generic-library/ffmpeg
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity
3f4c0dad16
ffmpeg/libavcodec/x86/rv40dsp.asm

197 lines
4.4 KiB
NASM
Raw Normal View History

rv40: x86 SIMD for biweight Provide MMX, SSE2 and SSSE3 versions, with a fast-path when the weights are multiples of 512 (which is often the case when the values round up nicely). *_TIMER report for the 16x16 and 8x8 cases: C: 9015 decicycles in 16, 524257 runs, 31 skips 2656 decicycles in 8, 524271 runs, 17 skips MMX: 4156 decicycles in 16, 262090 runs, 54 skips 1206 decicycles in 8, 262131 runs, 13 skips MMX on fast-path: 2760 decicycles in 16, 524222 runs, 66 skips 995 decicycles in 8, 524252 runs, 36 skips SSE2: 2163 decicycles in 16, 262131 runs, 13 skips 832 decicycles in 8, 262137 runs, 7 skips SSE2 with fast path: 1783 decicycles in 16, 524276 runs, 12 skips 711 decicycles in 8, 524283 runs, 5 skips SSSE3: 2117 decicycles in 16, 262136 runs, 8 skips 814 decicycles in 8, 262143 runs, 1 skips SSSE3 with fast path: 1315 decicycles in 16, 524285 runs, 3 skips 578 decicycles in 8, 524286 runs, 2 skips This means around a 4% speedup for some sequences. Signed-off-by: Diego Biurrun <diego@biurrun.de>
2012-01-12 00:11:15 +01:00
;******************************************************************************
;* MMX/SSE2-optimized functions for the RV40 decoder
;* 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
;******************************************************************************
%include "x86inc.asm"
%include "x86util.asm"
SECTION_RODATA
align 16
shift_round: times 8 dw 1 << (16 - 6)
cextern pw_16
SECTION .text
; %1=5bits weights?, %2=dst %3=src1 %4=src3 %5=stride if sse2
%macro RV40_WCORE 4-5
rv40dsp x86: use only one register, for both increment and loop counter Around 10 cycles faster for luma. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2012-03-20 16:13:55 +01:00
movh m4, [%3 + r6 + 0]
movh m5, [%4 + r6 + 0]
rv40: x86 SIMD for biweight Provide MMX, SSE2 and SSSE3 versions, with a fast-path when the weights are multiples of 512 (which is often the case when the values round up nicely). *_TIMER report for the 16x16 and 8x8 cases: C: 9015 decicycles in 16, 524257 runs, 31 skips 2656 decicycles in 8, 524271 runs, 17 skips MMX: 4156 decicycles in 16, 262090 runs, 54 skips 1206 decicycles in 8, 262131 runs, 13 skips MMX on fast-path: 2760 decicycles in 16, 524222 runs, 66 skips 995 decicycles in 8, 524252 runs, 36 skips SSE2: 2163 decicycles in 16, 262131 runs, 13 skips 832 decicycles in 8, 262137 runs, 7 skips SSE2 with fast path: 1783 decicycles in 16, 524276 runs, 12 skips 711 decicycles in 8, 524283 runs, 5 skips SSSE3: 2117 decicycles in 16, 262136 runs, 8 skips 814 decicycles in 8, 262143 runs, 1 skips SSSE3 with fast path: 1315 decicycles in 16, 524285 runs, 3 skips 578 decicycles in 8, 524286 runs, 2 skips This means around a 4% speedup for some sequences. Signed-off-by: Diego Biurrun <diego@biurrun.de>
2012-01-12 00:11:15 +01:00
%if %0 == 4
rv40dsp x86: use only one register, for both increment and loop counter Around 10 cycles faster for luma. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2012-03-20 16:13:55 +01:00
%define OFFSET r6 + mmsize / 2
rv40: x86 SIMD for biweight Provide MMX, SSE2 and SSSE3 versions, with a fast-path when the weights are multiples of 512 (which is often the case when the values round up nicely). *_TIMER report for the 16x16 and 8x8 cases: C: 9015 decicycles in 16, 524257 runs, 31 skips 2656 decicycles in 8, 524271 runs, 17 skips MMX: 4156 decicycles in 16, 262090 runs, 54 skips 1206 decicycles in 8, 262131 runs, 13 skips MMX on fast-path: 2760 decicycles in 16, 524222 runs, 66 skips 995 decicycles in 8, 524252 runs, 36 skips SSE2: 2163 decicycles in 16, 262131 runs, 13 skips 832 decicycles in 8, 262137 runs, 7 skips SSE2 with fast path: 1783 decicycles in 16, 524276 runs, 12 skips 711 decicycles in 8, 524283 runs, 5 skips SSSE3: 2117 decicycles in 16, 262136 runs, 8 skips 814 decicycles in 8, 262143 runs, 1 skips SSSE3 with fast path: 1315 decicycles in 16, 524285 runs, 3 skips 578 decicycles in 8, 524286 runs, 2 skips This means around a 4% speedup for some sequences. Signed-off-by: Diego Biurrun <diego@biurrun.de>
2012-01-12 00:11:15 +01:00
%else
; 8x8 block and sse2, stride was provided
rv40dsp x86: use only one register, for both increment and loop counter Around 10 cycles faster for luma. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2012-03-20 16:13:55 +01:00
%define OFFSET r6
add r6, r5
rv40: x86 SIMD for biweight Provide MMX, SSE2 and SSSE3 versions, with a fast-path when the weights are multiples of 512 (which is often the case when the values round up nicely). *_TIMER report for the 16x16 and 8x8 cases: C: 9015 decicycles in 16, 524257 runs, 31 skips 2656 decicycles in 8, 524271 runs, 17 skips MMX: 4156 decicycles in 16, 262090 runs, 54 skips 1206 decicycles in 8, 262131 runs, 13 skips MMX on fast-path: 2760 decicycles in 16, 524222 runs, 66 skips 995 decicycles in 8, 524252 runs, 36 skips SSE2: 2163 decicycles in 16, 262131 runs, 13 skips 832 decicycles in 8, 262137 runs, 7 skips SSE2 with fast path: 1783 decicycles in 16, 524276 runs, 12 skips 711 decicycles in 8, 524283 runs, 5 skips SSSE3: 2117 decicycles in 16, 262136 runs, 8 skips 814 decicycles in 8, 262143 runs, 1 skips SSSE3 with fast path: 1315 decicycles in 16, 524285 runs, 3 skips 578 decicycles in 8, 524286 runs, 2 skips This means around a 4% speedup for some sequences. Signed-off-by: Diego Biurrun <diego@biurrun.de>
2012-01-12 00:11:15 +01:00
%endif
movh m6, [%3 + OFFSET]
movh m7, [%4 + OFFSET]
%if %1 == 0
; 14bits weights
punpcklbw m4, m0
punpcklbw m5, m0
punpcklbw m6, m0
punpcklbw m7, m0
psllw m4, 7
psllw m5, 7
psllw m6, 7
psllw m7, 7
pmulhw m4, m3
pmulhw m5, m2
pmulhw m6, m3
pmulhw m7, m2
paddw m4, m5
paddw m6, m7
%else
; 5bits weights
%if cpuflag(ssse3)
punpcklbw m4, m5
punpcklbw m6, m7
pmaddubsw m4, m3
pmaddubsw m6, m3
%else
punpcklbw m4, m0
punpcklbw m5, m0
punpcklbw m6, m0
punpcklbw m7, m0
pmullw m4, m3
pmullw m5, m2
pmullw m6, m3
pmullw m7, m2
paddw m4, m5
paddw m6, m7
%endif
%endif
; bias and shift down
%if cpuflag(ssse3)
pmulhrsw m4, m1
pmulhrsw m6, m1
%else
paddw m4, m1
paddw m6, m1
psrlw m4, 5
psrlw m6, 5
%endif
packuswb m4, m6
%if %0 == 5
; Only called for 8x8 blocks and sse2
rv40dsp x86: use only one register, for both increment and loop counter Around 10 cycles faster for luma. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2012-03-20 16:13:55 +01:00
sub r6, r5
movh [%2 + r6], m4
add r6, r5
movhps [%2 + r6], m4
rv40: x86 SIMD for biweight Provide MMX, SSE2 and SSSE3 versions, with a fast-path when the weights are multiples of 512 (which is often the case when the values round up nicely). *_TIMER report for the 16x16 and 8x8 cases: C: 9015 decicycles in 16, 524257 runs, 31 skips 2656 decicycles in 8, 524271 runs, 17 skips MMX: 4156 decicycles in 16, 262090 runs, 54 skips 1206 decicycles in 8, 262131 runs, 13 skips MMX on fast-path: 2760 decicycles in 16, 524222 runs, 66 skips 995 decicycles in 8, 524252 runs, 36 skips SSE2: 2163 decicycles in 16, 262131 runs, 13 skips 832 decicycles in 8, 262137 runs, 7 skips SSE2 with fast path: 1783 decicycles in 16, 524276 runs, 12 skips 711 decicycles in 8, 524283 runs, 5 skips SSSE3: 2117 decicycles in 16, 262136 runs, 8 skips 814 decicycles in 8, 262143 runs, 1 skips SSSE3 with fast path: 1315 decicycles in 16, 524285 runs, 3 skips 578 decicycles in 8, 524286 runs, 2 skips This means around a 4% speedup for some sequences. Signed-off-by: Diego Biurrun <diego@biurrun.de>
2012-01-12 00:11:15 +01:00
%else
rv40dsp x86: use only one register, for both increment and loop counter Around 10 cycles faster for luma. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2012-03-20 16:13:55 +01:00
mova [%2 + r6], m4
rv40: x86 SIMD for biweight Provide MMX, SSE2 and SSSE3 versions, with a fast-path when the weights are multiples of 512 (which is often the case when the values round up nicely). *_TIMER report for the 16x16 and 8x8 cases: C: 9015 decicycles in 16, 524257 runs, 31 skips 2656 decicycles in 8, 524271 runs, 17 skips MMX: 4156 decicycles in 16, 262090 runs, 54 skips 1206 decicycles in 8, 262131 runs, 13 skips MMX on fast-path: 2760 decicycles in 16, 524222 runs, 66 skips 995 decicycles in 8, 524252 runs, 36 skips SSE2: 2163 decicycles in 16, 262131 runs, 13 skips 832 decicycles in 8, 262137 runs, 7 skips SSE2 with fast path: 1783 decicycles in 16, 524276 runs, 12 skips 711 decicycles in 8, 524283 runs, 5 skips SSSE3: 2117 decicycles in 16, 262136 runs, 8 skips 814 decicycles in 8, 262143 runs, 1 skips SSSE3 with fast path: 1315 decicycles in 16, 524285 runs, 3 skips 578 decicycles in 8, 524286 runs, 2 skips This means around a 4% speedup for some sequences. Signed-off-by: Diego Biurrun <diego@biurrun.de>
2012-01-12 00:11:15 +01:00
%endif
%endmacro
%macro MAIN_LOOP 2
%if mmsize == 8
RV40_WCORE %2, r0, r1, r2
%if %1 == 16
RV40_WCORE %2, r0 + 8, r1 + 8, r2 + 8
%endif
; Prepare for next loop
rv40dsp x86: use only one register, for both increment and loop counter Around 10 cycles faster for luma. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2012-03-20 16:13:55 +01:00
add r6, r5
rv40: x86 SIMD for biweight Provide MMX, SSE2 and SSSE3 versions, with a fast-path when the weights are multiples of 512 (which is often the case when the values round up nicely). *_TIMER report for the 16x16 and 8x8 cases: C: 9015 decicycles in 16, 524257 runs, 31 skips 2656 decicycles in 8, 524271 runs, 17 skips MMX: 4156 decicycles in 16, 262090 runs, 54 skips 1206 decicycles in 8, 262131 runs, 13 skips MMX on fast-path: 2760 decicycles in 16, 524222 runs, 66 skips 995 decicycles in 8, 524252 runs, 36 skips SSE2: 2163 decicycles in 16, 262131 runs, 13 skips 832 decicycles in 8, 262137 runs, 7 skips SSE2 with fast path: 1783 decicycles in 16, 524276 runs, 12 skips 711 decicycles in 8, 524283 runs, 5 skips SSSE3: 2117 decicycles in 16, 262136 runs, 8 skips 814 decicycles in 8, 262143 runs, 1 skips SSSE3 with fast path: 1315 decicycles in 16, 524285 runs, 3 skips 578 decicycles in 8, 524286 runs, 2 skips This means around a 4% speedup for some sequences. Signed-off-by: Diego Biurrun <diego@biurrun.de>
2012-01-12 00:11:15 +01:00
%else
%ifidn %1, 8
RV40_WCORE %2, r0, r1, r2, r5
; Prepare 2 next lines
rv40dsp x86: use only one register, for both increment and loop counter Around 10 cycles faster for luma. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2012-03-20 16:13:55 +01:00
add r6, r5
rv40: x86 SIMD for biweight Provide MMX, SSE2 and SSSE3 versions, with a fast-path when the weights are multiples of 512 (which is often the case when the values round up nicely). *_TIMER report for the 16x16 and 8x8 cases: C: 9015 decicycles in 16, 524257 runs, 31 skips 2656 decicycles in 8, 524271 runs, 17 skips MMX: 4156 decicycles in 16, 262090 runs, 54 skips 1206 decicycles in 8, 262131 runs, 13 skips MMX on fast-path: 2760 decicycles in 16, 524222 runs, 66 skips 995 decicycles in 8, 524252 runs, 36 skips SSE2: 2163 decicycles in 16, 262131 runs, 13 skips 832 decicycles in 8, 262137 runs, 7 skips SSE2 with fast path: 1783 decicycles in 16, 524276 runs, 12 skips 711 decicycles in 8, 524283 runs, 5 skips SSSE3: 2117 decicycles in 16, 262136 runs, 8 skips 814 decicycles in 8, 262143 runs, 1 skips SSSE3 with fast path: 1315 decicycles in 16, 524285 runs, 3 skips 578 decicycles in 8, 524286 runs, 2 skips This means around a 4% speedup for some sequences. Signed-off-by: Diego Biurrun <diego@biurrun.de>
2012-01-12 00:11:15 +01:00
%else
RV40_WCORE %2, r0, r1, r2
; Prepare single next line
rv40dsp x86: use only one register, for both increment and loop counter Around 10 cycles faster for luma. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2012-03-20 16:13:55 +01:00
add r6, r5
rv40: x86 SIMD for biweight Provide MMX, SSE2 and SSSE3 versions, with a fast-path when the weights are multiples of 512 (which is often the case when the values round up nicely). *_TIMER report for the 16x16 and 8x8 cases: C: 9015 decicycles in 16, 524257 runs, 31 skips 2656 decicycles in 8, 524271 runs, 17 skips MMX: 4156 decicycles in 16, 262090 runs, 54 skips 1206 decicycles in 8, 262131 runs, 13 skips MMX on fast-path: 2760 decicycles in 16, 524222 runs, 66 skips 995 decicycles in 8, 524252 runs, 36 skips SSE2: 2163 decicycles in 16, 262131 runs, 13 skips 832 decicycles in 8, 262137 runs, 7 skips SSE2 with fast path: 1783 decicycles in 16, 524276 runs, 12 skips 711 decicycles in 8, 524283 runs, 5 skips SSSE3: 2117 decicycles in 16, 262136 runs, 8 skips 814 decicycles in 8, 262143 runs, 1 skips SSSE3 with fast path: 1315 decicycles in 16, 524285 runs, 3 skips 578 decicycles in 8, 524286 runs, 2 skips This means around a 4% speedup for some sequences. Signed-off-by: Diego Biurrun <diego@biurrun.de>
2012-01-12 00:11:15 +01:00
%endif
%endif
%endmacro
; rv40_weight_func_%1(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w1, int w2, int stride)
; %1=size %2=num of xmm regs
rv40dsp: implement prescaled versions for biweight. Quite often, the original weights are multiple of 512. By prescaling them by 1/512 when they are computed (once per frame), no intermediate shifting is needed, and no prescaling on each call either. The x86 code already used that trick. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2012-03-19 22:46:28 +01:00
; The weights are FP0.14 notation of fractions depending on pts.
; For timebases without rounding error (i.e. PAL), the fractions
; can be simplified, and several operations can be avoided.
; Therefore, we check here whether they are multiples of 2^9 for
; those simplifications to occur.
%macro RV40_WEIGHT 3
rv40dsp x86: use only one register, for both increment and loop counter Around 10 cycles faster for luma. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2012-03-20 16:13:55 +01:00
cglobal rv40_weight_func_%1_%2, 6, 7, 8
rv40: x86 SIMD for biweight Provide MMX, SSE2 and SSSE3 versions, with a fast-path when the weights are multiples of 512 (which is often the case when the values round up nicely). *_TIMER report for the 16x16 and 8x8 cases: C: 9015 decicycles in 16, 524257 runs, 31 skips 2656 decicycles in 8, 524271 runs, 17 skips MMX: 4156 decicycles in 16, 262090 runs, 54 skips 1206 decicycles in 8, 262131 runs, 13 skips MMX on fast-path: 2760 decicycles in 16, 524222 runs, 66 skips 995 decicycles in 8, 524252 runs, 36 skips SSE2: 2163 decicycles in 16, 262131 runs, 13 skips 832 decicycles in 8, 262137 runs, 7 skips SSE2 with fast path: 1783 decicycles in 16, 524276 runs, 12 skips 711 decicycles in 8, 524283 runs, 5 skips SSSE3: 2117 decicycles in 16, 262136 runs, 8 skips 814 decicycles in 8, 262143 runs, 1 skips SSSE3 with fast path: 1315 decicycles in 16, 524285 runs, 3 skips 578 decicycles in 8, 524286 runs, 2 skips This means around a 4% speedup for some sequences. Signed-off-by: Diego Biurrun <diego@biurrun.de>
2012-01-12 00:11:15 +01:00
%if cpuflag(ssse3)
mova m1, [shift_round]
%else
mova m1, [pw_16]
%endif
pxor m0, m0
; Set loop counter and increments
rv40dsp x86: use only one register, for both increment and loop counter Around 10 cycles faster for luma. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2012-03-20 16:13:55 +01:00
mov r6, r5
shl r6, %3
add r0, r6
add r1, r6
add r2, r6
neg r6
rv40: x86 SIMD for biweight Provide MMX, SSE2 and SSSE3 versions, with a fast-path when the weights are multiples of 512 (which is often the case when the values round up nicely). *_TIMER report for the 16x16 and 8x8 cases: C: 9015 decicycles in 16, 524257 runs, 31 skips 2656 decicycles in 8, 524271 runs, 17 skips MMX: 4156 decicycles in 16, 262090 runs, 54 skips 1206 decicycles in 8, 262131 runs, 13 skips MMX on fast-path: 2760 decicycles in 16, 524222 runs, 66 skips 995 decicycles in 8, 524252 runs, 36 skips SSE2: 2163 decicycles in 16, 262131 runs, 13 skips 832 decicycles in 8, 262137 runs, 7 skips SSE2 with fast path: 1783 decicycles in 16, 524276 runs, 12 skips 711 decicycles in 8, 524283 runs, 5 skips SSSE3: 2117 decicycles in 16, 262136 runs, 8 skips 814 decicycles in 8, 262143 runs, 1 skips SSSE3 with fast path: 1315 decicycles in 16, 524285 runs, 3 skips 578 decicycles in 8, 524286 runs, 2 skips This means around a 4% speedup for some sequences. Signed-off-by: Diego Biurrun <diego@biurrun.de>
2012-01-12 00:11:15 +01:00
Fix NASM compilation. movd needs explicit register size prefix for NASM. Signed-off-by: Reimar Döffinger <Reimar.Doeffinger@gmx.de>
2012-02-03 20:34:56 +01:00
movd m2, r3d
movd m3, r4d
rv40dsp: implement prescaled versions for biweight. Quite often, the original weights are multiple of 512. By prescaling them by 1/512 when they are computed (once per frame), no intermediate shifting is needed, and no prescaling on each call either. The x86 code already used that trick. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2012-03-19 22:46:28 +01:00
%ifidn %1,rnd
%define RND 0
rv40: x86 SIMD for biweight Provide MMX, SSE2 and SSSE3 versions, with a fast-path when the weights are multiples of 512 (which is often the case when the values round up nicely). *_TIMER report for the 16x16 and 8x8 cases: C: 9015 decicycles in 16, 524257 runs, 31 skips 2656 decicycles in 8, 524271 runs, 17 skips MMX: 4156 decicycles in 16, 262090 runs, 54 skips 1206 decicycles in 8, 262131 runs, 13 skips MMX on fast-path: 2760 decicycles in 16, 524222 runs, 66 skips 995 decicycles in 8, 524252 runs, 36 skips SSE2: 2163 decicycles in 16, 262131 runs, 13 skips 832 decicycles in 8, 262137 runs, 7 skips SSE2 with fast path: 1783 decicycles in 16, 524276 runs, 12 skips 711 decicycles in 8, 524283 runs, 5 skips SSSE3: 2117 decicycles in 16, 262136 runs, 8 skips 814 decicycles in 8, 262143 runs, 1 skips SSSE3 with fast path: 1315 decicycles in 16, 524285 runs, 3 skips 578 decicycles in 8, 524286 runs, 2 skips This means around a 4% speedup for some sequences. Signed-off-by: Diego Biurrun <diego@biurrun.de>
2012-01-12 00:11:15 +01:00
SPLATW m2, m2
rv40dsp: implement prescaled versions for biweight. Quite often, the original weights are multiple of 512. By prescaling them by 1/512 when they are computed (once per frame), no intermediate shifting is needed, and no prescaling on each call either. The x86 code already used that trick. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2012-03-19 22:46:28 +01:00
%else
%define RND 1
rv40: x86 SIMD for biweight Provide MMX, SSE2 and SSSE3 versions, with a fast-path when the weights are multiples of 512 (which is often the case when the values round up nicely). *_TIMER report for the 16x16 and 8x8 cases: C: 9015 decicycles in 16, 524257 runs, 31 skips 2656 decicycles in 8, 524271 runs, 17 skips MMX: 4156 decicycles in 16, 262090 runs, 54 skips 1206 decicycles in 8, 262131 runs, 13 skips MMX on fast-path: 2760 decicycles in 16, 524222 runs, 66 skips 995 decicycles in 8, 524252 runs, 36 skips SSE2: 2163 decicycles in 16, 262131 runs, 13 skips 832 decicycles in 8, 262137 runs, 7 skips SSE2 with fast path: 1783 decicycles in 16, 524276 runs, 12 skips 711 decicycles in 8, 524283 runs, 5 skips SSSE3: 2117 decicycles in 16, 262136 runs, 8 skips 814 decicycles in 8, 262143 runs, 1 skips SSSE3 with fast path: 1315 decicycles in 16, 524285 runs, 3 skips 578 decicycles in 8, 524286 runs, 2 skips This means around a 4% speedup for some sequences. Signed-off-by: Diego Biurrun <diego@biurrun.de>
2012-01-12 00:11:15 +01:00
%if cpuflag(ssse3)
punpcklbw m3, m2
%else
SPLATW m2, m2
%endif
rv40dsp: implement prescaled versions for biweight. Quite often, the original weights are multiple of 512. By prescaling them by 1/512 when they are computed (once per frame), no intermediate shifting is needed, and no prescaling on each call either. The x86 code already used that trick. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2012-03-19 22:46:28 +01:00
%endif
SPLATW m3, m3
rv40: x86 SIMD for biweight Provide MMX, SSE2 and SSSE3 versions, with a fast-path when the weights are multiples of 512 (which is often the case when the values round up nicely). *_TIMER report for the 16x16 and 8x8 cases: C: 9015 decicycles in 16, 524257 runs, 31 skips 2656 decicycles in 8, 524271 runs, 17 skips MMX: 4156 decicycles in 16, 262090 runs, 54 skips 1206 decicycles in 8, 262131 runs, 13 skips MMX on fast-path: 2760 decicycles in 16, 524222 runs, 66 skips 995 decicycles in 8, 524252 runs, 36 skips SSE2: 2163 decicycles in 16, 262131 runs, 13 skips 832 decicycles in 8, 262137 runs, 7 skips SSE2 with fast path: 1783 decicycles in 16, 524276 runs, 12 skips 711 decicycles in 8, 524283 runs, 5 skips SSSE3: 2117 decicycles in 16, 262136 runs, 8 skips 814 decicycles in 8, 262143 runs, 1 skips SSSE3 with fast path: 1315 decicycles in 16, 524285 runs, 3 skips 578 decicycles in 8, 524286 runs, 2 skips This means around a 4% speedup for some sequences. Signed-off-by: Diego Biurrun <diego@biurrun.de>
2012-01-12 00:11:15 +01:00
rv40dsp: implement prescaled versions for biweight. Quite often, the original weights are multiple of 512. By prescaling them by 1/512 when they are computed (once per frame), no intermediate shifting is needed, and no prescaling on each call either. The x86 code already used that trick. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2012-03-19 22:46:28 +01:00
.loop:
MAIN_LOOP %2, RND
jnz .loop
REP_RET
rv40: x86 SIMD for biweight Provide MMX, SSE2 and SSSE3 versions, with a fast-path when the weights are multiples of 512 (which is often the case when the values round up nicely). *_TIMER report for the 16x16 and 8x8 cases: C: 9015 decicycles in 16, 524257 runs, 31 skips 2656 decicycles in 8, 524271 runs, 17 skips MMX: 4156 decicycles in 16, 262090 runs, 54 skips 1206 decicycles in 8, 262131 runs, 13 skips MMX on fast-path: 2760 decicycles in 16, 524222 runs, 66 skips 995 decicycles in 8, 524252 runs, 36 skips SSE2: 2163 decicycles in 16, 262131 runs, 13 skips 832 decicycles in 8, 262137 runs, 7 skips SSE2 with fast path: 1783 decicycles in 16, 524276 runs, 12 skips 711 decicycles in 8, 524283 runs, 5 skips SSSE3: 2117 decicycles in 16, 262136 runs, 8 skips 814 decicycles in 8, 262143 runs, 1 skips SSSE3 with fast path: 1315 decicycles in 16, 524285 runs, 3 skips 578 decicycles in 8, 524286 runs, 2 skips This means around a 4% speedup for some sequences. Signed-off-by: Diego Biurrun <diego@biurrun.de>
2012-01-12 00:11:15 +01:00
%endmacro
INIT_MMX mmx
rv40dsp: implement prescaled versions for biweight. Quite often, the original weights are multiple of 512. By prescaling them by 1/512 when they are computed (once per frame), no intermediate shifting is needed, and no prescaling on each call either. The x86 code already used that trick. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2012-03-19 22:46:28 +01:00
RV40_WEIGHT rnd, 8, 3
RV40_WEIGHT rnd, 16, 4
RV40_WEIGHT nornd, 8, 3
RV40_WEIGHT nornd, 16, 4
rv40: x86 SIMD for biweight Provide MMX, SSE2 and SSSE3 versions, with a fast-path when the weights are multiples of 512 (which is often the case when the values round up nicely). *_TIMER report for the 16x16 and 8x8 cases: C: 9015 decicycles in 16, 524257 runs, 31 skips 2656 decicycles in 8, 524271 runs, 17 skips MMX: 4156 decicycles in 16, 262090 runs, 54 skips 1206 decicycles in 8, 262131 runs, 13 skips MMX on fast-path: 2760 decicycles in 16, 524222 runs, 66 skips 995 decicycles in 8, 524252 runs, 36 skips SSE2: 2163 decicycles in 16, 262131 runs, 13 skips 832 decicycles in 8, 262137 runs, 7 skips SSE2 with fast path: 1783 decicycles in 16, 524276 runs, 12 skips 711 decicycles in 8, 524283 runs, 5 skips SSSE3: 2117 decicycles in 16, 262136 runs, 8 skips 814 decicycles in 8, 262143 runs, 1 skips SSSE3 with fast path: 1315 decicycles in 16, 524285 runs, 3 skips 578 decicycles in 8, 524286 runs, 2 skips This means around a 4% speedup for some sequences. Signed-off-by: Diego Biurrun <diego@biurrun.de>
2012-01-12 00:11:15 +01:00
INIT_XMM sse2
rv40dsp: implement prescaled versions for biweight. Quite often, the original weights are multiple of 512. By prescaling them by 1/512 when they are computed (once per frame), no intermediate shifting is needed, and no prescaling on each call either. The x86 code already used that trick. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2012-03-19 22:46:28 +01:00
RV40_WEIGHT rnd, 8, 3
RV40_WEIGHT rnd, 16, 4
RV40_WEIGHT nornd, 8, 3
RV40_WEIGHT nornd, 16, 4
rv40: x86 SIMD for biweight Provide MMX, SSE2 and SSSE3 versions, with a fast-path when the weights are multiples of 512 (which is often the case when the values round up nicely). *_TIMER report for the 16x16 and 8x8 cases: C: 9015 decicycles in 16, 524257 runs, 31 skips 2656 decicycles in 8, 524271 runs, 17 skips MMX: 4156 decicycles in 16, 262090 runs, 54 skips 1206 decicycles in 8, 262131 runs, 13 skips MMX on fast-path: 2760 decicycles in 16, 524222 runs, 66 skips 995 decicycles in 8, 524252 runs, 36 skips SSE2: 2163 decicycles in 16, 262131 runs, 13 skips 832 decicycles in 8, 262137 runs, 7 skips SSE2 with fast path: 1783 decicycles in 16, 524276 runs, 12 skips 711 decicycles in 8, 524283 runs, 5 skips SSSE3: 2117 decicycles in 16, 262136 runs, 8 skips 814 decicycles in 8, 262143 runs, 1 skips SSSE3 with fast path: 1315 decicycles in 16, 524285 runs, 3 skips 578 decicycles in 8, 524286 runs, 2 skips This means around a 4% speedup for some sequences. Signed-off-by: Diego Biurrun <diego@biurrun.de>
2012-01-12 00:11:15 +01:00
INIT_XMM ssse3
rv40dsp: implement prescaled versions for biweight. Quite often, the original weights are multiple of 512. By prescaling them by 1/512 when they are computed (once per frame), no intermediate shifting is needed, and no prescaling on each call either. The x86 code already used that trick. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2012-03-19 22:46:28 +01:00
RV40_WEIGHT rnd, 8, 3
RV40_WEIGHT rnd, 16, 4
RV40_WEIGHT nornd, 8, 3
RV40_WEIGHT nornd, 16, 4
Reference in New Issue Copy Permalink