SSE optimized vector_clipf(). 10% faster TwinVQ decoding.

Originally committed as revision 19728 to svn://svn.ffmpeg.org/ffmpeg/trunk

4 files changed, 84 insertions, 3 deletions

diff --git a/libavcodec/dsputil.c b/libavcodec/dsputil.c
index 92d613a93d..758f613c11 100644
--- a/libavcodec/dsputil.c
+++ b/libavcodec/dsputil.c
@@ -4093,6 +4093,51 @@ static void int32_to_float_fmul_scalar_c(float *dst, const int *src, float mul,
         dst[i] = src[i] * mul;
 }
 
+static inline uint32_t clipf_c_one(uint32_t a, uint32_t mini,
+                   uint32_t maxi, uint32_t maxisign)
+{
+
+    if(a > mini) return mini;
+    else if((a^(1<<31)) > maxisign) return maxi;
+    else return a;
+}
+
+static void vector_clipf_c_opposite_sign(float *dst, float *src, float *min, float *max, int len){
+    int i;
+    uint32_t mini = *(uint32_t*)min;
+    uint32_t maxi = *(uint32_t*)max;
+    uint32_t maxisign = maxi ^ (1<<31);
+    uint32_t *dsti = (uint32_t*)dst;
+    uint32_t *srci = (uint32_t*)src;
+    for(i=0; i<len; i+=8) {
+        dsti[i + 0] = clipf_c_one(srci[i + 0], mini, maxi, maxisign);
+        dsti[i + 1] = clipf_c_one(srci[i + 1], mini, maxi, maxisign);
+        dsti[i + 2] = clipf_c_one(srci[i + 2], mini, maxi, maxisign);
+        dsti[i + 3] = clipf_c_one(srci[i + 3], mini, maxi, maxisign);
+        dsti[i + 4] = clipf_c_one(srci[i + 4], mini, maxi, maxisign);
+        dsti[i + 5] = clipf_c_one(srci[i + 5], mini, maxi, maxisign);
+        dsti[i + 6] = clipf_c_one(srci[i + 6], mini, maxi, maxisign);
+        dsti[i + 7] = clipf_c_one(srci[i + 7], mini, maxi, maxisign);
+    }
+}
+static void vector_clipf_c(float *dst, float *src, float min, float max, int len){
+    int i;
+    if(min < 0 && max > 0) {
+        vector_clipf_c_opposite_sign(dst, src, &min, &max, len);
+    } else {
+        for(i=0; i < len; i+=8) {
+            dst[i    ] = av_clipf(src[i    ], min, max);
+            dst[i + 1] = av_clipf(src[i + 1], min, max);
+            dst[i + 2] = av_clipf(src[i + 2], min, max);
+            dst[i + 3] = av_clipf(src[i + 3], min, max);
+            dst[i + 4] = av_clipf(src[i + 4], min, max);
+            dst[i + 5] = av_clipf(src[i + 5], min, max);
+            dst[i + 6] = av_clipf(src[i + 6], min, max);
+            dst[i + 7] = av_clipf(src[i + 7], min, max);
+        }
+    }
+}
+
 static av_always_inline int float_to_int16_one(const float *src){
     int_fast32_t tmp = *(const int32_t*)src;
     if(tmp & 0xf0000){
@@ -4669,6 +4714,7 @@ void dsputil_init(DSPContext* c, AVCodecContext *avctx)
     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->vector_clipf = vector_clipf_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;
diff --git a/libavcodec/dsputil.h b/libavcodec/dsputil.h
index 2d15bd3538..032dd1d43c 100644
--- a/libavcodec/dsputil.h
+++ b/libavcodec/dsputil.h
@@ -396,6 +396,7 @@ typedef struct DSPContext {
     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);
+    void (*vector_clipf)(float *dst /* align 16 */, float *src /* align 16 */, float min, float max, int len /* align 16 */);
 
     /* 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 */
diff --git a/libavcodec/twinvq.c b/libavcodec/twinvq.c
index 1c05238efc..ba38ddcc78 100644
--- a/libavcodec/twinvq.c
+++ b/libavcodec/twinvq.c
@@ -825,7 +825,6 @@ static int twin_decode_frame(AVCodecContext * avctx, void *data,
     float *out = data;
     enum FrameType ftype;
     int window_type;
-    int i;
     static const enum FrameType wtype_to_ftype_table[] = {
         FT_LONG,   FT_LONG, FT_SHORT, FT_LONG,
         FT_MEDIUM, FT_LONG, FT_LONG,  FT_MEDIUM, FT_MEDIUM
@@ -860,8 +859,8 @@ static int twin_decode_frame(AVCodecContext * avctx, void *data,
         return buf_size;
     }
 
-    for (i=0; i < avctx->channels * mtab->size; i++)
-        out[i] = av_clipf(out[i], -32700./(1<<15), 32700./(1<<15));
+    tctx->dsp.vector_clipf(out, out, -32700./(1<<15), 32700./(1<<15),
+                           mtab->size);
 
     *data_size = mtab->size*avctx->channels*4;
 
diff --git a/libavcodec/x86/dsputil_mmx.c b/libavcodec/x86/dsputil_mmx.c
index da0f76412b..374aa64107 100644
--- a/libavcodec/x86/dsputil_mmx.c
+++ b/libavcodec/x86/dsputil_mmx.c
@@ -2346,6 +2346,40 @@ static void int32_to_float_fmul_scalar_sse2(float *dst, const int *src, float mu
     );
 }
 
+static void vector_clipf_sse(float *dst, float *src, float min, float max,
+                             int len)
+{
+    x86_reg i = (len-16)*4;
+    __asm__ volatile(
+        "movss  %3, %%xmm4 \n"
+        "movss  %4, %%xmm5 \n"
+        "shufps $0, %%xmm4, %%xmm4 \n"
+        "shufps $0, %%xmm5, %%xmm5 \n"
+        "1: \n\t"
+        "movaps    (%2,%0), %%xmm0 \n\t" // 3/1 on intel
+        "movaps  16(%2,%0), %%xmm1 \n\t"
+        "movaps  32(%2,%0), %%xmm2 \n\t"
+        "movaps  48(%2,%0), %%xmm3 \n\t"
+        "maxps      %%xmm4, %%xmm0 \n\t"
+        "maxps      %%xmm4, %%xmm1 \n\t"
+        "maxps      %%xmm4, %%xmm2 \n\t"
+        "maxps      %%xmm4, %%xmm3 \n\t"
+        "minps      %%xmm5, %%xmm0 \n\t"
+        "minps      %%xmm5, %%xmm1 \n\t"
+        "minps      %%xmm5, %%xmm2 \n\t"
+        "minps      %%xmm5, %%xmm3 \n\t"
+        "movaps  %%xmm0,   (%1,%0) \n\t"
+        "movaps  %%xmm1, 16(%1,%0) \n\t"
+        "movaps  %%xmm2, 32(%1,%0) \n\t"
+        "movaps  %%xmm3, 48(%1,%0) \n\t"
+        "sub  $64, %0 \n\t"
+        "jge 1b \n\t"
+        :"+r"(i)
+        :"r"(dst), "r"(src), "m"(min), "m"(max)
+        :"memory"
+    );
+}
+
 static void float_to_int16_3dnow(int16_t *dst, const float *src, long len){
     x86_reg reglen = len;
     // not bit-exact: pf2id uses different rounding than C and SSE
@@ -3046,6 +3080,7 @@ void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx)
             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->vector_clipf = vector_clipf_sse;
             c->float_to_int16 = float_to_int16_sse;
             c->float_to_int16_interleave = float_to_int16_interleave_sse;
         }