VC1: merge idct8x8, coeff adjustments and put_pixels.

Merging these functions allows merging some loops, which makes the
results (particularly after SIMD optimizations) much faster.
(cherry picked from commit f8bed30d)

libavcodec/ppc/vc1dsp_altivec.c

@@ -130,7 +130,8 @@ do { \
 
 /** Do inverse transform on 8x8 block
 */
-static void vc1_inv_trans_8x8_altivec(DCTELEM block[64])
+static void vc1_inv_trans_8x8_altivec(DCTELEM block[64],
+                                      int sign, int rangered)
 {
     vector signed short src0, src1, src2, src3, src4, src5, src6, src7;
     vector signed int s0, s1, s2, s3, s4, s5, s6, s7;
@@ -144,7 +145,9 @@ static void vc1_inv_trans_8x8_altivec(DCTELEM block[64])
     const vector unsigned int vec_2 = vec_splat_u32(2);
     const vector  signed int vec_1s = vec_splat_s32(1);
     const vector unsigned int vec_1 = vec_splat_u32(1);
-
+    const vector unsigned short rangered_shift = vec_splat_u16(1);
+    const vector   signed short signed_bias = vec_sl(vec_splat_u16(4),
+                                                     vec_splat_u16(4));
 
     src0 = vec_ld(  0, block);
     src1 = vec_ld( 16, block);
@@ -214,6 +217,27 @@ static void vc1_inv_trans_8x8_altivec(DCTELEM block[64])
     src6 = vec_pack(sE, s6);
     src7 = vec_pack(sF, s7);
 
+    if (rangered) {
+        if (!sign) {
+            vec_sub(src0, signed_bias);
+            vec_sub(src1, signed_bias);
+            vec_sub(src2, signed_bias);
+            vec_sub(src3, signed_bias);
+            vec_sub(src4, signed_bias);
+            vec_sub(src5, signed_bias);
+            vec_sub(src6, signed_bias);
+            vec_sub(src7, signed_bias);
+        }
+        vec_sl(src0, rangered_shift);
+        vec_sl(src1, rangered_shift);
+        vec_sl(src2, rangered_shift);
+        vec_sl(src3, rangered_shift);
+        vec_sl(src4, rangered_shift);
+        vec_sl(src5, rangered_shift);
+        vec_sl(src6, rangered_shift);
+        vec_sl(src7, rangered_shift);
+    }
+
     vec_st(src0,  0, block);
     vec_st(src1, 16, block);
     vec_st(src2, 32, block);
@@ -224,6 +248,36 @@ static void vc1_inv_trans_8x8_altivec(DCTELEM block[64])
     vec_st(src7,112, block);
 }
 
+static void vc1_inv_trans_8x8_add_altivec(uint8_t *dest, int stride, DCTELEM *b)
+{
+    vc1_inv_trans_8x8_altivec(b, 0, 0);
+    ff_add_pixels_clamped_c(b, dest, stride);
+}
+
+static void vc1_inv_trans_8x8_put_signed_altivec(uint8_t *dest, int stride, DCTELEM *b)
+{
+    vc1_inv_trans_8x8_altivec(b, 1, 0);
+    ff_put_signed_pixels_clamped_c(b, dest, stride);
+}
+
+static void vc1_inv_trans_8x8_put_signed_rangered_altivec(uint8_t *dest, int stride, DCTELEM *b)
+{
+    vc1_inv_trans_8x8_altivec(b, 1, 1);
+    ff_put_signed_pixels_clamped_c(b, dest, stride);
+}
+
+static void vc1_inv_trans_8x8_put_altivec(uint8_t *dest, int stride, DCTELEM *b)
+{
+    vc1_inv_trans_8x8_altivec(b, 0, 0);
+    ff_put_pixels_clamped_c(b, dest, stride);
+}
+
+static void vc1_inv_trans_8x8_put_rangered_altivec(uint8_t *dest, int stride, DCTELEM *b)
+{
+    vc1_inv_trans_8x8_altivec(b, 0, 1);
+    ff_put_pixels_clamped_c(b, dest, stride);
+}
+
 /** Do inverse transform on 8x4 part of block
 */
 static void vc1_inv_trans_8x4_altivec(uint8_t *dest, int stride, DCTELEM *block)
@@ -342,7 +396,11 @@ void ff_vc1dsp_init_altivec(VC1DSPContext* dsp)
     if (!(av_get_cpu_flags() & AV_CPU_FLAG_ALTIVEC))
         return;
 
-    dsp->vc1_inv_trans_8x8 = vc1_inv_trans_8x8_altivec;
+    dsp->vc1_inv_trans_8x8_add = vc1_inv_trans_8x8_add_altivec;
+    dsp->vc1_inv_trans_8x8_put_signed[0] = vc1_inv_trans_8x8_put_signed_altivec;
+    dsp->vc1_inv_trans_8x8_put_signed[1] = vc1_inv_trans_8x8_put_signed_rangered_altivec;
+    dsp->vc1_inv_trans_8x8_put[0] = vc1_inv_trans_8x8_put_altivec;
+    dsp->vc1_inv_trans_8x8_put[1] = vc1_inv_trans_8x8_put_rangered_altivec;
     dsp->vc1_inv_trans_8x4 = vc1_inv_trans_8x4_altivec;
     dsp->put_no_rnd_vc1_chroma_pixels_tab[0] = put_no_rnd_vc1_chroma_mc8_altivec;
     dsp->avg_no_rnd_vc1_chroma_pixels_tab[0] = avg_no_rnd_vc1_chroma_mc8_altivec;

libavcodec/vc1.c

@@ -280,6 +280,28 @@ static int vop_dquant_decoding(VC1Context *v)
 
 static int decode_sequence_header_adv(VC1Context *v, GetBitContext *gb);
 
+static void simple_idct_put_rangered(uint8_t *dest, int line_size, DCTELEM *block)
+{
+    int i;
+    ff_simple_idct(block);
+    for (i = 0; i < 64; i++) block[i] = (block[i] - 64) << 1;
+    ff_put_pixels_clamped_c(block, dest, line_size);
+}
+
+static void simple_idct_put_signed(uint8_t *dest, int line_size, DCTELEM *block)
+{
+    ff_simple_idct(block);
+    ff_put_signed_pixels_clamped_c(block, dest, line_size);
+}
+
+static void simple_idct_put_signed_rangered(uint8_t *dest, int line_size, DCTELEM *block)
+{
+    int i;
+    ff_simple_idct(block);
+    for (i = 0; i < 64; i++) block[i] <<= 1;
+    ff_put_signed_pixels_clamped_c(block, dest, line_size);
+}
+
 /**
  * Decode Simple/Main Profiles sequence header
  * @see Figure 7-8, p16-17
@@ -337,7 +359,11 @@ int vc1_decode_sequence_header(AVCodecContext *avctx, VC1Context *v, GetBitConte
     v->res_fasttx = get_bits1(gb);
     if (!v->res_fasttx)
     {
-        v->vc1dsp.vc1_inv_trans_8x8 = ff_simple_idct;
+        v->vc1dsp.vc1_inv_trans_8x8_add = ff_simple_idct_add;
+        v->vc1dsp.vc1_inv_trans_8x8_put[0] = ff_simple_idct_put;
+        v->vc1dsp.vc1_inv_trans_8x8_put[1] = simple_idct_put_rangered;
+        v->vc1dsp.vc1_inv_trans_8x8_put_signed[0] = simple_idct_put_signed;
+        v->vc1dsp.vc1_inv_trans_8x8_put_signed[1] = simple_idct_put_signed_rangered;
         v->vc1dsp.vc1_inv_trans_8x4 = ff_simple_idct84_add;
         v->vc1dsp.vc1_inv_trans_4x8 = ff_simple_idct48_add;
         v->vc1dsp.vc1_inv_trans_4x4 = ff_simple_idct44_add;

libavcodec/vc1dec.c

@@ -2009,8 +2009,7 @@ static int vc1_decode_p_block(VC1Context *v, DCTELEM block[64], int n, int mquan
             if(i==1)
                 v->vc1dsp.vc1_inv_trans_8x8_dc(dst, linesize, block);
             else{
-                v->vc1dsp.vc1_inv_trans_8x8(block);
-                s->dsp.add_pixels_clamped(block, dst, linesize);
+                v->vc1dsp.vc1_inv_trans_8x8_add(dst, linesize, block);
             }
             if(apply_filter && cbp_top  & 0xC)
                 v->vc1dsp.vc1_v_loop_filter8(dst, linesize, v->pq);
@@ -2117,7 +2116,7 @@ static int vc1_decode_p_mb(VC1Context *v)
 {
     MpegEncContext *s = &v->s;
     GetBitContext *gb = &s->gb;
-    int i, j;
+    int i;
     int mb_pos = s->mb_x + s->mb_y * s->mb_stride;
     int cbp; /* cbp decoding stuff */
     int mqdiff, mquant; /* MB quantization */
@@ -2149,6 +2148,8 @@ static int vc1_decode_p_mb(VC1Context *v)
     {
         if (!skipped)
         {
+            vc1_idct_func idct8x8_fn;
+
             GET_MVDATA(dmv_x, dmv_y);
 
             if (s->mb_intra) {
@@ -2183,6 +2184,7 @@ static int vc1_decode_p_mb(VC1Context *v)
                                 VC1_TTMB_VLC_BITS, 2);
             if(!s->mb_intra) vc1_mc_1mv(v, 0);
             dst_idx = 0;
+            idct8x8_fn = v->vc1dsp.vc1_inv_trans_8x8_put_signed[!!v->rangeredfrm];
             for (i=0; i<6; i++)
             {
                 s->dc_val[0][s->block_index[i]] = 0;
@@ -2200,9 +2202,9 @@ static int vc1_decode_p_mb(VC1Context *v)
 
                     vc1_decode_intra_block(v, s->block[i], i, val, mquant, (i&4)?v->codingset2:v->codingset);
                     if((i>3) && (s->flags & CODEC_FLAG_GRAY)) continue;
-                    v->vc1dsp.vc1_inv_trans_8x8(s->block[i]);
-                    if(v->rangeredfrm) for(j = 0; j < 64; j++) s->block[i][j] <<= 1;
-                    s->dsp.put_signed_pixels_clamped(s->block[i], s->dest[dst_idx] + off, i & 4 ? s->uvlinesize : s->linesize);
+                    idct8x8_fn(s->dest[dst_idx] + off,
+                               i & 4 ? s->uvlinesize : s->linesize,
+                               s->block[i]);
                     if(v->pq >= 9 && v->overlap) {
                         if(v->c_avail)
                             v->vc1dsp.vc1_h_overlap(s->dest[dst_idx] + off, i & 4 ? s->uvlinesize : s->linesize);
@@ -2267,6 +2269,7 @@ static int vc1_decode_p_mb(VC1Context *v)
         {
             int intra_count = 0, coded_inter = 0;
             int is_intra[6], is_coded[6];
+            vc1_idct_func idct8x8_fn;
             /* Get CBPCY */
             cbp = get_vlc2(&v->s.gb, v->cbpcy_vlc->table, VC1_CBPCY_P_VLC_BITS, 2);
             for (i=0; i<6; i++)
@@ -2316,6 +2319,7 @@ static int vc1_decode_p_mb(VC1Context *v)
             }
             if (!v->ttmbf && coded_inter)
                 ttmb = get_vlc2(gb, ff_vc1_ttmb_vlc[v->tt_index].table, VC1_TTMB_VLC_BITS, 2);
+            idct8x8_fn = v->vc1dsp.vc1_inv_trans_8x8_put_signed[!!v->rangeredfrm];
             for (i=0; i<6; i++)
             {
                 dst_idx += i >> 2;
@@ -2331,9 +2335,9 @@ static int vc1_decode_p_mb(VC1Context *v)
 
                     vc1_decode_intra_block(v, s->block[i], i, is_coded[i], mquant, (i&4)?v->codingset2:v->codingset);
                     if((i>3) && (s->flags & CODEC_FLAG_GRAY)) continue;
-                    v->vc1dsp.vc1_inv_trans_8x8(s->block[i]);
-                    if(v->rangeredfrm) for(j = 0; j < 64; j++) s->block[i][j] <<= 1;
-                    s->dsp.put_signed_pixels_clamped(s->block[i], s->dest[dst_idx] + off, (i&4)?s->uvlinesize:s->linesize);
+                    idct8x8_fn(s->dest[dst_idx] + off,
+                               (i&4)?s->uvlinesize:s->linesize,
+                               s->block[i]);
                     if(v->pq >= 9 && v->overlap) {
                         if(v->c_avail)
                             v->vc1dsp.vc1_h_overlap(s->dest[dst_idx] + off, i & 4 ? s->uvlinesize : s->linesize);
@@ -2409,7 +2413,7 @@ static void vc1_decode_b_mb(VC1Context *v)
 {
     MpegEncContext *s = &v->s;
     GetBitContext *gb = &s->gb;
-    int i, j;
+    int i;
     int mb_pos = s->mb_x + s->mb_y * s->mb_stride;
     int cbp = 0; /* cbp decoding stuff */
     int mqdiff, mquant; /* MB quantization */
@@ -2422,6 +2426,7 @@ static void vc1_decode_b_mb(VC1Context *v)
     int skipped, direct;
     int dmv_x[2], dmv_y[2];
     int bmvtype = BMV_TYPE_BACKWARD;
+    vc1_idct_func idct8x8_fn;
 
     mquant = v->pq; /* Loosy initialization */
     s->mb_intra = 0;
@@ -2519,6 +2524,7 @@ static void vc1_decode_b_mb(VC1Context *v)
         }
     }
     dst_idx = 0;
+    idct8x8_fn = v->vc1dsp.vc1_inv_trans_8x8_put_signed[!!v->rangeredfrm];
     for (i=0; i<6; i++)
     {
         s->dc_val[0][s->block_index[i]] = 0;
@@ -2536,9 +2542,9 @@ static void vc1_decode_b_mb(VC1Context *v)
 
             vc1_decode_intra_block(v, s->block[i], i, val, mquant, (i&4)?v->codingset2:v->codingset);
             if((i>3) && (s->flags & CODEC_FLAG_GRAY)) continue;
-            v->vc1dsp.vc1_inv_trans_8x8(s->block[i]);
-            if(v->rangeredfrm) for(j = 0; j < 64; j++) s->block[i][j] <<= 1;
-            s->dsp.put_signed_pixels_clamped(s->block[i], s->dest[dst_idx] + off, i & 4 ? s->uvlinesize : s->linesize);
+            idct8x8_fn(s->dest[dst_idx] + off,
+                       i & 4 ? s->uvlinesize : s->linesize,
+                       s->block[i]);
         } else if(val) {
             vc1_decode_p_block(v, s->block[i], i, mquant, ttmb, first_block, s->dest[dst_idx] + off, (i&4)?s->uvlinesize:s->linesize, (i&4) && (s->flags & CODEC_FLAG_GRAY), 0, 0, 0);
             if(!v->ttmbf && ttmb < 8) ttmb = -1;
@@ -2551,11 +2557,12 @@ static void vc1_decode_b_mb(VC1Context *v)
  */
 static void vc1_decode_i_blocks(VC1Context *v)
 {
-    int k, j;
+    int k;
     MpegEncContext *s = &v->s;
     int cbp, val;
     uint8_t *coded_val;
     int mb_pos;
+    vc1_idct_func idct8x8_fn;
 
     /* select codingmode used for VLC tables selection */
     switch(v->y_ac_table_index){
@@ -2590,6 +2597,10 @@ static void vc1_decode_i_blocks(VC1Context *v)
     s->mb_x = s->mb_y = 0;
     s->mb_intra = 1;
     s->first_slice_line = 1;
+    if(v->pq >= 9 && v->overlap) {
+        idct8x8_fn = v->vc1dsp.vc1_inv_trans_8x8_put_signed[!!v->rangeredfrm];
+    } else
+        idct8x8_fn = v->vc1dsp.vc1_inv_trans_8x8_put[!!v->rangeredfrm];
     for(s->mb_y = 0; s->mb_y < s->mb_height; s->mb_y++) {
         s->mb_x = 0;
         ff_init_block_index(s);
@@ -2626,14 +2637,9 @@ static void vc1_decode_i_blocks(VC1Context *v)
                 vc1_decode_i_block(v, s->block[k], k, val, (k<4)? v->codingset : v->codingset2);
 
                 if (k > 3 && (s->flags & CODEC_FLAG_GRAY)) continue;
-                v->vc1dsp.vc1_inv_trans_8x8(s->block[k]);
-                if(v->pq >= 9 && v->overlap) {
-                    if (v->rangeredfrm) for(j = 0; j < 64; j++) s->block[k][j] <<= 1;
-                    s->dsp.put_signed_pixels_clamped(s->block[k], dst[k], k & 4 ? s->uvlinesize : s->linesize);
-                } else {
-                    if (v->rangeredfrm) for(j = 0; j < 64; j++) s->block[k][j] = (s->block[k][j] - 64) << 1;
-                    s->dsp.put_pixels_clamped(s->block[k], dst[k], k & 4 ? s->uvlinesize : s->linesize);
-                }
+                idct8x8_fn(dst[k],
+                           k & 4 ? s->uvlinesize : s->linesize,
+                           s->block[k]);
             }
 
             if(v->pq >= 9 && v->overlap) {
@@ -2691,6 +2697,7 @@ static void vc1_decode_i_blocks_adv(VC1Context *v)
     int mqdiff;
     int overlap;
     GetBitContext *gb = &s->gb;
+    vc1_idct_func idct8x8_fn;
 
     /* select codingmode used for VLC tables selection */
     switch(v->y_ac_table_index){
@@ -2721,6 +2728,7 @@ static void vc1_decode_i_blocks_adv(VC1Context *v)
     s->mb_x = s->mb_y = 0;
     s->mb_intra = 1;
     s->first_slice_line = 1;
+    idct8x8_fn = v->vc1dsp.vc1_inv_trans_8x8_put_signed[0];
     for(s->mb_y = 0; s->mb_y < s->mb_height; s->mb_y++) {
         s->mb_x = 0;
         ff_init_block_index(s);
@@ -2777,9 +2785,9 @@ static void vc1_decode_i_blocks_adv(VC1Context *v)
                 vc1_decode_i_block_adv(v, s->block[k], k, val, (k<4)? v->codingset : v->codingset2, mquant);
 
                 if (k > 3 && (s->flags & CODEC_FLAG_GRAY)) continue;
-                v->vc1dsp.vc1_inv_trans_8x8(s->block[k]);
-                s->dsp.put_signed_pixels_clamped(s->block[k], dst[k],
-                                                 k & 4 ? s->uvlinesize : s->linesize);
+                idct8x8_fn(dst[k],
+                           k & 4 ? s->uvlinesize : s->linesize,
+                           s->block[k]);
             }
 
             if(overlap) {

libavcodec/vc1dsp.c

@@ -199,7 +199,7 @@ static void vc1_inv_trans_8x8_dc_c(uint8_t *dest, int linesize, DCTELEM *block)
     }
 }
 
-static void vc1_inv_trans_8x8_c(DCTELEM block[64])
+static av_always_inline void vc1_inv_trans_8x8_c(DCTELEM block[64], int shl, int sub)
 {
     int i;
     register int t1,t2,t3,t4,t5,t6,t7,t8;
@@ -254,20 +254,50 @@ static void vc1_inv_trans_8x8_c(DCTELEM block[64])
         t3 =  9 * src[ 8] - 16 * src[24] +  4 * src[40] + 15 * src[56];
         t4 =  4 * src[ 8] -  9 * src[24] + 15 * src[40] - 16 * src[56];
 
-        dst[ 0] = (t5 + t1) >> 7;
-        dst[ 8] = (t6 + t2) >> 7;
-        dst[16] = (t7 + t3) >> 7;
-        dst[24] = (t8 + t4) >> 7;
-        dst[32] = (t8 - t4 + 1) >> 7;
-        dst[40] = (t7 - t3 + 1) >> 7;
-        dst[48] = (t6 - t2 + 1) >> 7;
-        dst[56] = (t5 - t1 + 1) >> 7;
+        dst[ 0] = (((t5 + t1    ) >> 7) - sub) << shl;
+        dst[ 8] = (((t6 + t2    ) >> 7) - sub) << shl;
+        dst[16] = (((t7 + t3    ) >> 7) - sub) << shl;
+        dst[24] = (((t8 + t4    ) >> 7) - sub) << shl;
+        dst[32] = (((t8 - t4 + 1) >> 7) - sub) << shl;
+        dst[40] = (((t7 - t3 + 1) >> 7) - sub) << shl;
+        dst[48] = (((t6 - t2 + 1) >> 7) - sub) << shl;
+        dst[56] = (((t5 - t1 + 1) >> 7) - sub) << shl;
 
         src++;
         dst++;
     }
 }
 
+static void vc1_inv_trans_8x8_add_c(uint8_t *dest, int linesize, DCTELEM *block)
+{
+    vc1_inv_trans_8x8_c(block, 0, 0);
+    ff_add_pixels_clamped_c(block, dest, linesize);
+}
+
+static void vc1_inv_trans_8x8_put_signed_c(uint8_t *dest, int linesize, DCTELEM *block)
+{
+    vc1_inv_trans_8x8_c(block, 0, 0);
+    ff_put_signed_pixels_clamped_c(block, dest, linesize);
+}
+
+static void vc1_inv_trans_8x8_put_signed_rangered_c(uint8_t *dest, int linesize, DCTELEM *block)
+{
+    vc1_inv_trans_8x8_c(block, 1, 0);
+    ff_put_signed_pixels_clamped_c(block, dest, linesize);
+}
+
+static void vc1_inv_trans_8x8_put_c(uint8_t *dest, int linesize, DCTELEM *block)
+{
+    vc1_inv_trans_8x8_c(block, 0, 0);
+    ff_put_pixels_clamped_c(block, dest, linesize);
+}
+
+static void vc1_inv_trans_8x8_put_rangered_c(uint8_t *dest, int linesize, DCTELEM *block)
+{
+    vc1_inv_trans_8x8_c(block, 1, 64);
+    ff_put_pixels_clamped_c(block, dest, linesize);
+}
+
 /** Do inverse transform on 8x4 part of block
 */
 static void vc1_inv_trans_8x4_dc_c(uint8_t *dest, int linesize, DCTELEM *block)
@@ -662,7 +692,11 @@ static void avg_no_rnd_vc1_chroma_mc8_c(uint8_t *dst/*align 8*/, uint8_t *src/*a
 }
 
 av_cold void ff_vc1dsp_init(VC1DSPContext* dsp) {
-    dsp->vc1_inv_trans_8x8 = vc1_inv_trans_8x8_c;
+    dsp->vc1_inv_trans_8x8_add = vc1_inv_trans_8x8_add_c;
+    dsp->vc1_inv_trans_8x8_put_signed[0] = vc1_inv_trans_8x8_put_signed_c;
+    dsp->vc1_inv_trans_8x8_put_signed[1] = vc1_inv_trans_8x8_put_signed_rangered_c;
+    dsp->vc1_inv_trans_8x8_put[0] = vc1_inv_trans_8x8_put_c;
+    dsp->vc1_inv_trans_8x8_put[1] = vc1_inv_trans_8x8_put_rangered_c;
     dsp->vc1_inv_trans_4x8 = vc1_inv_trans_4x8_c;
     dsp->vc1_inv_trans_8x4 = vc1_inv_trans_8x4_c;
     dsp->vc1_inv_trans_4x4 = vc1_inv_trans_4x4_c;

libavcodec/vc1dsp.h

@@ -30,9 +30,13 @@
 
 #include "dsputil.h"
 
+typedef void (*vc1_idct_func)(uint8_t *dest, int line_size, DCTELEM *block);
+
 typedef struct VC1DSPContext {
     /* vc1 functions */
-    void (*vc1_inv_trans_8x8)(DCTELEM *b);
+    vc1_idct_func vc1_inv_trans_8x8_add;
+    vc1_idct_func vc1_inv_trans_8x8_put_signed[2];
+    vc1_idct_func vc1_inv_trans_8x8_put[2];
     void (*vc1_inv_trans_8x4)(uint8_t *dest, int line_size, DCTELEM *block);
     void (*vc1_inv_trans_4x8)(uint8_t *dest, int line_size, DCTELEM *block);
     void (*vc1_inv_trans_4x4)(uint8_t *dest, int line_size, DCTELEM *block);