simd int->float

20% faster ac3 if downmixing, 15% if not

Originally committed as revision 14743 to svn://svn.ffmpeg.org/ffmpeg/trunk

libavcodec/ac3dec.c

@@ -994,9 +994,7 @@ static int decode_audio_block(AC3DecodeContext *s, int blk)
         } else {
             gain *= s->dynamic_range[0];
         }
-        for(i=0; i<256; i++) {
-            s->transform_coeffs[ch][i] = s->fixed_coeffs[ch][i] * gain;
-        }
+        s->dsp.int32_to_float_fmul_scalar(s->transform_coeffs[ch], s->fixed_coeffs[ch], gain, 256);
     }
 
     /* downmix and MDCT. order depends on whether block switching is used for

libavcodec/ac3dec.h

@@ -158,7 +158,7 @@ typedef struct {
     float mul_bias;                         ///< scaling for float_to_int16 conversion
 ///@}
 
-    int fixed_coeffs[AC3_MAX_CHANNELS][AC3_MAX_COEFS];  ///> fixed-point transform coefficients
+    DECLARE_ALIGNED_16(int, fixed_coeffs[AC3_MAX_CHANNELS][AC3_MAX_COEFS]);  ///> fixed-point transform coefficients
 
 ///@defgroup arrays aligned arrays
     DECLARE_ALIGNED_16(float, transform_coeffs[AC3_MAX_CHANNELS][AC3_MAX_COEFS]);   ///< transform coefficients

libavcodec/dsputil.c

@@ -3948,6 +3948,12 @@ void ff_vector_fmul_window_c(float *dst, const float *src0, const float *src1, c
     }
 }
 
+static void int32_to_float_fmul_scalar_c(float *dst, const int *src, float mul, int len){
+    int i;
+    for(i=0; i<len; i++)
+        dst[i] = src[i] * mul;
+}
+
 static av_always_inline int float_to_int16_one(const float *src){
     int_fast32_t tmp = *(const int32_t*)src;
     if(tmp & 0xf0000){
@@ -4489,6 +4495,7 @@ void dsputil_init(DSPContext* c, AVCodecContext *avctx)
     c->vector_fmul_reverse = vector_fmul_reverse_c;
     c->vector_fmul_add_add = ff_vector_fmul_add_add_c;
     c->vector_fmul_window = ff_vector_fmul_window_c;
+    c->int32_to_float_fmul_scalar = int32_to_float_fmul_scalar_c;
     c->float_to_int16 = ff_float_to_int16_c;
     c->float_to_int16_interleave = ff_float_to_int16_interleave_c;
     c->add_int16 = add_int16_c;

libavcodec/dsputil.h

@@ -370,6 +370,8 @@ typedef struct DSPContext {
     void (*vector_fmul_add_add)(float *dst, const float *src0, const float *src1, const float *src2, int src3, int len, int step);
     /* assume len is a multiple of 4, and arrays are 16-byte aligned */
     void (*vector_fmul_window)(float *dst, const float *src0, const float *src1, const float *win, float add_bias, int len);
+    /* assume len is a multiple of 8, and arrays are 16-byte aligned */
+    void (*int32_to_float_fmul_scalar)(float *dst, const int *src, float mul, int len);
 
     /* C version: convert floats from the range [384.0,386.0] to ints in [-32768,32767]
      * simd versions: convert floats from [-32768.0,32767.0] without rescaling and arrays are 16byte aligned */

libavcodec/i386/dsputil_mmx.c

@@ -2192,6 +2192,50 @@ static void vector_fmul_window_sse(float *dst, const float *src0, const float *s
         ff_vector_fmul_window_c(dst, src0, src1, win, add_bias, len);
 }
 
+static void int32_to_float_fmul_scalar_sse(float *dst, const int *src, float mul, int len)
+{
+    x86_reg i = -4*len;
+    asm volatile(
+        "movss  %3, %%xmm4 \n"
+        "shufps $0, %%xmm4, %%xmm4 \n"
+        "1: \n"
+        "cvtpi2ps   (%2,%0), %%xmm0 \n"
+        "cvtpi2ps  8(%2,%0), %%xmm1 \n"
+        "cvtpi2ps 16(%2,%0), %%xmm2 \n"
+        "cvtpi2ps 24(%2,%0), %%xmm3 \n"
+        "movlhps  %%xmm1,    %%xmm0 \n"
+        "movlhps  %%xmm3,    %%xmm2 \n"
+        "mulps    %%xmm4,    %%xmm0 \n"
+        "mulps    %%xmm4,    %%xmm2 \n"
+        "movaps   %%xmm0,   (%1,%0) \n"
+        "movaps   %%xmm2, 16(%1,%0) \n"
+        "add $32, %0 \n"
+        "jl 1b \n"
+        :"+r"(i)
+        :"r"(dst+len), "r"(src+len), "xm"(mul)
+    );
+}
+
+static void int32_to_float_fmul_scalar_sse2(float *dst, const int *src, float mul, int len)
+{
+    x86_reg i = -4*len;
+    asm volatile(
+        "movss  %3, %%xmm4 \n"
+        "shufps $0, %%xmm4, %%xmm4 \n"
+        "1: \n"
+        "cvtdq2ps   (%2,%0), %%xmm0 \n"
+        "cvtdq2ps 16(%2,%0), %%xmm1 \n"
+        "mulps    %%xmm4,    %%xmm0 \n"
+        "mulps    %%xmm4,    %%xmm1 \n"
+        "movaps   %%xmm0,   (%1,%0) \n"
+        "movaps   %%xmm1, 16(%1,%0) \n"
+        "add $32, %0 \n"
+        "jl 1b \n"
+        :"+r"(i)
+        :"r"(dst+len), "r"(src+len), "xm"(mul)
+    );
+}
+
 static void float_to_int16_3dnow(int16_t *dst, const float *src, long len){
     // not bit-exact: pf2id uses different rounding than C and SSE
     asm volatile(
@@ -2786,12 +2830,14 @@ void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx)
             c->vector_fmul_reverse = vector_fmul_reverse_sse;
             c->vector_fmul_add_add = vector_fmul_add_add_sse;
             c->vector_fmul_window = vector_fmul_window_sse;
+            c->int32_to_float_fmul_scalar = int32_to_float_fmul_scalar_sse;
             c->float_to_int16 = float_to_int16_sse;
             c->float_to_int16_interleave = float_to_int16_interleave_sse;
         }
         if(mm_flags & MM_3DNOW)
             c->vector_fmul_add_add = vector_fmul_add_add_3dnow; // faster than sse
         if(mm_flags & MM_SSE2){
+            c->int32_to_float_fmul_scalar = int32_to_float_fmul_scalar_sse2;
             c->float_to_int16 = float_to_int16_sse2;
             c->float_to_int16_interleave = float_to_int16_interleave_sse2;
             c->add_int16 = add_int16_sse2;