Merge remote-tracking branch 'qatar/master'

* qatar/master:
  4xm: fix invalid array indexing
  rv34dsp: factorize a multiplication in the noround inverse transform
  rv40: perform bitwise checks in loop filter
  rv34: remove inline keyword from rv34_decode_block().
  rv40: change a logical test into a bitwise one.
  rv34: remove constant parameter
  rv40: don't always do the full prev_type search
  dsputil x86: revert a test back to its previous value
  rv34dsp x86: implement MMX2 inverse transform

Merged-by: Michael Niedermayer <michaelni@gmx.at>

7 files changed, 132 insertions, 51 deletions

diff --git a/libavcodec/4xm.c b/libavcodec/4xm.c
index 600cf81ccd..e186870590 100644
--- a/libavcodec/4xm.c
+++ b/libavcodec/4xm.c
@@ -240,15 +240,18 @@ static void idct(DCTELEM block[64])
 
 static av_cold void init_vlcs(FourXContext *f)
 {
-    static VLC_TYPE table[8][32][2];
-    int i;
-
-    for (i = 0; i < 8; i++) {
-        block_type_vlc[0][i].table           = table[i];
-        block_type_vlc[0][i].table_allocated = 32;
-        init_vlc(&block_type_vlc[0][i], BLOCK_TYPE_VLC_BITS, 7,
-                 &block_type_tab[0][i][0][1], 2, 1,
-                 &block_type_tab[0][i][0][0], 2, 1, INIT_VLC_USE_NEW_STATIC);
+    static VLC_TYPE table[2][4][32][2];
+    int i, j;
+
+    for (i = 0; i < 2; i++) {
+        for (j = 0; j < 4; j++) {
+            block_type_vlc[i][j].table           = table[i][j];
+            block_type_vlc[i][j].table_allocated = 32;
+            init_vlc(&block_type_vlc[i][j], BLOCK_TYPE_VLC_BITS, 7,
+                     &block_type_tab[i][j][0][1], 2, 1,
+                     &block_type_tab[i][j][0][0], 2, 1,
+                     INIT_VLC_USE_NEW_STATIC);
+        }
     }
 }
 
diff --git a/libavcodec/rv34.c b/libavcodec/rv34.c
index 263ad94b17..62619b3203 100644
--- a/libavcodec/rv34.c
+++ b/libavcodec/rv34.c
@@ -260,20 +260,15 @@ static inline void decode_subblock1(DCTELEM *dst, int code, GetBitContext *gb, V
     decode_coeff(dst, coeff, 3, gb, vlc, q);
 }
 
-static inline void decode_subblock3(DCTELEM *dst, int code, const int is_block2, GetBitContext *gb, VLC *vlc,
+static inline void decode_subblock3(DCTELEM *dst, int code, GetBitContext *gb, VLC *vlc,
                                     int q_dc, int q_ac1, int q_ac2)
 {
     int flags = modulo_three_table[code];
 
-    decode_coeff(    dst+0*4+0, (flags >> 6)    , 3, gb, vlc, q_dc);
-    if(is_block2){
-        decode_coeff(dst+1*4+0, (flags >> 4) & 3, 2, gb, vlc, q_ac1);
-        decode_coeff(dst+0*4+1, (flags >> 2) & 3, 2, gb, vlc, q_ac1);
-    }else{
-        decode_coeff(dst+0*4+1, (flags >> 4) & 3, 2, gb, vlc, q_ac1);
-        decode_coeff(dst+1*4+0, (flags >> 2) & 3, 2, gb, vlc, q_ac1);
-    }
-    decode_coeff(    dst+1*4+1, (flags >> 0) & 3, 2, gb, vlc, q_ac2);
+    decode_coeff(dst+0*4+0, (flags >> 6)    , 3, gb, vlc, q_dc);
+    decode_coeff(dst+0*4+1, (flags >> 4) & 3, 2, gb, vlc, q_ac1);
+    decode_coeff(dst+1*4+0, (flags >> 2) & 3, 2, gb, vlc, q_ac1);
+    decode_coeff(dst+1*4+1, (flags >> 0) & 3, 2, gb, vlc, q_ac2);
 }
 
 /**
@@ -287,7 +282,7 @@ static inline void decode_subblock3(DCTELEM *dst, int code, const int is_block2,
  *  o--o
  */
 
-static inline int rv34_decode_block(DCTELEM *dst, GetBitContext *gb, RV34VLC *rvlc, int fc, int sc, int q_dc, int q_ac1, int q_ac2)
+static int rv34_decode_block(DCTELEM *dst, GetBitContext *gb, RV34VLC *rvlc, int fc, int sc, int q_dc, int q_ac1, int q_ac2)
 {
     int code, pattern, has_ac = 1;
 
@@ -298,7 +293,7 @@ static inline int rv34_decode_block(DCTELEM *dst, GetBitContext *gb, RV34VLC *rv
     code >>= 3;
 
     if (modulo_three_table[code] & 0x3F) {
-        decode_subblock3(dst, code, 0, gb, &rvlc->coefficient, q_dc, q_ac1, q_ac2);
+        decode_subblock3(dst, code, gb, &rvlc->coefficient, q_dc, q_ac1, q_ac2);
     } else {
         decode_subblock1(dst, code, gb, &rvlc->coefficient, q_dc);
         if (!pattern)
@@ -318,7 +313,7 @@ static inline int rv34_decode_block(DCTELEM *dst, GetBitContext *gb, RV34VLC *rv
         code = get_vlc2(gb, rvlc->third_pattern[sc].table, 9, 2);
         decode_subblock(dst + 4*2+2, code, 0, gb, &rvlc->coefficient, q_ac2);
     }
-    return has_ac || pattern;
+    return has_ac | pattern;
 }
 
 /**
diff --git a/libavcodec/rv34dsp.c b/libavcodec/rv34dsp.c
index 4145c4dd85..1ddcea4c0f 100644
--- a/libavcodec/rv34dsp.c
+++ b/libavcodec/rv34dsp.c
@@ -88,15 +88,15 @@ static void rv34_inv_transform_noround_c(DCTELEM *block){
     rv34_row_transform(temp, block);
 
     for(i = 0; i < 4; i++){
-        const int z0 = 13*(temp[4*0+i] +    temp[4*2+i]);
-        const int z1 = 13*(temp[4*0+i] -    temp[4*2+i]);
-        const int z2 =  7* temp[4*1+i] - 17*temp[4*3+i];
-        const int z3 = 17* temp[4*1+i] +  7*temp[4*3+i];
-
-        block[i*4+0] = ((z0 + z3) * 3) >> 11;
-        block[i*4+1] = ((z1 + z2) * 3) >> 11;
-        block[i*4+2] = ((z1 - z2) * 3) >> 11;
-        block[i*4+3] = ((z0 - z3) * 3) >> 11;
+        const int z0 = 39*(temp[4*0+i] +    temp[4*2+i]);
+        const int z1 = 39*(temp[4*0+i] -    temp[4*2+i]);
+        const int z2 = 21* temp[4*1+i] - 51*temp[4*3+i];
+        const int z3 = 51* temp[4*1+i] + 21*temp[4*3+i];
+
+        block[i*4+0] = (z0 + z3) >> 11;
+        block[i*4+1] = (z1 + z2) >> 11;
+        block[i*4+2] = (z1 - z2) >> 11;
+        block[i*4+3] = (z0 - z3) >> 11;
     }
 }
 
diff --git a/libavcodec/rv40.c b/libavcodec/rv40.c
index e7390fb72a..a0940b32e1 100644
--- a/libavcodec/rv40.c
+++ b/libavcodec/rv40.c
@@ -228,8 +228,6 @@ static int rv40_decode_mb_info(RV34DecContext *r)
     int q, i;
     int prev_type = 0;
     int mb_pos = s->mb_x + s->mb_y * s->mb_stride;
-    int blocks[RV34_MB_TYPES] = {0};
-    int count = 0;
 
     if(!r->s.mb_skip_run) {
         r->s.mb_skip_run = svq3_get_ue_golomb(gb) + 1;
@@ -240,22 +238,27 @@ static int rv40_decode_mb_info(RV34DecContext *r)
     if(--r->s.mb_skip_run)
          return RV34_MB_SKIP;
 
-    if(r->avail_cache[6-1])
-        blocks[r->mb_type[mb_pos - 1]]++;
     if(r->avail_cache[6-4]){
+        int blocks[RV34_MB_TYPES] = {0};
+        int count = 0;
+        if(r->avail_cache[6-1])
+            blocks[r->mb_type[mb_pos - 1]]++;
         blocks[r->mb_type[mb_pos - s->mb_stride]]++;
         if(r->avail_cache[6-2])
             blocks[r->mb_type[mb_pos - s->mb_stride + 1]]++;
         if(r->avail_cache[6-5])
             blocks[r->mb_type[mb_pos - s->mb_stride - 1]]++;
-    }
-
-    for(i = 0; i < RV34_MB_TYPES; i++){
-        if(blocks[i] > count){
-            count = blocks[i];
-            prev_type = i;
+        for(i = 0; i < RV34_MB_TYPES; i++){
+            if(blocks[i] > count){
+                count = blocks[i];
+                prev_type = i;
+                if(count>1)
+                    break;
+            }
         }
-    }
+    } else if (r->avail_cache[6-1])
+        prev_type = r->mb_type[mb_pos - 1];
+
     if(s->pict_type == AV_PICTURE_TYPE_P){
         prev_type = block_num_to_ptype_vlc_num[prev_type];
         q = get_vlc2(gb, ptype_vlc[prev_type].table, PTYPE_VLC_BITS, 1);
@@ -431,7 +434,7 @@ static void rv40_loop_filter(RV34DecContext *r, int row)
             y_v_deblock &= ~MASK_Y_LEFT_COL;
         if(!row)
             y_h_deblock &= ~MASK_Y_TOP_ROW;
-        if(row == s->mb_height - 1 || (mb_strong[POS_CUR] || mb_strong[POS_BOTTOM]))
+        if(row == s->mb_height - 1 || (mb_strong[POS_CUR] | mb_strong[POS_BOTTOM]))
             y_h_deblock &= ~(MASK_Y_TOP_ROW << 16);
         /* Calculating chroma patterns is similar and easier since there is
          * no motion vector pattern for them.
@@ -448,7 +451,7 @@ static void rv40_loop_filter(RV34DecContext *r, int row)
                 c_v_deblock[i] &= ~MASK_C_LEFT_COL;
             if(!row)
                 c_h_deblock[i] &= ~MASK_C_TOP_ROW;
-            if(row == s->mb_height - 1 || mb_strong[POS_CUR] || mb_strong[POS_BOTTOM])
+            if(row == s->mb_height - 1 || (mb_strong[POS_CUR] | mb_strong[POS_BOTTOM]))
                 c_h_deblock[i] &= ~(MASK_C_TOP_ROW << 4);
         }
 
@@ -469,7 +472,7 @@ static void rv40_loop_filter(RV34DecContext *r, int row)
                                               0, 0, 0);
                 }
                 // filter left block edge in ordinary mode (with low filtering strength)
-                if(y_v_deblock & (MASK_CUR << ij) && (i || !(mb_strong[POS_CUR] || mb_strong[POS_LEFT]))){
+                if(y_v_deblock & (MASK_CUR << ij) && (i || !(mb_strong[POS_CUR] | mb_strong[POS_LEFT]))){
                     if(!i)
                         clip_left = mvmasks[POS_LEFT] & (MASK_RIGHT << j) ? clip[POS_LEFT] : 0;
                     else
@@ -480,14 +483,14 @@ static void rv40_loop_filter(RV34DecContext *r, int row)
                                               alpha, beta, betaY, 0, 0, 1);
                 }
                 // filter top edge of the current macroblock when filtering strength is high
-                if(!j && y_h_deblock & (MASK_CUR << i) && (mb_strong[POS_CUR] || mb_strong[POS_TOP])){
+                if(!j && y_h_deblock & (MASK_CUR << i) && (mb_strong[POS_CUR] | mb_strong[POS_TOP])){
                     rv40_adaptive_loop_filter(&r->rdsp, Y, s->linesize, dither,
                                        clip_cur,
                                        mvmasks[POS_TOP] & (MASK_TOP << i) ? clip[POS_TOP] : 0,
                                        alpha, beta, betaY, 0, 1, 0);
                 }
                 // filter left block edge in edge mode (with high filtering strength)
-                if(y_v_deblock & (MASK_CUR << ij) && !i && (mb_strong[POS_CUR] || mb_strong[POS_LEFT])){
+                if(y_v_deblock & (MASK_CUR << ij) && !i && (mb_strong[POS_CUR] | mb_strong[POS_LEFT])){
                     clip_left = mvmasks[POS_LEFT] & (MASK_RIGHT << j) ? clip[POS_LEFT] : 0;
                     rv40_adaptive_loop_filter(&r->rdsp, Y, s->linesize, dither,
                                        clip_cur,
@@ -509,7 +512,7 @@ static void rv40_loop_filter(RV34DecContext *r, int row)
                                            clip_cur,
                                            alpha, beta, betaC, 1, 0, 0);
                     }
-                    if((c_v_deblock[k] & (MASK_CUR << ij)) && (i || !(mb_strong[POS_CUR] || mb_strong[POS_LEFT]))){
+                    if((c_v_deblock[k] & (MASK_CUR << ij)) && (i || !(mb_strong[POS_CUR] | mb_strong[POS_LEFT]))){
                         if(!i)
                             clip_left = uvcbp[POS_LEFT][k] & (MASK_CUR << (2*j+1)) ? clip[POS_LEFT] : 0;
                         else
@@ -519,14 +522,14 @@ static void rv40_loop_filter(RV34DecContext *r, int row)
                                            clip_left,
                                            alpha, beta, betaC, 1, 0, 1);
                     }
-                    if(!j && c_h_deblock[k] & (MASK_CUR << ij) && (mb_strong[POS_CUR] || mb_strong[POS_TOP])){
+                    if(!j && c_h_deblock[k] & (MASK_CUR << ij) && (mb_strong[POS_CUR] | mb_strong[POS_TOP])){
                         int clip_top = uvcbp[POS_TOP][k] & (MASK_CUR << (ij+2)) ? clip[POS_TOP] : 0;
                         rv40_adaptive_loop_filter(&r->rdsp, C, s->uvlinesize, i*8,
                                            clip_cur,
                                            clip_top,
                                            alpha, beta, betaC, 1, 1, 0);
                     }
-                    if(c_v_deblock[k] & (MASK_CUR << ij) && !i && (mb_strong[POS_CUR] || mb_strong[POS_LEFT])){
+                    if(c_v_deblock[k] & (MASK_CUR << ij) && !i && (mb_strong[POS_CUR] | mb_strong[POS_LEFT])){
                         clip_left = uvcbp[POS_LEFT][k] & (MASK_CUR << (2*j+1)) ? clip[POS_LEFT] : 0;
                         rv40_adaptive_loop_filter(&r->rdsp, C, s->uvlinesize, j*8,
                                            clip_cur,
diff --git a/libavcodec/x86/dsputil_mmx.c b/libavcodec/x86/dsputil_mmx.c
index 4c96d62525..68df954897 100644
--- a/libavcodec/x86/dsputil_mmx.c
+++ b/libavcodec/x86/dsputil_mmx.c
@@ -3043,7 +3043,7 @@ static void dsputil_init_sse2(DSPContext *c, AVCodecContext *avctx,
     const int bit_depth      = avctx->bits_per_raw_sample;
     const int high_bit_depth = bit_depth > 8;
 
-    if (mm_flags & AV_CPU_FLAG_3DNOW) {
+    if (!(mm_flags & AV_CPU_FLAG_SSE2SLOW)) {
         // these functions are slower than mmx on AMD, but faster on Intel
         if (!high_bit_depth) {
             c->put_pixels_tab[0][0]        = put_pixels16_sse2;
diff --git a/libavcodec/x86/rv34dsp.asm b/libavcodec/x86/rv34dsp.asm
index 1e152b7d68..ae04e82c2d 100644
--- a/libavcodec/x86/rv34dsp.asm
+++ b/libavcodec/x86/rv34dsp.asm
@@ -22,6 +22,16 @@
 %include "x86inc.asm"
 %include "x86util.asm"
 
+SECTION_RODATA
+pw_row_coeffs:  times 4 dw 13
+                times 4 dw 17
+                times 4 dw  7
+pd_512: times 2 dd 0x200
+pw_col_coeffs:  dw 13,  13,  13, -13
+                dw 17,   7,   7, -17
+                dw 13, -13,  13,  13
+                dw -7,  17, -17,  -7
+
 SECTION .text
 
 %macro IDCT_DC_NOROUND 1
@@ -88,6 +98,74 @@ cglobal rv34_idct_dc_add_mmx, 3, 3
     movh       [r2+r1], m5
     RET
 
+; Load coeffs and perform row transform
+; Output: coeffs in mm[0467], rounder in mm5
+%macro ROW_TRANSFORM  1
+    pxor        mm7, mm7
+    mova        mm0, [%1+ 0*8]
+    mova        mm1, [%1+ 1*8]
+    mova        mm2, [%1+ 2*8]
+    mova        mm3, [%1+ 3*8]
+    mova  [%1+ 0*8], mm7
+    mova  [%1+ 1*8], mm7
+    mova  [%1+ 2*8], mm7
+    mova  [%1+ 3*8], mm7
+    mova        mm4, mm0
+    mova        mm6, [pw_row_coeffs+ 0]
+    paddsw      mm0, mm2                ; b0 + b2
+    psubsw      mm4, mm2                ; b0 - b2
+    pmullw      mm0, mm6                ; *13 = z0
+    pmullw      mm4, mm6                ; *13 = z1
+    mova        mm5, mm1
+    pmullw      mm1, [pw_row_coeffs+ 8] ; b1*17
+    pmullw      mm5, [pw_row_coeffs+16] ; b1* 7
+    mova        mm7, mm3
+    pmullw      mm3, [pw_row_coeffs+ 8] ; b3*17
+    pmullw      mm7, [pw_row_coeffs+16] ; b3* 7
+    paddsw      mm1, mm7                ; z3 = b1*17 + b3* 7
+    psubsw      mm5, mm3                ; z2 = b1* 7 - b3*17
+    mova        mm7, mm0
+    mova        mm6, mm4
+    paddsw      mm0, mm1                ; z0 + z3
+    psubsw      mm7, mm1                ; z0 - z3
+    paddsw      mm4, mm5                ; z1 + z2
+    psubsw      mm6, mm5                ; z1 - z2
+    mova        mm5, [pd_512]           ; 0x200
+%endmacro
+
+; ff_rv34_idct_add_mmx2(uint8_t *dst, ptrdiff_t stride, DCTELEM *block);
+%macro COL_TRANSFORM  4
+    pshufw      mm3, %2, 0xDD        ; col. 1,3,1,3
+    pshufw       %2, %2, 0x88        ; col. 0,2,0,2
+    pmaddwd      %2, %3              ; 13*c0+13*c2 | 13*c0-13*c2 = z0 | z1
+    pmaddwd     mm3, %4              ; 17*c1+ 7*c3 |  7*c1-17*c3 = z3 | z2
+    paddd        %2, mm5
+    pshufw      mm1,  %2, 01001110b  ;    z1 | z0
+    pshufw      mm2, mm3, 01001110b  ;    z2 | z3
+    paddd        %2, mm3             ; z0+z3 | z1+z2
+    psubd       mm1, mm2             ; z1-z2 | z0-z3
+    movd        mm3, %1
+    psrad        %2, 10
+    pxor        mm2, mm2
+    psrad       mm1, 10
+    punpcklbw   mm3, mm2
+    packssdw     %2, mm1
+    paddw        %2, mm3
+    packuswb     %2, %2
+    movd         %1, %2
+%endmacro
+INIT_MMX mmx2
+cglobal rv34_idct_add, 3,3,0, d, s, b
+    ROW_TRANSFORM       bq
+    COL_TRANSFORM     [dq], mm0, [pw_col_coeffs+ 0], [pw_col_coeffs+ 8]
+    mova               mm0, [pw_col_coeffs+ 0]
+    COL_TRANSFORM  [dq+sq], mm4, mm0, [pw_col_coeffs+ 8]
+    mova               mm4, [pw_col_coeffs+ 8]
+    lea                 dq, [dq + 2*sq]
+    COL_TRANSFORM     [dq], mm6, mm0, mm4
+    COL_TRANSFORM  [dq+sq], mm7, mm0, mm4
+    ret
+
 ; ff_rv34_idct_dc_add_sse4(uint8_t *dst, int stride, int dc);
 INIT_XMM
 cglobal rv34_idct_dc_add_sse4, 3, 3, 6
diff --git a/libavcodec/x86/rv34dsp_init.c b/libavcodec/x86/rv34dsp_init.c
index 38831255c9..d91818c375 100644
--- a/libavcodec/x86/rv34dsp_init.c
+++ b/libavcodec/x86/rv34dsp_init.c
@@ -28,6 +28,7 @@ void ff_rv34_idct_dc_mmx2(DCTELEM *block);
 void ff_rv34_idct_dc_noround_mmx2(DCTELEM *block);
 void ff_rv34_idct_dc_add_mmx(uint8_t *dst, ptrdiff_t stride, int dc);
 void ff_rv34_idct_dc_add_sse4(uint8_t *dst, ptrdiff_t stride, int dc);
+void ff_rv34_idct_add_mmx2(uint8_t *dst, ptrdiff_t stride, DCTELEM *block);
 
 av_cold void ff_rv34dsp_init_x86(RV34DSPContext* c, DSPContext *dsp)
 {
@@ -38,6 +39,7 @@ av_cold void ff_rv34dsp_init_x86(RV34DSPContext* c, DSPContext *dsp)
         c->rv34_idct_dc_add = ff_rv34_idct_dc_add_mmx;
     if (mm_flags & AV_CPU_FLAG_MMX2) {
         c->rv34_inv_transform_dc = ff_rv34_idct_dc_noround_mmx2;
+        c->rv34_idct_add         = ff_rv34_idct_add_mmx2;
     }
     if (mm_flags & AV_CPU_FLAG_SSE4)
         c->rv34_idct_dc_add = ff_rv34_idct_dc_add_sse4;