[x265] [PATCH] asm: avx2 code for intra_pred_ang32x32 mode 4 & 32, improved 8709c->4618c, 47%

dnyaneshwar at multicorewareinc.com dnyaneshwar at multicorewareinc.com
Tue Jun 9 13:43:35 CEST 2015


# HG changeset patch
# User Dnyaneshwar G <dnyaneshwar at multicorewareinc.com>
# Date 1433849586 -19800
#      Tue Jun 09 17:03:06 2015 +0530
# Node ID 7cfd60ed1696d523444aa3cfbf9827c34471b25e
# Parent  05df60c805d3f423db573885eb9f27b17dbc12a7
asm: avx2 code for intra_pred_ang32x32 mode 4 & 32, improved 8709c->4618c, 47%

diff -r 05df60c805d3 -r 7cfd60ed1696 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Tue Jun 09 12:31:29 2015 +0530
+++ b/source/common/x86/asm-primitives.cpp	Tue Jun 09 17:03:06 2015 +0530
@@ -1258,6 +1258,8 @@
 
         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[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 05df60c805d3 -r 7cfd60ed1696 source/common/x86/intrapred.h
--- a/source/common/x86/intrapred.h	Tue Jun 09 12:31:29 2015 +0530
+++ b/source/common/x86/intrapred.h	Tue Jun 09 17:03:06 2015 +0530
@@ -277,6 +277,7 @@
 void x265_intra_pred_ang32_34_avx2(pixel* dst, intptr_t dstStride, const pixel* srcPix, int dirMode, int bFilter);
 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_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 05df60c805d3 -r 7cfd60ed1696 source/common/x86/intrapred16.asm
--- a/source/common/x86/intrapred16.asm	Tue Jun 09 12:31:29 2015 +0530
+++ b/source/common/x86/intrapred16.asm	Tue Jun 09 17:03:06 2015 +0530
@@ -13344,6 +13344,223 @@
 
     call        ang16_mode_3_33
     RET
+
+;; angle 32, modes 4 and 32
+cglobal ang32_mode_4_32
+    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 - 13 * 32]          ; [5]
+    paddd           m4, [pd_16]
+    psrld           m4, 5
+    pmaddwd         m5, m0, [r3 - 13 * 32]
+    paddd           m5, [pd_16]
+    psrld           m5, 5
+    packusdw        m4, m5
+
+    pmaddwd         m5, m3, [r3 + 8 * 32]           ; [26]
+    paddd           m5, [pd_16]
+    psrld           m5, 5
+    pmaddwd         m8, m0, [r3 + 8 * 32]
+    paddd           m8, [pd_16]
+    psrld           m8, 5
+    packusdw        m5, m8
+
+    palignr         m6, m0, m3, 4                   ; [14 13 13 12 12 11 11 10  6  5  5  4  4  3  3  2]
+    pmaddwd         m6, [r3 - 3 * 32]               ; [15]
+    paddd           m6, [pd_16]
+    psrld           m6, 5
+    palignr         m7, m2, m0, 4                   ; [18 17 17 16 16 15 15 14 10  9  9  8  8  7  7  6]
+    pmaddwd         m7, [r3 - 3 * 32]
+    paddd           m7, [pd_16]
+    psrld           m7, 5
+    packusdw        m6, m7
+
+    palignr         m8, m0, m3, 8                   ; [15 14 14 13 13 12 12 11  7  6  6  5  5  4  4  3]
+    pmaddwd         m7, m8, [r3 - 14 * 32]              ; [4]
+    paddd           m7, [pd_16]
+    psrld           m7, 5
+    palignr         m9, m2, m0, 8                   ; [19 18 18 17 17 16 16 15 11 10 10  9  9  8  8  7]
+    pmaddwd         m10, m9, [r3 - 14 * 32]
+    paddd           m10, [pd_16]
+    psrld           m10, 5
+    packusdw        m7, m10
+
+    pmaddwd         m8, [r3 + 7 * 32]               ; [25]
+    paddd           m8, [pd_16]
+    psrld           m8, 5
+    pmaddwd         m9, [r3 + 7 * 32]
+    paddd           m9, [pd_16]
+    psrld           m9, 5
+    packusdw        m8, m9
+
+    palignr         m9, m0, m3, 12
+    pmaddwd         m9, [r3 - 4 * 32]               ; [14]
+    paddd           m9, [pd_16]
+    psrld           m9, 5
+    palignr         m3, m2, m0, 12
+    pmaddwd         m3, [r3 - 4 * 32]
+    paddd           m3, [pd_16]
+    psrld           m3, 5
+    packusdw        m9, m3
+
+    pmaddwd         m10, m0, [r3 - 15 * 32]         ; [3]
+    paddd           m10, [pd_16]
+    psrld           m10, 5
+    pmaddwd         m3, m2, [r3 - 15 * 32]
+    paddd           m3, [pd_16]
+    psrld           m3, 5
+    packusdw        m10, m3
+
+    pmaddwd         m11, m0, [r3 + 6 * 32]          ; [24]
+    paddd           m11, [pd_16]
+    psrld           m11, 5
+    pmaddwd         m3, m2, [r3 + 6 * 32]
+    paddd           m3, [pd_16]
+    psrld           m3, 5
+    packusdw        m11, m3
+
+    TRANSPOSE_STORE_AVX2 4, 5, 6, 7, 8, 9, 10, 11, 12, 3, 0
+
+    palignr         m4, m2, m0, 4
+    pmaddwd         m4, [r3 - 5* 32]                ; [13]
+    paddd           m4, [pd_16]
+    psrld           m4, 5
+    palignr         m5, m1, m2, 4
+    pmaddwd         m5, [r3 - 5  * 32]
+    paddd           m5, [pd_16]
+    psrld           m5, 5
+    packusdw        m4, m5
+
+    palignr         m6, m2, m0, 8
+    pmaddwd         m5, m6, [r3 - 16 * 32]          ; [2]
+    paddd           m5, [pd_16]
+    psrld           m5, 5
+    palignr         m7, m1, m2, 8
+    pmaddwd         m8, m7, [r3 - 16 * 32]
+    paddd           m8, [pd_16]
+    psrld           m8, 5
+    packusdw        m5, m8
+
+    pmaddwd         m6, [r3 + 5 * 32]               ; [23]
+    paddd           m6, [pd_16]
+    psrld           m6, 5
+    pmaddwd         m7, [r3 + 5 * 32]
+    paddd           m7, [pd_16]
+    psrld           m7, 5
+    packusdw        m6, m7
+
+    palignr         m7, m2, m0, 12
+    pmaddwd         m7, [r3 - 6 * 32]               ; [12]
+    paddd           m7, [pd_16]
+    psrld           m7, 5
+    palignr         m8, m1, m2, 12
+    pmaddwd         m8, [r3 - 6 * 32]
+    paddd           m8, [pd_16]
+    psrld           m8, 5
+    packusdw        m7, m8
+
+    movu            m0, [r2 + 34]                   ; [32 31 30 29 28 27 26 25 24 23 22 21 20 19 18 17]
+    pmaddwd         m8, m2, [r3 - 17 * 32]          ; [1]
+    paddd           m8, [pd_16]
+    psrld           m8, 5
+    pmaddwd         m9, m1, [r3 - 17 * 32]
+    paddd           m9, [pd_16]
+    psrld           m9, 5
+    packusdw        m8, m9
+
+    palignr         m3, m0, m0, 2                   ; [ x 32 31 30 29 28 27 26  x 24 23 22 21 20 19 18]
+    punpcklwd       m0, m3                          ; [29 29 28 28 27 27 26 22 21 20 20 19 19 18 18 17]
+
+    pmaddwd         m9, m2, [r3 + 4 * 32]           ; [22]
+    paddd           m9, [pd_16]
+    psrld           m9, 5
+    pmaddwd         m3, m1, [r3 + 4 * 32]
+    paddd           m3, [pd_16]
+    psrld           m3, 5
+    packusdw        m9, m3
+
+    palignr         m10, m1, m2, 4
+    pmaddwd         m10, [r3 - 7 * 32]              ; [11]
+    paddd           m10, [pd_16]
+    psrld           m10, 5
+    palignr         m11, m0, m1, 4
+    pmaddwd         m11, [r3 - 7 * 32]
+    paddd           m11, [pd_16]
+    psrld           m11, 5
+    packusdw        m10, m11
+
+    palignr         m3, m1, m2, 8
+    pmaddwd         m3, [r3 - 18 * 32]              ; [0]
+    paddd           m3, [pd_16]
+    psrld           m3, 5
+    palignr         m0, m1, 8
+    pmaddwd         m0, [r3 - 18 * 32]
+    paddd           m0, [pd_16]
+    psrld           m0, 5
+    packusdw        m3, m0
+    TRANSPOSE_STORE_AVX2 4, 5, 6, 7, 8, 9, 10, 3, 0, 1, 16
+    ret
+
+cglobal intra_pred_ang32_4, 3,8,13
+    add         r2,        128
+    xor         r6d,       r6d
+    lea         r3,        [ang_table_avx2 + 18 * 32]
+    add         r1d,       r1d
+    lea         r4,        [r1 * 3]
+    lea         r7,        [r0 + 8 * r1]
+
+    call        ang16_mode_4_32
+
+    add         r2,        22
+    lea         r0,        [r0 + 32]
+
+    call        ang32_mode_4_32
+
+    add         r2,        10
+    lea         r0,        [r7 + 8 * r1]
+
+    call        ang16_mode_4_32
+
+    add         r2,        22
+    lea         r0,        [r0 + 32]
+
+    call        ang32_mode_4_32
+    RET
+
+cglobal intra_pred_ang32_32, 3,7,13
+    xor         r6d,       r6d
+    inc         r6d
+    lea         r3,        [ang_table_avx2 + 18 * 32]
+    add         r1d,       r1d
+    lea         r4,        [r1 * 3]
+    lea         r5,        [r0 + 32]
+
+    call        ang16_mode_4_32
+
+    add         r2,        22
+
+    call        ang32_mode_4_32
+
+    add         r2,        10
+    mov         r0,        r5
+
+    call        ang16_mode_4_32
+
+    add         r2,        22
+
+    call        ang32_mode_4_32
+    RET
 ;-------------------------------------------------------------------------------------------------------
 ; end of avx2 code for intra_pred_ang32 mode 2 to 34
 ;-------------------------------------------------------------------------------------------------------


More information about the x265-devel mailing list