[x265] [PATCH 4 of 6] asm: avx2 code for intra_pred_ang16x16 mode 7, 8, 28 & 29

dnyaneshwar at multicorewareinc.com dnyaneshwar at multicorewareinc.com
Thu May 28 08:36:21 CEST 2015


# HG changeset patch
# User Dnyaneshwar G <dnyaneshwar at multicorewareinc.com>
# Date 1432792824 -19800
#      Thu May 28 11:30:24 2015 +0530
# Node ID d5f08c2b2b765f64b10ca847b12bc7a715a528b7
# Parent  9b83bec91d48517119661cd5ee596115d6f17ae1
asm: avx2 code for intra_pred_ang16x16 mode 7,8,28 & 29

performance improvement over SSE:
intra_ang_16x16[ 7]     1851c->1206c, 35%
intra_ang_16x16[ 8]     1870c->1187c, 36%
intra_ang_16x16[28]     1422c->757c,  47%
intra_ang_16x16[29]     1415c->761c,  46%

diff -r 9b83bec91d48 -r d5f08c2b2b76 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Thu May 28 11:26:09 2015 +0530
+++ b/source/common/x86/asm-primitives.cpp	Thu May 28 11:30:24 2015 +0530
@@ -1201,6 +1201,10 @@
         p.cu[BLOCK_16x16].intra_pred[4]     = x265_intra_pred_ang16_4_avx2;
         p.cu[BLOCK_16x16].intra_pred[5]     = x265_intra_pred_ang16_5_avx2;
         p.cu[BLOCK_16x16].intra_pred[6]     = x265_intra_pred_ang16_6_avx2;
+        p.cu[BLOCK_16x16].intra_pred[7]     = x265_intra_pred_ang16_7_avx2;
+        p.cu[BLOCK_16x16].intra_pred[8]     = x265_intra_pred_ang16_8_avx2;
+        p.cu[BLOCK_16x16].intra_pred[28]    = x265_intra_pred_ang16_28_avx2;
+        p.cu[BLOCK_16x16].intra_pred[29]    = x265_intra_pred_ang16_29_avx2;
         p.cu[BLOCK_16x16].intra_pred[30]    = x265_intra_pred_ang16_30_avx2;
         p.cu[BLOCK_16x16].intra_pred[31]    = x265_intra_pred_ang16_31_avx2;
         p.cu[BLOCK_16x16].intra_pred[32]    = x265_intra_pred_ang16_32_avx2;
diff -r 9b83bec91d48 -r d5f08c2b2b76 source/common/x86/intrapred16.asm
--- a/source/common/x86/intrapred16.asm	Thu May 28 11:26:09 2015 +0530
+++ b/source/common/x86/intrapred16.asm	Thu May 28 11:30:24 2015 +0530
@@ -10962,6 +10962,312 @@
     TRANSPOSE_STORE_AVX2 4, 5, 6, 7, 8, 9, 10, 11, 0, 1, 16
     ret
 
+;; angle 16, modes 7 and 29
+cglobal ang16_mode_7_29
+    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            m2, [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, m4                          ; [21 20 20 19 19 18 18 17 13 12 12 11 11 10 10  9]
+
+    pmaddwd         m4, m3, [r3 - 8 * 32]           ; [9]
+    paddd           m4, [pd_16]
+    psrld           m4, 5
+    pmaddwd         m5, m0, [r3 - 8 * 32]
+    paddd           m5, [pd_16]
+    psrld           m5, 5
+    packusdw        m4, m5
+
+    pmaddwd         m5, m3, [r3 + 1 * 32]           ; [18]
+    paddd           m5, [pd_16]
+    psrld           m5, 5
+    pmaddwd         m8, m0, [r3 + 1 * 32]
+    paddd           m8, [pd_16]
+    psrld           m8, 5
+    packusdw        m5, m8
+
+    pmaddwd         m6, m3, [r3 + 10 * 32]          ; [27]
+    paddd           m6, [pd_16]
+    psrld           m6, 5
+    pmaddwd         m9, m0, [r3 + 10 * 32]
+    paddd           m9, [pd_16]
+    psrld           m9, 5
+    packusdw        m6, m9
+
+    palignr         m10, m0, m3, 4
+    pmaddwd         m7, m10, [r3 - 13 * 32]         ; [4]
+    paddd           m7, [pd_16]
+    psrld           m7, 5
+    palignr         m11, m2, m0, 4
+    pmaddwd         m8, m11, [r3 - 13 * 32]
+    paddd           m8, [pd_16]
+    psrld           m8, 5
+    packusdw        m7, m8
+
+    pmaddwd         m8, m10, [r3 - 4 * 32]          ; [13]
+    paddd           m8, [pd_16]
+    psrld           m8, 5
+    pmaddwd         m9, m11, [r3 - 4 * 32]
+    paddd           m9, [pd_16]
+    psrld           m9, 5
+    packusdw        m8, m9
+
+    pmaddwd         m9, m10, [r3 + 5 * 32]          ; [22]
+    paddd           m9, [pd_16]
+    psrld           m9, 5
+    pmaddwd         m12, m11, [r3 + 5 * 32]
+    paddd           m12, [pd_16]
+    psrld           m12, 5
+    packusdw        m9, m12
+
+    pmaddwd         m10, [r3 + 14 * 32]             ; [31]
+    paddd           m10, [pd_16]
+    psrld           m10, 5
+    pmaddwd         m11, [r3 + 14 * 32]
+    paddd           m11, [pd_16]
+    psrld           m11, 5
+    packusdw        m10, m11
+
+    palignr         m11, m0, m3, 8
+    pmaddwd         m11, [r3 - 9 * 32]              ; [8]
+    paddd           m11, [pd_16]
+    psrld           m11, 5
+    palignr         m12, m2, m0, 8
+    pmaddwd         m12, [r3 - 9 * 32]
+    paddd           m12, [pd_16]
+    psrld           m12, 5
+    packusdw        m11, m12
+
+    TRANSPOSE_STORE_AVX2 4, 5, 6, 7, 8, 9, 10, 11, 12, 1, 0
+
+    palignr         m5, m0, m3, 8
+    pmaddwd         m4, m5, [r3]                    ; [17]
+    paddd           m4, [pd_16]
+    psrld           m4, 5
+    palignr         m6, m2, m0, 8
+    pmaddwd         m7, m6, [r3]
+    paddd           m7, [pd_16]
+    psrld           m7, 5
+    packusdw        m4, m7
+
+    pmaddwd         m5, [r3 + 9 * 32]               ; [26]
+    paddd           m5, [pd_16]
+    psrld           m5, 5
+    pmaddwd         m6, [r3 + 9 * 32]
+    paddd           m6, [pd_16]
+    psrld           m6, 5
+    packusdw        m5, m6
+
+    palignr         m9, m0, m3, 12
+    pmaddwd         m6, m9, [r3 - 14 * 32]          ; [3]
+    paddd           m6, [pd_16]
+    psrld           m6, 5
+    palignr         m3, m2, m0, 12
+    pmaddwd         m7, m3, [r3 - 14 * 32]
+    paddd           m7, [pd_16]
+    psrld           m7, 5
+    packusdw        m6, m7
+
+    pmaddwd         m7, m9, [r3 - 5 * 32]           ; [12]
+    paddd           m7, [pd_16]
+    psrld           m7, 5
+    pmaddwd         m8, m3, [r3 - 5 * 32]
+    paddd           m8, [pd_16]
+    psrld           m8, 5
+    packusdw        m7, m8
+
+    pmaddwd         m8, m9, [r3 + 4 * 32]           ; [21]
+    paddd           m8, [pd_16]
+    psrld           m8, 5
+    pmaddwd         m10, m3, [r3 + 4 * 32]
+    paddd           m10, [pd_16]
+    psrld           m10, 5
+    packusdw        m8, m10
+
+    pmaddwd         m9, [r3 + 13 * 32]              ; [30]
+    paddd           m9, [pd_16]
+    psrld           m9, 5
+    pmaddwd         m3, [r3 + 13 * 32]
+    paddd           m3, [pd_16]
+    psrld           m3, 5
+    packusdw        m9, m3
+
+    pmaddwd         m10, m0, [r3 - 10 * 32]         ; [7]
+    paddd           m10, [pd_16]
+    psrld           m10, 5
+    pmaddwd         m12, m2, [r3 - 10 * 32]
+    paddd           m12, [pd_16]
+    psrld           m12, 5
+    packusdw        m10, m12
+
+    pmaddwd         m0, [r3 - 1 * 32]               ; [16]
+    paddd           m0, [pd_16]
+    psrld           m0, 5
+    pmaddwd         m2, [r3 - 1 * 32]
+    paddd           m2, [pd_16]
+    psrld           m2, 5
+    packusdw        m0, m2
+    TRANSPOSE_STORE_AVX2 4, 5, 6, 7, 8, 9, 10, 0, 1, 2, 16
+    ret
+
+;; angle 16, modes 8 and 28
+cglobal ang16_mode_8_28
+    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            m2, [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, m4                          ; [21 20 20 19 19 18 18 17 13 12 12 11 11 10 10  9]
+
+    pmaddwd         m4, m3, [r3 - 10 * 32]           ; [5]
+    paddd           m4, [pd_16]
+    psrld           m4, 5
+    pmaddwd         m5, m0, [r3 - 10 * 32]
+    paddd           m5, [pd_16]
+    psrld           m5, 5
+    packusdw        m4, m5
+
+    pmaddwd         m5, m3, [r3 - 5 * 32]           ; [10]
+    paddd           m5, [pd_16]
+    psrld           m5, 5
+    pmaddwd         m8, m0, [r3 - 5 * 32]
+    paddd           m8, [pd_16]
+    psrld           m8, 5
+    packusdw        m5, m8
+
+    pmaddwd         m6, m3, [r3]                    ; [15]
+    paddd           m6, [pd_16]
+    psrld           m6, 5
+    pmaddwd         m9, m0, [r3]
+    paddd           m9, [pd_16]
+    psrld           m9, 5
+    packusdw        m6, m9
+
+    pmaddwd         m7, m3, [r3 + 5 * 32]           ; [20]
+    paddd           m7, [pd_16]
+    psrld           m7, 5
+    pmaddwd         m8, m0, [r3 + 5 * 32]
+    paddd           m8, [pd_16]
+    psrld           m8, 5
+    packusdw        m7, m8
+
+    pmaddwd         m8, m3, [r3 + 10 * 32]          ; [25]
+    paddd           m8, [pd_16]
+    psrld           m8, 5
+    pmaddwd         m9, m0, [r3 + 10 * 32]
+    paddd           m9, [pd_16]
+    psrld           m9, 5
+    packusdw        m8, m9
+
+    pmaddwd         m9, m3, [r3 + 15 * 32]          ; [30]
+    paddd           m9, [pd_16]
+    psrld           m9, 5
+    pmaddwd         m10, m0, [r3 + 15 * 32]
+    paddd           m10, [pd_16]
+    psrld           m10, 5
+    packusdw        m9, m10
+
+    palignr         m11, m0, m3, 4
+    pmaddwd         m10, m11, [r3 - 12 * 32]        ; [3]
+    paddd           m10, [pd_16]
+    psrld           m10, 5
+    palignr         m1, m2, m0, 4
+    pmaddwd         m12, m1, [r3 - 12 * 32]
+    paddd           m12, [pd_16]
+    psrld           m12, 5
+    packusdw        m10, m12
+
+    pmaddwd         m11, [r3 - 7 * 32]              ; [8]
+    paddd           m11, [pd_16]
+    psrld           m11, 5
+    pmaddwd         m1, [r3 - 7 * 32]
+    paddd           m1, [pd_16]
+    psrld           m1, 5
+    packusdw        m11, m1
+
+    TRANSPOSE_STORE_AVX2 4, 5, 6, 7, 8, 9, 10, 11, 12, 1, 0
+
+    palignr         m7, m0, m3, 4
+    pmaddwd         m4, m7, [r3 - 2 * 32]           ; [13]
+    paddd           m4, [pd_16]
+    psrld           m4, 5
+    palignr         m1, m2, m0, 4
+    pmaddwd         m5, m1, [r3 - 2 * 32]
+    paddd           m5, [pd_16]
+    psrld           m5, 5
+    packusdw        m4, m5
+
+    pmaddwd         m5, m7, [r3 + 3 * 32]           ; [18]
+    paddd           m5, [pd_16]
+    psrld           m5, 5
+    pmaddwd         m6, m1, [r3 + 3 * 32]
+    paddd           m6, [pd_16]
+    psrld           m6, 5
+    packusdw        m5, m6
+
+    pmaddwd         m6, m7, [r3 + 8 * 32]           ; [23]
+    paddd           m6, [pd_16]
+    psrld           m6, 5
+    pmaddwd         m8, m1, [r3 + 8 * 32]
+    paddd           m8, [pd_16]
+    psrld           m8, 5
+    packusdw        m6, m8
+
+    pmaddwd         m7, [r3 + 13 * 32]              ; [28]
+    paddd           m7, [pd_16]
+    psrld           m7, 5
+    pmaddwd         m1, [r3 + 13 * 32]
+    paddd           m1, [pd_16]
+    psrld           m1, 5
+    packusdw        m7, m1
+
+    palignr         m1, m0, m3, 8
+    pmaddwd         m8, m1, [r3 - 14 * 32]          ; [1]
+    paddd           m8, [pd_16]
+    psrld           m8, 5
+    palignr         m2, m0, 8
+    pmaddwd         m9, m2, [r3 - 14 * 32]
+    paddd           m9, [pd_16]
+    psrld           m9, 5
+    packusdw        m8, m9
+
+    pmaddwd         m9, m1, [r3 - 9 * 32]           ; [6]
+    paddd           m9, [pd_16]
+    psrld           m9, 5
+    pmaddwd         m3, m2, [r3 - 9 * 32]
+    paddd           m3, [pd_16]
+    psrld           m3, 5
+    packusdw        m9, m3
+
+    pmaddwd         m3, m1, [r3 - 4 * 32]           ; [11]
+    paddd           m3, [pd_16]
+    psrld           m3, 5
+    pmaddwd         m0, m2, [r3 - 4 * 32]
+    paddd           m0, [pd_16]
+    psrld           m0, 5
+    packusdw        m3, m0
+
+    pmaddwd         m1, [r3 + 1 * 32]               ; [16]
+    paddd           m1, [pd_16]
+    psrld           m1, 5
+    pmaddwd         m2, [r3 + 1 * 32]
+    paddd           m2, [pd_16]
+    psrld           m2, 5
+    packusdw        m1, m2
+    TRANSPOSE_STORE_AVX2 4, 5, 6, 7, 8, 9, 3, 1, 0, 2, 16
+    ret
+
 cglobal intra_pred_ang16_3, 3,7,13
     add         r2,        64
     xor         r6d,       r6d
@@ -11052,6 +11358,46 @@
     call        ang16_mode_6_30
     RET
 
+cglobal intra_pred_ang16_7, 3,7,13
+    add         r2,        64
+    xor         r6d,       r6d
+    lea         r3,        [ang_table_avx2 + 17 * 32]
+    add         r1d,       r1d
+    lea         r4,        [r1 * 3]
+
+    call        ang16_mode_7_29
+    RET
+
+cglobal intra_pred_ang16_29, 3,7,13
+    xor         r6d,       r6d
+    inc         r6d
+    lea         r3,        [ang_table_avx2 + 17 * 32]
+    add         r1d,       r1d
+    lea         r4,        [r1 * 3]
+
+    call        ang16_mode_7_29
+    RET
+
+cglobal intra_pred_ang16_8, 3,7,13
+    add         r2,        64
+    xor         r6d,       r6d
+    lea         r3,        [ang_table_avx2 + 15 * 32]
+    add         r1d,       r1d
+    lea         r4,        [r1 * 3]
+
+    call        ang16_mode_8_28
+    RET
+
+cglobal intra_pred_ang16_28, 3,7,13
+    xor         r6d,       r6d
+    inc         r6d
+    lea         r3,        [ang_table_avx2 + 15 * 32]
+    add         r1d,       r1d
+    lea         r4,        [r1 * 3]
+
+    call        ang16_mode_8_28
+    RET
+
 ;-------------------------------------------------------------------------------------------------------
 ; end of avx2 code for intra_pred_ang16 mode 2 to 34
 ;-------------------------------------------------------------------------------------------------------


More information about the x265-devel mailing list