generic-library/vpx
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity
b00057c88a
vpx/vp9/common/x86/vp9_subpixel_8t_ssse3.asm

1042 lines
25 KiB
NASM
Raw Normal View History

Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
;
; Copyright (c) 2010 The WebM project authors. All Rights Reserved.
;
; Use of this source code is governed by a BSD-style license
; that can be found in the LICENSE file in the root of the source
; tree. An additional intellectual property rights grant can be found
; in the file PATENTS. All contributing project authors may
; be found in the AUTHORS file in the root of the source tree.
;
%include "vpx_ports/x86_abi_support.asm"
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
%macro VERTx4 1
Bug fix: ssse3 version of subpixel did not match C code A 16 bit overflow condition occurs when using the EIGHTTAP_SMOOTH filters. (vp9_sub_pel_filters_8lp) Changed the order of the adds to fix this problem. Also added ssse3 support for 4x4 subpixel filtering. Change-Id: I475eaadae920794c2de5e01e9735c059a856518e
2013-02-10 00:15:14 +01:00
mov rdx, arg(5) ;filter ptr
mov rsi, arg(0) ;src_ptr
mov rdi, arg(2) ;output_ptr
mov rcx, 0x0400040
movdqa xmm4, [rdx] ;load filters
movd xmm5, rcx
packsswb xmm4, xmm4
pshuflw xmm0, xmm4, 0b ;k0_k1
pshuflw xmm1, xmm4, 01010101b ;k2_k3
pshuflw xmm2, xmm4, 10101010b ;k4_k5
pshuflw xmm3, xmm4, 11111111b ;k6_k7
punpcklqdq xmm0, xmm0
punpcklqdq xmm1, xmm1
punpcklqdq xmm2, xmm2
punpcklqdq xmm3, xmm3
movdqa k0k1, xmm0
movdqa k2k3, xmm1
pshufd xmm5, xmm5, 0
movdqa k4k5, xmm2
movdqa k6k7, xmm3
movdqa krd, xmm5
movsxd rdx, DWORD PTR arg(1) ;pixels_per_line
%if ABI_IS_32BIT=0
movsxd r8, DWORD PTR arg(3) ;out_pitch
%endif
mov rax, rsi
movsxd rcx, DWORD PTR arg(4) ;output_height
add rax, rdx
lea rbx, [rdx + rdx*4]
add rbx, rdx ;pitch * 6
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
.loop:
Bug fix: ssse3 version of subpixel did not match C code A 16 bit overflow condition occurs when using the EIGHTTAP_SMOOTH filters. (vp9_sub_pel_filters_8lp) Changed the order of the adds to fix this problem. Also added ssse3 support for 4x4 subpixel filtering. Change-Id: I475eaadae920794c2de5e01e9735c059a856518e
2013-02-10 00:15:14 +01:00
movd xmm0, [rsi] ;A
movd xmm1, [rsi + rdx] ;B
movd xmm2, [rsi + rdx * 2] ;C
movd xmm3, [rax + rdx * 2] ;D
movd xmm4, [rsi + rdx * 4] ;E
movd xmm5, [rax + rdx * 4] ;F
punpcklbw xmm0, xmm1 ;A B
punpcklbw xmm2, xmm3 ;C D
punpcklbw xmm4, xmm5 ;E F
movd xmm6, [rsi + rbx] ;G
movd xmm7, [rax + rbx] ;H
pmaddubsw xmm0, k0k1
pmaddubsw xmm2, k2k3
punpcklbw xmm6, xmm7 ;G H
pmaddubsw xmm4, k4k5
pmaddubsw xmm6, k6k7
Fix decoder mismatch with ssse3 enabled This patch fixed issue 661: "Decoder produces mismatched outputs with ssse3 enabled and disabled." In sub-pixel filters, a pixel value was multiplied by a filter coefficient, and the results were added up. The order of adding up these multiplications had to be arranged carefully to prevent incorrect overflowing. Change-Id: Id08af4200fea9e1b896fc40157b8651c2c7e80f2
2013-11-19 23:29:25 +01:00
movdqa xmm1, xmm2
convolve: support larger blocks, fix asm saturation bug Updates the common convoloution code to support blocks larger than 16x16, and rectangular blocks. This uncovered a bug in the SSSE3 filtering routines due to the order of application of saturation. This commit fixes that bug, adjusts the unit test to bias its random values towards the extremes, and adds a test to ensure that all filters conform to the expected pairwise addition structure. Change-Id: I81f69668b1de0de5a8ed43f0643845641525c8f0
2013-04-18 22:05:38 +02:00
paddsw xmm0, xmm6
Fix decoder mismatch with ssse3 enabled This patch fixed issue 661: "Decoder produces mismatched outputs with ssse3 enabled and disabled." In sub-pixel filters, a pixel value was multiplied by a filter coefficient, and the results were added up. The order of adding up these multiplications had to be arranged carefully to prevent incorrect overflowing. Change-Id: Id08af4200fea9e1b896fc40157b8651c2c7e80f2
2013-11-19 23:29:25 +01:00
pmaxsw xmm2, xmm4
pminsw xmm4, xmm1
Bug fix: ssse3 version of subpixel did not match C code A 16 bit overflow condition occurs when using the EIGHTTAP_SMOOTH filters. (vp9_sub_pel_filters_8lp) Changed the order of the adds to fix this problem. Also added ssse3 support for 4x4 subpixel filtering. Change-Id: I475eaadae920794c2de5e01e9735c059a856518e
2013-02-10 00:15:14 +01:00
paddsw xmm0, xmm4
Fix decoder mismatch with ssse3 enabled This patch fixed issue 661: "Decoder produces mismatched outputs with ssse3 enabled and disabled." In sub-pixel filters, a pixel value was multiplied by a filter coefficient, and the results were added up. The order of adding up these multiplications had to be arranged carefully to prevent incorrect overflowing. Change-Id: Id08af4200fea9e1b896fc40157b8651c2c7e80f2
2013-11-19 23:29:25 +01:00
paddsw xmm0, xmm2
Bug fix: ssse3 version of subpixel did not match C code A 16 bit overflow condition occurs when using the EIGHTTAP_SMOOTH filters. (vp9_sub_pel_filters_8lp) Changed the order of the adds to fix this problem. Also added ssse3 support for 4x4 subpixel filtering. Change-Id: I475eaadae920794c2de5e01e9735c059a856518e
2013-02-10 00:15:14 +01:00
Fix decoder mismatch with ssse3 enabled This patch fixed issue 661: "Decoder produces mismatched outputs with ssse3 enabled and disabled." In sub-pixel filters, a pixel value was multiplied by a filter coefficient, and the results were added up. The order of adding up these multiplications had to be arranged carefully to prevent incorrect overflowing. Change-Id: Id08af4200fea9e1b896fc40157b8651c2c7e80f2
2013-11-19 23:29:25 +01:00
paddsw xmm0, krd
Bug fix: ssse3 version of subpixel did not match C code A 16 bit overflow condition occurs when using the EIGHTTAP_SMOOTH filters. (vp9_sub_pel_filters_8lp) Changed the order of the adds to fix this problem. Also added ssse3 support for 4x4 subpixel filtering. Change-Id: I475eaadae920794c2de5e01e9735c059a856518e
2013-02-10 00:15:14 +01:00
psraw xmm0, 7
packuswb xmm0, xmm0
add rsi, rdx
add rax, rdx
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
%if %1
movd xmm1, [rdi]
pavgb xmm0, xmm1
%endif
Bug fix: ssse3 version of subpixel did not match C code A 16 bit overflow condition occurs when using the EIGHTTAP_SMOOTH filters. (vp9_sub_pel_filters_8lp) Changed the order of the adds to fix this problem. Also added ssse3 support for 4x4 subpixel filtering. Change-Id: I475eaadae920794c2de5e01e9735c059a856518e
2013-02-10 00:15:14 +01:00
movd [rdi], xmm0
%if ABI_IS_32BIT
add rdi, DWORD PTR arg(3) ;out_pitch
%else
add rdi, r8
%endif
dec rcx
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
jnz .loop
%endm
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
%macro VERTx8 1
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
mov rdx, arg(5) ;filter ptr
mov rsi, arg(0) ;src_ptr
mov rdi, arg(2) ;output_ptr
mov rcx, 0x0400040
movdqa xmm4, [rdx] ;load filters
Fix --as=nasm compatibility for new asm code. s/movd/movq/ Change-Id: Id1a56de91551f8dc796f14f1056c565dfc1ba626
2013-02-27 18:51:59 +01:00
movq xmm5, rcx
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
packsswb xmm4, xmm4
pshuflw xmm0, xmm4, 0b ;k0_k1
pshuflw xmm1, xmm4, 01010101b ;k2_k3
pshuflw xmm2, xmm4, 10101010b ;k4_k5
pshuflw xmm3, xmm4, 11111111b ;k6_k7
punpcklqdq xmm0, xmm0
punpcklqdq xmm1, xmm1
punpcklqdq xmm2, xmm2
punpcklqdq xmm3, xmm3
movdqa k0k1, xmm0
movdqa k2k3, xmm1
pshufd xmm5, xmm5, 0
movdqa k4k5, xmm2
movdqa k6k7, xmm3
movdqa krd, xmm5
movsxd rdx, DWORD PTR arg(1) ;pixels_per_line
%if ABI_IS_32BIT=0
movsxd r8, DWORD PTR arg(3) ;out_pitch
%endif
mov rax, rsi
movsxd rcx, DWORD PTR arg(4) ;output_height
add rax, rdx
lea rbx, [rdx + rdx*4]
add rbx, rdx ;pitch * 6
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
.loop:
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
movq xmm0, [rsi] ;A
movq xmm1, [rsi + rdx] ;B
movq xmm2, [rsi + rdx * 2] ;C
movq xmm3, [rax + rdx * 2] ;D
movq xmm4, [rsi + rdx * 4] ;E
movq xmm5, [rax + rdx * 4] ;F
punpcklbw xmm0, xmm1 ;A B
punpcklbw xmm2, xmm3 ;C D
punpcklbw xmm4, xmm5 ;E F
movq xmm6, [rsi + rbx] ;G
movq xmm7, [rax + rbx] ;H
pmaddubsw xmm0, k0k1
pmaddubsw xmm2, k2k3
punpcklbw xmm6, xmm7 ;G H
pmaddubsw xmm4, k4k5
pmaddubsw xmm6, k6k7
convolve: support larger blocks, fix asm saturation bug Updates the common convoloution code to support blocks larger than 16x16, and rectangular blocks. This uncovered a bug in the SSSE3 filtering routines due to the order of application of saturation. This commit fixes that bug, adjusts the unit test to bias its random values towards the extremes, and adds a test to ensure that all filters conform to the expected pairwise addition structure. Change-Id: I81f69668b1de0de5a8ed43f0643845641525c8f0
2013-04-18 22:05:38 +02:00
paddsw xmm0, xmm6
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
paddsw xmm0, xmm2
paddsw xmm0, xmm4
convolve: support larger blocks, fix asm saturation bug Updates the common convoloution code to support blocks larger than 16x16, and rectangular blocks. This uncovered a bug in the SSSE3 filtering routines due to the order of application of saturation. This commit fixes that bug, adjusts the unit test to bias its random values towards the extremes, and adds a test to ensure that all filters conform to the expected pairwise addition structure. Change-Id: I81f69668b1de0de5a8ed43f0643845641525c8f0
2013-04-18 22:05:38 +02:00
paddsw xmm0, krd
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
psraw xmm0, 7
packuswb xmm0, xmm0
add rsi, rdx
add rax, rdx
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
%if %1
movq xmm1, [rdi]
pavgb xmm0, xmm1
%endif
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
movq [rdi], xmm0
%if ABI_IS_32BIT
add rdi, DWORD PTR arg(3) ;out_pitch
%else
add rdi, r8
%endif
dec rcx
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
jnz .loop
%endm
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
%macro VERTx16 1
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
mov rdx, arg(5) ;filter ptr
mov rsi, arg(0) ;src_ptr
mov rdi, arg(2) ;output_ptr
mov rcx, 0x0400040
movdqa xmm4, [rdx] ;load filters
Fix --as=nasm compatibility for new asm code. s/movd/movq/ Change-Id: Id1a56de91551f8dc796f14f1056c565dfc1ba626
2013-02-27 18:51:59 +01:00
movq xmm5, rcx
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
packsswb xmm4, xmm4
pshuflw xmm0, xmm4, 0b ;k0_k1
pshuflw xmm1, xmm4, 01010101b ;k2_k3
pshuflw xmm2, xmm4, 10101010b ;k4_k5
pshuflw xmm3, xmm4, 11111111b ;k6_k7
punpcklqdq xmm0, xmm0
punpcklqdq xmm1, xmm1
punpcklqdq xmm2, xmm2
punpcklqdq xmm3, xmm3
movdqa k0k1, xmm0
movdqa k2k3, xmm1
pshufd xmm5, xmm5, 0
movdqa k4k5, xmm2
movdqa k6k7, xmm3
movdqa krd, xmm5
movsxd rdx, DWORD PTR arg(1) ;pixels_per_line
%if ABI_IS_32BIT=0
movsxd r8, DWORD PTR arg(3) ;out_pitch
%endif
mov rax, rsi
movsxd rcx, DWORD PTR arg(4) ;output_height
add rax, rdx
lea rbx, [rdx + rdx*4]
add rbx, rdx ;pitch * 6
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
.loop:
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
movq xmm0, [rsi] ;A
movq xmm1, [rsi + rdx] ;B
movq xmm2, [rsi + rdx * 2] ;C
movq xmm3, [rax + rdx * 2] ;D
movq xmm4, [rsi + rdx * 4] ;E
movq xmm5, [rax + rdx * 4] ;F
punpcklbw xmm0, xmm1 ;A B
punpcklbw xmm2, xmm3 ;C D
punpcklbw xmm4, xmm5 ;E F
movq xmm6, [rsi + rbx] ;G
movq xmm7, [rax + rbx] ;H
pmaddubsw xmm0, k0k1
pmaddubsw xmm2, k2k3
punpcklbw xmm6, xmm7 ;G H
pmaddubsw xmm4, k4k5
pmaddubsw xmm6, k6k7
convolve: support larger blocks, fix asm saturation bug Updates the common convoloution code to support blocks larger than 16x16, and rectangular blocks. This uncovered a bug in the SSSE3 filtering routines due to the order of application of saturation. This commit fixes that bug, adjusts the unit test to bias its random values towards the extremes, and adds a test to ensure that all filters conform to the expected pairwise addition structure. Change-Id: I81f69668b1de0de5a8ed43f0643845641525c8f0
2013-04-18 22:05:38 +02:00
paddsw xmm0, xmm6
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
paddsw xmm0, xmm2
paddsw xmm0, xmm4
convolve: support larger blocks, fix asm saturation bug Updates the common convoloution code to support blocks larger than 16x16, and rectangular blocks. This uncovered a bug in the SSSE3 filtering routines due to the order of application of saturation. This commit fixes that bug, adjusts the unit test to bias its random values towards the extremes, and adds a test to ensure that all filters conform to the expected pairwise addition structure. Change-Id: I81f69668b1de0de5a8ed43f0643845641525c8f0
2013-04-18 22:05:38 +02:00
paddsw xmm0, krd
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
psraw xmm0, 7
packuswb xmm0, xmm0
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
%if %1
movq xmm1, [rdi]
pavgb xmm0, xmm1
%endif
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
movq [rdi], xmm0
movq xmm0, [rsi + 8] ;A
movq xmm1, [rsi + rdx + 8] ;B
movq xmm2, [rsi + rdx * 2 + 8] ;C
movq xmm3, [rax + rdx * 2 + 8] ;D
movq xmm4, [rsi + rdx * 4 + 8] ;E
movq xmm5, [rax + rdx * 4 + 8] ;F
punpcklbw xmm0, xmm1 ;A B
punpcklbw xmm2, xmm3 ;C D
punpcklbw xmm4, xmm5 ;E F
movq xmm6, [rsi + rbx + 8] ;G
movq xmm7, [rax + rbx + 8] ;H
punpcklbw xmm6, xmm7 ;G H
pmaddubsw xmm0, k0k1
pmaddubsw xmm2, k2k3
pmaddubsw xmm4, k4k5
pmaddubsw xmm6, k6k7
convolve: support larger blocks, fix asm saturation bug Updates the common convoloution code to support blocks larger than 16x16, and rectangular blocks. This uncovered a bug in the SSSE3 filtering routines due to the order of application of saturation. This commit fixes that bug, adjusts the unit test to bias its random values towards the extremes, and adds a test to ensure that all filters conform to the expected pairwise addition structure. Change-Id: I81f69668b1de0de5a8ed43f0643845641525c8f0
2013-04-18 22:05:38 +02:00
paddsw xmm0, xmm6
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
paddsw xmm0, xmm2
paddsw xmm0, xmm4
convolve: support larger blocks, fix asm saturation bug Updates the common convoloution code to support blocks larger than 16x16, and rectangular blocks. This uncovered a bug in the SSSE3 filtering routines due to the order of application of saturation. This commit fixes that bug, adjusts the unit test to bias its random values towards the extremes, and adds a test to ensure that all filters conform to the expected pairwise addition structure. Change-Id: I81f69668b1de0de5a8ed43f0643845641525c8f0
2013-04-18 22:05:38 +02:00
paddsw xmm0, krd
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
psraw xmm0, 7
packuswb xmm0, xmm0
add rsi, rdx
add rax, rdx
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
%if %1
movq xmm1, [rdi+8]
pavgb xmm0, xmm1
%endif
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
movq [rdi+8], xmm0
%if ABI_IS_32BIT
add rdi, DWORD PTR arg(3) ;out_pitch
%else
add rdi, r8
%endif
dec rcx
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
jnz .loop
%endm
;void vp9_filter_block1d8_v8_ssse3
;(
; unsigned char *src_ptr,
; unsigned int src_pitch,
; unsigned char *output_ptr,
; unsigned int out_pitch,
; unsigned int output_height,
; short *filter
;)
global sym(vp9_filter_block1d4_v8_ssse3) PRIVATE
sym(vp9_filter_block1d4_v8_ssse3):
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 6
SAVE_XMM 7
push rsi
push rdi
push rbx
; end prolog
ALIGN_STACK 16, rax
sub rsp, 16*5
%define k0k1 [rsp + 16*0]
%define k2k3 [rsp + 16*1]
%define k4k5 [rsp + 16*2]
%define k6k7 [rsp + 16*3]
%define krd [rsp + 16*4]
VERTx4 0
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
add rsp, 16*5
pop rsp
pop rbx
; begin epilog
pop rdi
pop rsi
RESTORE_XMM
UNSHADOW_ARGS
pop rbp
ret
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
;void vp9_filter_block1d8_v8_ssse3
Bug fix: ssse3 version of subpixel did not match C code A 16 bit overflow condition occurs when using the EIGHTTAP_SMOOTH filters. (vp9_sub_pel_filters_8lp) Changed the order of the adds to fix this problem. Also added ssse3 support for 4x4 subpixel filtering. Change-Id: I475eaadae920794c2de5e01e9735c059a856518e
2013-02-10 00:15:14 +01:00
;(
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
; unsigned char *src_ptr,
; unsigned int src_pitch,
; unsigned char *output_ptr,
; unsigned int out_pitch,
; unsigned int output_height,
Bug fix: ssse3 version of subpixel did not match C code A 16 bit overflow condition occurs when using the EIGHTTAP_SMOOTH filters. (vp9_sub_pel_filters_8lp) Changed the order of the adds to fix this problem. Also added ssse3 support for 4x4 subpixel filtering. Change-Id: I475eaadae920794c2de5e01e9735c059a856518e
2013-02-10 00:15:14 +01:00
; short *filter
;)
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
global sym(vp9_filter_block1d8_v8_ssse3) PRIVATE
sym(vp9_filter_block1d8_v8_ssse3):
Bug fix: ssse3 version of subpixel did not match C code A 16 bit overflow condition occurs when using the EIGHTTAP_SMOOTH filters. (vp9_sub_pel_filters_8lp) Changed the order of the adds to fix this problem. Also added ssse3 support for 4x4 subpixel filtering. Change-Id: I475eaadae920794c2de5e01e9735c059a856518e
2013-02-10 00:15:14 +01:00
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 6
SAVE_XMM 7
push rsi
push rdi
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
push rbx
Bug fix: ssse3 version of subpixel did not match C code A 16 bit overflow condition occurs when using the EIGHTTAP_SMOOTH filters. (vp9_sub_pel_filters_8lp) Changed the order of the adds to fix this problem. Also added ssse3 support for 4x4 subpixel filtering. Change-Id: I475eaadae920794c2de5e01e9735c059a856518e
2013-02-10 00:15:14 +01:00
; end prolog
ALIGN_STACK 16, rax
sub rsp, 16*5
%define k0k1 [rsp + 16*0]
%define k2k3 [rsp + 16*1]
%define k4k5 [rsp + 16*2]
%define k6k7 [rsp + 16*3]
%define krd [rsp + 16*4]
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
VERTx8 0
Bug fix: ssse3 version of subpixel did not match C code A 16 bit overflow condition occurs when using the EIGHTTAP_SMOOTH filters. (vp9_sub_pel_filters_8lp) Changed the order of the adds to fix this problem. Also added ssse3 support for 4x4 subpixel filtering. Change-Id: I475eaadae920794c2de5e01e9735c059a856518e
2013-02-10 00:15:14 +01:00
add rsp, 16*5
pop rsp
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
pop rbx
Bug fix: ssse3 version of subpixel did not match C code A 16 bit overflow condition occurs when using the EIGHTTAP_SMOOTH filters. (vp9_sub_pel_filters_8lp) Changed the order of the adds to fix this problem. Also added ssse3 support for 4x4 subpixel filtering. Change-Id: I475eaadae920794c2de5e01e9735c059a856518e
2013-02-10 00:15:14 +01:00
; begin epilog
pop rdi
pop rsi
RESTORE_XMM
UNSHADOW_ARGS
pop rbp
ret
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
;void vp9_filter_block1d16_v8_ssse3
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
;(
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
; unsigned char *src_ptr,
; unsigned int src_pitch,
; unsigned char *output_ptr,
; unsigned int out_pitch,
; unsigned int output_height,
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
; short *filter
;)
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
global sym(vp9_filter_block1d16_v8_ssse3) PRIVATE
sym(vp9_filter_block1d16_v8_ssse3):
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 6
SAVE_XMM 7
push rsi
push rdi
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
push rbx
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
; end prolog
ALIGN_STACK 16, rax
sub rsp, 16*5
%define k0k1 [rsp + 16*0]
%define k2k3 [rsp + 16*1]
%define k4k5 [rsp + 16*2]
%define k6k7 [rsp + 16*3]
%define krd [rsp + 16*4]
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
VERTx16 0
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
add rsp, 16*5
pop rsp
pop rbx
; begin epilog
pop rdi
pop rsi
RESTORE_XMM
UNSHADOW_ARGS
pop rbp
ret
;~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
global sym(vp9_filter_block1d4_v8_avg_ssse3) PRIVATE
sym(vp9_filter_block1d4_v8_avg_ssse3):
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 6
SAVE_XMM 7
push rsi
push rdi
push rbx
; end prolog
ALIGN_STACK 16, rax
sub rsp, 16*5
%define k0k1 [rsp + 16*0]
%define k2k3 [rsp + 16*1]
%define k4k5 [rsp + 16*2]
%define k6k7 [rsp + 16*3]
%define krd [rsp + 16*4]
VERTx4 1
add rsp, 16*5
pop rsp
pop rbx
; begin epilog
pop rdi
pop rsi
RESTORE_XMM
UNSHADOW_ARGS
pop rbp
ret
global sym(vp9_filter_block1d8_v8_avg_ssse3) PRIVATE
sym(vp9_filter_block1d8_v8_avg_ssse3):
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 6
SAVE_XMM 7
push rsi
push rdi
push rbx
; end prolog
ALIGN_STACK 16, rax
sub rsp, 16*5
%define k0k1 [rsp + 16*0]
%define k2k3 [rsp + 16*1]
%define k4k5 [rsp + 16*2]
%define k6k7 [rsp + 16*3]
%define krd [rsp + 16*4]
VERTx8 1
add rsp, 16*5
pop rsp
pop rbx
; begin epilog
pop rdi
pop rsi
RESTORE_XMM
UNSHADOW_ARGS
pop rbp
ret
global sym(vp9_filter_block1d16_v8_avg_ssse3) PRIVATE
sym(vp9_filter_block1d16_v8_avg_ssse3):
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 6
SAVE_XMM 7
push rsi
push rdi
push rbx
; end prolog
ALIGN_STACK 16, rax
sub rsp, 16*5
%define k0k1 [rsp + 16*0]
%define k2k3 [rsp + 16*1]
%define k4k5 [rsp + 16*2]
%define k6k7 [rsp + 16*3]
%define krd [rsp + 16*4]
VERTx16 1
add rsp, 16*5
pop rsp
pop rbx
; begin epilog
pop rdi
pop rsi
RESTORE_XMM
UNSHADOW_ARGS
pop rbp
ret
;~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Rewrite HORIZx4 and HORIZx8 in subpixel filter functions In subpixel filters, prefetched source data, unrolled loops, and interleaved instructions. In HORIZx4, integrated the idea in Scott's CL (commit: d22a504d11a15dc3eab666859db0046b5a7d75c5), which was suggested by Erik/Tamar from Intel. Further tweaking was done to combine row 0, 2, and row 1, 3 in registers to do more 2-row-in-1 operations until the last add. Test showed a ~2% decoder speedup. Change-Id: Ib53d04ede8166c38c3dc744da8c6f737ce26a0e3
2013-10-03 02:26:01 +02:00
%macro HORIZx4_ROW 2
movdqa %2, %1
pshufb %1, [GLOBAL(shuf_t0t1)]
pshufb %2, [GLOBAL(shuf_t2t3)]
Fix decoder mismatch with ssse3 enabled This patch fixed issue 661: "Decoder produces mismatched outputs with ssse3 enabled and disabled." In sub-pixel filters, a pixel value was multiplied by a filter coefficient, and the results were added up. The order of adding up these multiplications had to be arranged carefully to prevent incorrect overflowing. Change-Id: Id08af4200fea9e1b896fc40157b8651c2c7e80f2
2013-11-19 23:29:25 +01:00
pmaddubsw %1, k0k1k4k5
pmaddubsw %2, k2k3k6k7
Rewrite HORIZx4 and HORIZx8 in subpixel filter functions In subpixel filters, prefetched source data, unrolled loops, and interleaved instructions. In HORIZx4, integrated the idea in Scott's CL (commit: d22a504d11a15dc3eab666859db0046b5a7d75c5), which was suggested by Erik/Tamar from Intel. Further tweaking was done to combine row 0, 2, and row 1, 3 in registers to do more 2-row-in-1 operations until the last add. Test showed a ~2% decoder speedup. Change-Id: Ib53d04ede8166c38c3dc744da8c6f737ce26a0e3
2013-10-03 02:26:01 +02:00
Fix decoder mismatch with ssse3 enabled This patch fixed issue 661: "Decoder produces mismatched outputs with ssse3 enabled and disabled." In sub-pixel filters, a pixel value was multiplied by a filter coefficient, and the results were added up. The order of adding up these multiplications had to be arranged carefully to prevent incorrect overflowing. Change-Id: Id08af4200fea9e1b896fc40157b8651c2c7e80f2
2013-11-19 23:29:25 +01:00
movdqa xmm4, %1
movdqa xmm5, %2
psrldq %1, 8
Rewrite HORIZx4 and HORIZx8 in subpixel filter functions In subpixel filters, prefetched source data, unrolled loops, and interleaved instructions. In HORIZx4, integrated the idea in Scott's CL (commit: d22a504d11a15dc3eab666859db0046b5a7d75c5), which was suggested by Erik/Tamar from Intel. Further tweaking was done to combine row 0, 2, and row 1, 3 in registers to do more 2-row-in-1 operations until the last add. Test showed a ~2% decoder speedup. Change-Id: Ib53d04ede8166c38c3dc744da8c6f737ce26a0e3
2013-10-03 02:26:01 +02:00
psrldq %2, 8
Fix decoder mismatch with ssse3 enabled This patch fixed issue 661: "Decoder produces mismatched outputs with ssse3 enabled and disabled." In sub-pixel filters, a pixel value was multiplied by a filter coefficient, and the results were added up. The order of adding up these multiplications had to be arranged carefully to prevent incorrect overflowing. Change-Id: Id08af4200fea9e1b896fc40157b8651c2c7e80f2
2013-11-19 23:29:25 +01:00
movdqa xmm6, xmm5
paddsw xmm4, %2
pmaxsw xmm5, %1
pminsw %1, xmm6
paddsw %1, xmm4
paddsw %1, xmm5
paddsw %1, krd
Rewrite HORIZx4 and HORIZx8 in subpixel filter functions In subpixel filters, prefetched source data, unrolled loops, and interleaved instructions. In HORIZx4, integrated the idea in Scott's CL (commit: d22a504d11a15dc3eab666859db0046b5a7d75c5), which was suggested by Erik/Tamar from Intel. Further tweaking was done to combine row 0, 2, and row 1, 3 in registers to do more 2-row-in-1 operations until the last add. Test showed a ~2% decoder speedup. Change-Id: Ib53d04ede8166c38c3dc744da8c6f737ce26a0e3
2013-10-03 02:26:01 +02:00
psraw %1, 7
packuswb %1, %1
%endm
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
%macro HORIZx4 1
mov rdx, arg(5) ;filter ptr
mov rsi, arg(0) ;src_ptr
mov rdi, arg(2) ;output_ptr
mov rcx, 0x0400040
movdqa xmm4, [rdx] ;load filters
Merge master branch into experimental Picks up some build system changes, compiler warning fixes, etc. Change-Id: I2712f99e653502818a101a72696ad54018152d4e
2013-03-01 01:25:38 +01:00
movq xmm5, rcx
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
packsswb xmm4, xmm4
Rewrite HORIZx4 and HORIZx8 in subpixel filter functions In subpixel filters, prefetched source data, unrolled loops, and interleaved instructions. In HORIZx4, integrated the idea in Scott's CL (commit: d22a504d11a15dc3eab666859db0046b5a7d75c5), which was suggested by Erik/Tamar from Intel. Further tweaking was done to combine row 0, 2, and row 1, 3 in registers to do more 2-row-in-1 operations until the last add. Test showed a ~2% decoder speedup. Change-Id: Ib53d04ede8166c38c3dc744da8c6f737ce26a0e3
2013-10-03 02:26:01 +02:00
pshuflw xmm6, xmm4, 0b ;k0_k1
pshufhw xmm6, xmm6, 10101010b ;k0_k1_k4_k5
pshuflw xmm7, xmm4, 01010101b ;k2_k3
pshufhw xmm7, xmm7, 11111111b ;k2_k3_k6_k7
pshufd xmm5, xmm5, 0 ;rounding
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
Fix decoder mismatch with ssse3 enabled This patch fixed issue 661: "Decoder produces mismatched outputs with ssse3 enabled and disabled." In sub-pixel filters, a pixel value was multiplied by a filter coefficient, and the results were added up. The order of adding up these multiplications had to be arranged carefully to prevent incorrect overflowing. Change-Id: Id08af4200fea9e1b896fc40157b8651c2c7e80f2
2013-11-19 23:29:25 +01:00
movdqa k0k1k4k5, xmm6
movdqa k2k3k6k7, xmm7
movdqa krd, xmm5
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
movsxd rax, dword ptr arg(1) ;src_pixels_per_line
movsxd rdx, dword ptr arg(3) ;output_pitch
movsxd rcx, dword ptr arg(4) ;output_height
Rewrite HORIZx4 and HORIZx8 in subpixel filter functions In subpixel filters, prefetched source data, unrolled loops, and interleaved instructions. In HORIZx4, integrated the idea in Scott's CL (commit: d22a504d11a15dc3eab666859db0046b5a7d75c5), which was suggested by Erik/Tamar from Intel. Further tweaking was done to combine row 0, 2, and row 1, 3 in registers to do more 2-row-in-1 operations until the last add. Test showed a ~2% decoder speedup. Change-Id: Ib53d04ede8166c38c3dc744da8c6f737ce26a0e3
2013-10-03 02:26:01 +02:00
shr rcx, 1
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
.loop:
Rewrite HORIZx4 and HORIZx8 in subpixel filter functions In subpixel filters, prefetched source data, unrolled loops, and interleaved instructions. In HORIZx4, integrated the idea in Scott's CL (commit: d22a504d11a15dc3eab666859db0046b5a7d75c5), which was suggested by Erik/Tamar from Intel. Further tweaking was done to combine row 0, 2, and row 1, 3 in registers to do more 2-row-in-1 operations until the last add. Test showed a ~2% decoder speedup. Change-Id: Ib53d04ede8166c38c3dc744da8c6f737ce26a0e3
2013-10-03 02:26:01 +02:00
;Do two rows once
movq xmm0, [rsi - 3] ;load src
movq xmm1, [rsi + 5]
movq xmm2, [rsi + rax - 3]
movq xmm3, [rsi + rax + 5]
punpcklqdq xmm0, xmm1
punpcklqdq xmm2, xmm3
HORIZx4_ROW xmm0, xmm1
HORIZx4_ROW xmm2, xmm3
%if %1
movd xmm1, [rdi]
pavgb xmm0, xmm1
movd xmm3, [rdi + rdx]
pavgb xmm2, xmm3
%endif
movd [rdi], xmm0
movd [rdi +rdx], xmm2
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
Rewrite HORIZx4 and HORIZx8 in subpixel filter functions In subpixel filters, prefetched source data, unrolled loops, and interleaved instructions. In HORIZx4, integrated the idea in Scott's CL (commit: d22a504d11a15dc3eab666859db0046b5a7d75c5), which was suggested by Erik/Tamar from Intel. Further tweaking was done to combine row 0, 2, and row 1, 3 in registers to do more 2-row-in-1 operations until the last add. Test showed a ~2% decoder speedup. Change-Id: Ib53d04ede8166c38c3dc744da8c6f737ce26a0e3
2013-10-03 02:26:01 +02:00
lea rsi, [rsi + rax]
prefetcht0 [rsi + 4 * rax - 3]
lea rsi, [rsi + rax]
lea rdi, [rdi + 2 * rdx]
prefetcht0 [rsi + 2 * rax - 3]
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
Rewrite HORIZx4 and HORIZx8 in subpixel filter functions In subpixel filters, prefetched source data, unrolled loops, and interleaved instructions. In HORIZx4, integrated the idea in Scott's CL (commit: d22a504d11a15dc3eab666859db0046b5a7d75c5), which was suggested by Erik/Tamar from Intel. Further tweaking was done to combine row 0, 2, and row 1, 3 in registers to do more 2-row-in-1 operations until the last add. Test showed a ~2% decoder speedup. Change-Id: Ib53d04ede8166c38c3dc744da8c6f737ce26a0e3
2013-10-03 02:26:01 +02:00
dec rcx
jnz .loop
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
Rewrite HORIZx4 and HORIZx8 in subpixel filter functions In subpixel filters, prefetched source data, unrolled loops, and interleaved instructions. In HORIZx4, integrated the idea in Scott's CL (commit: d22a504d11a15dc3eab666859db0046b5a7d75c5), which was suggested by Erik/Tamar from Intel. Further tweaking was done to combine row 0, 2, and row 1, 3 in registers to do more 2-row-in-1 operations until the last add. Test showed a ~2% decoder speedup. Change-Id: Ib53d04ede8166c38c3dc744da8c6f737ce26a0e3
2013-10-03 02:26:01 +02:00
; Do last row if output_height is odd
movsxd rcx, dword ptr arg(4) ;output_height
and rcx, 1
je .done
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
Rewrite HORIZx4 and HORIZx8 in subpixel filter functions In subpixel filters, prefetched source data, unrolled loops, and interleaved instructions. In HORIZx4, integrated the idea in Scott's CL (commit: d22a504d11a15dc3eab666859db0046b5a7d75c5), which was suggested by Erik/Tamar from Intel. Further tweaking was done to combine row 0, 2, and row 1, 3 in registers to do more 2-row-in-1 operations until the last add. Test showed a ~2% decoder speedup. Change-Id: Ib53d04ede8166c38c3dc744da8c6f737ce26a0e3
2013-10-03 02:26:01 +02:00
movq xmm0, [rsi - 3] ; load src
movq xmm1, [rsi + 5]
punpcklqdq xmm0, xmm1
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
Rewrite HORIZx4 and HORIZx8 in subpixel filter functions In subpixel filters, prefetched source data, unrolled loops, and interleaved instructions. In HORIZx4, integrated the idea in Scott's CL (commit: d22a504d11a15dc3eab666859db0046b5a7d75c5), which was suggested by Erik/Tamar from Intel. Further tweaking was done to combine row 0, 2, and row 1, 3 in registers to do more 2-row-in-1 operations until the last add. Test showed a ~2% decoder speedup. Change-Id: Ib53d04ede8166c38c3dc744da8c6f737ce26a0e3
2013-10-03 02:26:01 +02:00
HORIZx4_ROW xmm0, xmm1
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
%if %1
movd xmm1, [rdi]
pavgb xmm0, xmm1
%endif
movd [rdi], xmm0
Rewrite HORIZx4 and HORIZx8 in subpixel filter functions In subpixel filters, prefetched source data, unrolled loops, and interleaved instructions. In HORIZx4, integrated the idea in Scott's CL (commit: d22a504d11a15dc3eab666859db0046b5a7d75c5), which was suggested by Erik/Tamar from Intel. Further tweaking was done to combine row 0, 2, and row 1, 3 in registers to do more 2-row-in-1 operations until the last add. Test showed a ~2% decoder speedup. Change-Id: Ib53d04ede8166c38c3dc744da8c6f737ce26a0e3
2013-10-03 02:26:01 +02:00
.done
%endm
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
Rewrite HORIZx4 and HORIZx8 in subpixel filter functions In subpixel filters, prefetched source data, unrolled loops, and interleaved instructions. In HORIZx4, integrated the idea in Scott's CL (commit: d22a504d11a15dc3eab666859db0046b5a7d75c5), which was suggested by Erik/Tamar from Intel. Further tweaking was done to combine row 0, 2, and row 1, 3 in registers to do more 2-row-in-1 operations until the last add. Test showed a ~2% decoder speedup. Change-Id: Ib53d04ede8166c38c3dc744da8c6f737ce26a0e3
2013-10-03 02:26:01 +02:00
%macro HORIZx8_ROW 4
movdqa %2, %1
movdqa %3, %1
movdqa %4, %1
pshufb %1, [GLOBAL(shuf_t0t1)]
pshufb %2, [GLOBAL(shuf_t2t3)]
pshufb %3, [GLOBAL(shuf_t4t5)]
pshufb %4, [GLOBAL(shuf_t6t7)]
pmaddubsw %1, k0k1
pmaddubsw %2, k2k3
pmaddubsw %3, k4k5
pmaddubsw %4, k6k7
paddsw %1, %2
paddsw %1, %4
paddsw %1, %3
paddsw %1, krd
psraw %1, 7
packuswb %1, %1
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
%endm
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
%macro HORIZx8 1
mov rdx, arg(5) ;filter ptr
mov rsi, arg(0) ;src_ptr
mov rdi, arg(2) ;output_ptr
mov rcx, 0x0400040
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
movdqa xmm4, [rdx] ;load filters
movd xmm5, rcx
packsswb xmm4, xmm4
pshuflw xmm0, xmm4, 0b ;k0_k1
pshuflw xmm1, xmm4, 01010101b ;k2_k3
pshuflw xmm2, xmm4, 10101010b ;k4_k5
pshuflw xmm3, xmm4, 11111111b ;k6_k7
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
punpcklqdq xmm0, xmm0
punpcklqdq xmm1, xmm1
punpcklqdq xmm2, xmm2
punpcklqdq xmm3, xmm3
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
movdqa k0k1, xmm0
movdqa k2k3, xmm1
pshufd xmm5, xmm5, 0
movdqa k4k5, xmm2
movdqa k6k7, xmm3
movdqa krd, xmm5
movsxd rax, dword ptr arg(1) ;src_pixels_per_line
movsxd rdx, dword ptr arg(3) ;output_pitch
movsxd rcx, dword ptr arg(4) ;output_height
Rewrite HORIZx4 and HORIZx8 in subpixel filter functions In subpixel filters, prefetched source data, unrolled loops, and interleaved instructions. In HORIZx4, integrated the idea in Scott's CL (commit: d22a504d11a15dc3eab666859db0046b5a7d75c5), which was suggested by Erik/Tamar from Intel. Further tweaking was done to combine row 0, 2, and row 1, 3 in registers to do more 2-row-in-1 operations until the last add. Test showed a ~2% decoder speedup. Change-Id: Ib53d04ede8166c38c3dc744da8c6f737ce26a0e3
2013-10-03 02:26:01 +02:00
shr rcx, 1
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
.loop:
Rewrite HORIZx4 and HORIZx8 in subpixel filter functions In subpixel filters, prefetched source data, unrolled loops, and interleaved instructions. In HORIZx4, integrated the idea in Scott's CL (commit: d22a504d11a15dc3eab666859db0046b5a7d75c5), which was suggested by Erik/Tamar from Intel. Further tweaking was done to combine row 0, 2, and row 1, 3 in registers to do more 2-row-in-1 operations until the last add. Test showed a ~2% decoder speedup. Change-Id: Ib53d04ede8166c38c3dc744da8c6f737ce26a0e3
2013-10-03 02:26:01 +02:00
movq xmm0, [rsi - 3] ;load src
movq xmm3, [rsi + 5]
movq xmm4, [rsi + rax - 3]
movq xmm7, [rsi + rax + 5]
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
punpcklqdq xmm0, xmm3
Rewrite HORIZx4 and HORIZx8 in subpixel filter functions In subpixel filters, prefetched source data, unrolled loops, and interleaved instructions. In HORIZx4, integrated the idea in Scott's CL (commit: d22a504d11a15dc3eab666859db0046b5a7d75c5), which was suggested by Erik/Tamar from Intel. Further tweaking was done to combine row 0, 2, and row 1, 3 in registers to do more 2-row-in-1 operations until the last add. Test showed a ~2% decoder speedup. Change-Id: Ib53d04ede8166c38c3dc744da8c6f737ce26a0e3
2013-10-03 02:26:01 +02:00
punpcklqdq xmm4, xmm7
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
Rewrite HORIZx4 and HORIZx8 in subpixel filter functions In subpixel filters, prefetched source data, unrolled loops, and interleaved instructions. In HORIZx4, integrated the idea in Scott's CL (commit: d22a504d11a15dc3eab666859db0046b5a7d75c5), which was suggested by Erik/Tamar from Intel. Further tweaking was done to combine row 0, 2, and row 1, 3 in registers to do more 2-row-in-1 operations until the last add. Test showed a ~2% decoder speedup. Change-Id: Ib53d04ede8166c38c3dc744da8c6f737ce26a0e3
2013-10-03 02:26:01 +02:00
HORIZx8_ROW xmm0, xmm1, xmm2, xmm3
HORIZx8_ROW xmm4, xmm5, xmm6, xmm7
%if %1
movq xmm1, [rdi]
movq xmm2, [rdi + rdx]
pavgb xmm0, xmm1
pavgb xmm4, xmm2
%endif
movq [rdi], xmm0
movq [rdi + rdx], xmm4
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
Rewrite HORIZx4 and HORIZx8 in subpixel filter functions In subpixel filters, prefetched source data, unrolled loops, and interleaved instructions. In HORIZx4, integrated the idea in Scott's CL (commit: d22a504d11a15dc3eab666859db0046b5a7d75c5), which was suggested by Erik/Tamar from Intel. Further tweaking was done to combine row 0, 2, and row 1, 3 in registers to do more 2-row-in-1 operations until the last add. Test showed a ~2% decoder speedup. Change-Id: Ib53d04ede8166c38c3dc744da8c6f737ce26a0e3
2013-10-03 02:26:01 +02:00
lea rsi, [rsi + rax]
prefetcht0 [rsi + 4 * rax - 3]
lea rsi, [rsi + rax]
lea rdi, [rdi + 2 * rdx]
prefetcht0 [rsi + 2 * rax - 3]
dec rcx
jnz .loop
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
Rewrite HORIZx4 and HORIZx8 in subpixel filter functions In subpixel filters, prefetched source data, unrolled loops, and interleaved instructions. In HORIZx4, integrated the idea in Scott's CL (commit: d22a504d11a15dc3eab666859db0046b5a7d75c5), which was suggested by Erik/Tamar from Intel. Further tweaking was done to combine row 0, 2, and row 1, 3 in registers to do more 2-row-in-1 operations until the last add. Test showed a ~2% decoder speedup. Change-Id: Ib53d04ede8166c38c3dc744da8c6f737ce26a0e3
2013-10-03 02:26:01 +02:00
;Do last row if output_height is odd
movsxd rcx, dword ptr arg(4) ;output_height
and rcx, 1
je .done
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
Rewrite HORIZx4 and HORIZx8 in subpixel filter functions In subpixel filters, prefetched source data, unrolled loops, and interleaved instructions. In HORIZx4, integrated the idea in Scott's CL (commit: d22a504d11a15dc3eab666859db0046b5a7d75c5), which was suggested by Erik/Tamar from Intel. Further tweaking was done to combine row 0, 2, and row 1, 3 in registers to do more 2-row-in-1 operations until the last add. Test showed a ~2% decoder speedup. Change-Id: Ib53d04ede8166c38c3dc744da8c6f737ce26a0e3
2013-10-03 02:26:01 +02:00
movq xmm0, [rsi - 3]
movq xmm3, [rsi + 5]
punpcklqdq xmm0, xmm3
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
Rewrite HORIZx4 and HORIZx8 in subpixel filter functions In subpixel filters, prefetched source data, unrolled loops, and interleaved instructions. In HORIZx4, integrated the idea in Scott's CL (commit: d22a504d11a15dc3eab666859db0046b5a7d75c5), which was suggested by Erik/Tamar from Intel. Further tweaking was done to combine row 0, 2, and row 1, 3 in registers to do more 2-row-in-1 operations until the last add. Test showed a ~2% decoder speedup. Change-Id: Ib53d04ede8166c38c3dc744da8c6f737ce26a0e3
2013-10-03 02:26:01 +02:00
HORIZx8_ROW xmm0, xmm1, xmm2, xmm3
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
%if %1
movq xmm1, [rdi]
pavgb xmm0, xmm1
%endif
movq [rdi], xmm0
Rewrite HORIZx4 and HORIZx8 in subpixel filter functions In subpixel filters, prefetched source data, unrolled loops, and interleaved instructions. In HORIZx4, integrated the idea in Scott's CL (commit: d22a504d11a15dc3eab666859db0046b5a7d75c5), which was suggested by Erik/Tamar from Intel. Further tweaking was done to combine row 0, 2, and row 1, 3 in registers to do more 2-row-in-1 operations until the last add. Test showed a ~2% decoder speedup. Change-Id: Ib53d04ede8166c38c3dc744da8c6f737ce26a0e3
2013-10-03 02:26:01 +02:00
.done
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
%endm
%macro HORIZx16 1
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
mov rdx, arg(5) ;filter ptr
mov rsi, arg(0) ;src_ptr
mov rdi, arg(2) ;output_ptr
mov rcx, 0x0400040
movdqa xmm4, [rdx] ;load filters
Merge master branch into experimental Picks up some build system changes, compiler warning fixes, etc. Change-Id: I2712f99e653502818a101a72696ad54018152d4e
2013-03-01 01:25:38 +01:00
movq xmm5, rcx
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
packsswb xmm4, xmm4
pshuflw xmm0, xmm4, 0b ;k0_k1
pshuflw xmm1, xmm4, 01010101b ;k2_k3
pshuflw xmm2, xmm4, 10101010b ;k4_k5
pshuflw xmm3, xmm4, 11111111b ;k6_k7
punpcklqdq xmm0, xmm0
punpcklqdq xmm1, xmm1
punpcklqdq xmm2, xmm2
punpcklqdq xmm3, xmm3
movdqa k0k1, xmm0
movdqa k2k3, xmm1
pshufd xmm5, xmm5, 0
movdqa k4k5, xmm2
movdqa k6k7, xmm3
movdqa krd, xmm5
movsxd rax, dword ptr arg(1) ;src_pixels_per_line
movsxd rdx, dword ptr arg(3) ;output_pitch
movsxd rcx, dword ptr arg(4) ;output_height
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
.loop:
Modify HORIZx16 macro in subpixel filter functions Interleaved the instructions, reduced register dependency, and prefetched the source data. This improved the decoder speed by 0.6% - 2%. Change-Id: I568067aa0c629b2e58219326899c82aedf7eccca
2013-10-01 21:49:25 +02:00
prefetcht0 [rsi + 2 * rax -3]
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
Modify HORIZx16 macro in subpixel filter functions Interleaved the instructions, reduced register dependency, and prefetched the source data. This improved the decoder speed by 0.6% - 2%. Change-Id: I568067aa0c629b2e58219326899c82aedf7eccca
2013-10-01 21:49:25 +02:00
movq xmm0, [rsi - 3] ;load src data
movq xmm4, [rsi + 5]
movq xmm7, [rsi + 13]
punpcklqdq xmm0, xmm4
punpcklqdq xmm4, xmm7
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
movdqa xmm1, xmm0
Modify HORIZx16 macro in subpixel filter functions Interleaved the instructions, reduced register dependency, and prefetched the source data. This improved the decoder speed by 0.6% - 2%. Change-Id: I568067aa0c629b2e58219326899c82aedf7eccca
2013-10-01 21:49:25 +02:00
movdqa xmm2, xmm0
movdqa xmm3, xmm0
movdqa xmm5, xmm4
movdqa xmm6, xmm4
movdqa xmm7, xmm4
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
Modify HORIZx16 macro in subpixel filter functions Interleaved the instructions, reduced register dependency, and prefetched the source data. This improved the decoder speed by 0.6% - 2%. Change-Id: I568067aa0c629b2e58219326899c82aedf7eccca
2013-10-01 21:49:25 +02:00
pshufb xmm0, [GLOBAL(shuf_t0t1)]
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
pshufb xmm1, [GLOBAL(shuf_t2t3)]
pshufb xmm2, [GLOBAL(shuf_t4t5)]
Modify HORIZx16 macro in subpixel filter functions Interleaved the instructions, reduced register dependency, and prefetched the source data. This improved the decoder speed by 0.6% - 2%. Change-Id: I568067aa0c629b2e58219326899c82aedf7eccca
2013-10-01 21:49:25 +02:00
pshufb xmm3, [GLOBAL(shuf_t6t7)]
pshufb xmm4, [GLOBAL(shuf_t0t1)]
pshufb xmm5, [GLOBAL(shuf_t2t3)]
pshufb xmm6, [GLOBAL(shuf_t4t5)]
pshufb xmm7, [GLOBAL(shuf_t6t7)]
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
Modify HORIZx16 macro in subpixel filter functions Interleaved the instructions, reduced register dependency, and prefetched the source data. This improved the decoder speed by 0.6% - 2%. Change-Id: I568067aa0c629b2e58219326899c82aedf7eccca
2013-10-01 21:49:25 +02:00
pmaddubsw xmm0, k0k1
pmaddubsw xmm1, k2k3
pmaddubsw xmm2, k4k5
pmaddubsw xmm3, k6k7
pmaddubsw xmm4, k0k1
pmaddubsw xmm5, k2k3
pmaddubsw xmm6, k4k5
pmaddubsw xmm7, k6k7
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
paddsw xmm0, xmm1
Modify HORIZx16 macro in subpixel filter functions Interleaved the instructions, reduced register dependency, and prefetched the source data. This improved the decoder speed by 0.6% - 2%. Change-Id: I568067aa0c629b2e58219326899c82aedf7eccca
2013-10-01 21:49:25 +02:00
paddsw xmm0, xmm3
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
paddsw xmm0, xmm2
Modify HORIZx16 macro in subpixel filter functions Interleaved the instructions, reduced register dependency, and prefetched the source data. This improved the decoder speed by 0.6% - 2%. Change-Id: I568067aa0c629b2e58219326899c82aedf7eccca
2013-10-01 21:49:25 +02:00
paddsw xmm4, xmm5
paddsw xmm4, xmm7
paddsw xmm4, xmm6
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
paddsw xmm0, krd
Modify HORIZx16 macro in subpixel filter functions Interleaved the instructions, reduced register dependency, and prefetched the source data. This improved the decoder speed by 0.6% - 2%. Change-Id: I568067aa0c629b2e58219326899c82aedf7eccca
2013-10-01 21:49:25 +02:00
paddsw xmm4, krd
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
psraw xmm0, 7
Modify HORIZx16 macro in subpixel filter functions Interleaved the instructions, reduced register dependency, and prefetched the source data. This improved the decoder speed by 0.6% - 2%. Change-Id: I568067aa0c629b2e58219326899c82aedf7eccca
2013-10-01 21:49:25 +02:00
psraw xmm4, 7
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
packuswb xmm0, xmm0
Modify HORIZx16 macro in subpixel filter functions Interleaved the instructions, reduced register dependency, and prefetched the source data. This improved the decoder speed by 0.6% - 2%. Change-Id: I568067aa0c629b2e58219326899c82aedf7eccca
2013-10-01 21:49:25 +02:00
packuswb xmm4, xmm4
punpcklqdq xmm0, xmm4
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
%if %1
movdqa xmm1, [rdi]
pavgb xmm0, xmm1
%endif
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
lea rsi, [rsi + rax]
movdqa [rdi], xmm0
lea rdi, [rdi + rdx]
dec rcx
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
jnz .loop
%endm
;void vp9_filter_block1d4_h8_ssse3
;(
; unsigned char *src_ptr,
; unsigned int src_pixels_per_line,
; unsigned char *output_ptr,
; unsigned int output_pitch,
; unsigned int output_height,
; short *filter
;)
global sym(vp9_filter_block1d4_h8_ssse3) PRIVATE
sym(vp9_filter_block1d4_h8_ssse3):
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 6
SAVE_XMM 7
GET_GOT rbx
push rsi
push rdi
; end prolog
Fix decoder mismatch with ssse3 enabled This patch fixed issue 661: "Decoder produces mismatched outputs with ssse3 enabled and disabled." In sub-pixel filters, a pixel value was multiplied by a filter coefficient, and the results were added up. The order of adding up these multiplications had to be arranged carefully to prevent incorrect overflowing. Change-Id: Id08af4200fea9e1b896fc40157b8651c2c7e80f2
2013-11-19 23:29:25 +01:00
ALIGN_STACK 16, rax
sub rsp, 16 * 3
%define k0k1k4k5 [rsp + 16 * 0]
%define k2k3k6k7 [rsp + 16 * 1]
%define krd [rsp + 16 * 2]
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
HORIZx4 0
Fix decoder mismatch with ssse3 enabled This patch fixed issue 661: "Decoder produces mismatched outputs with ssse3 enabled and disabled." In sub-pixel filters, a pixel value was multiplied by a filter coefficient, and the results were added up. The order of adding up these multiplications had to be arranged carefully to prevent incorrect overflowing. Change-Id: Id08af4200fea9e1b896fc40157b8651c2c7e80f2
2013-11-19 23:29:25 +01:00
add rsp, 16 * 3
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
; begin epilog
pop rdi
pop rsi
RESTORE_GOT
RESTORE_XMM
UNSHADOW_ARGS
pop rbp
ret
;void vp9_filter_block1d8_h8_ssse3
;(
; unsigned char *src_ptr,
; unsigned int src_pixels_per_line,
; unsigned char *output_ptr,
; unsigned int output_pitch,
; unsigned int output_height,
; short *filter
;)
global sym(vp9_filter_block1d8_h8_ssse3) PRIVATE
sym(vp9_filter_block1d8_h8_ssse3):
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 6
SAVE_XMM 7
GET_GOT rbx
push rsi
push rdi
; end prolog
ALIGN_STACK 16, rax
sub rsp, 16*5
%define k0k1 [rsp + 16*0]
%define k2k3 [rsp + 16*1]
%define k4k5 [rsp + 16*2]
%define k6k7 [rsp + 16*3]
%define krd [rsp + 16*4]
HORIZx8 0
add rsp, 16*5
pop rsp
; begin epilog
pop rdi
pop rsi
RESTORE_GOT
RESTORE_XMM
UNSHADOW_ARGS
pop rbp
ret
;void vp9_filter_block1d16_h8_ssse3
;(
; unsigned char *src_ptr,
; unsigned int src_pixels_per_line,
; unsigned char *output_ptr,
; unsigned int output_pitch,
; unsigned int output_height,
; short *filter
;)
global sym(vp9_filter_block1d16_h8_ssse3) PRIVATE
sym(vp9_filter_block1d16_h8_ssse3):
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 6
SAVE_XMM 7
GET_GOT rbx
push rsi
push rdi
; end prolog
ALIGN_STACK 16, rax
sub rsp, 16*5
%define k0k1 [rsp + 16*0]
%define k2k3 [rsp + 16*1]
%define k4k5 [rsp + 16*2]
%define k6k7 [rsp + 16*3]
%define krd [rsp + 16*4]
HORIZx16 0
add rsp, 16*5
pop rsp
; begin epilog
pop rdi
pop rsi
RESTORE_GOT
RESTORE_XMM
UNSHADOW_ARGS
pop rbp
ret
global sym(vp9_filter_block1d4_h8_avg_ssse3) PRIVATE
sym(vp9_filter_block1d4_h8_avg_ssse3):
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 6
SAVE_XMM 7
GET_GOT rbx
push rsi
push rdi
; end prolog
Fix decoder mismatch with ssse3 enabled This patch fixed issue 661: "Decoder produces mismatched outputs with ssse3 enabled and disabled." In sub-pixel filters, a pixel value was multiplied by a filter coefficient, and the results were added up. The order of adding up these multiplications had to be arranged carefully to prevent incorrect overflowing. Change-Id: Id08af4200fea9e1b896fc40157b8651c2c7e80f2
2013-11-19 23:29:25 +01:00
ALIGN_STACK 16, rax
sub rsp, 16 * 3
%define k0k1k4k5 [rsp + 16 * 0]
%define k2k3k6k7 [rsp + 16 * 1]
%define krd [rsp + 16 * 2]
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
HORIZx4 1
Fix decoder mismatch with ssse3 enabled This patch fixed issue 661: "Decoder produces mismatched outputs with ssse3 enabled and disabled." In sub-pixel filters, a pixel value was multiplied by a filter coefficient, and the results were added up. The order of adding up these multiplications had to be arranged carefully to prevent incorrect overflowing. Change-Id: Id08af4200fea9e1b896fc40157b8651c2c7e80f2
2013-11-19 23:29:25 +01:00
add rsp, 16 * 3
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
; begin epilog
pop rdi
pop rsi
RESTORE_GOT
RESTORE_XMM
UNSHADOW_ARGS
pop rbp
ret
global sym(vp9_filter_block1d8_h8_avg_ssse3) PRIVATE
sym(vp9_filter_block1d8_h8_avg_ssse3):
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 6
SAVE_XMM 7
GET_GOT rbx
push rsi
push rdi
; end prolog
ALIGN_STACK 16, rax
sub rsp, 16*5
%define k0k1 [rsp + 16*0]
%define k2k3 [rsp + 16*1]
%define k4k5 [rsp + 16*2]
%define k6k7 [rsp + 16*3]
%define krd [rsp + 16*4]
HORIZx8 1
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
add rsp, 16*5
pop rsp
; begin epilog
pop rdi
pop rsi
RESTORE_GOT
RESTORE_XMM
UNSHADOW_ARGS
pop rbp
ret
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
global sym(vp9_filter_block1d16_h8_avg_ssse3) PRIVATE
sym(vp9_filter_block1d16_h8_avg_ssse3):
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 6
SAVE_XMM 7
GET_GOT rbx
push rsi
push rdi
; end prolog
ALIGN_STACK 16, rax
sub rsp, 16*5
%define k0k1 [rsp + 16*0]
%define k2k3 [rsp + 16*1]
%define k4k5 [rsp + 16*2]
%define k6k7 [rsp + 16*3]
%define krd [rsp + 16*4]
HORIZx16 1
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
WIP: ssse3 version of convolve avg functions Initial ssse3 convolve avg functions and is one step closer to using x86inc.asm. The decoder performance improved by 8% for the test clip used. This should be revisited later to see if averaging outside the loop is better than having many similar filter functions. Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
2013-02-13 18:15:38 +01:00
add rsp, 16*5
pop rsp
; begin epilog
pop rdi
pop rsi
RESTORE_GOT
RESTORE_XMM
UNSHADOW_ARGS
pop rbp
ret
Faster 8t filtering Quickly modified the ssse3 sixtap filters to support eight taps. For the test clip used, a 23+% boost in decoder performance was seen. We can revisit later and improve further. Change-Id: I5f59860459e80d6fa23e6cc0fd91296a969f5240
2012-10-26 02:24:50 +02:00
SECTION_RODATA
align 16
shuf_t0t1:
db 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8
align 16
shuf_t2t3:
db 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10
align 16
shuf_t4t5:
db 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12
align 16
shuf_t6t7:
db 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14
Reference in New Issue Copy Permalink