Merge remote branch 'origin/master' into experimental

Change-Id: I1a58ce4643377bae4cc6bf9c89320251f724ca66

8 files changed, 114 insertions, 97 deletions

diff --git a/vp8/encoder/encodeframe.c b/vp8/encoder/encodeframe.c
index d82a822e5..3a0ee5c31 100644
--- a/vp8/encoder/encodeframe.c
+++ b/vp8/encoder/encodeframe.c
@@ -1147,7 +1147,7 @@ static void sum_intra_stats(VP8_COMP *cpi, MACROBLOCK *x)
 }
 int vp8cx_encode_intra_macro_block(VP8_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t)
 {
-    int Error4x4, Error16x16, error_uv;
+    int Error4x4, Error16x16;
     int rate4x4, rate16x16, rateuv;
     int dist4x4, dist16x16, distuv;
     int rate = 0;
@@ -1160,7 +1160,7 @@ int vp8cx_encode_intra_macro_block(VP8_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t)
 #if !(CONFIG_REALTIME_ONLY)
     if (cpi->sf.RD && cpi->compressor_speed != 2)
     {
-        error_uv = vp8_rd_pick_intra_mbuv_mode(cpi, x, &rateuv, &rateuv_tokenonly, &distuv);
+        vp8_rd_pick_intra_mbuv_mode(cpi, x, &rateuv, &rateuv_tokenonly, &distuv);
         rate += rateuv;
 
         Error16x16 = vp8_rd_pick_intra16x16mby_mode(cpi, x, &rate16x16, &rate16x16_tokenonly, &dist16x16);
@@ -1231,7 +1231,6 @@ int vp8cx_encode_inter_macroblock
 )
 {
     MACROBLOCKD *const xd = &x->e_mbd;
-    int inter_error;
     int intra_error = 0;
     int rate;
     int distortion;
@@ -1258,7 +1257,7 @@ int vp8cx_encode_inter_macroblock
              * do not recalculate */
             cpi->zbin_mode_boost_enabled = 0;
         }
-        inter_error = vp8_rd_pick_inter_mode(cpi, x, recon_yoffset, recon_uvoffset, &rate, &distortion, &intra_error);
+        vp8_rd_pick_inter_mode(cpi, x, recon_yoffset, recon_uvoffset, &rate, &distortion, &intra_error);
 
         /* switch back to the regular quantizer for the encode */
         if (cpi->sf.improved_quant)
@@ -1272,10 +1271,9 @@ int vp8cx_encode_inter_macroblock
     }
     else
 #endif
-        inter_error = vp8_pick_inter_mode(cpi, x, recon_yoffset, recon_uvoffset, &rate, &distortion, &intra_error);
+        vp8_pick_inter_mode(cpi, x, recon_yoffset, recon_uvoffset, &rate, &distortion, &intra_error);
 
-
-    cpi->prediction_error += inter_error;
+    cpi->prediction_error += distortion;
     cpi->intra_error += intra_error;
 
 #if 0
diff --git a/vp8/encoder/pickinter.c b/vp8/encoder/pickinter.c
index 0edd806a2..0759e2d5b 100644
--- a/vp8/encoder/pickinter.c
+++ b/vp8/encoder/pickinter.c
@@ -263,7 +263,7 @@ int vp8_pick_intra4x4mby_modes(const VP8_ENCODER_RTCD *rtcd, MACROBLOCK *mb, int
     return error;
 }
 
-int vp8_pick_intra_mbuv_mode(MACROBLOCK *mb)
+void vp8_pick_intra_mbuv_mode(MACROBLOCK *mb)
 {
 
     MACROBLOCKD *x = &mb->e_mbd;
@@ -408,11 +408,10 @@ int vp8_pick_intra_mbuv_mode(MACROBLOCK *mb)
 
 
     mb->e_mbd.mode_info_context->mbmi.uv_mode = best_mode;
-    return best_error;
 
 }
 
-int vp8_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset, int recon_uvoffset, int *returnrate, int *returndistortion, int *returnintra)
+void vp8_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset, int recon_uvoffset, int *returnrate, int *returndistortion, int *returnintra)
 {
     BLOCK *b = &x->block[0];
     BLOCKD *d = &x->e_mbd.block[0];
@@ -504,7 +503,7 @@ int vp8_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset, int rec
 
     cpi->mbs_tested_so_far++;          // Count of the number of MBs tested so far this frame
 
-    *returnintra = best_intra_rd;
+    *returnintra = INT_MAX;
     x->skip = 0;
 
     ref_frame_cost[INTRA_FRAME]   = vp8_cost_zero(cpi->prob_intra_coded);
@@ -649,7 +648,7 @@ int vp8_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset, int rec
                 if (this_rd < best_intra_rd)
                 {
                     best_intra_rd = this_rd;
-                    *returnintra = best_intra_rd ;
+                    *returnintra = distortion2;
                 }
             }
 
@@ -673,9 +672,8 @@ int vp8_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset, int rec
             if (this_rd < best_intra_rd)
             {
                 best_intra_rd = this_rd;
-                *returnintra = best_intra_rd ;
+                *returnintra = distortion2;
             }
-
             break;
 
         case NEWMV:
@@ -934,8 +932,6 @@ int vp8_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset, int rec
         }
 
         x->e_mbd.mode_info_context->mbmi.mv.as_int = 0;
-
-        return best_rd;
     }
 
 
@@ -955,6 +951,4 @@ int vp8_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset, int rec
     }
 
     x->e_mbd.mode_info_context->mbmi.mv.as_mv = x->e_mbd.block[15].bmi.mv.as_mv;
-
-    return best_rd;
 }
diff --git a/vp8/encoder/pickinter.h b/vp8/encoder/pickinter.h
index af060d701..8fea98397 100644
--- a/vp8/encoder/pickinter.h
+++ b/vp8/encoder/pickinter.h
@@ -16,6 +16,6 @@
 
 #define RD_ESTIMATE(RM,DM,R,D) ( ((128+(R)*(RM)) >> 8) + (DM)*(D) )
 extern int vp8_pick_intra4x4mby_modes(const VP8_ENCODER_RTCD *, MACROBLOCK *mb, int *Rate, int *Distortion);
-extern int vp8_pick_intra_mbuv_mode(MACROBLOCK *mb);
-extern int vp8_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset, int recon_uvoffset, int *returnrate, int *returndistortion, int *returnintra);
+extern void vp8_pick_intra_mbuv_mode(MACROBLOCK *mb);
+extern void vp8_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset, int recon_uvoffset, int *returnrate, int *returndistortion, int *returnintra);
 #endif
diff --git a/vp8/encoder/rdopt.c b/vp8/encoder/rdopt.c
index cd8ac85ba..078837595 100644
--- a/vp8/encoder/rdopt.c
+++ b/vp8/encoder/rdopt.c
@@ -864,7 +864,7 @@ static int vp8_rd_inter_uv(VP8_COMP *cpi, MACROBLOCK *x, int *rate, int *distort
     return RDCOST(x->rdmult, x->rddiv, *rate, *distortion);
 }
 
-int vp8_rd_pick_intra_mbuv_mode(VP8_COMP *cpi, MACROBLOCK *x, int *rate, int *rate_tokenonly, int *distortion)
+void vp8_rd_pick_intra_mbuv_mode(VP8_COMP *cpi, MACROBLOCK *x, int *rate, int *rate_tokenonly, int *distortion)
 {
     MB_PREDICTION_MODE mode;
     MB_PREDICTION_MODE UNINITIALIZED_IS_SAFE(mode_selected);
@@ -907,7 +907,6 @@ int vp8_rd_pick_intra_mbuv_mode(VP8_COMP *cpi, MACROBLOCK *x, int *rate, int *ra
     *distortion = d;
 
     x->e_mbd.mode_info_context->mbmi.uv_mode = mode_selected;
-    return best_rd;
 }
 #endif
 
@@ -1256,7 +1255,12 @@ static void rd_check_segment(VP8_COMP *cpi, MACROBLOCK *x,
                     // Should we do a full search (best quality only)
                     if ((cpi->compressor_speed == 0) && (bestsme >> sseshift) > 4000)
                     {
-                        thissme = cpi->full_search_sad(x, c, e, bsi->mvp,
+                        MV full_mvp;
+
+                        full_mvp.row = bsi->mvp->row >>3;
+                        full_mvp.col = bsi->mvp->col >>3;
+
+                        thissme = cpi->full_search_sad(x, c, e, &full_mvp,
                                                        sadpb / 4, 16, v_fn_ptr, x->mvcost, bsi->ref_mv);
 
                         if (thissme < bestsme)
@@ -1787,7 +1791,7 @@ void vp8_cal_sad(VP8_COMP *cpi, MACROBLOCKD *xd, MACROBLOCK *x, int recon_yoffse
 }
 
 #if !(CONFIG_REALTIME_ONLY)
-int vp8_rd_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset, int recon_uvoffset, int *returnrate, int *returndistortion, int *returnintra)
+void vp8_rd_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset, int recon_uvoffset, int *returnrate, int *returndistortion, int *returnintra)
 {
     BLOCK *b = &x->block[0];
     BLOCKD *d = &x->e_mbd.block[0];
@@ -1806,7 +1810,8 @@ int vp8_rd_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset, int
     int mdcounts[4];
     int rate;
     int distortion;
-    int best_rd = INT_MAX; // 1 << 30;
+    int best_rd = INT_MAX;
+    int best_intra_rd = INT_MAX;
     int ref_frame_cost[MAX_REF_FRAMES];
     int rate2, distortion2;
     int uv_intra_rate, uv_intra_distortion, uv_intra_rate_tokenonly;
@@ -2423,9 +2428,12 @@ int vp8_rd_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset, int
         //all_rates[mode_index] = rate2;
         //all_dist[mode_index] = distortion2;
 
-        if ((x->e_mbd.mode_info_context->mbmi.ref_frame == INTRA_FRAME)  && (this_rd < *returnintra))
+        // Keep record of best intra distortion
+        if ((x->e_mbd.mode_info_context->mbmi.ref_frame == INTRA_FRAME) &&
+            (this_rd < best_intra_rd) )
         {
-            *returnintra = this_rd ;
+            best_intra_rd = this_rd;
+            *returnintra = distortion2 ;
         }
 
         // Did this mode help.. i.i is it the new best mode
@@ -2541,8 +2549,6 @@ int vp8_rd_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset, int
         }
 
         x->e_mbd.mode_info_context->mbmi.mv.as_int = 0;
-
-        return best_rd;
     }
 
 
@@ -2565,7 +2571,5 @@ int vp8_rd_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset, int
     }
 
     x->e_mbd.mode_info_context->mbmi.mv.as_mv = x->e_mbd.block[15].bmi.mv.as_mv;
-
-    return best_rd;
 }
 #endif
diff --git a/vp8/encoder/rdopt.h b/vp8/encoder/rdopt.h
index 72ba9a0b5..1d1be11a4 100644
--- a/vp8/encoder/rdopt.h
+++ b/vp8/encoder/rdopt.h
@@ -11,11 +11,11 @@
 
 #ifndef __INC_RDOPT_H
 #define __INC_RDOPT_H
-void vp8_initialize_rd_consts(VP8_COMP *cpi, int Qvalue);
-int vp8_rd_pick_intra4x4mby_modes(VP8_COMP *cpi, MACROBLOCK *mb, int *rate, int *rate_to, int *distortion, int best_rd);
-int vp8_rd_pick_intra16x16mby_mode(VP8_COMP *cpi, MACROBLOCK *x, int *returnrate, int *rate_to, int *returndistortion);
-int vp8_rd_pick_intra_mbuv_mode(VP8_COMP *cpi, MACROBLOCK *x, int *rate, int *rate_to, int *distortion);
-extern int vp8_rd_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset, int recon_uvoffset, int *returnrate, int *returndistortion, int *returnintra);
+extern void vp8_initialize_rd_consts(VP8_COMP *cpi, int Qvalue);
+extern int vp8_rd_pick_intra4x4mby_modes(VP8_COMP *cpi, MACROBLOCK *mb, int *rate, int *rate_to, int *distortion, int best_rd);
+extern int vp8_rd_pick_intra16x16mby_mode(VP8_COMP *cpi, MACROBLOCK *x, int *returnrate, int *rate_to, int *returndistortion);
+extern void vp8_rd_pick_intra_mbuv_mode(VP8_COMP *cpi, MACROBLOCK *x, int *rate, int *rate_to, int *distortion);
+extern void vp8_rd_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset, int recon_uvoffset, int *returnrate, int *returndistortion, int *returnintra);
 
 extern void vp8_mv_pred
 (
diff --git a/vp8/encoder/x86/quantize_ssse3.asm b/vp8/encoder/x86/quantize_ssse3.asm
index 2f33199e5..912007e02 100644
--- a/vp8/encoder/x86/quantize_ssse3.asm
+++ b/vp8/encoder/x86/quantize_ssse3.asm
@@ -9,38 +9,62 @@
 
 
 %include "vpx_ports/x86_abi_support.asm"
+%include "asm_enc_offsets.asm"
 
 
-;int vp8_fast_quantize_b_impl_ssse3(short *coeff_ptr
-;               short *qcoeff_ptr,short *dequant_ptr,
-;               short *round_ptr,
-;               short *quant_ptr, short *dqcoeff_ptr);
+; void vp8_fast_quantize_b_ssse3 | arg
+;  (BLOCK  *b,                   |  0
+;   BLOCKD *d)                   |  1
 ;
-global sym(vp8_fast_quantize_b_impl_ssse3)
-sym(vp8_fast_quantize_b_impl_ssse3):
+
+global sym(vp8_fast_quantize_b_ssse3)
+sym(vp8_fast_quantize_b_ssse3):
     push        rbp
     mov         rbp, rsp
-    SHADOW_ARGS_TO_STACK 6
     GET_GOT     rbx
+
+%if ABI_IS_32BIT
+    push        rdi
     push        rsi
+%else
+  %ifidn __OUTPUT_FORMAT__,x64
     push        rdi
+    push        rsi
+  %endif
+%endif
     ; end prolog
 
-    mov         rdx, arg(0)                 ;coeff_ptr
-    mov         rdi, arg(3)                 ;round_ptr
-    mov         rsi, arg(4)                 ;quant_ptr
-
-    movdqa      xmm0, [rdx]
-    movdqa      xmm4, [rdx + 16]
-
-    movdqa      xmm2, [rdi]                 ;round lo
-    movdqa      xmm3, [rdi + 16]            ;round hi
+%if ABI_IS_32BIT
+    mov         rdi, arg(0)                 ; BLOCK *b
+    mov         rsi, arg(1)                 ; BLOCKD *d
+%else
+  %ifidn __OUTPUT_FORMAT__,x64
+    mov         rdi, rcx                    ; BLOCK *b
+    mov         rsi, rdx                    ; BLOCKD *d
+  %else
+    ;mov         rdi, rdi                    ; BLOCK *b
+    ;mov         rsi, rsi                    ; BLOCKD *d
+  %endif
+%endif
+
+    mov         rax, [rdi + vp8_block_coeff]
+    mov         rcx, [rdi + vp8_block_round]
+    mov         rdx, [rdi + vp8_block_quant_fast]
+
+    ; coeff
+    movdqa      xmm0, [rax]
+    movdqa      xmm4, [rax + 16]
+
+    ; round
+    movdqa      xmm2, [rcx]
+    movdqa      xmm3, [rcx + 16]
 
     movdqa      xmm1, xmm0
     movdqa      xmm5, xmm4
 
-    psraw       xmm0, 15                    ;sign of z (aka sz)
-    psraw       xmm4, 15                    ;sign of z (aka sz)
+    ; sz = z >> 15
+    psraw       xmm0, 15
+    psraw       xmm4, 15
 
     pabsw       xmm1, xmm1
     pabsw       xmm5, xmm5
@@ -48,23 +72,24 @@ sym(vp8_fast_quantize_b_impl_ssse3):
     paddw       xmm1, xmm2
     paddw       xmm5, xmm3
 
-    pmulhw      xmm1, [rsi]
-    pmulhw      xmm5, [rsi + 16]
+    ; quant_fast
+    pmulhw      xmm1, [rdx]
+    pmulhw      xmm5, [rdx + 16]
 
-    mov         rdi, arg(1)                 ;qcoeff_ptr
-    mov         rcx, arg(2)                 ;dequant_ptr
-    mov         rsi, arg(5)                 ;dqcoeff_ptr
+    mov         rax, [rsi + vp8_blockd_qcoeff]
+    mov         rdi, [rsi + vp8_blockd_dequant]
+    mov         rcx, [rsi + vp8_blockd_dqcoeff]
 
     pxor        xmm1, xmm0
     pxor        xmm5, xmm4
     psubw       xmm1, xmm0
     psubw       xmm5, xmm4
 
-    movdqa      [rdi], xmm1
-    movdqa      [rdi + 16], xmm5
+    movdqa      [rax], xmm1
+    movdqa      [rax + 16], xmm5
 
-    movdqa      xmm2, [rcx]
-    movdqa      xmm3, [rcx + 16]
+    movdqa      xmm2, [rdi]
+    movdqa      xmm3, [rdi + 16]
 
     pxor        xmm4, xmm4
     pmullw      xmm2, xmm1
@@ -73,38 +98,37 @@ sym(vp8_fast_quantize_b_impl_ssse3):
     pcmpeqw     xmm1, xmm4                  ;non zero mask
     pcmpeqw     xmm5, xmm4                  ;non zero mask
     packsswb    xmm1, xmm5
-    pshufb      xmm1, [ GLOBAL(zz_shuf)]
+    pshufb      xmm1, [GLOBAL(zz_shuf)]
 
     pmovmskb    edx, xmm1
 
-;    xor         ecx, ecx
-;    mov         eax, -1
-;find_eob_loop:
-;    shr         edx, 1
-;    jc          fq_skip
-;    mov         eax, ecx
-;fq_skip:
-;    inc         ecx
-;    cmp         ecx, 16
-;    jne         find_eob_loop
     xor         rdi, rdi
     mov         eax, -1
     xor         dx, ax                      ;flip the bits for bsr
     bsr         eax, edx
 
-    movdqa      [rsi], xmm2                 ;store dqcoeff
-    movdqa      [rsi + 16], xmm3            ;store dqcoeff
+    movdqa      [rcx], xmm2                 ;store dqcoeff
+    movdqa      [rcx + 16], xmm3            ;store dqcoeff
 
     sub         edi, edx                    ;check for all zeros in bit mask
     sar         edi, 31                     ;0 or -1
     add         eax, 1
     and         eax, edi                    ;if the bit mask was all zero,
                                             ;then eob = 0
+    mov         [rsi + vp8_blockd_eob], eax
+
     ; begin epilog
+%if ABI_IS_32BIT
+    pop         rsi
     pop         rdi
+%else
+  %ifidn __OUTPUT_FORMAT__,x64
     pop         rsi
+    pop         rdi
+  %endif
+%endif
+
     RESTORE_GOT
-    UNSHADOW_ARGS
     pop         rbp
     ret
 
diff --git a/vp8/encoder/x86/quantize_x86.h b/vp8/encoder/x86/quantize_x86.h
index df2e0bc39..f09358061 100644
--- a/vp8/encoder/x86/quantize_x86.h
+++ b/vp8/encoder/x86/quantize_x86.h
@@ -19,7 +19,7 @@
  */
 #if HAVE_MMX
 
-#endif
+#endif /* HAVE_MMX */
 
 
 #if HAVE_SSE2
@@ -34,9 +34,21 @@ extern prototype_quantize_block(vp8_fast_quantize_b_sse2);
 #undef vp8_quantize_fastquantb
 #define vp8_quantize_fastquantb vp8_fast_quantize_b_sse2
 
-#endif
+#endif /* !CONFIG_RUNTIME_CPU_DETECT */
 
-#endif
+#endif /* HAVE_SSE2 */
 
 
-#endif
+#if HAVE_SSSE3
+extern prototype_quantize_block(vp8_fast_quantize_b_ssse3);
+
+#if !CONFIG_RUNTIME_CPU_DETECT
+
+#undef vp8_quantize_fastquantb
+#define vp8_quantize_fastquantb vp8_fast_quantize_b_ssse3
+
+#endif /* !CONFIG_RUNTIME_CPU_DETECT */
+
+#endif /* HAVE_SSSE3 */
+
+#endif /* QUANTIZE_X86_H */
diff --git a/vp8/encoder/x86/x86_csystemdependent.c b/vp8/encoder/x86/x86_csystemdependent.c
index 8bceecec4..8f2774b7a 100644
--- a/vp8/encoder/x86/x86_csystemdependent.c
+++ b/vp8/encoder/x86/x86_csystemdependent.c
@@ -112,21 +112,6 @@ static void subtract_b_sse2(BLOCK *be, BLOCKD *bd, int pitch)
 #endif
 
 #if HAVE_SSSE3
-int vp8_fast_quantize_b_impl_ssse3(short *coeff_ptr,
-                                 short *qcoeff_ptr, short *dequant_ptr,
-                                 short *round_ptr,
-                                 short *quant_ptr, short *dqcoeff_ptr);
-static void fast_quantize_b_ssse3(BLOCK *b, BLOCKD *d)
-{
-    d->eob = vp8_fast_quantize_b_impl_ssse3(
-                    b->coeff,
-                    d->qcoeff,
-                    d->dequant,
-                    b->round,
-                    b->quant_fast,
-                    d->dqcoeff
-               );
-}
 #if CONFIG_PSNR
 #if ARCH_X86_64
 typedef void ssimpf
@@ -307,7 +292,7 @@ void vp8_arch_x86_encoder_init(VP8_COMP *cpi)
         cpi->rtcd.variance.subpixvar16x8         = vp8_sub_pixel_variance16x8_ssse3;
         cpi->rtcd.variance.subpixvar16x16        = vp8_sub_pixel_variance16x16_ssse3;
 
-        cpi->rtcd.quantize.fastquantb            = fast_quantize_b_ssse3;
+        cpi->rtcd.quantize.fastquantb            = vp8_fast_quantize_b_ssse3;
 
 #if CONFIG_PSNR
 #if ARCH_X86_64