x86: fmtconvert: add special asm for float_to_int16_interleave_misc_*

This gets rid of a variable-length array and a for loop in C code.

Signed-off-by: Martin Storsjö <martin@martin.st>

2 files changed, 85 insertions, 5 deletions

diff --git a/libavcodec/x86/fmtconvert.asm b/libavcodec/x86/fmtconvert.asm
index 63befc94f6..4916e7af33 100644
--- a/libavcodec/x86/fmtconvert.asm
+++ b/libavcodec/x86/fmtconvert.asm
@@ -115,6 +115,84 @@ FLOAT_TO_INT16 sse, 0
 FLOAT_TO_INT16 3dnow, 0
 %undef cvtps2pi
 
+;------------------------------------------------------------------------------
+; void ff_float_to_int16_step(int16_t *dst, const float *src, long len, long step);
+;------------------------------------------------------------------------------
+%macro FLOAT_TO_INT16_STEP 2
+cglobal float_to_int16_step_%1, 4,7,%2, dst, src, len, step, step3, v1, v2
+    add       lenq, lenq
+    lea       srcq, [srcq+2*lenq]
+    lea     step3q, [stepq*3]
+    neg       lenq
+.loop:
+%ifidn %1, sse2
+    cvtps2dq    m0, [srcq+2*lenq   ]
+    cvtps2dq    m1, [srcq+2*lenq+16]
+    packssdw    m0, m1
+    movd       v1d, m0
+    psrldq      m0, 4
+    movd       v2d, m0
+    psrldq      m0, 4
+    mov     [dstq], v1w
+    mov  [dstq+stepq*4], v2w
+    shr        v1d, 16
+    shr        v2d, 16
+    mov  [dstq+stepq*2], v1w
+    mov  [dstq+step3q*2], v2w
+    lea       dstq, [dstq+stepq*8]
+    movd       v1d, m0
+    psrldq      m0, 4
+    movd       v2d, m0
+    mov     [dstq], v1w
+    mov  [dstq+stepq*4], v2w
+    shr        v1d, 16
+    shr        v2d, 16
+    mov  [dstq+stepq*2], v1w
+    mov  [dstq+step3q*2], v2w
+    lea       dstq, [dstq+stepq*8]
+%else
+    cvtps2pi    m0, [srcq+2*lenq   ]
+    cvtps2pi    m1, [srcq+2*lenq+ 8]
+    cvtps2pi    m2, [srcq+2*lenq+16]
+    cvtps2pi    m3, [srcq+2*lenq+24]
+    packssdw    m0, m1
+    packssdw    m2, m3
+    movd       v1d, m0
+    psrlq       m0, 32
+    movd       v2d, m0
+    mov     [dstq], v1w
+    mov  [dstq+stepq*4], v2w
+    shr        v1d, 16
+    shr        v2d, 16
+    mov  [dstq+stepq*2], v1w
+    mov  [dstq+step3q*2], v2w
+    lea       dstq, [dstq+stepq*8]
+    movd       v1d, m2
+    psrlq       m2, 32
+    movd       v2d, m2
+    mov     [dstq], v1w
+    mov  [dstq+stepq*4], v2w
+    shr        v1d, 16
+    shr        v2d, 16
+    mov  [dstq+stepq*2], v1w
+    mov  [dstq+step3q*2], v2w
+    lea       dstq, [dstq+stepq*8]
+%endif
+    add       lenq, 16
+    js .loop
+%ifnidn %1, sse2
+    emms
+%endif
+    REP_RET
+%endmacro
+
+INIT_XMM
+FLOAT_TO_INT16_STEP sse2, 2
+INIT_MMX
+FLOAT_TO_INT16_STEP sse, 0
+%define cvtps2pi pf2id
+FLOAT_TO_INT16_STEP 3dnow, 0
+%undef cvtps2pi
 
 ;-------------------------------------------------------------------------------
 ; void ff_float_to_int16_interleave2(int16_t *dst, const float **src, long len);
diff --git a/libavcodec/x86/fmtconvert_mmx.c b/libavcodec/x86/fmtconvert_mmx.c
index 42cb0bc85b..aaf634d37f 100644
--- a/libavcodec/x86/fmtconvert_mmx.c
+++ b/libavcodec/x86/fmtconvert_mmx.c
@@ -25,6 +25,7 @@
 #include "libavutil/cpu.h"
 #include "libavutil/x86_cpu.h"
 #include "libavcodec/fmtconvert.h"
+#include "libavcodec/dsputil.h"
 
 #if HAVE_YASM
 
@@ -35,6 +36,10 @@ void ff_float_to_int16_3dnow(int16_t *dst, const float *src, long len);
 void ff_float_to_int16_sse  (int16_t *dst, const float *src, long len);
 void ff_float_to_int16_sse2 (int16_t *dst, const float *src, long len);
 
+void ff_float_to_int16_step_3dnow(int16_t *dst, const float *src, long len, long step);
+void ff_float_to_int16_step_sse  (int16_t *dst, const float *src, long len, long step);
+void ff_float_to_int16_step_sse2 (int16_t *dst, const float *src, long len, long step);
+
 void ff_float_to_int16_interleave2_3dnow(int16_t *dst, const float **src, long len);
 void ff_float_to_int16_interleave2_sse  (int16_t *dst, const float **src, long len);
 void ff_float_to_int16_interleave2_sse2 (int16_t *dst, const float **src, long len);
@@ -48,12 +53,9 @@ void ff_float_to_int16_interleave6_3dn2(int16_t *dst, const float **src, int len
 #define FLOAT_TO_INT16_INTERLEAVE(cpu) \
 /* gcc pessimizes register allocation if this is in the same function as float_to_int16_interleave_sse2*/\
 static av_noinline void float_to_int16_interleave_misc_##cpu(int16_t *dst, const float **src, long len, int channels){\
-    DECLARE_ALIGNED(16, int16_t, tmp)[len];\
-    int i,j,c;\
+    int c;\
     for(c=0; c<channels; c++){\
-        ff_float_to_int16_##cpu(tmp, src[c], len);\
-        for(i=0, j=c; i<len; i++, j+=channels)\
-            dst[j] = tmp[i];\
+        ff_float_to_int16_step_##cpu(dst+c, src[c], len, channels);\
     }\
 }\
 \