ref: f65170ea842dc8031bc431e9c5d37230820d3eed
parent: 9a780fa7db79b709787a9ca56fc324a118158da7
author: Jingning Han <[email protected]>
date: Thu Jan 12 10:06:30 EST 2017
Rework 8x8 transpose SSSE3 for inverse 2D-DCT Use same transpose process as inv_txfm_sse2 does. Change-Id: Ic4827825bd174cba57a0a80e19bf458a648e7d94
--- a/vpx_dsp/x86/inv_txfm_ssse3_x86_64.asm
+++ b/vpx_dsp/x86/inv_txfm_ssse3_x86_64.asm
@@ -10,10 +10,6 @@
%include "third_party/x86inc/x86inc.asm"
-; This file provides SSSE3 version of the inverse transformation. Part
-; of the functions are originally derived from the ffmpeg project.
-; Note that the current version applies to x86 64-bit only.
-
SECTION_RODATA
pw_11585x2: times 8 dw 23170
@@ -141,31 +137,44 @@
packssdw m%2, m%6
%endmacro
-; matrix transpose
-%macro INTERLEAVE_2X 4
- punpckh%1 m%4, m%2, m%3
- punpckl%1 m%2, m%3
- SWAP %3, %4
-%endmacro
+; 8x8 transpose. This follows same operations as SSE2 does.
+%macro TRANSPOSE8X8 10
+ ; stage 1
+ punpcklwd m%9, m%1, m%2
+ punpcklwd m%10, m%3, m%4
+ punpckhwd m%1, m%2
+ punpckhwd m%3, m%4
-%macro TRANSPOSE8X8 9
- INTERLEAVE_2X wd, %1, %2, %9
- INTERLEAVE_2X wd, %3, %4, %9
- INTERLEAVE_2X wd, %5, %6, %9
- INTERLEAVE_2X wd, %7, %8, %9
+ punpcklwd m%2, m%5, m%6
+ punpcklwd m%4, m%7, m%8
+ punpckhwd m%5, m%6
+ punpckhwd m%7, m%8
- INTERLEAVE_2X dq, %1, %3, %9
- INTERLEAVE_2X dq, %2, %4, %9
- INTERLEAVE_2X dq, %5, %7, %9
- INTERLEAVE_2X dq, %6, %8, %9
+ ; stage 2
+ punpckldq m%6, m%9, m%10
+ punpckldq m%8, m%1, m%3
+ punpckhdq m%9, m%10
+ punpckhdq m%1, m%3
- INTERLEAVE_2X qdq, %1, %5, %9
- INTERLEAVE_2X qdq, %3, %7, %9
- INTERLEAVE_2X qdq, %2, %6, %9
- INTERLEAVE_2X qdq, %4, %8, %9
+ punpckldq m%10, m%2, m%4
+ punpckldq m%3, m%5, m%7
+ punpckhdq m%2, m%4
+ punpckhdq m%5, m%7
- SWAP %2, %5
- SWAP %4, %7
+ ; stage 3
+ punpckhqdq m%4, m%9, m%2 ; out3
+ punpcklqdq m%9, m%2 ; out2
+ punpcklqdq m%7, m%1, m%5 ; out6
+ punpckhqdq m%1, m%5 ; out7
+
+ punpckhqdq m%2, m%6, m%10 ; out1
+ punpcklqdq m%6, m%10 ; out0
+ punpcklqdq m%5, m%8, m%3 ; out4
+ punpckhqdq m%8, m%3 ; out5
+
+ SWAP %6, %1
+ SWAP %3, %9
+ SWAP %8, %6
%endmacro
%macro IDCT8_1D 0
@@ -247,9 +256,9 @@
mova m6, [inputq + 96]
mova m7, [inputq + 112]
%endif
- TRANSPOSE8X8 0, 1, 2, 3, 4, 5, 6, 7, 9
+ TRANSPOSE8X8 0, 1, 2, 3, 4, 5, 6, 7, 9, 10
IDCT8_1D
- TRANSPOSE8X8 0, 1, 2, 3, 4, 5, 6, 7, 9
+ TRANSPOSE8X8 0, 1, 2, 3, 4, 5, 6, 7, 9, 10
IDCT8_1D
pxor m12, m12
@@ -821,7 +830,7 @@
mova m7, [r3 + 16 * 28]
%endif
- TRANSPOSE8X8 0, 1, 2, 3, 4, 5, 6, 7, 9
+ TRANSPOSE8X8 0, 1, 2, 3, 4, 5, 6, 7, 9, 10
IDCT32X32_34 16*0, 16*32, 16*64, 16*96
lea stp, [stp + 16 * 8]
@@ -843,7 +852,7 @@
mova m6, [r3 + 16 * 6]
mova m7, [r3 + 16 * 7]
- TRANSPOSE8X8 0, 1, 2, 3, 4, 5, 6, 7, 9
+ TRANSPOSE8X8 0, 1, 2, 3, 4, 5, 6, 7, 9, 10
IDCT32X32_34 16*0, 16*8, 16*16, 16*24
@@ -1250,7 +1259,7 @@
mova m6, [r3 + 16 * 24]
mova m7, [r3 + 16 * 28]
%endif
- TRANSPOSE8X8 0, 1, 2, 3, 4, 5, 6, 7, 9
+ TRANSPOSE8X8 0, 1, 2, 3, 4, 5, 6, 7, 9, 10
mova [r4 + 0], m0
mova [r4 + 16 * 1], m1
@@ -1299,7 +1308,7 @@
mova m6, [r3 + 16 * 6]
mova m7, [r3 + 16 * 7]
- TRANSPOSE8X8 0, 1, 2, 3, 4, 5, 6, 7, 9
+ TRANSPOSE8X8 0, 1, 2, 3, 4, 5, 6, 7, 9, 10
mova [r4 + 0], m0
mova [r4 + 16 * 1], m1
@@ -1715,7 +1724,7 @@
mova m7, [r3 + 16 * 28]
%endif
- TRANSPOSE8X8 0, 1, 2, 3, 4, 5, 6, 7, 9
+ TRANSPOSE8X8 0, 1, 2, 3, 4, 5, 6, 7, 9, 10
mova [r4 + 0], m0
mova [r4 + 16 * 1], m1
@@ -1764,7 +1773,7 @@
mova m6, [r3 + 16 * 6]
mova m7, [r3 + 16 * 7]
- TRANSPOSE8X8 0, 1, 2, 3, 4, 5, 6, 7, 9
+ TRANSPOSE8X8 0, 1, 2, 3, 4, 5, 6, 7, 9, 10
mova [r4 + 0], m0
mova [r4 + 16 * 1], m1