x86/fmtconvert: add ff_int32_to_float_fmul_array8_{sse,sse2}

About two times faster than the c wrapper.
Reviewed-by: Michael Niedermayer <michaelni@gmx.at>
Signed-off-by: James Almer <jamrial@gmail.com>

libavcodec/x86/fmtconvert.asm

@@ -77,6 +77,44 @@ INT32_TO_FLOAT_FMUL_SCALAR 5
 INIT_XMM sse2
 INT32_TO_FLOAT_FMUL_SCALAR 3
 
+;------------------------------------------------------------------------------
+; void ff_int32_to_float_fmul_array8(FmtConvertContext *c, float *dst, const int32_t *src,
+;                                    const float *mul, int len);
+;------------------------------------------------------------------------------
+%macro INT32_TO_FLOAT_FMUL_ARRAY8 0
+cglobal int32_to_float_fmul_array8, 5, 5, 5, c, dst, src, mul, len
+    shl     lenq, 2
+    add     srcq, lenq
+    add     dstq, lenq
+    neg     lenq
+.loop:
+    movss     m0, [mulq]
+    SPLATD    m0
+%if cpuflag(sse2)
+    cvtdq2ps  m1, [srcq+lenq   ]
+    cvtdq2ps  m2, [srcq+lenq+16]
+%else
+    cvtpi2ps  m1, [srcq+lenq   ]
+    cvtpi2ps  m3, [srcq+lenq+ 8]
+    cvtpi2ps  m2, [srcq+lenq+16]
+    cvtpi2ps  m4, [srcq+lenq+24]
+    movlhps   m1, m3
+    movlhps   m2, m4
+%endif
+    mulps     m1, m0
+    mulps     m2, m0
+    mova  [dstq+lenq   ], m1
+    mova  [dstq+lenq+16], m2
+    add     mulq, 4
+    add     lenq, 32
+    jl .loop
+    REP_RET
+%endmacro
+
+INIT_XMM sse
+INT32_TO_FLOAT_FMUL_ARRAY8
+INIT_XMM sse2
+INT32_TO_FLOAT_FMUL_ARRAY8
 
 ;------------------------------------------------------------------------------
 ; void ff_float_to_int16(int16_t *dst, const float *src, long len);

libavcodec/x86/fmtconvert_init.c

@@ -32,6 +32,10 @@
 
 void ff_int32_to_float_fmul_scalar_sse (float *dst, const int32_t *src, float mul, int len);
 void ff_int32_to_float_fmul_scalar_sse2(float *dst, const int32_t *src, float mul, int len);
+void ff_int32_to_float_fmul_array8_sse (FmtConvertContext *c, float *dst, const int32_t *src,
+                                        const float *mul, int len);
+void ff_int32_to_float_fmul_array8_sse2(FmtConvertContext *c, float *dst, const int32_t *src,
+                                        const float *mul, int len);
 
 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);
@@ -134,12 +138,14 @@ av_cold void ff_fmt_convert_init_x86(FmtConvertContext *c, AVCodecContext *avctx
     }
     if (EXTERNAL_SSE(cpu_flags)) {
         c->int32_to_float_fmul_scalar = ff_int32_to_float_fmul_scalar_sse;
+        c->int32_to_float_fmul_array8 = ff_int32_to_float_fmul_array8_sse;
         c->float_to_int16             = ff_float_to_int16_sse;
         c->float_to_int16_interleave  = float_to_int16_interleave_sse;
         c->float_interleave           = float_interleave_sse;
     }
     if (EXTERNAL_SSE2(cpu_flags)) {
         c->int32_to_float_fmul_scalar = ff_int32_to_float_fmul_scalar_sse2;
+        c->int32_to_float_fmul_array8 = ff_int32_to_float_fmul_array8_sse2;
         c->float_to_int16             = ff_float_to_int16_sse2;
         c->float_to_int16_interleave  = float_to_int16_interleave_sse2;
     }