x86: AVX-512 pixel_var2_8x8 and 8x16

common/common.h

@@ -778,8 +778,8 @@ struct x264_t
             /* space for p_fenc and p_fdec */
 #define FENC_STRIDE 16
 #define FDEC_STRIDE 32
-            ALIGNED_32( pixel fenc_buf[48*FENC_STRIDE] );
-            ALIGNED_32( pixel fdec_buf[52*FDEC_STRIDE] );
+            ALIGNED_64( pixel fenc_buf[48*FENC_STRIDE] );
+            ALIGNED_64( pixel fdec_buf[52*FDEC_STRIDE] );
 
             /* i4x4 and i8x8 backup data, for skipping the encode stage when possible */
             ALIGNED_16( pixel i4x4_fdec_buf[16*16] );

common/macroblock.c

@@ -532,16 +532,17 @@ void x264_macroblock_thread_init( x264_t *h )
     h->mb.pic.p_fenc[0] = h->mb.pic.fenc_buf;
     h->mb.pic.p_fdec[0] = h->mb.pic.fdec_buf + 2*FDEC_STRIDE;
     h->mb.pic.p_fenc[1] = h->mb.pic.fenc_buf + 16*FENC_STRIDE;
-    h->mb.pic.p_fdec[1] = h->mb.pic.fdec_buf + 19*FDEC_STRIDE;
     if( CHROMA444 )
     {
         h->mb.pic.p_fenc[2] = h->mb.pic.fenc_buf + 32*FENC_STRIDE;
+        h->mb.pic.p_fdec[1] = h->mb.pic.fdec_buf + 19*FDEC_STRIDE;
         h->mb.pic.p_fdec[2] = h->mb.pic.fdec_buf + 36*FDEC_STRIDE;
     }
     else
     {
         h->mb.pic.p_fenc[2] = h->mb.pic.fenc_buf + 16*FENC_STRIDE + 8;
-        h->mb.pic.p_fdec[2] = h->mb.pic.fdec_buf + 19*FDEC_STRIDE + 16;
+        h->mb.pic.p_fdec[1] = h->mb.pic.fdec_buf + 20*FDEC_STRIDE;
+        h->mb.pic.p_fdec[2] = h->mb.pic.fdec_buf + 20*FDEC_STRIDE + 16;
     }
 }
 

common/pixel.c

@@ -1049,6 +1049,8 @@ void x264_pixel_init( int cpu, x264_pixel_function_t *pixf )
     {
         pixf->var[PIXEL_8x16]  = x264_pixel_var_8x16_avx512;
         pixf->var[PIXEL_16x16] = x264_pixel_var_16x16_avx512;
+        pixf->var2[PIXEL_8x8]  = x264_pixel_var2_8x8_avx512;
+        pixf->var2[PIXEL_8x16] = x264_pixel_var2_8x16_avx512;
     }
 #endif // HAVE_MMX
 #else // !HIGH_BIT_DEPTH
@@ -1351,6 +1353,8 @@ void x264_pixel_init( int cpu, x264_pixel_function_t *pixf )
         pixf->var[PIXEL_8x8]   = x264_pixel_var_8x8_avx512;
         pixf->var[PIXEL_8x16]  = x264_pixel_var_8x16_avx512;
         pixf->var[PIXEL_16x16] = x264_pixel_var_16x16_avx512;
+        pixf->var2[PIXEL_8x8]  = x264_pixel_var2_8x8_avx512;
+        pixf->var2[PIXEL_8x16] = x264_pixel_var2_8x16_avx512;
     }
 #endif //HAVE_MMX
 

common/x86/pixel-a.asm

@@ -1128,10 +1128,17 @@ VAR2_8x8_SSSE3 16, 7
 
 %macro VAR2_AVX2_LOAD 3 ; offset_reg, row1_offset, row2_offset
 %if HIGH_BIT_DEPTH
+%if mmsize == 64
+    mova        m2, [r1+2*%1+%2*FDEC_STRIDEB]
+    vshufi32x4  m2, [r1+2*%1+%2*FDEC_STRIDEB+64], q2020
+    mova        m3, [r1+2*%1+%3*FDEC_STRIDEB]
+    vshufi32x4  m3, [r1+2*%1+%3*FDEC_STRIDEB+64], q2020
+%else
     mova       xm2, [r1+2*%1+%2*FDEC_STRIDEB]
     vinserti128 m2, [r1+2*%1+%2*FDEC_STRIDEB+32], 1
     mova       xm3, [r1+2*%1+%3*FDEC_STRIDEB]
     vinserti128 m3, [r1+2*%1+%3*FDEC_STRIDEB+32], 1
+%endif
     psubw       m2, [r0+1*%1+%2*FENC_STRIDEB]
     psubw       m3, [r0+1*%1+%3*FENC_STRIDEB]
 %else
@@ -1174,6 +1181,44 @@ INIT_YMM avx2
 VAR2_8x8_AVX2  8, 6
 VAR2_8x8_AVX2 16, 7
 
+%macro VAR2_AVX512_END 1 ; shift
+    vbroadcasti32x4 m2, [pw_1]
+    pmaddwd         m0, m2
+    SBUTTERFLY     qdq, 0, 1, 2
+    paddd           m0, m1
+    vextracti32x8  ym1, m0, 1
+    paddd          ym0, ym1
+    psrlq          ym1, ym0, 32
+    paddd          ym0, ym1
+    vpmovqd       xmm0, ym0 ; sum_u, sqr_u, sum_v, sqr_v
+    VAR2_END      xmm0, xmm1, %1
+%endmacro
+
+INIT_ZMM avx512
+cglobal pixel_var2_8x8, 2,3
+%if HIGH_BIT_DEPTH == 0
+    pxor          xm6, xm6
+%endif
+    VAR2_AVX2_LOAD  0, 0, 2
+    VAR2_CORE      m2, m3, 0
+    VAR2_AVX2_LOAD  0, 4, 6
+    VAR2_CORE      m2, m3, 1
+    VAR2_AVX512_END 6
+
+cglobal pixel_var2_8x16, 2,3
+%if HIGH_BIT_DEPTH == 0
+    pxor          xm6, xm6
+%endif
+    mov           t0d, 10*FENC_STRIDEB
+    VAR2_AVX2_LOAD  0, 14, 12
+    VAR2_CORE      m2, m3, 0
+.loop:
+    VAR2_AVX2_LOAD t0, 0, -2
+    VAR2_CORE      m2, m3, 1
+    sub           t0d, 4*FENC_STRIDEB
+    jg .loop
+    VAR2_AVX512_END 7
+
 ;=============================================================================
 ; SATD
 ;=============================================================================

common/x86/pixel.h

@@ -169,9 +169,11 @@ float x264_pixel_ssim_end4_avx ( int sum0[5][4], int sum1[5][4], int width );
 int  x264_pixel_var2_8x8_sse2   ( pixel   *fenc, pixel   *fdec, int ssd[2] );
 int  x264_pixel_var2_8x8_ssse3  ( uint8_t *fenc, uint8_t *fdec, int ssd[2] );
 int  x264_pixel_var2_8x8_avx2   ( pixel   *fenc, pixel   *fdec, int ssd[2] );
+int  x264_pixel_var2_8x8_avx512 ( pixel   *fenc, pixel   *fdec, int ssd[2] );
 int  x264_pixel_var2_8x16_sse2  ( pixel   *fenc, pixel   *fdec, int ssd[2] );
 int  x264_pixel_var2_8x16_ssse3 ( uint8_t *fenc, uint8_t *fdec, int ssd[2] );
 int  x264_pixel_var2_8x16_avx2  ( pixel   *fenc, pixel   *fdec, int ssd[2] );
+int  x264_pixel_var2_8x16_avx512( pixel   *fenc, pixel   *fdec, int ssd[2] );
 int  x264_pixel_vsad_mmx2 ( pixel *src, intptr_t stride, int height );
 int  x264_pixel_vsad_sse2 ( pixel *src, intptr_t stride, int height );
 int  x264_pixel_vsad_ssse3( pixel *src, intptr_t stride, int height );