3 files changed, 12 insertions, 4 deletions
diff --git a/ffmpeg/libswresample/x86/audio_convert.asm b/ffmpeg/libswresample/x86/audio_convert.asm
index ad46977..b6e9e5d 100644
--- a/ffmpeg/libswresample/x86/audio_convert.asm
+++ b/ffmpeg/libswresample/x86/audio_convert.asm
@@ -20,8 +20,7 @@
 
 %include "libavutil/x86/x86util.asm"
 
-SECTION_RODATA
-align 32
+SECTION_RODATA 32
 flt2pm31: times 8 dd 4.6566129e-10
 flt2p31 : times 8 dd 2147483648.0
 flt2p15 : times 8 dd 32768.0
@@ -195,7 +194,12 @@ cglobal %2_to_%1_%3, 3, 3, 6, dst, src, len
     add lenq, 2*mmsize/(1<<%4)
 %endif
         jl .next
+%if mmsize == 8
+    emms
+    RET
+%else
     REP_RET
+%endif
 %endmacro
 
 %macro PACK_6CH 5-7
diff --git a/ffmpeg/libswresample/x86/rematrix.asm b/ffmpeg/libswresample/x86/rematrix.asm
index 84448e8..f0ae959 100644
--- a/ffmpeg/libswresample/x86/rematrix.asm
+++ b/ffmpeg/libswresample/x86/rematrix.asm
@@ -21,8 +21,7 @@
 %include "libavutil/x86/x86util.asm"
 
 
-SECTION_RODATA
-align 32
+SECTION_RODATA 32
 dw1: times 8  dd 1
 w1 : times 16 dw 1
 
diff --git a/ffmpeg/libswresample/x86/swresample_x86.c b/ffmpeg/libswresample/x86/swresample_x86.c
index e18f0c5..581dc17 100644
--- a/ffmpeg/libswresample/x86/swresample_x86.c
+++ b/ffmpeg/libswresample/x86/swresample_x86.c
@@ -169,6 +169,7 @@ av_cold void swri_rematrix_init_x86(struct SwrContext *s){
             s->mix_2_1_simd = ff_mix_2_1_a_int16_sse2;
         }
         s->native_simd_matrix = av_mallocz(2 * num * sizeof(int16_t));
+        s->native_simd_one    = av_mallocz(2 * sizeof(int16_t));
         for(i=0; i<nb_out; i++){
             int sh = 0;
             for(j=0; j<nb_in; j++)
@@ -180,6 +181,8 @@ av_cold void swri_rematrix_init_x86(struct SwrContext *s){
                     ((((int*)s->native_matrix)[i * nb_in + j]) + (1<<sh>>1)) >> sh;
             }
         }
+        ((int16_t*)s->native_simd_one)[1] = 14;
+        ((int16_t*)s->native_simd_one)[0] = 16384;
     } else if(s->midbuf.fmt == AV_SAMPLE_FMT_FLTP){
         if(mm_flags & AV_CPU_FLAG_SSE) {
             s->mix_1_1_simd = ff_mix_1_1_a_float_sse;
@@ -191,5 +194,7 @@ av_cold void swri_rematrix_init_x86(struct SwrContext *s){
         }
         s->native_simd_matrix = av_mallocz(num * sizeof(float));
         memcpy(s->native_simd_matrix, s->native_matrix, num * sizeof(float));
+        s->native_simd_one = av_mallocz(sizeof(float));
+        memcpy(s->native_simd_one, s->native_one, sizeof(float));
     }
 }