x86: AVX-512 mbtree_propagate_cost

Also make the AVX and AVX2 implementations slightly faster.

diff --git a/common/frame.c b/common/frame.c
index 2cbcf1e5ccdcdee0176a77e7e234183c7f7b6a91..a81e9b1071c50caa99304c47d814c81dce0a517e 100644 (file)
--- a/common/frame.c
+++ b/common/frame.c
@@ -223,11 +223,13 @@ static x264_frame_t *x264_frame_new( x264_t *h, int b_fdec )
                     PREALLOC( frame->lowres_mvs[j][i], 2*h->mb.i_mb_count*sizeof(int16_t) );
                     PREALLOC( frame->lowres_mv_costs[j][i], h->mb.i_mb_count*sizeof(int) );
                 }
-            PREALLOC( frame->i_propagate_cost, (i_mb_count+7) * sizeof(uint16_t) );
+            PREALLOC( frame->i_propagate_cost, i_mb_count * sizeof(uint16_t) );
             for( int j = 0; j <= h->param.i_bframe+1; j++ )
                 for( int i = 0; i <= h->param.i_bframe+1; i++ )
-                    PREALLOC( frame->lowres_costs[j][i], (i_mb_count+3) * sizeof(uint16_t) );
+                    PREALLOC( frame->lowres_costs[j][i], i_mb_count * sizeof(uint16_t) );
 
+            /* mbtree asm can overread the input buffers, make sure we don't read outside of allocated memory. */
+            prealloc_size += NATIVE_ALIGN;
         }
         if( h->param.rc.i_aq_mode )
         {

diff --git a/common/macroblock.c b/common/macroblock.c
index e5097a6d2559b843ec90ef8925179898eb6eeaea..0aa09de4ef34a314720718b554dc2e283b923d18 100644 (file)
--- a/common/macroblock.c
+++ b/common/macroblock.c
@@ -388,7 +388,7 @@ int x264_macroblock_thread_allocate( x264_t *h, int b_lookahead )
             ((me_range*2+24) * sizeof(int16_t) + (me_range+4) * (me_range+1) * 4 * sizeof(mvsad_t));
         scratch_size = X264_MAX3( buf_hpel, buf_ssim, buf_tesa );
     }
-    int buf_mbtree = h->param.rc.b_mb_tree * ((h->mb.i_mb_width+7)&~7) * sizeof(int16_t);
+    int buf_mbtree = h->param.rc.b_mb_tree * ((h->mb.i_mb_width+15)&~15) * sizeof(int16_t);
     scratch_size = X264_MAX( scratch_size, buf_mbtree );
     if( scratch_size )
         CHECKED_MALLOC( h->scratch_buffer, scratch_size );

diff --git a/common/x86/mc-a2.asm b/common/x86/mc-a2.asm
index b3cb46341bd92e6128d33907fda7f59c49ae83ab..e0b4a0cd818563b924ea0160f3db7d3a371b4471 100644 (file)
--- a/common/x86/mc-a2.asm
+++ b/common/x86/mc-a2.asm
@@ -2147,13 +2147,13 @@ MBTREE
 cglobal mbtree_propagate_cost, 6,6,8-2*cpuflag(avx2)
     vbroadcastss m5, [r5]
     mov         r5d, r6m
-    lea          r0, [r0+r5*2]
+    lea          r2, [r2+r5*2]
     add         r5d, r5d
-    add          r1, r5
-    add          r2, r5
-    add          r3, r5
     add          r4, r5
     neg          r5
+    sub          r1, r5
+    sub          r3, r5
+    sub          r0, r5
     mova        xm4, [pw_3fff]
 %if notcpuflag(avx2)
     pxor        xm7, xm7
@@ -2165,9 +2165,8 @@ cglobal mbtree_propagate_cost, 6,6,8-2*cpuflag(avx2)
     pmovzxwd     m2, [r1+r5]      ; prop
     pand        xm3, xm4, [r3+r5] ; inter
     pmovzxwd     m3, xm3
-    pminsd       m3, m0
     pmaddwd      m1, m0
-    psubd        m3, m0, m3
+    psubusw      m3, m0, m3
     cvtdq2ps     m0, m0
     cvtdq2ps     m1, m1
     cvtdq2ps     m2, m2
@@ -2184,7 +2183,7 @@ cglobal mbtree_propagate_cost, 6,6,8-2*cpuflag(avx2)
     movu        xm1, [r4+r5]
     movu        xm2, [r1+r5]
     pand        xm3, xm4, [r3+r5]
-    pminsw      xm3, xm0
+    psubusw     xm3, xm0, xm3
     INT16_UNPACK 0
     INT16_UNPACK 1
     INT16_UNPACK 2
@@ -2194,7 +2193,6 @@ cglobal mbtree_propagate_cost, 6,6,8-2*cpuflag(avx2)
     cvtdq2ps     m2, m2
     cvtdq2ps     m3, m3
     mulps        m1, m0
-    subps        m3, m0, m3
     mulps        m1, m5         ; intra*invq*fps_factor>>8
     addps        m1, m2         ; prop + (intra*invq*fps_factor>>8)
     rcpps        m2, m0         ; 1 / intra 1st approximation
@@ -2205,7 +2203,7 @@ cglobal mbtree_propagate_cost, 6,6,8-2*cpuflag(avx2)
     subps        m2, m0         ; 2nd approximation for 1/intra
     mulps        m1, m2         ; / intra
 %endif
-    vcvtps2dq    m1, m1
+    cvtps2dq     m1, m1
     vextractf128 xm2, m1, 1
     packssdw    xm1, xm2
     mova    [r0+r5], xm1
@@ -2219,6 +2217,39 @@ MBTREE_AVX
 INIT_YMM avx2
 MBTREE_AVX
 
+INIT_ZMM avx512
+cglobal mbtree_propagate_cost, 6,6
+    vbroadcastss  m5, [r5]
+    mov          r5d, 0x3fff3fff
+    vpbroadcastd ym4, r5d
+    mov          r5d, r6m
+    lea           r2, [r2+r5*2]
+    add          r5d, r5d
+    add           r1, r5
+    neg           r5
+    sub           r4, r5
+    sub           r3, r5
+    sub           r0, r5
+.loop:
+    pmovzxwd      m0, [r2+r5]      ; intra
+    pmovzxwd      m1, [r1+r5]      ; prop
+    pmovzxwd      m2, [r4+r5]      ; invq
+    pand         ym3, ym4, [r3+r5] ; inter
+    pmovzxwd      m3, ym3
+    psubusw       m3, m0, m3
+    cvtdq2ps      m0, m0
+    cvtdq2ps      m1, m1
+    cvtdq2ps      m2, m2
+    cvtdq2ps      m3, m3
+    vdivps        m1, m0, {rn-sae}
+    fmaddps       m1, m2, m5, m1
+    mulps         m1, m3
+    cvtps2dq      m1, m1
+    vpmovsdw [r0+r5], m1
+    add           r5, 32
+    jl .loop
+    RET
+
 %macro MBTREE_PROPAGATE_LIST 0
 ;-----------------------------------------------------------------------------
 ; void mbtree_propagate_list_internal( int16_t (*mvs)[2], int16_t *propagate_amount, uint16_t *lowres_costs,

diff --git a/common/x86/mc-c.c b/common/x86/mc-c.c
index 90a6cc198ca31e960e0d58e60c8eaf0b0f192b3c..f6e349a987df4498e6739abd8f5348150e07ee7f 100644 (file)
--- a/common/x86/mc-c.c
+++ b/common/x86/mc-c.c
@@ -160,14 +160,16 @@ void x264_integral_init4v_avx2( uint16_t *sum8, uint16_t *sum4, intptr_t stride
 void x264_integral_init8v_mmx ( uint16_t *sum8, intptr_t stride );
 void x264_integral_init8v_sse2( uint16_t *sum8, intptr_t stride );
 void x264_integral_init8v_avx2( uint16_t *sum8, intptr_t stride );
-void x264_mbtree_propagate_cost_sse2( int16_t *dst, uint16_t *propagate_in, uint16_t *intra_costs,
-                                      uint16_t *inter_costs, uint16_t *inv_qscales, float *fps_factor, int len );
-void x264_mbtree_propagate_cost_avx ( int16_t *dst, uint16_t *propagate_in, uint16_t *intra_costs,
-                                      uint16_t *inter_costs, uint16_t *inv_qscales, float *fps_factor, int len );
-void x264_mbtree_propagate_cost_fma4( int16_t *dst, uint16_t *propagate_in, uint16_t *intra_costs,
-                                      uint16_t *inter_costs, uint16_t *inv_qscales, float *fps_factor, int len );
-void x264_mbtree_propagate_cost_avx2( int16_t *dst, uint16_t *propagate_in, uint16_t *intra_costs,
-                                      uint16_t *inter_costs, uint16_t *inv_qscales, float *fps_factor, int len );
+void x264_mbtree_propagate_cost_sse2  ( int16_t *dst, uint16_t *propagate_in, uint16_t *intra_costs,
+                                        uint16_t *inter_costs, uint16_t *inv_qscales, float *fps_factor, int len );
+void x264_mbtree_propagate_cost_avx   ( int16_t *dst, uint16_t *propagate_in, uint16_t *intra_costs,
+                                        uint16_t *inter_costs, uint16_t *inv_qscales, float *fps_factor, int len );
+void x264_mbtree_propagate_cost_fma4  ( int16_t *dst, uint16_t *propagate_in, uint16_t *intra_costs,
+                                        uint16_t *inter_costs, uint16_t *inv_qscales, float *fps_factor, int len );
+void x264_mbtree_propagate_cost_avx2  ( int16_t *dst, uint16_t *propagate_in, uint16_t *intra_costs,
+                                        uint16_t *inter_costs, uint16_t *inv_qscales, float *fps_factor, int len );
+void x264_mbtree_propagate_cost_avx512( int16_t *dst, uint16_t *propagate_in, uint16_t *intra_costs,
+                                        uint16_t *inter_costs, uint16_t *inv_qscales, float *fps_factor, int len );
 void x264_mbtree_fix8_pack_ssse3( uint16_t *dst, float *src, int count );
 void x264_mbtree_fix8_pack_avx2 ( uint16_t *dst, float *src, int count );
 void x264_mbtree_fix8_unpack_ssse3( float *dst, uint16_t *src, int count );
@@ -864,4 +866,8 @@ void x264_mc_init_mmx( int cpu, x264_mc_functions_t *pf )
     pf->mbtree_propagate_list = x264_mbtree_propagate_list_avx2;
     pf->mbtree_fix8_pack      = x264_mbtree_fix8_pack_avx2;
     pf->mbtree_fix8_unpack    = x264_mbtree_fix8_unpack_avx2;
+
+    if( !(cpu&X264_CPU_AVX512) )
+        return;
+    pf->mbtree_propagate_cost = x264_mbtree_propagate_cost_avx512;
 }

diff --git a/tools/checkasm.c b/tools/checkasm.c
index 7575d3f4190d8c0267ffbe009aea77ff57d2a947..7ada27911fa16c724584d9a95dc3dd94b811d4ae 100644 (file)
--- a/tools/checkasm.c
+++ b/tools/checkasm.c
@@ -1743,7 +1743,7 @@ static int check_mc( int cpu_ref, int cpu_new )
             {
                 ok &= abs( dstc[j]-dsta[j] ) <= 1 || fabs( (double)dstc[j]/dsta[j]-1 ) < 1e-4;
                 if( !ok )
-                    fprintf( stderr, "mbtree_propagate_cost FAILED: %f !~= %f\n", (double)dstc[j], (double)dsta[j] );
+                    fprintf( stderr, "mbtree_propagate_cost FAILED: %d !~= %d\n", dstc[j], dsta[j] );
             }
         }
     }