diff options
author | Johann <johannkoenig@google.com> | 2010-07-23 13:42:30 -0400 |
---|---|---|
committer | Fritz Koenig <frkoenig@google.com> | 2010-07-26 08:55:19 -0400 |
commit | 56f5a9a060d4c89a71616a90207327e6c544f543 (patch) | |
tree | 173d002c3bde5a79e1b5e893b4a778712efd8f65 /vp8/decoder/arm/neon | |
parent | 98fcccfe9751894ace9693a39ba0609fe5ea904d (diff) | |
download | libvpx-56f5a9a060d4c89a71616a90207327e6c544f543.tar libvpx-56f5a9a060d4c89a71616a90207327e6c544f543.tar.gz libvpx-56f5a9a060d4c89a71616a90207327e6c544f543.tar.bz2 libvpx-56f5a9a060d4c89a71616a90207327e6c544f543.zip |
update arm idct functions
Jeff Muizelaar posted some changes to the idct/reconstruction c code.
This is the equivalent update for the arm assembly.
This shows a good boost on v6, and a minor boost on neon.
Here are some numbers for highway in qcif, 2641 frames:
HEAD neon: ~161 fps
new neon: ~162 fps
HEAD v6: ~102 fps
new v6: ~106 fps
The following functions have been updated for armv6 and neon:
vp8_dc_only_idct_add
vp8_dequant_idct_add
vp8_dequant_dc_idct_add
Conflicts:
vp8/decoder/arm/armv6/dequantdcidct_v6.asm
vp8/decoder/arm/armv6/dequantidct_v6.asm
Resolved by removing these files. When I rewrote the functions, I also
moved the files to dequant_dc_idct_v6.asm/dequant_idct_v6.asm
Change-Id: Ie3300df824d52474eca1a5134cf22d8b7809a5d4
Diffstat (limited to 'vp8/decoder/arm/neon')
-rw-r--r-- | vp8/decoder/arm/neon/dequant_dc_idct_neon.asm (renamed from vp8/decoder/arm/neon/dequantdcidct_neon.asm) | 72 | ||||
-rw-r--r-- | vp8/decoder/arm/neon/dequant_idct_neon.asm (renamed from vp8/decoder/arm/neon/dequantidct_neon.asm) | 67 |
2 files changed, 71 insertions, 68 deletions
diff --git a/vp8/decoder/arm/neon/dequantdcidct_neon.asm b/vp8/decoder/arm/neon/dequant_dc_idct_neon.asm index ae126f23f..ddb324068 100644 --- a/vp8/decoder/arm/neon/dequantdcidct_neon.asm +++ b/vp8/decoder/arm/neon/dequant_dc_idct_neon.asm @@ -9,31 +9,43 @@ ; - EXPORT |vp8_dequant_dc_idct_neon| + EXPORT |vp8_dequant_dc_idct_add_neon| ARM REQUIRE8 PRESERVE8 AREA ||.text||, CODE, READONLY, ALIGN=2 -;void vp8_dequant_dc_idct_c(short *input, short *dq, short *output, int pitch, int Dc); +;void vp8_dequant_dc_idct_add_neon(short *input, short *dq, unsigned char *pred, +; unsigned char *dest, int pitch, int stride, +; int Dc); ; r0 short *input, ; r1 short *dq, -; r2 short *output, -; r3 int pitch, -; (stack) int Dc -|vp8_dequant_dc_idct_neon| PROC +; r2 unsigned char *pred +; r3 unsigned char *dest +; sp int pitch +; sp+4 int stride +; sp+8 int Dc +|vp8_dequant_dc_idct_add_neon| PROC vld1.16 {q3, q4}, [r0] vld1.16 {q5, q6}, [r1] - ldr r1, [sp] ;load Dc from stack + ldr r1, [sp, #8] ;load Dc from stack - ldr r12, _dcidct_coeff_ + ldr r12, _CONSTANTS_ vmul.i16 q1, q3, q5 ;input for short_idct4x4llm_neon vmul.i16 q2, q4, q6 vmov.16 d2[0], r1 + ldr r1, [sp] ; pitch + vld1.32 {d14[0]}, [r2], r1 + vld1.32 {d14[1]}, [r2], r1 + vld1.32 {d15[0]}, [r2], r1 + vld1.32 {d15[1]}, [r2] + + ldr r1, [sp, #4] ; stride + ;|short_idct4x4llm_neon| PROC vld1.16 {d0}, [r12] vswp d3, d4 ;q2(vp[4] vp[12]) @@ -47,14 +59,9 @@ vshr.s16 q3, q3, #1 vshr.s16 q4, q4, #1 - vqadd.s16 q3, q3, q2 ;modify since sinpi8sqrt2 > 65536/2 (negtive number) + vqadd.s16 q3, q3, q2 vqadd.s16 q4, q4, q2 - ;d6 - c1:temp1 - ;d7 - d1:temp2 - ;d8 - d1:temp1 - ;d9 - c1:temp2 - vqsub.s16 d10, d6, d9 ;c1 vqadd.s16 d11, d7, d8 ;d1 @@ -83,7 +90,7 @@ vshr.s16 q3, q3, #1 vshr.s16 q4, q4, #1 - vqadd.s16 q3, q3, q2 ;modify since sinpi8sqrt2 > 65536/2 (negtive number) + vqadd.s16 q3, q3, q2 vqadd.s16 q4, q4, q2 vqsub.s16 d10, d6, d9 ;c1 @@ -101,34 +108,29 @@ vrshr.s16 d4, d4, #3 vrshr.s16 d5, d5, #3 - add r1, r2, r3 - add r12, r1, r3 - add r0, r12, r3 - vtrn.32 d2, d4 vtrn.32 d3, d5 vtrn.16 d2, d3 vtrn.16 d4, d5 - vst1.16 {d2}, [r2] - vst1.16 {d3}, [r1] - vst1.16 {d4}, [r12] - vst1.16 {d5}, [r0] + vaddw.u8 q1, q1, d14 + vaddw.u8 q2, q2, d15 - bx lr + vqmovun.s16 d0, q1 + vqmovun.s16 d1, q2 + + vst1.32 {d0[0]}, [r3], r1 + vst1.32 {d0[1]}, [r3], r1 + vst1.32 {d1[0]}, [r3], r1 + vst1.32 {d1[1]}, [r3] - ENDP + bx lr -;----------------- - AREA dcidct4x4_dat, DATA, READWRITE ;read/write by default -;Data section with name data_area is specified. DCD reserves space in memory for 48 data. -;One word each is reserved. Label filter_coeff can be used to access the data. -;Data address: filter_coeff, filter_coeff+4, filter_coeff+8 ... -_dcidct_coeff_ - DCD dcidct_coeff -dcidct_coeff - DCD 0x4e7b4e7b, 0x8a8c8a8c + ENDP ; |vp8_dequant_dc_idct_add_neon| -;20091, 20091, 35468, 35468 +; Constant Pool +_CONSTANTS_ DCD cospi8sqrt2minus1 +cospi8sqrt2minus1 DCD 0x4e7b4e7b +sinpi8sqrt2 DCD 0x8a8c8a8c END diff --git a/vp8/decoder/arm/neon/dequantidct_neon.asm b/vp8/decoder/arm/neon/dequant_idct_neon.asm index e0888ed35..5c60dd690 100644 --- a/vp8/decoder/arm/neon/dequantidct_neon.asm +++ b/vp8/decoder/arm/neon/dequant_idct_neon.asm @@ -9,22 +9,33 @@ ; - EXPORT |vp8_dequant_idct_neon| + EXPORT |vp8_dequant_idct_add_neon| ARM REQUIRE8 PRESERVE8 AREA ||.text||, CODE, READONLY, ALIGN=2 -;void vp8_dequant_idct_c(short *input, short *dq, short *output, int pitch); +;void vp8_dequant_idct_neon(short *input, short *dq, unsigned char *pred, +; unsigned char *dest, int pitch, int stride) ; r0 short *input, ; r1 short *dq, -; r2 short *output, -; r3 int pitch, -|vp8_dequant_idct_neon| PROC +; r2 unsigned char *pred +; r3 unsigned char *dest +; sp int pitch +; sp+4 int stride + +|vp8_dequant_idct_add_neon| PROC vld1.16 {q3, q4}, [r0] vld1.16 {q5, q6}, [r1] + ldr r1, [sp] ; pitch + vld1.32 {d14[0]}, [r2], r1 + vld1.32 {d14[1]}, [r2], r1 + vld1.32 {d15[0]}, [r2], r1 + vld1.32 {d15[1]}, [r2] + + ldr r1, [sp, #4] ; stride - ldr r12, _didct_coeff_ + ldr r12, _CONSTANTS_ vmul.i16 q1, q3, q5 ;input for short_idct4x4llm_neon vmul.i16 q2, q4, q6 @@ -42,14 +53,9 @@ vshr.s16 q3, q3, #1 vshr.s16 q4, q4, #1 - vqadd.s16 q3, q3, q2 ;modify since sinpi8sqrt2 > 65536/2 (negtive number) + vqadd.s16 q3, q3, q2 vqadd.s16 q4, q4, q2 - ;d6 - c1:temp1 - ;d7 - d1:temp2 - ;d8 - d1:temp1 - ;d9 - c1:temp2 - vqsub.s16 d10, d6, d9 ;c1 vqadd.s16 d11, d7, d8 ;d1 @@ -78,7 +84,7 @@ vshr.s16 q3, q3, #1 vshr.s16 q4, q4, #1 - vqadd.s16 q3, q3, q2 ;modify since sinpi8sqrt2 > 65536/2 (negtive number) + vqadd.s16 q3, q3, q2 vqadd.s16 q4, q4, q2 vqsub.s16 d10, d6, d9 ;c1 @@ -96,34 +102,29 @@ vrshr.s16 d4, d4, #3 vrshr.s16 d5, d5, #3 - add r1, r2, r3 - add r12, r1, r3 - add r0, r12, r3 - vtrn.32 d2, d4 vtrn.32 d3, d5 vtrn.16 d2, d3 vtrn.16 d4, d5 - vst1.16 {d2}, [r2] - vst1.16 {d3}, [r1] - vst1.16 {d4}, [r12] - vst1.16 {d5}, [r0] + vaddw.u8 q1, q1, d14 + vaddw.u8 q2, q2, d15 - bx lr + vqmovun.s16 d0, q1 + vqmovun.s16 d1, q2 + + vst1.32 {d0[0]}, [r3], r1 + vst1.32 {d0[1]}, [r3], r1 + vst1.32 {d1[0]}, [r3], r1 + vst1.32 {d1[1]}, [r3] - ENDP + bx lr -;----------------- - AREA didct4x4_dat, DATA, READWRITE ;read/write by default -;Data section with name data_area is specified. DCD reserves space in memory for 48 data. -;One word each is reserved. Label filter_coeff can be used to access the data. -;Data address: filter_coeff, filter_coeff+4, filter_coeff+8 ... -_didct_coeff_ - DCD didct_coeff -didct_coeff - DCD 0x4e7b4e7b, 0x8a8c8a8c + ENDP ; |vp8_dequant_idct_add_neon| -;20091, 20091, 35468, 35468 +; Constant Pool +_CONSTANTS_ DCD cospi8sqrt2minus1 +cospi8sqrt2minus1 DCD 0x4e7b4e7b +sinpi8sqrt2 DCD 0x8a8c8a8c END |