[x265] [PATCH] asm: 10bpp avx2 code for intra_pred_ang32x32 mode 5 & 31

dnyaneshwar at multicorewareinc.com dnyaneshwar at multicorewareinc.com
Wed Jun 10 10:36:25 CEST 2015


# HG changeset patch
# User Dnyaneshwar G <dnyaneshwar at multicorewareinc.com>
# Date 1433914676 -19800
#      Wed Jun 10 11:07:56 2015 +0530
# Node ID 2188ac83f8f59c9e68bc90c540f3951b74f2a32a
# Parent  6245476add8f0562e3ccb657f572ff94fe96adf0
asm: 10bpp avx2 code for intra_pred_ang32x32 mode 5 & 31

performance improvement over SSE:
intra_ang_32x32[ 5]     8772c->4360c, 48%
intra_ang_32x32[31]     6308c->2923c, 53%

diff -r 6245476add8f -r 2188ac83f8f5 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Wed Jun 10 11:54:27 2015 +0530
+++ b/source/common/x86/asm-primitives.cpp	Wed Jun 10 11:07:56 2015 +0530
@@ -1259,6 +1259,8 @@
         p.cu[BLOCK_32x32].intra_pred[2]     = x265_intra_pred_ang32_2_avx2;
         p.cu[BLOCK_32x32].intra_pred[3]     = x265_intra_pred_ang32_3_avx2;
         p.cu[BLOCK_32x32].intra_pred[4]     = x265_intra_pred_ang32_4_avx2;
+        p.cu[BLOCK_32x32].intra_pred[5]     = x265_intra_pred_ang32_5_avx2;
+        p.cu[BLOCK_32x32].intra_pred[31]    = x265_intra_pred_ang32_31_avx2;
         p.cu[BLOCK_32x32].intra_pred[32]    = x265_intra_pred_ang32_32_avx2;
         p.cu[BLOCK_32x32].intra_pred[33]    = x265_intra_pred_ang32_33_avx2;
         p.cu[BLOCK_32x32].intra_pred[34]    = x265_intra_pred_ang32_2_avx2;
diff -r 6245476add8f -r 2188ac83f8f5 source/common/x86/intrapred.h
--- a/source/common/x86/intrapred.h	Wed Jun 10 11:54:27 2015 +0530
+++ b/source/common/x86/intrapred.h	Wed Jun 10 11:07:56 2015 +0530
@@ -278,6 +278,7 @@
 void x265_intra_pred_ang32_2_avx2(pixel* dst, intptr_t dstStride, const pixel* srcPix, int dirMode, int bFilter);
 void x265_intra_pred_ang32_3_avx2(pixel* dst, intptr_t dstStride, const pixel* srcPix, int dirMode, int bFilter);
 void x265_intra_pred_ang32_4_avx2(pixel* dst, intptr_t dstStride, const pixel* srcPix, int dirMode, int bFilter);
+void x265_intra_pred_ang32_5_avx2(pixel* dst, intptr_t dstStride, const pixel* srcPix, int dirMode, int bFilter);
 void x265_intra_pred_ang32_26_avx2(pixel* dst, intptr_t dstStride, const pixel* srcPix, int dirMode, int bFilter);
 void x265_intra_pred_ang32_27_avx2(pixel* dst, intptr_t dstStride, const pixel* srcPix, int dirMode, int bFilter);
 void x265_intra_pred_ang32_28_avx2(pixel* dst, intptr_t dstStride, const pixel* srcPix, int dirMode, int bFilter);
diff -r 6245476add8f -r 2188ac83f8f5 source/common/x86/intrapred16.asm
--- a/source/common/x86/intrapred16.asm	Wed Jun 10 11:54:27 2015 +0530
+++ b/source/common/x86/intrapred16.asm	Wed Jun 10 11:07:56 2015 +0530
@@ -13561,6 +13561,215 @@
 
     call        ang32_mode_4_32
     RET
+
+;; angle 32, modes 5 and 31
+cglobal ang32_mode_5_31
+    test            r6d, r6d
+
+    movu            m0, [r2 + 2]                    ; [16 15 14 13 12 11 10  9  8  7  6  5  4  3  2  1]
+    movu            m1, [r2 + 4]                    ; [17 16 15 14 13 12 11 10  9  8  7  6  5  4  3  2]
+
+    punpcklwd       m3, m0, m1                      ; [13 12 12 11 11 10 10  9  5  4  4  3  3  2  2  1]
+    punpckhwd       m0, m1                          ; [17 16 16 15 15 14 14 13  9  8  8  7  7  6  6  5]
+
+    movu            m1, [r2 + 18]                   ; [24 23 22 21 20 19 18 17 16 15 14 13 12 11 10  9]
+    movu            m4, [r2 + 20]                   ; [25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10]
+    punpcklwd       m2, m1, m4                      ; [21 20 20 19 19 18 18 17 13 12 12 11 11 10 10  9]
+    punpckhwd       m1, m4                          ; [25 24 24 23 23 22 22 21 17 16 16 15 15 14 14 13]
+
+    pmaddwd         m4, m3, [r3 - 15 * 32]          ; [1]
+    paddd           m4, [pd_16]
+    psrld           m4, 5
+    pmaddwd         m5, m0, [r3 - 15 * 32]
+    paddd           m5, [pd_16]
+    psrld           m5, 5
+    packusdw        m4, m5
+
+    pmaddwd         m5, m3, [r3 + 2 * 32]           ; [18]
+    paddd           m5, [pd_16]
+    psrld           m5, 5
+    pmaddwd         m8, m0, [r3 + 2 * 32]
+    paddd           m8, [pd_16]
+    psrld           m8, 5
+    packusdw        m5, m8
+
+    palignr         m7, m0, m3, 4
+    pmaddwd         m6, m7, [r3 - 13 * 32]          ; [3]
+    paddd           m6, [pd_16]
+    psrld           m6, 5
+    palignr         m8, m2, m0, 4
+    pmaddwd         m9, m8, [r3 - 13 * 32]
+    paddd           m9, [pd_16]
+    psrld           m9, 5
+    packusdw        m6, m9
+
+    pmaddwd         m7, [r3 + 4 * 32]               ; [20]
+    paddd           m7, [pd_16]
+    psrld           m7, 5
+    pmaddwd         m8, [r3 + 4 * 32]
+    paddd           m8, [pd_16]
+    psrld           m8, 5
+    packusdw        m7, m8
+
+    palignr         m9, m0, m3, 8
+    pmaddwd         m8, m9, [r3 - 11 * 32]          ; [5]
+    paddd           m8, [pd_16]
+    psrld           m8, 5
+    palignr         m10, m2, m0, 8
+    pmaddwd         m11, m10, [r3 - 11 * 32]
+    paddd           m11, [pd_16]
+    psrld           m11, 5
+    packusdw        m8, m11
+
+    pmaddwd         m9, [r3 + 6 * 32]               ; [22]
+    paddd           m9, [pd_16]
+    psrld           m9, 5
+    pmaddwd         m10, [r3 + 6 * 32]
+    paddd           m10, [pd_16]
+    psrld           m10, 5
+    packusdw        m9, m10
+
+    palignr         m11, m0, m3, 12
+    pmaddwd         m10, m11, [r3 - 9 * 32]         ; [7]
+    paddd           m10, [pd_16]
+    psrld           m10, 5
+    palignr         m12, m2, m0, 12
+    pmaddwd         m3, m12, [r3 - 9 * 32]
+    paddd           m3, [pd_16]
+    psrld           m3, 5
+    packusdw        m10, m3
+
+    pmaddwd         m11, [r3 + 8 * 32]              ; [24]
+    paddd           m11, [pd_16]
+    psrld           m11, 5
+    pmaddwd         m12, [r3 + 8 * 32]
+    paddd           m12, [pd_16]
+    psrld           m12, 5
+    packusdw        m11, m12
+
+    TRANSPOSE_STORE_AVX2 4, 5, 6, 7, 8, 9, 10, 11, 12, 3, 0
+
+    pmaddwd         m4, m0, [r3 - 7 * 32]           ; [9]
+    paddd           m4, [pd_16]
+    psrld           m4, 5
+    pmaddwd         m5, m2, [r3 - 7  * 32]
+    paddd           m5, [pd_16]
+    psrld           m5, 5
+    packusdw        m4, m5
+
+    pmaddwd         m5, m0, [r3 + 10 * 32]          ; [26]
+    paddd           m5, [pd_16]
+    psrld           m5, 5
+    pmaddwd         m3, m2, [r3 + 10 * 32]
+    paddd           m3, [pd_16]
+    psrld           m3, 5
+    packusdw        m5, m3
+
+    palignr         m7, m2, m0, 4
+    pmaddwd         m6, m7, [r3 - 5 * 32]           ; [11]
+    paddd           m6, [pd_16]
+    psrld           m6, 5
+    palignr         m8, m1, m2, 4
+    pmaddwd         m9, m8, [r3 - 5 * 32]
+    paddd           m9, [pd_16]
+    psrld           m9, 5
+    packusdw        m6, m9
+
+    pmaddwd         m7, [r3 + 12 * 32]              ; [28]
+    paddd           m7, [pd_16]
+    psrld           m7, 5
+    pmaddwd         m8, [r3 + 12 * 32]
+    paddd           m8, [pd_16]
+    psrld           m8, 5
+    packusdw        m7, m8
+
+    palignr         m9, m2, m0, 8
+    pmaddwd         m8, m9, [r3 - 3 * 32]           ; [13]
+    paddd           m8, [pd_16]
+    psrld           m8, 5
+    palignr         m3, m1, m2, 8
+    pmaddwd         m10, m3, [r3 - 3 * 32]
+    paddd           m10, [pd_16]
+    psrld           m10, 5
+    packusdw        m8, m10
+
+    pmaddwd         m9, [r3 + 14 * 32]              ; [30]
+    paddd           m9, [pd_16]
+    psrld           m9, 5
+    pmaddwd         m3, [r3 + 14 * 32]
+    paddd           m3, [pd_16]
+    psrld           m3, 5
+    packusdw        m9, m3
+
+    palignr         m10, m2, m0, 12
+    pmaddwd         m10, [r3 - 1 * 32]              ; [15]
+    paddd           m10, [pd_16]
+    psrld           m10, 5
+    palignr         m11, m1, m2, 12
+    pmaddwd         m11, [r3 - 1 * 32]
+    paddd           m11, [pd_16]
+    psrld           m11, 5
+    packusdw        m10, m11
+
+    pmaddwd         m2, [r3 - 16 * 32]              ; [0]
+    paddd           m2, [pd_16]
+    psrld           m2, 5
+    pmaddwd         m1, [r3 - 16 * 32]
+    paddd           m1, [pd_16]
+    psrld           m1, 5
+    packusdw        m2, m1
+    TRANSPOSE_STORE_AVX2 4, 5, 6, 7, 8, 9, 10, 2, 0, 1, 16
+    ret
+
+cglobal intra_pred_ang32_5, 3,8,13
+    add         r2,        128
+    xor         r6d,       r6d
+    lea         r3,        [ang_table_avx2 + 16 * 32]
+    add         r1d,       r1d
+    lea         r4,        [r1 * 3]
+    lea         r7,        [r0 + 8 * r1]
+
+    call        ang16_mode_5_31
+
+    add         r2,        18
+    lea         r0,        [r0 + 32]
+
+    call        ang32_mode_5_31
+
+    add         r2,        14
+    lea         r0,        [r7 + 8 * r1]
+
+    call        ang16_mode_5_31
+
+    add         r2,        18
+    lea         r0,        [r0 + 32]
+
+    call        ang32_mode_5_31
+    RET
+
+cglobal intra_pred_ang32_31, 3,7,13
+    xor         r6d,       r6d
+    inc         r6d
+    lea         r3,        [ang_table_avx2 + 16 * 32]
+    add         r1d,       r1d
+    lea         r4,        [r1 * 3]
+    lea         r5,        [r0 + 32]
+
+    call        ang16_mode_5_31
+
+    add         r2,        18
+
+    call        ang32_mode_5_31
+
+    add         r2,        14
+    mov         r0,        r5
+
+    call        ang16_mode_5_31
+
+    add         r2,        18
+
+    call        ang32_mode_5_31
+    RET
 ;-------------------------------------------------------------------------------------------------------
 ; end of avx2 code for intra_pred_ang32 mode 2 to 34
 ;-------------------------------------------------------------------------------------------------------


More information about the x265-devel mailing list