generic-library/ffmpeg
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity
1c6bb81328
ffmpeg/libavcodec/x86/sbrdsp.asm

306 lines
8.6 KiB
NASM
Raw Normal View History

SBR DSP x86: implement SSE sbr_sum_square_sse The 32bits targets have been compiled with -mfpmath=sse for proper reference. sbr_sum_square C /32bits: 82c (unrolled)/102c C /64bits: 69c (unrolled)/82c SSE/32bits: 42c SSE/64bits: 31c Use of SSE4.1 dpps to perform the final sum is slower. Not unrolling to perform 8 operations in a loop yields 10 more cycles. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2012-02-23 19:48:58 +01:00
;******************************************************************************
;* AAC Spectral Band Replication decoding functions
;* Copyright (C) 2012 Christophe Gisquet <christophe.gisquet@gmail.com>
;*
;* This file is part of Libav.
;*
;* Libav is free software; you can redistribute it and/or
;* modify it under the terms of the GNU Lesser General Public
;* License as published by the Free Software Foundation; either
;* version 2.1 of the License, or (at your option) any later version.
;*
;* Libav is distributed in the hope that it will be useful,
;* but WITHOUT ANY WARRANTY; without even the implied warranty of
;* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
;* Lesser General Public License for more details.
;*
;* You should have received a copy of the GNU Lesser General Public
;* License along with Libav; if not, write to the Free Software
;* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
;******************************************************************************
x86: yasm: Use complete source path for macro helper %includes This is more consistent with the way we handle C #includes and it simplifies the build system.
2012-07-15 11:48:21 +02:00
%include "libavutil/x86/x86util.asm"
SBR DSP x86: implement SSE sbr_sum_square_sse The 32bits targets have been compiled with -mfpmath=sse for proper reference. sbr_sum_square C /32bits: 82c (unrolled)/102c C /64bits: 69c (unrolled)/82c SSE/32bits: 42c SSE/64bits: 31c Use of SSE4.1 dpps to perform the final sum is slower. Not unrolling to perform 8 operations in a loop yields 10 more cycles. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2012-02-23 19:48:58 +01:00
SBR DSP x86: implement SSE sbr_hf_gen Start and end index are multiple of 2, therefore guaranteeing aligned access. Also, this allows to generate 4 floats per loop, keeping the alignment all along. Timing: - 32 bits: 326c -> 172c - 64 bits: 323c -> 156c Signed-off-by: Diego Biurrun <diego@biurrun.de>
2012-02-24 22:11:19 +01:00
SECTION_RODATA
; mask equivalent for multiply by -1.0 1.0
ps_mask times 2 dd 1<<31, 0
x86: sbrdsp: Implement SSE neg_odd_64 Timing on Arrandale: C SSE Win32: 57 44 Win64: 47 38 Unrolling and not storing mask both save some cycles. Signed-off-by: Diego Biurrun <diego@biurrun.de>
2012-12-21 18:11:06 +01:00
ps_mask2 times 2 dd 0, 1<<31
x86: sbrdsp: Implement SSE qmf_post_shuffle 255 to 174 cycles on Arrandale / Win64. Unrolling yields no gain. Signed-off-by: Diego Biurrun <diego@biurrun.de>
2012-11-25 07:55:28 +01:00
ps_neg times 4 dd 1<<31
SBR DSP x86: implement SSE sbr_hf_gen Start and end index are multiple of 2, therefore guaranteeing aligned access. Also, this allows to generate 4 floats per loop, keeping the alignment all along. Timing: - 32 bits: 326c -> 172c - 64 bits: 323c -> 156c Signed-off-by: Diego Biurrun <diego@biurrun.de>
2012-02-24 22:11:19 +01:00
SECTION_TEXT
SBR DSP x86: implement SSE sbr_sum_square_sse The 32bits targets have been compiled with -mfpmath=sse for proper reference. sbr_sum_square C /32bits: 82c (unrolled)/102c C /64bits: 69c (unrolled)/82c SSE/32bits: 42c SSE/64bits: 31c Use of SSE4.1 dpps to perform the final sum is slower. Not unrolling to perform 8 operations in a loop yields 10 more cycles. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2012-02-23 19:48:58 +01:00
INIT_XMM sse
cglobal sbr_sum_square, 2, 3, 6
mov r2, r1
xorps m0, m0
xorps m1, m1
sar r2, 3
jz .prepare
.loop:
movu m2, [r0 + 0]
movu m3, [r0 + 16]
movu m4, [r0 + 32]
movu m5, [r0 + 48]
mulps m2, m2
mulps m3, m3
mulps m4, m4
mulps m5, m5
addps m0, m2
addps m1, m3
addps m0, m4
addps m1, m5
add r0, 64
dec r2
jnz .loop
.prepare:
and r1, 7
sar r1, 1
jz .end
; len is a multiple of 2, thus there are at least 4 elements to process
.endloop:
movu m2, [r0]
add r0, 16
mulps m2, m2
dec r1
addps m0, m2
jnz .endloop
.end:
addps m0, m1
movhlps m2, m0
addps m0, m2
movss m1, m0
shufps m0, m0, 1
addss m0, m1
%if ARCH_X86_64 == 0
dsputil x86: use SSE float instruction instead of SSE2 integer equivalent All the more required since the users are pure SSE functions. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2012-03-21 20:46:29 +01:00
movss r0m, m0
SBR DSP x86: implement SSE sbr_sum_square_sse The 32bits targets have been compiled with -mfpmath=sse for proper reference. sbr_sum_square C /32bits: 82c (unrolled)/102c C /64bits: 69c (unrolled)/82c SSE/32bits: 42c SSE/64bits: 31c Use of SSE4.1 dpps to perform the final sum is slower. Not unrolling to perform 8 operations in a loop yields 10 more cycles. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2012-02-23 19:48:58 +01:00
fld dword r0m
%endif
RET
SBR DSP x86: implement SSE sbr_hf_g_filt Unrolling the main loop to process, instead of 4 elements: - 8: minor gain of 2 cycles (not worth the extra object size) - 2: loss of 8 cycles. Assigning STEP to a register is a loss. Output address (Y) is almost always unaligned. Timings: - C (32/64 bits): 117/109 cycles - SSE: 57 cycles Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2012-02-23 20:12:39 +01:00
%define STEP 40*4*2
cglobal sbr_hf_g_filt, 5, 6, 5
lea r1, [r1 + 8*r4] ; offset by ixh elements into X_high
mov r5, r3
and r3, 0xFC
lea r2, [r2 + r3*4]
lea r0, [r0 + r3*8]
neg r3
aacsbr: handle m_max values smaller than 4. Prevents a signflip in the counter, and a subsequent crash because of overreads/overwrites. Found-by: Mateusz "j00ru" Jurczyk and Gynvael Coldwind CC: libav-stable@libav.org
2012-03-22 19:50:48 +01:00
jz .loop1
SBR DSP x86: implement SSE sbr_hf_g_filt Unrolling the main loop to process, instead of 4 elements: - 8: minor gain of 2 cycles (not worth the extra object size) - 2: loss of 8 cycles. Assigning STEP to a register is a loss. Output address (Y) is almost always unaligned. Timings: - C (32/64 bits): 117/109 cycles - SSE: 57 cycles Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2012-02-23 20:12:39 +01:00
.loop4:
sbrdsp.asm: convert all instructions to float/SSE ones. Since the values are floats, using the float operations makes sense, improves performance on some CPUs and makes the code SSE compatible instead of needing SSE2. Based on suggestion by Jason. Signed-off-by: Reimar Döffinger <Reimar.Doeffinger@gmx.de> Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2012-03-07 21:35:13 +01:00
movlps m0, [r2 + 4*r3 + 0]
movlps m1, [r2 + 4*r3 + 8]
movlps m2, [r1 + 0*STEP]
movlps m3, [r1 + 2*STEP]
SBR DSP x86: implement SSE sbr_hf_g_filt Unrolling the main loop to process, instead of 4 elements: - 8: minor gain of 2 cycles (not worth the extra object size) - 2: loss of 8 cycles. Assigning STEP to a register is a loss. Output address (Y) is almost always unaligned. Timings: - C (32/64 bits): 117/109 cycles - SSE: 57 cycles Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2012-02-23 20:12:39 +01:00
movhps m2, [r1 + 1*STEP]
movhps m3, [r1 + 3*STEP]
sbrdsp.asm: convert all instructions to float/SSE ones. Since the values are floats, using the float operations makes sense, improves performance on some CPUs and makes the code SSE compatible instead of needing SSE2. Based on suggestion by Jason. Signed-off-by: Reimar Döffinger <Reimar.Doeffinger@gmx.de> Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2012-03-07 21:35:13 +01:00
unpcklps m0, m0
unpcklps m1, m1
SBR DSP x86: implement SSE sbr_hf_g_filt Unrolling the main loop to process, instead of 4 elements: - 8: minor gain of 2 cycles (not worth the extra object size) - 2: loss of 8 cycles. Assigning STEP to a register is a loss. Output address (Y) is almost always unaligned. Timings: - C (32/64 bits): 117/109 cycles - SSE: 57 cycles Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2012-02-23 20:12:39 +01:00
mulps m0, m2
mulps m1, m3
movu [r0 + 8*r3 + 0], m0
movu [r0 + 8*r3 + 16], m1
add r1, 4*STEP
add r3, 4
jnz .loop4
and r5, 3 ; number of single element loops
jz .end
.loop1: ; element 0 and 1 can be computed at the same time
movss m0, [r2]
sbrdsp.asm: convert all instructions to float/SSE ones. Since the values are floats, using the float operations makes sense, improves performance on some CPUs and makes the code SSE compatible instead of needing SSE2. Based on suggestion by Jason. Signed-off-by: Reimar Döffinger <Reimar.Doeffinger@gmx.de> Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2012-03-07 21:35:13 +01:00
movlps m2, [r1]
unpcklps m0, m0
SBR DSP x86: implement SSE sbr_hf_g_filt Unrolling the main loop to process, instead of 4 elements: - 8: minor gain of 2 cycles (not worth the extra object size) - 2: loss of 8 cycles. Assigning STEP to a register is a loss. Output address (Y) is almost always unaligned. Timings: - C (32/64 bits): 117/109 cycles - SSE: 57 cycles Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2012-02-23 20:12:39 +01:00
mulps m2, m0
SBR DSP: fix SSE code to not use SSE2 instructions. movq from SSE register _to_ memory is an SSE2 instruction. Use the SSE movlps function instead that does the same thing. Signed-off-by: Reimar Döffinger <Reimar.Doeffinger@gmx.de> Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2012-03-06 22:11:30 +01:00
movlps [r0], m2
SBR DSP x86: implement SSE sbr_hf_g_filt Unrolling the main loop to process, instead of 4 elements: - 8: minor gain of 2 cycles (not worth the extra object size) - 2: loss of 8 cycles. Assigning STEP to a register is a loss. Output address (Y) is almost always unaligned. Timings: - C (32/64 bits): 117/109 cycles - SSE: 57 cycles Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2012-02-23 20:12:39 +01:00
add r0, 8
add r2, 4
add r1, STEP
dec r5
jnz .loop1
.end:
RET
SBR DSP x86: implement SSE sbr_hf_gen Start and end index are multiple of 2, therefore guaranteeing aligned access. Also, this allows to generate 4 floats per loop, keeping the alignment all along. Timing: - 32 bits: 326c -> 172c - 64 bits: 323c -> 156c Signed-off-by: Diego Biurrun <diego@biurrun.de>
2012-02-24 22:11:19 +01:00
x86: Make function prototype comments in assembly code consistent This helps grepping for functions, among other things.
2014-01-28 20:35:58 +01:00
; void ff_sbr_hf_gen_sse(float (*X_high)[2], const float (*X_low)[2],
; const float alpha0[2], const float alpha1[2],
; float bw, int start, int end)
SBR DSP x86: implement SSE sbr_hf_gen Start and end index are multiple of 2, therefore guaranteeing aligned access. Also, this allows to generate 4 floats per loop, keeping the alignment all along. Timing: - 32 bits: 326c -> 172c - 64 bits: 323c -> 156c Signed-off-by: Diego Biurrun <diego@biurrun.de>
2012-02-24 22:11:19 +01:00
;
cglobal sbr_hf_gen, 4,4,8, X_high, X_low, alpha0, alpha1, BW, S, E
; load alpha factors
%define bw m0
%if ARCH_X86_64 == 0 || WIN64
movss bw, BWm
%endif
movlps m2, [alpha1q]
movlps m1, [alpha0q]
shufps bw, bw, 0
mulps m2, bw ; (a1[0] a1[1])*bw
mulps m1, bw ; (a0[0] a0[1])*bw = (a2 a3)
mulps m2, bw ; (a1[0] a1[1])*bw*bw = (a0 a1)
mova m3, m1
mova m4, m2
mova m7, [ps_mask]
; Set pointers
%if ARCH_X86_64 == 0 || WIN64
; start and end 6th and 7th args on stack
mov r2d, Sm
mov r3d, Em
%define start r2q
%define end r3q
%else
; BW does not actually occupy a register, so shift by 1
%define start BWq
%define end Sq
%endif
sub start, end ; neg num of loops
lea X_highq, [X_highq + end*2*4]
lea X_lowq, [X_lowq + end*2*4 - 2*2*4]
shl start, 3 ; offset from num loops
mova m0, [X_lowq + start]
movlhps m1, m1 ; (a2 a3 a2 a3)
movlhps m2, m2 ; (a0 a1 a0 a1)
shufps m3, m3, q0101 ; (a3 a2 a3 a2)
shufps m4, m4, q0101 ; (a1 a0 a1 a0)
xorps m3, m7 ; (-a3 a2 -a3 a2)
xorps m4, m7 ; (-a1 a0 -a1 a0)
.loop2:
mova m5, m0
mova m6, m0
shufps m0, m0, q2200 ; {Xl[-2][0],",Xl[-1][0],"}
shufps m5, m5, q3311 ; {Xl[-2][1],",Xl[-1][1],"}
mulps m0, m2
mulps m5, m4
mova m7, m6
addps m5, m0
mova m0, [X_lowq + start + 2*2*4]
shufps m6, m0, q0022 ; {Xl[-1][0],",Xl[0][0],"}
shufps m7, m0, q1133 ; {Xl[-1][1],",Xl[1][1],"}
mulps m6, m1
mulps m7, m3
addps m5, m6
addps m7, m0
addps m5, m7
mova [X_highq + start], m5
add start, 16
jnz .loop2
RET
x86: sbrdsp: Implement SSE sum64x5 698 to 174 cycles on Arrandale. Unrolling is a 6 cycles gain. Signed-off-by: Diego Biurrun <diego@biurrun.de>
2012-11-24 23:15:07 +01:00
cglobal sbr_sum64x5, 1,2,4,z
lea r1q, [zq+ 256]
.loop:
mova m0, [zq+ 0]
mova m2, [zq+ 16]
mova m1, [zq+ 256]
mova m3, [zq+ 272]
addps m0, [zq+ 512]
addps m2, [zq+ 528]
addps m1, [zq+ 768]
addps m3, [zq+ 784]
addps m0, [zq+1024]
addps m2, [zq+1040]
addps m0, m1
addps m2, m3
mova [zq], m0
mova [zq+16], m2
add zq, 32
cmp zq, r1q
jne .loop
REP_RET
x86: sbrdsp: Implement SSE qmf_post_shuffle 255 to 174 cycles on Arrandale / Win64. Unrolling yields no gain. Signed-off-by: Diego Biurrun <diego@biurrun.de>
2012-11-25 07:55:28 +01:00
INIT_XMM sse
cglobal sbr_qmf_post_shuffle, 2,3,4,W,z
lea r2q, [zq + (64-4)*4]
mova m3, [ps_neg]
.loop:
mova m1, [zq]
xorps m0, m3, [r2q]
shufps m0, m0, m0, q0123
unpcklps m2, m0, m1
unpckhps m0, m0, m1
mova [Wq + 0], m2
mova [Wq + 16], m0
add Wq, 32
sub r2q, 16
add zq, 16
cmp zq, r2q
jl .loop
REP_RET
x86: sbrdsp: Implement SSE neg_odd_64 Timing on Arrandale: C SSE Win32: 57 44 Win64: 47 38 Unrolling and not storing mask both save some cycles. Signed-off-by: Diego Biurrun <diego@biurrun.de>
2012-12-21 18:11:06 +01:00
INIT_XMM sse
cglobal sbr_neg_odd_64, 1,2,4,z
lea r1q, [zq+256]
.loop:
mova m0, [zq+ 0]
mova m1, [zq+16]
mova m2, [zq+32]
mova m3, [zq+48]
xorps m0, [ps_mask2]
xorps m1, [ps_mask2]
xorps m2, [ps_mask2]
xorps m3, [ps_mask2]
mova [zq+ 0], m0
mova [zq+16], m1
mova [zq+32], m2
mova [zq+48], m3
add zq, 64
cmp zq, r1q
jne .loop
REP_RET
x86: sbrdsp: Implement SSE2 qmf_deint_bfly Sandybridge: 47 cycles Having a loop counter is a 7 cycle gain. Unrolling is another 7 cycle gain. Working in reverse scan is another 6 cycles. Signed-off-by: Diego Biurrun <diego@biurrun.de>
2013-04-09 23:57:07 +02:00
INIT_XMM sse2
x86: Make function prototype comments in assembly code consistent This helps grepping for functions, among other things.
2014-01-28 20:35:58 +01:00
; void ff_sbr_qmf_deint_bfly_sse2(float *v, const float *src0, const float *src1)
x86: sbrdsp: Implement SSE2 qmf_deint_bfly Sandybridge: 47 cycles Having a loop counter is a 7 cycle gain. Unrolling is another 7 cycle gain. Working in reverse scan is another 6 cycles. Signed-off-by: Diego Biurrun <diego@biurrun.de>
2013-04-09 23:57:07 +02:00
cglobal sbr_qmf_deint_bfly, 3,5,8, v,src0,src1,vrev,c
mov cq, 64*4-2*mmsize
lea vrevq, [vq + 64*4]
.loop:
mova m0, [src0q+cq]
mova m1, [src1q]
mova m2, [src0q+cq+mmsize]
mova m3, [src1q+mmsize]
pshufd m4, m0, q0123
pshufd m5, m1, q0123
pshufd m6, m2, q0123
pshufd m7, m3, q0123
addps m3, m4
subps m0, m7
addps m1, m6
subps m2, m5
mova [vrevq], m1
mova [vrevq+mmsize], m3
mova [vq+cq], m0
mova [vq+cq+mmsize], m2
add src1q, 2*mmsize
add vrevq, 2*mmsize
sub cq, 2*mmsize
jge .loop
REP_RET
x86: sbrdsp: implement SSE2 qmf_pre_shuffle From 253 to 51 cycles on Arrandale and Win64. 44 cycles on SandyBridge. Signed-off-by: Anton Khirnov <anton@khirnov.net>
2012-11-25 09:10:36 +01:00
INIT_XMM sse2
cglobal sbr_qmf_pre_shuffle, 1,4,6,z
%define OFFSET (32*4-2*mmsize)
mov r3q, OFFSET
lea r1q, [zq + (32+1)*4]
lea r2q, [zq + 64*4]
mova m5, [ps_neg]
.loop:
movu m0, [r1q]
movu m2, [r1q + mmsize]
movu m1, [zq + r3q + 4 + mmsize]
movu m3, [zq + r3q + 4]
pxor m2, m5
pxor m0, m5
pshufd m2, m2, q0123
pshufd m0, m0, q0123
SBUTTERFLY dq, 2, 3, 4
SBUTTERFLY dq, 0, 1, 4
mova [r2q + 2*r3q + 0*mmsize], m2
mova [r2q + 2*r3q + 1*mmsize], m3
mova [r2q + 2*r3q + 2*mmsize], m0
mova [r2q + 2*r3q + 3*mmsize], m1
add r1q, 2*mmsize
sub r3q, 2*mmsize
jge .loop
movq m2, [zq]
movq [r2q], m2
REP_RET
Reference in New Issue Copy Permalink