sse2 version of compute_autocorr().

4x faster than c (somehow, even though doubles only allow 2x simd).
overal flac encoding: 15-50% faster on core2, 4-11% on k8, 3-13% on p4.

Originally committed as revision 10621 to svn://svn.ffmpeg.org/ffmpeg/trunk

libavcodec/dsputil.c

@@ -41,6 +41,9 @@ void ff_spatial_dwt(int *buffer, int width, int height, int stride, int type, in
 /* vorbis.c */
 void vorbis_inverse_coupling(float *mag, float *ang, int blocksize);
 
+/* flacenc.c */
+void ff_flac_compute_autocorr(const int32_t *data, int len, int lag, double *autoc);
+
 uint8_t ff_cropTbl[256 + 2 * MAX_NEG_CROP] = {0, };
 uint32_t ff_squareTbl[512] = {0, };
 
@@ -4131,6 +4134,9 @@ void dsputil_init(DSPContext* c, AVCodecContext *avctx)
 
 #ifdef CONFIG_VORBIS_DECODER
     c->vorbis_inverse_coupling = vorbis_inverse_coupling;
+#endif
+#ifdef CONFIG_FLAC_ENCODER
+    c->flac_compute_autocorr = ff_flac_compute_autocorr;
 #endif
     c->vector_fmul = vector_fmul_c;
     c->vector_fmul_reverse = vector_fmul_reverse_c;

libavcodec/dsputil.h

@@ -328,6 +328,8 @@ typedef struct DSPContext {
 
     /* assume len is a multiple of 4, and arrays are 16-byte aligned */
     void (*vorbis_inverse_coupling)(float *mag, float *ang, int blocksize);
+    /* no alignment needed */
+    void (*flac_compute_autocorr)(const int32_t *data, int len, int lag, double *autoc);
     /* assume len is a multiple of 8, and arrays are 16-byte aligned */
     void (*vector_fmul)(float *dst, const float *src, int len);
     void (*vector_fmul_reverse)(float *dst, const float *src0, const float *src1, int len);

libavcodec/flacenc.c

@@ -22,6 +22,7 @@
 #include "avcodec.h"
 #include "bitstream.h"
 #include "crc.h"
+#include "dsputil.h"
 #include "golomb.h"
 #include "lls.h"
 
@@ -107,6 +108,7 @@ typedef struct FlacEncodeContext {
     FlacFrame frame;
     CompressionOptions options;
     AVCodecContext *avctx;
+    DSPContext dsp;
 } FlacEncodeContext;
 
 static const int flac_samplerates[16] = {
@@ -177,6 +179,8 @@ static int flac_encode_init(AVCodecContext *avctx)
 
     s->avctx = avctx;
 
+    dsputil_init(&s->dsp, avctx);
+
     if(avctx->sample_fmt != SAMPLE_FMT_S16) {
         return -1;
     }
@@ -604,8 +608,8 @@ static void apply_welch_window(const int32_t *data, int len, double *w_data)
  * Calculates autocorrelation data from audio samples
  * A Welch window function is applied before calculation.
  */
-static void compute_autocorr(const int32_t *data, int len, int lag,
+void ff_flac_compute_autocorr(const int32_t *data, int len, int lag,
-                             double *autoc)
+                              double *autoc)
 {
     int i, j;
     double tmp[len + lag + 1];
@@ -747,7 +751,8 @@ static int estimate_best_order(double *ref, int max_order)
 /**
  * Calculate LPC coefficients for multiple orders
  */
-static int lpc_calc_coefs(const int32_t *samples, int blocksize, int max_order,
+static int lpc_calc_coefs(FlacEncodeContext *s,
+                          const int32_t *samples, int blocksize, int max_order,
                           int precision, int32_t coefs[][MAX_LPC_ORDER],
                           int *shift, int use_lpc, int omethod)
 {
@@ -760,7 +765,7 @@ static int lpc_calc_coefs(const int32_t *samples, int blocksize, int max_order,
     assert(max_order >= MIN_LPC_ORDER && max_order <= MAX_LPC_ORDER);
 
     if(use_lpc == 1){
-        compute_autocorr(samples, blocksize, max_order, autoc);
+        s->dsp.flac_compute_autocorr(samples, blocksize, max_order, autoc);
 
         compute_lpc_coefs(autoc, max_order, lpc, ref);
     }else{
@@ -1017,7 +1022,7 @@ static int encode_residual(FlacEncodeContext *ctx, int ch)
     }
 
     /* LPC */
-    opt_order = lpc_calc_coefs(smp, n, max_order, precision, coefs, shift, ctx->options.use_lpc, omethod);
+    opt_order = lpc_calc_coefs(ctx, smp, n, max_order, precision, coefs, shift, ctx->options.use_lpc, omethod);
 
     if(omethod == ORDER_METHOD_2LEVEL ||
        omethod == ORDER_METHOD_4LEVEL ||

libavcodec/i386/dsputil_mmx.c

@@ -65,6 +65,9 @@ static const uint64_t ff_pb_A1 attribute_used __attribute__ ((aligned(8))) = 0xA
 static const uint64_t ff_pb_5F attribute_used __attribute__ ((aligned(8))) = 0x5F5F5F5F5F5F5F5FULL;
 static const uint64_t ff_pb_FC attribute_used __attribute__ ((aligned(8))) = 0xFCFCFCFCFCFCFCFCULL;
 
+static const double ff_pd_1[2] attribute_used __attribute__ ((aligned(16))) = { 1.0, 1.0 };
+static const double ff_pd_2[2] attribute_used __attribute__ ((aligned(16))) = { 2.0, 2.0 };
+
 #define JUMPALIGN() __asm __volatile (ASMALIGN(3)::)
 #define MOVQ_ZERO(regd)  __asm __volatile ("pxor %%" #regd ", %%" #regd ::)
 
@@ -2958,6 +2961,125 @@ static void vorbis_inverse_coupling_sse(float *mag, float *ang, int blocksize)
     }
 }
 
+#ifdef CONFIG_ENCODERS
+static void apply_welch_window_sse2(const int32_t *data, int len, double *w_data)
+{
+    double c = 2.0 / (len-1.0);
+    int n2 = len>>1;
+    long i = -n2*sizeof(int32_t);
+    long j =  n2*sizeof(int32_t);
+    asm volatile(
+        "movsd   %0,     %%xmm7 \n\t"
+        "movapd  %1,     %%xmm6 \n\t"
+        "movapd  %2,     %%xmm5 \n\t"
+        "movlhps %%xmm7, %%xmm7 \n\t"
+        "subpd   %%xmm5, %%xmm7 \n\t"
+        "addsd   %%xmm6, %%xmm7 \n\t"
+        ::"m"(c), "m"(*ff_pd_1), "m"(*ff_pd_2)
+    );
+#define WELCH(MOVPD)\
+    asm volatile(\
+        "1:                         \n\t"\
+        "movapd   %%xmm7,  %%xmm1   \n\t"\
+        "mulpd    %%xmm1,  %%xmm1   \n\t"\
+        "movapd   %%xmm6,  %%xmm0   \n\t"\
+        "subpd    %%xmm1,  %%xmm0   \n\t"\
+        "pshufd   $0x4e,   %%xmm0, %%xmm1 \n\t"\
+        "cvtpi2pd (%4,%0), %%xmm2   \n\t"\
+        "cvtpi2pd (%5,%1), %%xmm3   \n\t"\
+        "mulpd    %%xmm0,  %%xmm2   \n\t"\
+        "mulpd    %%xmm1,  %%xmm3   \n\t"\
+        "movapd   %%xmm2, (%2,%0,2) \n\t"\
+        MOVPD"    %%xmm3, (%3,%1,2) \n\t"\
+        "subpd    %%xmm5,  %%xmm7   \n\t"\
+        "sub      $8,      %1       \n\t"\
+        "add      $8,      %0       \n\t"\
+        "jl 1b                      \n\t"\
+        :"+&r"(i), "+&r"(j)\
+        :"r"(w_data+n2), "r"(w_data+len-2-n2),\
+         "r"(data+n2), "r"(data+len-2-n2)\
+    );
+    if(len&1)
+        WELCH("movupd")
+    else
+        WELCH("movapd")
+#undef WELCH
+}
+
+static void flac_compute_autocorr_sse2(const int32_t *data, int len, int lag,
+                                       double *autoc)
+{
+    double tmp[len + lag + 2];
+    double *data1 = tmp + lag;
+    int j;
+
+    if((long)data1 & 15)
+        data1++;
+
+    apply_welch_window_sse2(data, len, data1);
+
+    for(j=0; j<lag; j++)
+        data1[j-lag]= 0.0;
+    data1[len] = 0.0;
+
+    for(j=0; j<lag; j+=2){
+        long i = -len*sizeof(double);
+        if(j == lag-2) {
+            asm volatile(
+                "movsd     %6,     %%xmm0 \n\t"
+                "movsd     %6,     %%xmm1 \n\t"
+                "movsd     %6,     %%xmm2 \n\t"
+                "1:                       \n\t"
+                "movapd   (%4,%0), %%xmm3 \n\t"
+                "movupd -8(%5,%0), %%xmm4 \n\t"
+                "movapd   (%5,%0), %%xmm5 \n\t"
+                "mulpd     %%xmm3, %%xmm4 \n\t"
+                "mulpd     %%xmm3, %%xmm5 \n\t"
+                "mulpd -16(%5,%0), %%xmm3 \n\t"
+                "addpd     %%xmm4, %%xmm1 \n\t"
+                "addpd     %%xmm5, %%xmm0 \n\t"
+                "addpd     %%xmm3, %%xmm2 \n\t"
+                "add       $16,    %0     \n\t"
+                "jl 1b                    \n\t"
+                "movhlps   %%xmm0, %%xmm3 \n\t"
+                "movhlps   %%xmm1, %%xmm4 \n\t"
+                "movhlps   %%xmm2, %%xmm5 \n\t"
+                "addsd     %%xmm3, %%xmm0 \n\t"
+                "addsd     %%xmm4, %%xmm1 \n\t"
+                "addsd     %%xmm5, %%xmm2 \n\t"
+                "movsd     %%xmm0, %1     \n\t"
+                "movsd     %%xmm1, %2     \n\t"
+                "movsd     %%xmm2, %3     \n\t"
+                :"+&r"(i), "=m"(autoc[j]), "=m"(autoc[j+1]), "=m"(autoc[j+2])
+                :"r"(data1+len), "r"(data1+len-j), "m"(*ff_pd_1)
+            );
+        } else {
+            asm volatile(
+                "movsd     %5,     %%xmm0 \n\t"
+                "movsd     %5,     %%xmm1 \n\t"
+                "1:                       \n\t"
+                "movapd   (%3,%0), %%xmm3 \n\t"
+                "movupd -8(%4,%0), %%xmm4 \n\t"
+                "mulpd     %%xmm3, %%xmm4 \n\t"
+                "mulpd    (%4,%0), %%xmm3 \n\t"
+                "addpd     %%xmm4, %%xmm1 \n\t"
+                "addpd     %%xmm3, %%xmm0 \n\t"
+                "add       $16,    %0     \n\t"
+                "jl 1b                    \n\t"
+                "movhlps   %%xmm0, %%xmm3 \n\t"
+                "movhlps   %%xmm1, %%xmm4 \n\t"
+                "addsd     %%xmm3, %%xmm0 \n\t"
+                "addsd     %%xmm4, %%xmm1 \n\t"
+                "movsd     %%xmm0, %1     \n\t"
+                "movsd     %%xmm1, %2     \n\t"
+                :"+&r"(i), "=m"(autoc[j]), "=m"(autoc[j+1])
+                :"r"(data1+len), "r"(data1+len-j), "m"(*ff_pd_1)
+            );
+        }
+    }
+}
+#endif // CONFIG_ENCODERS
+
 static void vector_fmul_3dnow(float *dst, const float *src, int len){
     long i = (len-4)*4;
     asm volatile(
@@ -3605,6 +3727,7 @@ void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx)
             c->sum_abs_dctelem= sum_abs_dctelem_sse2;
             c->hadamard8_diff[0]= hadamard8_diff16_sse2;
             c->hadamard8_diff[1]= hadamard8_diff_sse2;
+            c->flac_compute_autocorr = flac_compute_autocorr_sse2;
         }
 
 #ifdef HAVE_SSSE3