++++-----------
 2 files changed, 33 insertions(+), 21 deletions(-)

diff --git a/libswresample/x86/audio_convert.asm b/libswresample/x86/audio_convert.asm
index b6e9e5d..d77e934 100644
--- a/libswresample/x86/audio_convert.asm
+++ b/libswresample/x86/audio_convert.asm
@@ -245,15 +245,27 @@ pack_6ch_%2_to_%1_u_int %+ SUFFIX
     mov%3     m4, [srcq+src4q]
     mov%3     m5, [srcq+src5q]
     %7 x,x,x,x,m7,x
-%if cpuflag(sse4)
+%if cpuflag(sse)
     SBUTTERFLYPS 0, 1, 6
     SBUTTERFLYPS 2, 3, 6
     SBUTTERFLYPS 4, 5, 6
 
+%if cpuflag(avx)
     blendps   m6, m4, m0, 1100b
+%else
+    movaps    m6, m4
+    shufps    m4, m0, q3210
+    SWAP 4,6
+%endif
     movlhps   m0, m2
     movhlps   m4, m2
+%if cpuflag(avx)
     blendps   m2, m5, m1, 1100b
+%else
+    movaps    m2, m5
+    shufps    m5, m1, q3210
+    SWAP 2,5
+%endif
     movlhps   m1, m3
     movhlps   m5, m3
 
@@ -380,6 +392,10 @@ CONV int16, int32, a, 1, 2, INT32_TO_INT16_N, NOP_N
 PACK_6CH float, float, u, 2, 2, NOP_N, NOP_N
 PACK_6CH float, float, a, 2, 2, NOP_N, NOP_N
 
+INIT_XMM sse
+PACK_6CH float, float, u, 2, 2, NOP_N, NOP_N
+PACK_6CH float, float, a, 2, 2, NOP_N, NOP_N
+
 INIT_XMM sse2
 CONV int32, int16, u, 2, 1, INT16_TO_INT32_N, NOP_N
 CONV int32, int16, a, 2, 1, INT16_TO_INT32_N, NOP_N
@@ -431,6 +447,10 @@ UNPACK_2CH float, int16, a, 2, 1, INT16_TO_FLOAT_N, INT16_TO_FLOAT_INIT
 UNPACK_2CH int16, float, u, 1, 2, FLOAT_TO_INT16_N, FLOAT_TO_INT16_INIT
 UNPACK_2CH int16, float, a, 1, 2, FLOAT_TO_INT16_N, FLOAT_TO_INT16_INIT
 
+PACK_6CH float, int32, u, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT
+PACK_6CH float, int32, a, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT
+PACK_6CH int32, float, u, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT
+PACK_6CH int32, float, a, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT
 
 INIT_XMM ssse3
 UNPACK_2CH int16, int16, u, 1, 1, NOP_N, NOP_N
@@ -440,15 +460,6 @@ UNPACK_2CH int32, int16, a, 2, 1, INT16_TO_INT32_N, NOP_N
 UNPACK_2CH float, int16, u, 2, 1, INT16_TO_FLOAT_N, INT16_TO_FLOAT_INIT
 UNPACK_2CH float, int16, a, 2, 1, INT16_TO_FLOAT_N, INT16_TO_FLOAT_INIT
 
-INIT_XMM sse4
-PACK_6CH float, float, u, 2, 2, NOP_N, NOP_N
-PACK_6CH float, float, a, 2, 2, NOP_N, NOP_N
-
-PACK_6CH float, int32, u, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT
-PACK_6CH float, int32, a, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT
-PACK_6CH int32, float, u, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT
-PACK_6CH int32, float, a, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT
-
 %if HAVE_AVX_EXTERNAL
 INIT_XMM avx
 PACK_6CH float, float, u, 2, 2, NOP_N, NOP_N
diff --git a/libswresample/x86/audio_convert_init.c b/libswresample/x86/audio_convert_init.c
index a26cdf6..769575d 100644
--- a/libswresample/x86/audio_convert_init.c
+++ b/libswresample/x86/audio_convert_init.c
@@ -58,7 +58,12 @@ MULTI_CAPS_FUNC(SSE2, sse2)
                 ac->simd_f =  ff_pack_6ch_float_to_float_a_mmx;
         }
     }
-
+    if(EXTERNAL_SSE(mm_flags)) {
+        if(channels == 6) {
+            if(   out_fmt == AV_SAMPLE_FMT_FLT  && in_fmt == AV_SAMPLE_FMT_FLTP || out_fmt == AV_SAMPLE_FMT_S32 && in_fmt == AV_SAMPLE_FMT_S32P)
+                ac->simd_f =  ff_pack_6ch_float_to_float_a_sse;
+        }
+    }
     if(EXTERNAL_SSE2(mm_flags)) {
         if(   out_fmt == AV_SAMPLE_FMT_FLT  && in_fmt == AV_SAMPLE_FMT_S32 || out_fmt == AV_SAMPLE_FMT_FLTP && in_fmt == AV_SAMPLE_FMT_S32P)
             ac->simd_f =  ff_int32_to_float_a_sse2;
@@ -105,6 +110,12 @@ MULTI_CAPS_FUNC(SSE2, sse2)
             if(   out_fmt == AV_SAMPLE_FMT_S16P  && in_fmt == AV_SAMPLE_FMT_FLT)
                 ac->simd_f =  ff_unpack_2ch_float_to_int16_a_sse2;
         }
+        if(channels == 6) {
+            if(   out_fmt == AV_SAMPLE_FMT_FLT  && in_fmt == AV_SAMPLE_FMT_S32P)
+                ac->simd_f =  ff_pack_6ch_int32_to_float_a_sse2;
+            if(   out_fmt == AV_SAMPLE_FMT_S32  && in_fmt == AV_SAMPLE_FMT_FLTP)
+                ac->simd_f =  ff_pack_6ch_float_to_int32_a_sse2;
+        }
     }
     if(EXTERNAL_SSSE3(mm_flags)) {
         if(channels == 2) {
@@ -116,16 +127,6 @@ MULTI_CAPS_FUNC(SSE2, sse2)
                 ac->simd_f =  ff_unpack_2ch_int16_to_float_a_ssse3;
         }
     }
-    if(EXTERNAL_SSE4(mm_flags)) {
-        if(channels == 6) {
-            if(   out_fmt == AV_SAMPLE_FMT_FLT  && in_fmt == AV_SAMPLE_FMT_FLTP || out_fmt == AV_SAMPLE_FMT_S32 && in_fmt == AV_SAMPLE_FMT_S32P)
-                ac->simd_f =  ff_pack_6ch_float_to_float_a_sse4;
-            if(   out_fmt == AV_SAMPLE_FMT_FLT  && in_fmt == AV_SAMPLE_FMT_S32P)
-                ac->simd_f =  ff_pack_6ch_int32_to_float_a_sse4;
-            if(   out_fmt == AV_SAMPLE_FMT_S32  && in_fmt == AV_SAMPLE_FMT_FLTP)
-                ac->simd_f =  ff_pack_6ch_float_to_int32_a_sse4;
-        }
-    }
     if(EXTERNAL_AVX(mm_flags)) {
         if(   out_fmt == AV_SAMPLE_FMT_FLT  && in_fmt == AV_SAMPLE_FMT_S32 || out_fmt == AV_SAMPLE_FMT_FLTP && in_fmt == AV_SAMPLE_FMT_S32P)
             ac->simd_f =  ff_int32_to_float_a_avx;
-- 
2.0.4