Chiebot-Mirror
/
FFmpeg
mirror of https://github.com/FFmpeg/FFmpeg.git
8
0
Fork 0
Code Issues Projects Releases Wiki Activity
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
71122 Commits
37 Branches
400 Tags
474 MiB
Tag: Branch: Tree: 048b6331e0
Branches Tags
${ item.name }
Create tag ${ searchTerm }
Create branch ${ searchTerm }
from '048b6331e0'
${ noResults }
FFmpeg/libavcodec/x86/hevc_res_add.asm

389 lines
11 KiB
Raw Normal View History Unescape

x86: hevc: adding transform_add Reviewed-by: James Almer <jamrial@gmail.com> Approved-by: Ronald S. Bultje Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
; /*
; * Provide SIMD optimizations for transform_add functions for HEVC decoding
; * Copyright (c) 2014 Pierre-Edouard LEPERE
; *
; * This file is part of FFmpeg.
; *
; * FFmpeg 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.
; *
; * FFmpeg 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 FFmpeg; if not, write to the Free Software
; * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
; */
%include "libavutil/x86/x86util.asm"
x86: lavc: share more constant through defines Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
10 years ago
SECTION .text
x86: hevc: adding transform_add Reviewed-by: James Almer <jamrial@gmail.com> Approved-by: Ronald S. Bultje Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
x86: lavc: share more constant through defines Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
10 years ago
cextern pw_1023
%define max_pixels_10 pw_1023
x86: hevc: adding transform_add Reviewed-by: James Almer <jamrial@gmail.com> Approved-by: Ronald S. Bultje Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
;the tr_add macros and functions were largely inspired by x264 project's code in the h264_idct.asm file
%macro TR_ADD_MMX_4_8 0
mova m2, [r1]
mova m4, [r1+8]
pxor m3, m3
psubw m3, m2
packuswb m2, m2
packuswb m3, m3
pxor m5, m5
psubw m5, m4
packuswb m4, m4
packuswb m5, m5
movh m0, [r0 ]
movh m1, [r0+r2 ]
paddusb m0, m2
paddusb m1, m4
psubusb m0, m3
psubusb m1, m5
movh [r0 ], m0
movh [r0+r2 ], m1
%endmacro
INIT_MMX mmxext
; void ff_hevc_tranform_add_8_mmxext(uint8_t *dst, int16_t *coeffs, ptrdiff_t stride)
cglobal hevc_transform_add4_8, 3, 4, 6
TR_ADD_MMX_4_8
add r1, 16
lea r0, [r0+r2*2]
TR_ADD_MMX_4_8
RET
%macro TR_ADD_SSE_8_8 0
pxor m3, m3
mova m4, [r1]
mova m6, [r1+16]
mova m0, [r1+32]
mova m2, [r1+48]
psubw m5, m3, m4
psubw m7, m3, m6
psubw m1, m3, m0
packuswb m4, m0
packuswb m5, m1
psubw m3, m2
packuswb m6, m2
packuswb m7, m3
movq m0, [r0 ]
movq m1, [r0+r2 ]
movhps m0, [r0+r2*2]
movhps m1, [r0+r3 ]
paddusb m0, m4
paddusb m1, m6
psubusb m0, m5
psubusb m1, m7
movq [r0 ], m0
movq [r0+r2 ], m1
movhps [r0+2*r2], m0
movhps [r0+r3 ], m1
%endmacro
x86/hevc_res_add: refactor ff_hevc_transform_add{16,32}_8 * Reduced xmm register count to 7 (As such they are now enabled for x86_32). * Removed four movdqa (affects the sse2 version only). * pxor is now used to clear m0 only once. ~5% faster. Reviewed-by: Christophe Gisquet <christophe.gisquet@gmail.com> Signed-off-by: James Almer <jamrial@gmail.com>
11 years ago
%macro TR_ADD_SSE_16_32_8 3
x86/hevc_res_add: add ff_hevc_transform_add32_8_avx2 ~20% faster than AVX. Reviewed-by: Michael Niedermayer <michaelni@gmx.at> Signed-off-by: James Almer <jamrial@gmail.com>
10 years ago
mova xm2, [r1+%1 ]
mova xm6, [r1+%1+16]
%if cpuflag(avx2)
vinserti128 m2, m2, [r1+%1+32], 1
vinserti128 m6, m6, [r1+%1+48], 1
%endif
x86/hevc_res_add: refactor ff_hevc_transform_add{16,32}_8 * Reduced xmm register count to 7 (As such they are now enabled for x86_32). * Removed four movdqa (affects the sse2 version only). * pxor is now used to clear m0 only once. ~5% faster. Reviewed-by: Christophe Gisquet <christophe.gisquet@gmail.com> Signed-off-by: James Almer <jamrial@gmail.com>
11 years ago
%if cpuflag(avx)
psubw m1, m0, m2
psubw m5, m0, m6
%else
mova m1, m0
mova m5, m0
psubw m1, m2
psubw m5, m6
%endif
packuswb m2, m6
packuswb m1, m5
x86/hevc_res_add: add ff_hevc_transform_add32_8_avx2 ~20% faster than AVX. Reviewed-by: Michael Niedermayer <michaelni@gmx.at> Signed-off-by: James Almer <jamrial@gmail.com>
10 years ago
mova xm4, [r1+%1+mmsize*2 ]
mova xm6, [r1+%1+mmsize*2+16]
%if cpuflag(avx2)
vinserti128 m4, m4, [r1+%1+96 ], 1
vinserti128 m6, m6, [r1+%1+112], 1
%endif
x86/hevc_res_add: refactor ff_hevc_transform_add{16,32}_8 * Reduced xmm register count to 7 (As such they are now enabled for x86_32). * Removed four movdqa (affects the sse2 version only). * pxor is now used to clear m0 only once. ~5% faster. Reviewed-by: Christophe Gisquet <christophe.gisquet@gmail.com> Signed-off-by: James Almer <jamrial@gmail.com>
11 years ago
%if cpuflag(avx)
psubw m3, m0, m4
psubw m5, m0, m6
%else
mova m3, m0
mova m5, m0
psubw m3, m4
psubw m5, m6
%endif
packuswb m4, m6
packuswb m3, m5
paddusb m2, [%2]
paddusb m4, [%3]
psubusb m2, m1
psubusb m4, m3
mova [%2], m2
mova [%3], m4
x86: hevc: adding transform_add Reviewed-by: James Almer <jamrial@gmail.com> Approved-by: Ronald S. Bultje Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
%endmacro
x86/hecv_res_add: add ff_hevc_transform_add{8,16,32}_8_avx ~15% faster than sse2 Reviewed-by: Mickaël Raulet <mraulet@gmail.com> Reviewed-by: Christophe Gisquet <christophe.gisquet@gmail.com> Signed-off-by: James Almer <jamrial@gmail.com>
11 years ago
%macro TRANSFORM_ADD_8 0
; void ff_hevc_transform_add8_8_<opt>(uint8_t *dst, int16_t *coeffs, ptrdiff_t stride)
x86: hevc: adding transform_add Reviewed-by: James Almer <jamrial@gmail.com> Approved-by: Ronald S. Bultje Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
cglobal hevc_transform_add8_8, 3, 4, 8
lea r3, [r2*3]
TR_ADD_SSE_8_8
add r1, 64
lea r0, [r0+r2*4]
TR_ADD_SSE_8_8
RET
x86/hecv_res_add: add ff_hevc_transform_add{8,16,32}_8_avx ~15% faster than sse2 Reviewed-by: Mickaël Raulet <mraulet@gmail.com> Reviewed-by: Christophe Gisquet <christophe.gisquet@gmail.com> Signed-off-by: James Almer <jamrial@gmail.com>
11 years ago
; void ff_hevc_transform_add16_8_<opt>(uint8_t *dst, int16_t *coeffs, ptrdiff_t stride)
x86/hevc_res_add: refactor ff_hevc_transform_add{16,32}_8 * Reduced xmm register count to 7 (As such they are now enabled for x86_32). * Removed four movdqa (affects the sse2 version only). * pxor is now used to clear m0 only once. ~5% faster. Reviewed-by: Christophe Gisquet <christophe.gisquet@gmail.com> Signed-off-by: James Almer <jamrial@gmail.com>
11 years ago
cglobal hevc_transform_add16_8, 3, 4, 7
pxor m0, m0
x86: hevc: adding transform_add Reviewed-by: James Almer <jamrial@gmail.com> Approved-by: Ronald S. Bultje Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
lea r3, [r2*3]
x86/hevc_res_add: refactor ff_hevc_transform_add{16,32}_8 * Reduced xmm register count to 7 (As such they are now enabled for x86_32). * Removed four movdqa (affects the sse2 version only). * pxor is now used to clear m0 only once. ~5% faster. Reviewed-by: Christophe Gisquet <christophe.gisquet@gmail.com> Signed-off-by: James Almer <jamrial@gmail.com>
11 years ago
TR_ADD_SSE_16_32_8 0, r0, r0+r2
TR_ADD_SSE_16_32_8 64, r0+r2*2, r0+r3
x86: hevc: adding transform_add Reviewed-by: James Almer <jamrial@gmail.com> Approved-by: Ronald S. Bultje Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
%rep 3
add r1, 128
lea r0, [r0+r2*4]
x86/hevc_res_add: refactor ff_hevc_transform_add{16,32}_8 * Reduced xmm register count to 7 (As such they are now enabled for x86_32). * Removed four movdqa (affects the sse2 version only). * pxor is now used to clear m0 only once. ~5% faster. Reviewed-by: Christophe Gisquet <christophe.gisquet@gmail.com> Signed-off-by: James Almer <jamrial@gmail.com>
11 years ago
TR_ADD_SSE_16_32_8 0, r0, r0+r2
TR_ADD_SSE_16_32_8 64, r0+r2*2, r0+r3
x86: hevc: adding transform_add Reviewed-by: James Almer <jamrial@gmail.com> Approved-by: Ronald S. Bultje Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
%endrep
RET
x86/hecv_res_add: add ff_hevc_transform_add{8,16,32}_8_avx ~15% faster than sse2 Reviewed-by: Mickaël Raulet <mraulet@gmail.com> Reviewed-by: Christophe Gisquet <christophe.gisquet@gmail.com> Signed-off-by: James Almer <jamrial@gmail.com>
11 years ago
; void ff_hevc_transform_add32_8_<opt>(uint8_t *dst, int16_t *coeffs, ptrdiff_t stride)
x86/hevc_res_add: refactor ff_hevc_transform_add{16,32}_8 * Reduced xmm register count to 7 (As such they are now enabled for x86_32). * Removed four movdqa (affects the sse2 version only). * pxor is now used to clear m0 only once. ~5% faster. Reviewed-by: Christophe Gisquet <christophe.gisquet@gmail.com> Signed-off-by: James Almer <jamrial@gmail.com>
11 years ago
cglobal hevc_transform_add32_8, 3, 4, 7
pxor m0, m0
TR_ADD_SSE_16_32_8 0, r0, r0+16
TR_ADD_SSE_16_32_8 64, r0+r2, r0+r2+16
x86: hevc: adding transform_add Reviewed-by: James Almer <jamrial@gmail.com> Approved-by: Ronald S. Bultje Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
%rep 15
add r1, 128
lea r0, [r0+r2*2]
x86/hevc_res_add: refactor ff_hevc_transform_add{16,32}_8 * Reduced xmm register count to 7 (As such they are now enabled for x86_32). * Removed four movdqa (affects the sse2 version only). * pxor is now used to clear m0 only once. ~5% faster. Reviewed-by: Christophe Gisquet <christophe.gisquet@gmail.com> Signed-off-by: James Almer <jamrial@gmail.com>
11 years ago
TR_ADD_SSE_16_32_8 0, r0, r0+16
TR_ADD_SSE_16_32_8 64, r0+r2, r0+r2+16
x86: hevc: adding transform_add Reviewed-by: James Almer <jamrial@gmail.com> Approved-by: Ronald S. Bultje Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
%endrep
RET
x86/hecv_res_add: add ff_hevc_transform_add{8,16,32}_8_avx ~15% faster than sse2 Reviewed-by: Mickaël Raulet <mraulet@gmail.com> Reviewed-by: Christophe Gisquet <christophe.gisquet@gmail.com> Signed-off-by: James Almer <jamrial@gmail.com>
11 years ago
%endmacro
INIT_XMM sse2
TRANSFORM_ADD_8
INIT_XMM avx
TRANSFORM_ADD_8
x86/hevc_res_add: add missing guards to hevc_transform_add32_8_avx2 Should fix compilation with old Yasm/Nasm versions. Signed-off-by: James Almer <jamrial@gmail.com>
10 years ago
%if HAVE_AVX2_EXTERNAL
x86/hevc_res_add: add ff_hevc_transform_add32_8_avx2 ~20% faster than AVX. Reviewed-by: Michael Niedermayer <michaelni@gmx.at> Signed-off-by: James Almer <jamrial@gmail.com>
10 years ago
INIT_YMM avx2
; void ff_hevc_transform_add32_8_avx2(uint8_t *dst, int16_t *coeffs, ptrdiff_t stride)
cglobal hevc_transform_add32_8, 3, 4, 7
pxor m0, m0
lea r3, [r2*3]
TR_ADD_SSE_16_32_8 0, r0, r0+r2
TR_ADD_SSE_16_32_8 128, r0+r2*2, r0+r3
%rep 7
add r1, 256
lea r0, [r0+r2*4]
TR_ADD_SSE_16_32_8 0, r0, r0+r2
TR_ADD_SSE_16_32_8 128, r0+r2*2, r0+r3
%endrep
RET
x86/hevc_res_add: add missing guards to hevc_transform_add32_8_avx2 Should fix compilation with old Yasm/Nasm versions. Signed-off-by: James Almer <jamrial@gmail.com>
10 years ago
%endif
x86/hevc_res_add: add ff_hevc_transform_add32_8_avx2 ~20% faster than AVX. Reviewed-by: Michael Niedermayer <michaelni@gmx.at> Signed-off-by: James Almer <jamrial@gmail.com>
10 years ago
x86: hevc: adding transform_add Reviewed-by: James Almer <jamrial@gmail.com> Approved-by: Ronald S. Bultje Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
;-----------------------------------------------------------------------------
; void ff_hevc_transform_add_10(pixel *dst, int16_t *block, int stride)
;-----------------------------------------------------------------------------
%macro TR_ADD_SSE_8_10 4
mova m0, [%4]
mova m1, [%4+16]
mova m2, [%4+32]
mova m3, [%4+48]
paddw m0, [%1+0 ]
paddw m1, [%1+%2 ]
paddw m2, [%1+%2*2]
paddw m3, [%1+%3 ]
CLIPW m0, m4, m5
CLIPW m1, m4, m5
CLIPW m2, m4, m5
CLIPW m3, m4, m5
mova [%1+0 ], m0
mova [%1+%2 ], m1
mova [%1+%2*2], m2
mova [%1+%3 ], m3
%endmacro
%macro TR_ADD_MMX4_10 3
mova m0, [%1+0 ]
mova m1, [%1+%2 ]
paddw m0, [%3]
paddw m1, [%3+8]
CLIPW m0, m2, m3
CLIPW m1, m2, m3
mova [%1+0 ], m0
mova [%1+%2 ], m1
%endmacro
%macro TRANS_ADD_SSE_16_10 3
mova m0, [%3]
mova m1, [%3+16]
mova m2, [%3+32]
mova m3, [%3+48]
paddw m0, [%1 ]
paddw m1, [%1+16 ]
paddw m2, [%1+%2 ]
paddw m3, [%1+%2+16]
CLIPW m0, m4, m5
CLIPW m1, m4, m5
CLIPW m2, m4, m5
CLIPW m3, m4, m5
mova [%1 ], m0
mova [%1+16 ], m1
mova [%1+%2 ], m2
mova [%1+%2+16], m3
%endmacro
%macro TRANS_ADD_SSE_32_10 2
mova m0, [%2]
mova m1, [%2+16]
mova m2, [%2+32]
mova m3, [%2+48]
paddw m0, [%1 ]
paddw m1, [%1+16]
paddw m2, [%1+32]
paddw m3, [%1+48]
CLIPW m0, m4, m5
CLIPW m1, m4, m5
CLIPW m2, m4, m5
CLIPW m3, m4, m5
mova [%1 ], m0
mova [%1+16], m1
mova [%1+32], m2
mova [%1+48], m3
%endmacro
%macro TRANS_ADD16_AVX2 4
mova m0, [%4]
mova m1, [%4+32]
mova m2, [%4+64]
mova m3, [%4+96]
paddw m0, [%1+0 ]
paddw m1, [%1+%2 ]
paddw m2, [%1+%2*2]
paddw m3, [%1+%3 ]
CLIPW m0, m4, m5
CLIPW m1, m4, m5
CLIPW m2, m4, m5
CLIPW m3, m4, m5
mova [%1+0 ], m0
mova [%1+%2 ], m1
mova [%1+%2*2], m2
mova [%1+%3 ], m3
%endmacro
%macro TRANS_ADD32_AVX2 3
mova m0, [%3]
mova m1, [%3+32]
mova m2, [%3+64]
mova m3, [%3+96]
paddw m0, [%1 ]
paddw m1, [%1+32 ]
paddw m2, [%1+%2 ]
paddw m3, [%1+%2+32]
CLIPW m0, m4, m5
CLIPW m1, m4, m5
CLIPW m2, m4, m5
CLIPW m3, m4, m5
mova [%1 ], m0
mova [%1+32 ], m1
mova [%1+%2 ], m2
mova [%1+%2+32], m3
%endmacro
INIT_MMX mmxext
cglobal hevc_transform_add4_10,3,4, 6
pxor m2, m2
mova m3, [max_pixels_10]
TR_ADD_MMX4_10 r0, r2, r1
add r1, 16
lea r0, [r0+2*r2]
TR_ADD_MMX4_10 r0, r2, r1
RET
;-----------------------------------------------------------------------------
; void ff_hevc_transform_add_10(pixel *dst, int16_t *block, int stride)
;-----------------------------------------------------------------------------
INIT_XMM sse2
cglobal hevc_transform_add8_10,3,4,6
pxor m4, m4
mova m5, [max_pixels_10]
lea r3, [r2*3]
TR_ADD_SSE_8_10 r0, r2, r3, r1
lea r0, [r0+r2*4]
add r1, 64
TR_ADD_SSE_8_10 r0, r2, r3, r1
RET
cglobal hevc_transform_add16_10,3,4,6
pxor m4, m4
mova m5, [max_pixels_10]
TRANS_ADD_SSE_16_10 r0, r2, r1
%rep 7
lea r0, [r0+r2*2]
add r1, 64
TRANS_ADD_SSE_16_10 r0, r2, r1
%endrep
RET
cglobal hevc_transform_add32_10,3,4,6
pxor m4, m4
mova m5, [max_pixels_10]
TRANS_ADD_SSE_32_10 r0, r1
%rep 31
lea r0, [r0+r2]
add r1, 64
TRANS_ADD_SSE_32_10 r0, r1
%endrep
RET
%if HAVE_AVX2_EXTERNAL
INIT_YMM avx2
x86/hevc_res_add: fix register count in hevc_transform_add{16,32}_10_avx2 Signed-off-by: James Almer <jamrial@gmail.com>
11 years ago
cglobal hevc_transform_add16_10,3,4,6
x86: hevc: adding transform_add Reviewed-by: James Almer <jamrial@gmail.com> Approved-by: Ronald S. Bultje Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
pxor m4, m4
mova m5, [max_pixels_10]
lea r3, [r2*3]
TRANS_ADD16_AVX2 r0, r2, r3, r1
%rep 3
lea r0, [r0+r2*4]
add r1, 128
TRANS_ADD16_AVX2 r0, r2, r3, r1
%endrep
RET
x86/hevc_res_add: fix register count in hevc_transform_add{16,32}_10_avx2 Signed-off-by: James Almer <jamrial@gmail.com>
11 years ago
cglobal hevc_transform_add32_10,3,4,6
x86: hevc: adding transform_add Reviewed-by: James Almer <jamrial@gmail.com> Approved-by: Ronald S. Bultje Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
pxor m4, m4
mova m5, [max_pixels_10]
TRANS_ADD32_AVX2 r0, r2, r1
%rep 15
lea r0, [r0+r2*2]
add r1, 128
TRANS_ADD32_AVX2 r0, r2, r1
%endrep
RET
%endif ;HAVE_AVX_EXTERNAL