[x265] [PATCH 220 of 307] x86: AVX512 interp_8tap_vert_sp_32xN

mythreyi at multicorewareinc.com mythreyi at multicorewareinc.com
Sat Apr 7 04:33:38 CEST 2018


# HG changeset patch
# User Vignesh Vijayakumar<vignesh at multicorewareinc.com>
# Date 1511948305 -19800
#      Wed Nov 29 15:08:25 2017 +0530
# Node ID 9f2c4a0d09f3405f9c28cd3ebf229617c2278681
# Parent  3e2058cec6c6f4ad49d92f9df7fbc110a54f4b4b
x86: AVX512 interp_8tap_vert_sp_32xN

Size  |  AVX2 performance | AVX512 performance
----------------------------------------------
32x8  |      11.19x       |      21.10x
32x16 |      12.71x       |      21.91x
32x24 |      11.99x       |      22.26x
32x32 |      12.52x       |      22.42x
32x64 |      12.61x       |      22.54x

diff -r 3e2058cec6c6 -r 9f2c4a0d09f3 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Wed Nov 29 14:51:30 2017 +0530
+++ b/source/common/x86/asm-primitives.cpp	Wed Nov 29 15:08:25 2017 +0530
@@ -2881,6 +2881,11 @@
         p.pu[LUMA_64x48].luma_vsp = PFX(interp_8tap_vert_sp_64x48_avx512);
         p.pu[LUMA_64x64].luma_vsp = PFX(interp_8tap_vert_sp_64x64_avx512);
         p.pu[LUMA_48x64].luma_vsp = PFX(interp_8tap_vert_sp_48x64_avx512);
+        p.pu[LUMA_32x64].luma_vsp = PFX(interp_8tap_vert_sp_32x64_avx512);
+        p.pu[LUMA_32x32].luma_vsp = PFX(interp_8tap_vert_sp_32x32_avx512);
+        p.pu[LUMA_32x24].luma_vsp = PFX(interp_8tap_vert_sp_32x24_avx512);
+        p.pu[LUMA_32x16].luma_vsp = PFX(interp_8tap_vert_sp_32x16_avx512);
+        p.pu[LUMA_32x8].luma_vsp = PFX(interp_8tap_vert_sp_32x8_avx512);
 
         p.cu[BLOCK_8x8].dct = PFX(dct8_avx512);
         p.cu[BLOCK_8x8].idct = PFX(idct8_avx512);
diff -r 3e2058cec6c6 -r 9f2c4a0d09f3 source/common/x86/ipfilter8.asm
--- a/source/common/x86/ipfilter8.asm	Wed Nov 29 14:51:30 2017 +0530
+++ b/source/common/x86/ipfilter8.asm	Wed Nov 29 15:08:25 2017 +0530
@@ -246,9 +246,8 @@
 const interp4_hps_shuf_avx512,  dq 0, 4, 1, 5, 2, 6, 3, 7
 const interp4_hps_store_16xN_avx512,  dq 0, 2, 1, 3, 4, 6, 5, 7
 const interp8_hps_store_avx512,  dq 0, 1, 4, 5, 2, 3, 6, 7
-
+const interp8_vsp_store_avx512,  dq 0, 2, 4, 6, 1, 3, 5, 7
 SECTION .text
-
 cextern pb_128
 cextern pw_1
 cextern pw_32
@@ -13357,8 +13356,7 @@
     PROCESS_LUMA_VERT_SS_24x8_AVX512
     RET
 %endif
-
-%macro PROCESS_LUMA_VERT_SS_32x2_AVX512 0
+%macro PROCESS_LUMA_VERT_S_32x2_AVX512 1
     movu                 m1,                  [r0]                           ;0 row
     movu                 m3,                  [r0 + r1]                      ;1 row
     punpcklwd            m0,                  m1,                     m3
@@ -13409,23 +13407,38 @@
     pmaddwd              m14,                 m18
     punpckhwd            m12,                 m13
     pmaddwd              m12,                 m18
-
     paddd                m8,                  m14
     paddd                m4,                  m12
-    paddd                m0,                  m8
-    paddd                m1,                  m4
-
     movu                 m12,                 [r6 + 4 * r1]                 ; 8 row
     punpcklwd            m14,                 m13,                    m12
     pmaddwd              m14,                 m18
     punpckhwd            m13,                 m12
     pmaddwd              m13,                 m18
-
     paddd                m10,                 m14
     paddd                m11,                 m13
+
+    paddd                m0,                  m8
+    paddd                m1,                  m4
     paddd                m2,                  m10
     paddd                m3,                  m11
-
+%ifidn %1, sp
+    paddd                m0,                  m19
+    paddd                m1,                  m19
+    paddd                m2,                  m19
+    paddd                m3,                  m19
+
+    psrad                m0,                  12
+    psrad                m1,                  12
+    psrad                m2,                  12
+    psrad                m3,                  12
+
+    packssdw             m0,                  m1
+    packssdw             m2,                  m3
+    packuswb             m0,                  m2
+    vpermq               m0,                  m20,                   m0
+    movu                 [r2],                ym0
+    vextracti32x8        [r2 + r3],           m0,                    1
+%else
     psrad                m0,                  6
     psrad                m1,                  6
     psrad                m2,                  6
@@ -13433,18 +13446,17 @@
 
     packssdw             m0,                  m1
     packssdw             m2,                  m3
-
     movu                 [r2],                m0
     movu                 [r2 + r3],           m2
+%endif
 %endmacro
 ;-----------------------------------------------------------------------------------------------------------------
 ; void interp_8tap_vert(int16_t *src, intptr_t srcStride, int16_t *dst, intptr_t dstStride, int coeffIdx)
 ;-----------------------------------------------------------------------------------------------------------------
-%macro FILTER_VER_SS_LUMA_32xN_AVX512 1
+%macro FILTER_VER_S_LUMA_32xN_AVX512 2
 INIT_ZMM avx512
-cglobal interp_8tap_vert_ss_32x%1, 5, 8, 19
+cglobal interp_8tap_vert_%1_32x%2, 5, 8, 21
     add                   r1d,                r1d
-    add                   r3d,                r3d
     lea                   r7,                 [3 * r1]
     sub                   r0,                 r7
     shl                   r4d,                8
@@ -13461,26 +13473,36 @@
     mova                  m17,                [r5 + 2 * mmsize]
     mova                  m18,                [r5 + 3 * mmsize]
 %endif
-
-%rep %1/2 - 1
-    PROCESS_LUMA_VERT_SS_32x2_AVX512
+%ifidn %1, sp
+    vbroadcasti32x4       m19,                [pd_526336]
+    mova                  m20,                [interp8_vsp_store_avx512]
+%else
+    add                   r3d,                r3d
+%endif
+
+%rep %2/2 - 1
+    PROCESS_LUMA_VERT_S_32x2_AVX512 %1
     lea                   r0,                 [r0 + 2 * r1]
     lea                   r2,                 [r2 + 2 * r3]
 %endrep
-    PROCESS_LUMA_VERT_SS_32x2_AVX512
+    PROCESS_LUMA_VERT_S_32x2_AVX512 %1
     RET
 %endmacro
 
 %if ARCH_X86_64
-    FILTER_VER_SS_LUMA_32xN_AVX512 8
-    FILTER_VER_SS_LUMA_32xN_AVX512 16
-    FILTER_VER_SS_LUMA_32xN_AVX512 32
-    FILTER_VER_SS_LUMA_32xN_AVX512 24
-    FILTER_VER_SS_LUMA_32xN_AVX512 64
-%endif
-
+    FILTER_VER_S_LUMA_32xN_AVX512 ss, 8
+    FILTER_VER_S_LUMA_32xN_AVX512 ss, 16
+    FILTER_VER_S_LUMA_32xN_AVX512 ss, 32
+    FILTER_VER_S_LUMA_32xN_AVX512 ss, 24
+    FILTER_VER_S_LUMA_32xN_AVX512 ss, 64
+    FILTER_VER_S_LUMA_32xN_AVX512 sp, 8
+    FILTER_VER_S_LUMA_32xN_AVX512 sp, 16
+    FILTER_VER_S_LUMA_32xN_AVX512 sp, 32
+    FILTER_VER_S_LUMA_32xN_AVX512 sp, 24
+    FILTER_VER_S_LUMA_32xN_AVX512 sp, 64
+%endif
 %macro PROCESS_LUMA_VERT_SS_48x4_AVX512 0
-    PROCESS_LUMA_VERT_SS_32x2_AVX512
+    PROCESS_LUMA_VERT_S_32x2_AVX512 ss
     movu                 m1,                  [r0 + 2 * r1]
     movu                 m3,                  [r0 + r7]
     punpcklwd            m0,                  m1,                     m3


More information about the x265-devel mailing list