h264_idct8_add_mmx

Originally committed as revision 5123 to svn://svn.ffmpeg.org/ffmpeg/trunk

3 files changed, 160 insertions, 6 deletions

diff --git a/libavcodec/h264.c b/libavcodec/h264.c
index a0a0ada235..b9f000f217 100644
--- a/libavcodec/h264.c
+++ b/libavcodec/h264.c
@@ -358,8 +358,12 @@ typedef struct H264Context{
 
     uint8_t zigzag_scan[16];
     uint8_t field_scan[16];
+    uint8_t zigzag_scan8x8[64];
+    uint8_t zigzag_scan8x8_cavlc[64];
     const uint8_t *zigzag_scan_q0;
     const uint8_t *field_scan_q0;
+    const uint8_t *zigzag_scan8x8_q0;
+    const uint8_t *zigzag_scan8x8_cavlc_q0;
 
     int x264_build;
 }H264Context;
@@ -2953,6 +2957,7 @@ static void free_tables(H264Context *h){
 
 static void init_dequant8_coeff_table(H264Context *h){
     int i,q,x;
+    const int transpose = (h->s.dsp.h264_idct8_add != ff_h264_idct8_add_c); //FIXME ugly
     h->dequant8_coeff[0] = h->dequant8_buffer[0];
     h->dequant8_coeff[1] = h->dequant8_buffer[1];
 
@@ -2966,8 +2971,9 @@ static void init_dequant8_coeff_table(H264Context *h){
             int shift = div6[q];
             int idx = rem6[q];
             for(x=0; x<64; x++)
-                h->dequant8_coeff[i][q][x] = ((uint32_t)dequant8_coeff_init[idx][
-                    dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] * h->pps.scaling_matrix8[i][x]) << shift;
+                h->dequant8_coeff[i][q][transpose ? (x>>3)|((x&7)<<3) : x] =
+                    ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
+                    h->pps.scaling_matrix8[i][x]) << shift;
         }
     }
 }
@@ -4317,14 +4323,31 @@ static int decode_slice_header(H264Context *h){
 #define T(x) (x>>2) | ((x<<2) & 0xF)
                 h->zigzag_scan[i] = T(zigzag_scan[i]);
                 h-> field_scan[i] = T( field_scan[i]);
+#undef T
+            }
+        }
+        if(s->dsp.h264_idct8_add == ff_h264_idct8_add_c){
+            memcpy(h->zigzag_scan8x8, zigzag_scan8x8, 64*sizeof(uint8_t));
+            memcpy(h->zigzag_scan8x8_cavlc, zigzag_scan8x8_cavlc, 64*sizeof(uint8_t));
+        }else{
+            int i;
+            for(i=0; i<64; i++){
+#define T(x) (x>>3) | ((x&7)<<3)
+                h->zigzag_scan8x8[i] = T(zigzag_scan8x8[i]);
+                h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
+#undef T
             }
         }
         if(h->sps.transform_bypass){ //FIXME same ugly
             h->zigzag_scan_q0 = zigzag_scan;
             h->field_scan_q0 = field_scan;
+            h->zigzag_scan8x8_q0 = zigzag_scan8x8;
+            h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
         }else{
             h->zigzag_scan_q0 = h->zigzag_scan;
             h->field_scan_q0 = h->field_scan;
+            h->zigzag_scan8x8_q0 = h->zigzag_scan8x8;
+            h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
         }
 
         alloc_tables(h);
@@ -5101,7 +5124,7 @@ decode_intra_mb:
         int i8x8, i4x4, chroma_idx;
         int chroma_qp, dquant;
         GetBitContext *gb= IS_INTRA(mb_type) ? h->intra_gb_ptr : h->inter_gb_ptr;
-        const uint8_t *scan, *dc_scan;
+        const uint8_t *scan, *scan8x8, *dc_scan;
 
 //        fill_non_zero_count_cache(h);
 
@@ -5112,6 +5135,7 @@ decode_intra_mb:
             scan= s->qscale ? h->zigzag_scan : h->zigzag_scan_q0;
             dc_scan= luma_dc_zigzag_scan;
         }
+        scan8x8= s->qscale ? h->zigzag_scan8x8_cavlc : h->zigzag_scan8x8_cavlc_q0;
 
         dquant= get_se_golomb(&s->gb);
 
@@ -5153,7 +5177,7 @@ decode_intra_mb:
                         DCTELEM *buf = &h->mb[64*i8x8];
                         uint8_t *nnz;
                         for(i4x4=0; i4x4<4; i4x4++){
-                            if( decode_residual(h, gb, buf, i4x4+4*i8x8, zigzag_scan8x8_cavlc+16*i4x4,
+                            if( decode_residual(h, gb, buf, i4x4+4*i8x8, scan8x8+16*i4x4,
                                                 h->dequant8_coeff[IS_INTRA( mb_type ) ? 0:1][s->qscale], 16) <0 )
                                 return -1;
                         }
@@ -6144,7 +6168,7 @@ decode_intra_mb:
     s->current_picture.mb_type[mb_xy]= mb_type;
 
     if( cbp || IS_INTRA16x16( mb_type ) ) {
-        const uint8_t *scan, *dc_scan;
+        const uint8_t *scan, *scan8x8, *dc_scan;
         int dqp;
 
         if(IS_INTERLACED(mb_type)){
@@ -6154,6 +6178,7 @@ decode_intra_mb:
             scan= s->qscale ? h->zigzag_scan : h->zigzag_scan_q0;
             dc_scan= luma_dc_zigzag_scan;
         }
+        scan8x8= s->qscale ? h->zigzag_scan8x8 : h->zigzag_scan8x8_q0;
 
         h->last_qscale_diff = dqp = decode_cabac_mb_dqp( h );
         if( dqp == INT_MIN ){
@@ -6187,7 +6212,7 @@ decode_intra_mb:
                 if( cbp & (1<<i8x8) ) {
                     if( IS_8x8DCT(mb_type) ) {
                         if( decode_cabac_residual(h, h->mb + 64*i8x8, 5, 4*i8x8,
-                                zigzag_scan8x8, h->dequant8_coeff[IS_INTRA( mb_type ) ? 0:1][s->qscale], 64) < 0 )
+                            scan8x8, h->dequant8_coeff[IS_INTRA( mb_type ) ? 0:1][s->qscale], 64) < 0 )
                             return -1;
                     } else
                     for( i4x4 = 0; i4x4 < 4; i4x4++ ) {
diff --git a/libavcodec/i386/dsputil_mmx.c b/libavcodec/i386/dsputil_mmx.c
index 3af3974c2a..1e8590e9d9 100644
--- a/libavcodec/i386/dsputil_mmx.c
+++ b/libavcodec/i386/dsputil_mmx.c
@@ -2734,6 +2734,8 @@ void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx)
 
         c->h264_idct_dc_add=
         c->h264_idct_add= ff_h264_idct_add_mmx;
+        c->h264_idct8_dc_add=
+        c->h264_idct8_add= ff_h264_idct8_add_mmx;
 
         if (mm_flags & MM_MMXEXT) {
             c->put_pixels_tab[0][1] = put_pixels16_x2_mmx2;
diff --git a/libavcodec/i386/h264dsp_mmx.c b/libavcodec/i386/h264dsp_mmx.c
index 65fa95d3e5..9a95895506 100644
--- a/libavcodec/i386/h264dsp_mmx.c
+++ b/libavcodec/i386/h264dsp_mmx.c
@@ -104,6 +104,133 @@ static void ff_h264_idct_add_mmx(uint8_t *dst, int16_t *block, int stride)
     );
 }
 
+static inline void h264_idct8_1d(int16_t *block)
+{
+    asm volatile(
+        "movq 112(%0), %%mm7  \n\t"
+        "movq  80(%0), %%mm5  \n\t"
+        "movq  48(%0), %%mm3  \n\t"
+        "movq  16(%0), %%mm1  \n\t"
+
+        "movq   %%mm7, %%mm4  \n\t"
+        "movq   %%mm3, %%mm6  \n\t"
+        "movq   %%mm5, %%mm0  \n\t"
+        "movq   %%mm7, %%mm2  \n\t"
+        "psraw  $1,    %%mm4  \n\t"
+        "psraw  $1,    %%mm6  \n\t"
+        "psubw  %%mm7, %%mm0  \n\t"
+        "psubw  %%mm6, %%mm2  \n\t"
+        "psubw  %%mm4, %%mm0  \n\t"
+        "psubw  %%mm3, %%mm2  \n\t"
+        "psubw  %%mm3, %%mm0  \n\t"
+        "paddw  %%mm1, %%mm2  \n\t"
+
+        "movq   %%mm5, %%mm4  \n\t"
+        "movq   %%mm1, %%mm6  \n\t"
+        "psraw  $1,    %%mm4  \n\t"
+        "psraw  $1,    %%mm6  \n\t"
+        "paddw  %%mm5, %%mm4  \n\t"
+        "paddw  %%mm1, %%mm6  \n\t"
+        "paddw  %%mm7, %%mm4  \n\t"
+        "paddw  %%mm5, %%mm6  \n\t"
+        "psubw  %%mm1, %%mm4  \n\t"
+        "paddw  %%mm3, %%mm6  \n\t"
+
+        "movq   %%mm0, %%mm1  \n\t"
+        "movq   %%mm4, %%mm3  \n\t"
+        "movq   %%mm2, %%mm5  \n\t"
+        "movq   %%mm6, %%mm7  \n\t"
+        "psraw  $2,    %%mm6  \n\t"
+        "psraw  $2,    %%mm3  \n\t"
+        "psraw  $2,    %%mm5  \n\t"
+        "psraw  $2,    %%mm0  \n\t"
+        "paddw  %%mm6, %%mm1  \n\t"
+        "paddw  %%mm2, %%mm3  \n\t"
+        "psubw  %%mm4, %%mm5  \n\t"
+        "psubw  %%mm0, %%mm7  \n\t"
+
+        "movq  32(%0), %%mm2  \n\t"
+        "movq  96(%0), %%mm6  \n\t"
+        "movq   %%mm2, %%mm4  \n\t"
+        "movq   %%mm6, %%mm0  \n\t"
+        "psraw  $1,    %%mm4  \n\t"
+        "psraw  $1,    %%mm6  \n\t"
+        "psubw  %%mm0, %%mm4  \n\t"
+        "paddw  %%mm2, %%mm6  \n\t"
+
+        "movq    (%0), %%mm2  \n\t"
+        "movq  64(%0), %%mm0  \n\t"
+        SUMSUB_BA( %%mm0, %%mm2 )
+        SUMSUB_BA( %%mm6, %%mm0 )
+        SUMSUB_BA( %%mm4, %%mm2 )
+        SUMSUB_BA( %%mm7, %%mm6 )
+        SUMSUB_BA( %%mm5, %%mm4 )
+        SUMSUB_BA( %%mm3, %%mm2 )
+        SUMSUB_BA( %%mm1, %%mm0 )
+        :: "r"(block)
+    );
+}
+
+static void ff_h264_idct8_add_mmx(uint8_t *dst, int16_t *block, int stride)
+{
+    int i;
+    int16_t __attribute__ ((aligned(8))) b2[64];
+
+    block[0] += 32;
+
+    for(i=0; i<2; i++){
+        uint64_t tmp;
+
+        h264_idct8_1d(block+4*i);
+
+        asm volatile(
+            "movq   %%mm7,    %0   \n\t"
+            TRANSPOSE4( %%mm0, %%mm2, %%mm4, %%mm6, %%mm7 )
+            "movq   %%mm0,  8(%1)  \n\t"
+            "movq   %%mm6, 24(%1)  \n\t"
+            "movq   %%mm7, 40(%1)  \n\t"
+            "movq   %%mm4, 56(%1)  \n\t"
+            "movq    %0,    %%mm7  \n\t"
+            TRANSPOSE4( %%mm7, %%mm5, %%mm3, %%mm1, %%mm0 )
+            "movq   %%mm7,   (%1)  \n\t"
+            "movq   %%mm1, 16(%1)  \n\t"
+            "movq   %%mm0, 32(%1)  \n\t"
+            "movq   %%mm3, 48(%1)  \n\t"
+            : "=m"(tmp)
+            : "r"(b2+32*i)
+            : "memory"
+        );
+    }
+
+    for(i=0; i<2; i++){
+        h264_idct8_1d(b2+4*i);
+
+        asm volatile(
+            "psraw     $6, %%mm7  \n\t"
+            "psraw     $6, %%mm6  \n\t"
+            "psraw     $6, %%mm5  \n\t"
+            "psraw     $6, %%mm4  \n\t"
+            "psraw     $6, %%mm3  \n\t"
+            "psraw     $6, %%mm2  \n\t"
+            "psraw     $6, %%mm1  \n\t"
+            "psraw     $6, %%mm0  \n\t"
+
+            "movq   %%mm7,    (%0)  \n\t"
+            "movq   %%mm5,  16(%0)  \n\t"
+            "movq   %%mm3,  32(%0)  \n\t"
+            "movq   %%mm1,  48(%0)  \n\t"
+            "movq   %%mm0,  64(%0)  \n\t"
+            "movq   %%mm2,  80(%0)  \n\t"
+            "movq   %%mm4,  96(%0)  \n\t"
+            "movq   %%mm6, 112(%0)  \n\t"
+            :: "r"(b2+4*i)
+            : "memory"
+        );
+    }
+
+    add_pixels_clamped_mmx(b2, dst, stride);
+}
+
 static void ff_h264_idct_dc_add_mmx2(uint8_t *dst, int16_t *block, int stride)
 {
     int dc = (block[0] + 32) >> 6;