virtualenv/FFmpeg
1
0
Fork 0
mirror of https://github.com/FFmpeg/FFmpeg.git synced 2024-11-26 19:01:44 +02:00
Code Issues Releases Activity
b3dc51dd57
FFmpeg/libavcodec/arm/rv34dsp_neon.S

157 lines
6.3 KiB
ArmAsm
Raw Normal View History

rv34: NEON optimised inverse transform functions Signed-off-by: Mans Rullgard <mans@mansr.com>
2011-09-24 14:05:55 +03:00
/*
* Copyright (c) 2011 Janne Grunau <janne-libav@jannau.net>
*
Reinstate proper FFmpeg license for all files.
2013-08-16 00:12:51 +03:00
* This file is part of FFmpeg.
rv34: NEON optimised inverse transform functions Signed-off-by: Mans Rullgard <mans@mansr.com>
2011-09-24 14:05:55 +03:00
*
Reinstate proper FFmpeg license for all files.
2013-08-16 00:12:51 +03:00
* FFmpeg is free software; you can redistribute it and/or
rv34: NEON optimised inverse transform functions Signed-off-by: Mans Rullgard <mans@mansr.com>
2011-09-24 14:05:55 +03:00
* 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.
*
Reinstate proper FFmpeg license for all files.
2013-08-16 00:12:51 +03:00
* FFmpeg is distributed in the hope that it will be useful,
rv34: NEON optimised inverse transform functions Signed-off-by: Mans Rullgard <mans@mansr.com>
2011-09-24 14:05:55 +03:00
* 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
Reinstate proper FFmpeg license for all files.
2013-08-16 00:12:51 +03:00
* License along with FFmpeg; if not, write to the Free Software
rv34: NEON optimised inverse transform functions Signed-off-by: Mans Rullgard <mans@mansr.com>
2011-09-24 14:05:55 +03:00
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
ARM: Move asm.S from libavcodec to libavutil This will allow for easier implementation of ARM-optimized functions in libraries other than libavcodec.
2012-05-21 22:46:23 +03:00
#include "libavutil/arm/asm.S"
rv34: add NEON rv34_idct_add Overall almost 4% faster, idct_add down from 350 to 85 cycles, idct_dc_add down from 83 to 30 cycles. squash: rv34 idct rearrange partial register loads
2012-01-15 21:16:02 +03:00
#include "neon.S"
rv34: NEON optimised inverse transform functions Signed-off-by: Mans Rullgard <mans@mansr.com>
2011-09-24 14:05:55 +03:00
rv34: add NEON rv34_idct_add Overall almost 4% faster, idct_add down from 350 to 85 cycles, idct_dc_add down from 83 to 30 cycles. squash: rv34 idct rearrange partial register loads
2012-01-15 21:16:02 +03:00
.macro rv34_inv_transform r0
vld1.16 {q14-q15}, [\r0,:128]
rv34: NEON optimised inverse transform functions Signed-off-by: Mans Rullgard <mans@mansr.com>
2011-09-24 14:05:55 +03:00
vmov.s16 d0, #13
vshll.s16 q12, d29, #3
vshll.s16 q13, d29, #4
vshll.s16 q9, d31, #3
vshll.s16 q1, d31, #4
vmull.s16 q10, d28, d0
vmlal.s16 q10, d30, d0
vmull.s16 q11, d28, d0
vmlsl.s16 q11, d30, d0
rv34: 1-pass inter MB reconstruction Implement 1-pass inverse transform and reconstruction for inter blocks.
2012-01-03 22:38:29 +03:00
vsubw.s16 q12, q12, d29 @ z2 = block[i+4*1]*7
vaddw.s16 q13, q13, d29 @ z3 = block[i+4*1]*17
rv34: NEON optimised inverse transform functions Signed-off-by: Mans Rullgard <mans@mansr.com>
2011-09-24 14:05:55 +03:00
vsubw.s16 q9, q9, d31
vaddw.s16 q1, q1, d31
rv34: 1-pass inter MB reconstruction Implement 1-pass inverse transform and reconstruction for inter blocks.
2012-01-03 22:38:29 +03:00
vadd.s32 q13, q13, q9 @ z3 = 17*block[i+4*1] + 7*block[i+4*3]
vsub.s32 q12, q12, q1 @ z2 = 7*block[i+4*1] - 17*block[i+4*3]
rv34: NEON optimised inverse transform functions Signed-off-by: Mans Rullgard <mans@mansr.com>
2011-09-24 14:05:55 +03:00
vadd.s32 q1, q10, q13 @ z0 + z3
vadd.s32 q2, q11, q12 @ z1 + z2
vsub.s32 q8, q10, q13 @ z0 - z3
vsub.s32 q3, q11, q12 @ z1 - z2
vtrn.32 q1, q2
vtrn.32 q3, q8
vswp d3, d6
vswp d5, d16
vmov.s32 d0, #13
vadd.s32 q10, q1, q3
vsub.s32 q11, q1, q3
vshl.s32 q12, q2, #3
vshl.s32 q9, q2, #4
vmul.s32 q13, q11, d0[0]
vshl.s32 q11, q8, #4
vadd.s32 q9, q9, q2
vshl.s32 q15, q8, #3
vsub.s32 q12, q12, q2
vadd.s32 q11, q11, q8
vmul.s32 q14, q10, d0[0]
vsub.s32 q8, q15, q8
vsub.s32 q12, q12, q11
vadd.s32 q9, q9, q8
vadd.s32 q2, q13, q12 @ z1 + z2
vadd.s32 q1, q14, q9 @ z0 + z3
vsub.s32 q3, q13, q12 @ z1 - z2
vsub.s32 q15, q14, q9 @ z0 - z3
.endm
Drop DCTELEM typedef It does not help as an abstraction and adds dsputil dependencies. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2013-01-20 03:02:29 +03:00
/* void rv34_idct_add_c(uint8_t *dst, int stride, int16_t *block) */
rv34: add NEON rv34_idct_add Overall almost 4% faster, idct_add down from 350 to 85 cycles, idct_dc_add down from 83 to 30 cycles. squash: rv34 idct rearrange partial register loads
2012-01-15 21:16:02 +03:00
function ff_rv34_idct_add_neon, export=1
mov r3, r0
rv34_inv_transform r2
vmov.i16 q12, #0
vrshrn.s32 d16, q1, #10 @ (z0 + z3) >> 10
vrshrn.s32 d17, q2, #10 @ (z1 + z2) >> 10
vrshrn.s32 d18, q3, #10 @ (z1 - z2) >> 10
vrshrn.s32 d19, q15, #10 @ (z0 - z3) >> 10
vld1.32 {d28[]}, [r0,:32], r1
vld1.32 {d29[]}, [r0,:32], r1
vtrn.32 q8, q9
vld1.32 {d28[1]}, [r0,:32], r1
vld1.32 {d29[1]}, [r0,:32], r1
vst1.16 {q12}, [r2,:128]! @ memset(block, 0, 16)
vst1.16 {q12}, [r2,:128] @ memset(block+16, 0, 16)
vtrn.16 d16, d17
vtrn.32 d28, d29
vtrn.16 d18, d19
vaddw.u8 q0, q8, d28
vaddw.u8 q1, q9, d29
vqmovun.s16 d28, q0
vqmovun.s16 d29, q1
vst1.32 {d28[0]}, [r3,:32], r1
vst1.32 {d28[1]}, [r3,:32], r1
vst1.32 {d29[0]}, [r3,:32], r1
vst1.32 {d29[1]}, [r3,:32], r1
bx lr
endfunc
Drop DCTELEM typedef It does not help as an abstraction and adds dsputil dependencies. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2013-01-20 03:02:29 +03:00
/* void rv34_inv_transform_noround_neon(int16_t *block); */
rv34: NEON optimised inverse transform functions Signed-off-by: Mans Rullgard <mans@mansr.com>
2011-09-24 14:05:55 +03:00
function ff_rv34_inv_transform_noround_neon, export=1
rv34: add NEON rv34_idct_add Overall almost 4% faster, idct_add down from 350 to 85 cycles, idct_dc_add down from 83 to 30 cycles. squash: rv34 idct rearrange partial register loads
2012-01-15 21:16:02 +03:00
rv34_inv_transform r0
rv34: NEON optimised inverse transform functions Signed-off-by: Mans Rullgard <mans@mansr.com>
2011-09-24 14:05:55 +03:00
vshl.s32 q11, q2, #1
vshl.s32 q10, q1, #1
vshl.s32 q12, q3, #1
vshl.s32 q13, q15, #1
vadd.s32 q11, q11, q2
vadd.s32 q10, q10, q1
vadd.s32 q12, q12, q3
vadd.s32 q13, q13, q15
vshrn.s32 d0, q10, #11 @ (z0 + z3)*3 >> 11
vshrn.s32 d1, q11, #11 @ (z1 + z2)*3 >> 11
vshrn.s32 d2, q12, #11 @ (z1 - z2)*3 >> 11
vshrn.s32 d3, q13, #11 @ (z0 - z3)*3 >> 11
rv34: 1-pass inter MB reconstruction Implement 1-pass inverse transform and reconstruction for inter blocks.
2012-01-03 22:38:29 +03:00
vst4.16 {d0[0], d1[0], d2[0], d3[0]}, [r0,:64]!
vst4.16 {d0[1], d1[1], d2[1], d3[1]}, [r0,:64]!
vst4.16 {d0[2], d1[2], d2[2], d3[2]}, [r0,:64]!
vst4.16 {d0[3], d1[3], d2[3], d3[3]}, [r0,:64]!
rv34: NEON optimised inverse transform functions Signed-off-by: Mans Rullgard <mans@mansr.com>
2011-09-24 14:05:55 +03:00
bx lr
endfunc
rv34: NEON optimised dc only inverse transform 30-50% faster than the C implementation, 0.5% overall speedup on bourne.rmvb.
2012-01-12 19:21:48 +03:00
rv34: add NEON rv34_idct_add Overall almost 4% faster, idct_add down from 350 to 85 cycles, idct_dc_add down from 83 to 30 cycles. squash: rv34 idct rearrange partial register loads
2012-01-15 21:16:02 +03:00
/* void ff_rv34_idct_dc_add_neon(uint8_t *dst, int stride, int dc) */
function ff_rv34_idct_dc_add_neon, export=1
mov r3, r0
vld1.32 {d28[]}, [r0,:32], r1
vld1.32 {d29[]}, [r0,:32], r1
vdup.16 d0, r2
vmov.s16 d1, #169
vld1.32 {d28[1]}, [r0,:32], r1
vmull.s16 q1, d0, d1 @ dc * 13 * 13
vld1.32 {d29[1]}, [r0,:32], r1
vrshrn.s32 d0, q1, #10 @ (dc * 13 * 13 + 0x200) >> 10
vmov d1, d0
vaddw.u8 q2, q0, d28
vaddw.u8 q3, q0, d29
vqmovun.s16 d28, q2
vqmovun.s16 d29, q3
vst1.32 {d28[0]}, [r3,:32], r1
vst1.32 {d29[0]}, [r3,:32], r1
vst1.32 {d28[1]}, [r3,:32], r1
vst1.32 {d29[1]}, [r3,:32], r1
bx lr
endfunc
rv34: NEON optimised dc only inverse transform 30-50% faster than the C implementation, 0.5% overall speedup on bourne.rmvb.
2012-01-12 19:21:48 +03:00
Drop DCTELEM typedef It does not help as an abstraction and adds dsputil dependencies. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2013-01-20 03:02:29 +03:00
/* void rv34_inv_transform_dc_noround_c(int16_t *block) */
rv34: NEON optimised dc only inverse transform 30-50% faster than the C implementation, 0.5% overall speedup on bourne.rmvb.
2012-01-12 19:21:48 +03:00
function ff_rv34_inv_transform_noround_dc_neon, export=1
ARM: rv34: fix asm syntax in dc transform functions Signed-off-by: Mans Rullgard <mans@mansr.com> Signed-off-by: Janne Grunau <janne-libav@jannau.net>
2012-01-12 22:37:31 +03:00
vld1.16 {d28[]}, [r0,:16] @ block[0]
rv34: NEON optimised dc only inverse transform 30-50% faster than the C implementation, 0.5% overall speedup on bourne.rmvb.
2012-01-12 19:21:48 +03:00
vmov.i16 d4, #251
ARM: rv34: fix asm syntax in dc transform functions Signed-off-by: Mans Rullgard <mans@mansr.com> Signed-off-by: Janne Grunau <janne-libav@jannau.net>
2012-01-12 22:37:31 +03:00
vorr.s16 d4, #256 @ 13^2 * 3
rv34: NEON optimised dc only inverse transform 30-50% faster than the C implementation, 0.5% overall speedup on bourne.rmvb.
2012-01-12 19:21:48 +03:00
vmull.s16 q3, d28, d4
vshrn.s32 d0, q3, #11
rv34: 1-pass inter MB reconstruction Implement 1-pass inverse transform and reconstruction for inter blocks.
2012-01-03 22:38:29 +03:00
vmov.i16 d1, d0
vst1.64 {q0}, [r0,:128]!
vst1.64 {q0}, [r0,:128]!
rv34: NEON optimised dc only inverse transform 30-50% faster than the C implementation, 0.5% overall speedup on bourne.rmvb.
2012-01-12 19:21:48 +03:00
bx lr
endfunc
Reference in New Issue Copy Permalink