generic-library/vpx
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity
b5b61c179d
vpx/vp8/common/arm/neon/vp8_subpixelvariance16x16s_neon.asm

573 lines
18 KiB
NASM
Raw Normal View History

Initial WebM release
2010-05-18 17:58:33 +02:00
;
Use WebM in copyright notice for consistency Changes 'The VP8 project' to 'The WebM project', for consistency with other webmproject.org repositories. Fixes issue #97. Change-Id: I37c13ed5fbdb9d334ceef71c6350e9febed9bbba
2010-09-09 14:16:39 +02:00
; Copyright (c) 2010 The WebM project authors. All Rights Reserved.
Initial WebM release
2010-05-18 17:58:33 +02:00
;
cosmetics: trim trailing whitespace When the license headers were updated, they accidentally contained trailing whitespace, so unfortunately we have to touch all the files again. Change-Id: I236c05fade06589e417179c0444cb39b09e4200d
2010-06-18 18:39:21 +02:00
; Use of this source code is governed by a BSD-style license
LICENSE: update with latest text Change-Id: Ieebea089095d9073b3a94932791099f614ce120c
2010-06-04 22:19:40 +02:00
; that can be found in the LICENSE file in the root of the source
; tree. An additional intellectual property rights grant can be found
cosmetics: trim trailing whitespace When the license headers were updated, they accidentally contained trailing whitespace, so unfortunately we have to touch all the files again. Change-Id: I236c05fade06589e417179c0444cb39b09e4200d
2010-06-18 18:39:21 +02:00
; in the file PATENTS. All contributing project authors may
LICENSE: update with latest text Change-Id: Ieebea089095d9073b3a94932791099f614ce120c
2010-06-04 22:19:40 +02:00
; be found in the AUTHORS file in the root of the source tree.
Initial WebM release
2010-05-18 17:58:33 +02:00
;
Add half-pixel variance RTCD functions NEON has optimized 16x16 half-pixel variance functions, but they were not part of the RTCD framework. Add these functions to RTCD, so that other platforms can make use of this optimization in the future and special-case ARM code can be removed. A number of functions were taking two variance functions as parameters. These functions were changed to take a single parameter, a pointer to a struct containing all the variance functions for that block size. This provides additional flexibility for calling additional variance functions (the half-pixel special case, for example) and by initializing the table for all block sizes, we don't have to construct this function pointer table for each macroblock. Change-Id: I78289ff36b2715f9a7aa04d5f6fbe3d23acdc29c
2010-10-26 21:34:16 +02:00
EXPORT |vp8_variance_halfpixvar16x16_h_neon|
EXPORT |vp8_variance_halfpixvar16x16_v_neon|
EXPORT |vp8_variance_halfpixvar16x16_hv_neon|
Initial WebM release
2010-05-18 17:58:33 +02:00
EXPORT |vp8_sub_pixel_variance16x16s_neon|
ARM
REQUIRE8
PRESERVE8
AREA ||.text||, CODE, READONLY, ALIGN=2
;================================================
Add half-pixel variance RTCD functions NEON has optimized 16x16 half-pixel variance functions, but they were not part of the RTCD framework. Add these functions to RTCD, so that other platforms can make use of this optimization in the future and special-case ARM code can be removed. A number of functions were taking two variance functions as parameters. These functions were changed to take a single parameter, a pointer to a struct containing all the variance functions for that block size. This provides additional flexibility for calling additional variance functions (the half-pixel special case, for example) and by initializing the table for all block sizes, we don't have to construct this function pointer table for each macroblock. Change-Id: I78289ff36b2715f9a7aa04d5f6fbe3d23acdc29c
2010-10-26 21:34:16 +02:00
;unsigned int vp8_variance_halfpixvar16x16_h_neon
Initial WebM release
2010-05-18 17:58:33 +02:00
;(
; unsigned char *src_ptr, r0
; int src_pixels_per_line, r1
; unsigned char *dst_ptr, r2
; int dst_pixels_per_line, r3
; unsigned int *sse
;);
;================================================
Add half-pixel variance RTCD functions NEON has optimized 16x16 half-pixel variance functions, but they were not part of the RTCD framework. Add these functions to RTCD, so that other platforms can make use of this optimization in the future and special-case ARM code can be removed. A number of functions were taking two variance functions as parameters. These functions were changed to take a single parameter, a pointer to a struct containing all the variance functions for that block size. This provides additional flexibility for calling additional variance functions (the half-pixel special case, for example) and by initializing the table for all block sizes, we don't have to construct this function pointer table for each macroblock. Change-Id: I78289ff36b2715f9a7aa04d5f6fbe3d23acdc29c
2010-10-26 21:34:16 +02:00
|vp8_variance_halfpixvar16x16_h_neon| PROC
Initial WebM release
2010-05-18 17:58:33 +02:00
push {lr}
mov r12, #4 ;loop counter
ldr lr, [sp, #4] ;load *sse from stack
vmov.i8 q8, #0 ;q8 - sum
vmov.i8 q9, #0 ;q9, q10 - sse
vmov.i8 q10, #0
;First Pass: output_height lines x output_width columns (16x16)
vp8_filt_fpo16x16s_4_0_loop_neon
vld1.u8 {d0, d1, d2, d3}, [r0], r1 ;load src data
vld1.8 {q11}, [r2], r3
vld1.u8 {d4, d5, d6, d7}, [r0], r1
vld1.8 {q12}, [r2], r3
vld1.u8 {d8, d9, d10, d11}, [r0], r1
vld1.8 {q13}, [r2], r3
vld1.u8 {d12, d13, d14, d15}, [r0], r1
;pld [r0]
;pld [r0, r1]
;pld [r0, r1, lsl #1]
vext.8 q1, q0, q1, #1 ;construct src_ptr[1]
vext.8 q3, q2, q3, #1
vext.8 q5, q4, q5, #1
vext.8 q7, q6, q7, #1
vrhadd.u8 q0, q0, q1 ;(src_ptr[0]+src_ptr[1])/round/shift right 1
vld1.8 {q14}, [r2], r3
vrhadd.u8 q1, q2, q3
vrhadd.u8 q2, q4, q5
vrhadd.u8 q3, q6, q7
vsubl.u8 q4, d0, d22 ;diff
vsubl.u8 q5, d1, d23
vsubl.u8 q6, d2, d24
vsubl.u8 q7, d3, d25
vsubl.u8 q0, d4, d26
vsubl.u8 q1, d5, d27
vsubl.u8 q2, d6, d28
vsubl.u8 q3, d7, d29
vpadal.s16 q8, q4 ;sum
vmlal.s16 q9, d8, d8 ;sse
vmlal.s16 q10, d9, d9
subs r12, r12, #1
vpadal.s16 q8, q5
vmlal.s16 q9, d10, d10
vmlal.s16 q10, d11, d11
vpadal.s16 q8, q6
vmlal.s16 q9, d12, d12
vmlal.s16 q10, d13, d13
vpadal.s16 q8, q7
vmlal.s16 q9, d14, d14
vmlal.s16 q10, d15, d15
vpadal.s16 q8, q0 ;sum
vmlal.s16 q9, d0, d0 ;sse
vmlal.s16 q10, d1, d1
vpadal.s16 q8, q1
vmlal.s16 q9, d2, d2
vmlal.s16 q10, d3, d3
vpadal.s16 q8, q2
vmlal.s16 q9, d4, d4
vmlal.s16 q10, d5, d5
vpadal.s16 q8, q3
vmlal.s16 q9, d6, d6
vmlal.s16 q10, d7, d7
bne vp8_filt_fpo16x16s_4_0_loop_neon
vadd.u32 q10, q9, q10 ;accumulate sse
vpaddl.s32 q0, q8 ;accumulate sum
vpaddl.u32 q1, q10
vadd.s64 d0, d0, d1
vadd.u64 d1, d2, d3
vmull.s32 q5, d0, d0
vst1.32 {d1[0]}, [lr] ;store sse
Fix variance overflow In the variance calculations the difference is summed and later squared. When the sum exceeds sqrt(2^31) the value is treated as a negative when it is shifted which gives incorrect results. To fix this we cast the result of the multiplication as unsigned. The alternative fix is to shift sum down by 4 before multiplying. However that will reduce precision. For 16x16 blocks the maximum sum is 65280 and sqrt(2^31) is 46340 (and change). PPC change is untested. Change-Id: I1bad27ea0720067def6d71a6da5f789508cec265
2012-02-09 21:38:31 +01:00
vshr.u32 d10, d10, #8
vsub.u32 d0, d1, d10
Initial WebM release
2010-05-18 17:58:33 +02:00
vmov.32 r0, d0[0] ;return
pop {pc}
ENDP
;================================================
Add half-pixel variance RTCD functions NEON has optimized 16x16 half-pixel variance functions, but they were not part of the RTCD framework. Add these functions to RTCD, so that other platforms can make use of this optimization in the future and special-case ARM code can be removed. A number of functions were taking two variance functions as parameters. These functions were changed to take a single parameter, a pointer to a struct containing all the variance functions for that block size. This provides additional flexibility for calling additional variance functions (the half-pixel special case, for example) and by initializing the table for all block sizes, we don't have to construct this function pointer table for each macroblock. Change-Id: I78289ff36b2715f9a7aa04d5f6fbe3d23acdc29c
2010-10-26 21:34:16 +02:00
;unsigned int vp8_variance_halfpixvar16x16_v_neon
Initial WebM release
2010-05-18 17:58:33 +02:00
;(
; unsigned char *src_ptr, r0
; int src_pixels_per_line, r1
; unsigned char *dst_ptr, r2
; int dst_pixels_per_line, r3
; unsigned int *sse
;);
;================================================
Add half-pixel variance RTCD functions NEON has optimized 16x16 half-pixel variance functions, but they were not part of the RTCD framework. Add these functions to RTCD, so that other platforms can make use of this optimization in the future and special-case ARM code can be removed. A number of functions were taking two variance functions as parameters. These functions were changed to take a single parameter, a pointer to a struct containing all the variance functions for that block size. This provides additional flexibility for calling additional variance functions (the half-pixel special case, for example) and by initializing the table for all block sizes, we don't have to construct this function pointer table for each macroblock. Change-Id: I78289ff36b2715f9a7aa04d5f6fbe3d23acdc29c
2010-10-26 21:34:16 +02:00
|vp8_variance_halfpixvar16x16_v_neon| PROC
Initial WebM release
2010-05-18 17:58:33 +02:00
push {lr}
mov r12, #4 ;loop counter
vld1.u8 {q0}, [r0], r1 ;load src data
ldr lr, [sp, #4] ;load *sse from stack
vmov.i8 q8, #0 ;q8 - sum
vmov.i8 q9, #0 ;q9, q10 - sse
vmov.i8 q10, #0
vp8_filt_spo16x16s_0_4_loop_neon
vld1.u8 {q2}, [r0], r1
vld1.8 {q1}, [r2], r3
vld1.u8 {q4}, [r0], r1
vld1.8 {q3}, [r2], r3
vld1.u8 {q6}, [r0], r1
vld1.8 {q5}, [r2], r3
vld1.u8 {q15}, [r0], r1
vrhadd.u8 q0, q0, q2
vld1.8 {q7}, [r2], r3
vrhadd.u8 q2, q2, q4
vrhadd.u8 q4, q4, q6
vrhadd.u8 q6, q6, q15
vsubl.u8 q11, d0, d2 ;diff
vsubl.u8 q12, d1, d3
vsubl.u8 q13, d4, d6
vsubl.u8 q14, d5, d7
vsubl.u8 q0, d8, d10
vsubl.u8 q1, d9, d11
vsubl.u8 q2, d12, d14
vsubl.u8 q3, d13, d15
vpadal.s16 q8, q11 ;sum
vmlal.s16 q9, d22, d22 ;sse
vmlal.s16 q10, d23, d23
subs r12, r12, #1
vpadal.s16 q8, q12
vmlal.s16 q9, d24, d24
vmlal.s16 q10, d25, d25
vpadal.s16 q8, q13
vmlal.s16 q9, d26, d26
vmlal.s16 q10, d27, d27
vpadal.s16 q8, q14
vmlal.s16 q9, d28, d28
vmlal.s16 q10, d29, d29
vpadal.s16 q8, q0 ;sum
vmlal.s16 q9, d0, d0 ;sse
vmlal.s16 q10, d1, d1
vpadal.s16 q8, q1
vmlal.s16 q9, d2, d2
vmlal.s16 q10, d3, d3
vpadal.s16 q8, q2
vmlal.s16 q9, d4, d4
vmlal.s16 q10, d5, d5
vmov q0, q15
vpadal.s16 q8, q3
vmlal.s16 q9, d6, d6
vmlal.s16 q10, d7, d7
bne vp8_filt_spo16x16s_0_4_loop_neon
vadd.u32 q10, q9, q10 ;accumulate sse
vpaddl.s32 q0, q8 ;accumulate sum
vpaddl.u32 q1, q10
vadd.s64 d0, d0, d1
vadd.u64 d1, d2, d3
vmull.s32 q5, d0, d0
vst1.32 {d1[0]}, [lr] ;store sse
Fix variance overflow In the variance calculations the difference is summed and later squared. When the sum exceeds sqrt(2^31) the value is treated as a negative when it is shifted which gives incorrect results. To fix this we cast the result of the multiplication as unsigned. The alternative fix is to shift sum down by 4 before multiplying. However that will reduce precision. For 16x16 blocks the maximum sum is 65280 and sqrt(2^31) is 46340 (and change). PPC change is untested. Change-Id: I1bad27ea0720067def6d71a6da5f789508cec265
2012-02-09 21:38:31 +01:00
vshr.u32 d10, d10, #8
vsub.u32 d0, d1, d10
Initial WebM release
2010-05-18 17:58:33 +02:00
vmov.32 r0, d0[0] ;return
pop {pc}
ENDP
;================================================
Add half-pixel variance RTCD functions NEON has optimized 16x16 half-pixel variance functions, but they were not part of the RTCD framework. Add these functions to RTCD, so that other platforms can make use of this optimization in the future and special-case ARM code can be removed. A number of functions were taking two variance functions as parameters. These functions were changed to take a single parameter, a pointer to a struct containing all the variance functions for that block size. This provides additional flexibility for calling additional variance functions (the half-pixel special case, for example) and by initializing the table for all block sizes, we don't have to construct this function pointer table for each macroblock. Change-Id: I78289ff36b2715f9a7aa04d5f6fbe3d23acdc29c
2010-10-26 21:34:16 +02:00
;unsigned int vp8_variance_halfpixvar16x16_hv_neon
Initial WebM release
2010-05-18 17:58:33 +02:00
;(
; unsigned char *src_ptr, r0
; int src_pixels_per_line, r1
; unsigned char *dst_ptr, r2
; int dst_pixels_per_line, r3
; unsigned int *sse
;);
;================================================
Add half-pixel variance RTCD functions NEON has optimized 16x16 half-pixel variance functions, but they were not part of the RTCD framework. Add these functions to RTCD, so that other platforms can make use of this optimization in the future and special-case ARM code can be removed. A number of functions were taking two variance functions as parameters. These functions were changed to take a single parameter, a pointer to a struct containing all the variance functions for that block size. This provides additional flexibility for calling additional variance functions (the half-pixel special case, for example) and by initializing the table for all block sizes, we don't have to construct this function pointer table for each macroblock. Change-Id: I78289ff36b2715f9a7aa04d5f6fbe3d23acdc29c
2010-10-26 21:34:16 +02:00
|vp8_variance_halfpixvar16x16_hv_neon| PROC
Initial WebM release
2010-05-18 17:58:33 +02:00
push {lr}
vld1.u8 {d0, d1, d2, d3}, [r0], r1 ;load src data
ldr lr, [sp, #4] ;load *sse from stack
vmov.i8 q13, #0 ;q8 - sum
vext.8 q1, q0, q1, #1 ;construct src_ptr[1]
vmov.i8 q14, #0 ;q9, q10 - sse
vmov.i8 q15, #0
mov r12, #4 ;loop counter
vrhadd.u8 q0, q0, q1 ;(src_ptr[0]+src_ptr[1])/round/shift right 1
;First Pass: output_height lines x output_width columns (17x16)
vp8_filt16x16s_4_4_loop_neon
vld1.u8 {d4, d5, d6, d7}, [r0], r1
vld1.u8 {d8, d9, d10, d11}, [r0], r1
vld1.u8 {d12, d13, d14, d15}, [r0], r1
vld1.u8 {d16, d17, d18, d19}, [r0], r1
;pld [r0]
;pld [r0, r1]
;pld [r0, r1, lsl #1]
vext.8 q3, q2, q3, #1 ;construct src_ptr[1]
vext.8 q5, q4, q5, #1
vext.8 q7, q6, q7, #1
vext.8 q9, q8, q9, #1
vrhadd.u8 q1, q2, q3 ;(src_ptr[0]+src_ptr[1])/round/shift right 1
vrhadd.u8 q2, q4, q5
vrhadd.u8 q3, q6, q7
vrhadd.u8 q4, q8, q9
vld1.8 {q5}, [r2], r3
vrhadd.u8 q0, q0, q1
vld1.8 {q6}, [r2], r3
vrhadd.u8 q1, q1, q2
vld1.8 {q7}, [r2], r3
vrhadd.u8 q2, q2, q3
vld1.8 {q8}, [r2], r3
vrhadd.u8 q3, q3, q4
vsubl.u8 q9, d0, d10 ;diff
vsubl.u8 q10, d1, d11
vsubl.u8 q11, d2, d12
vsubl.u8 q12, d3, d13
vsubl.u8 q0, d4, d14 ;diff
vsubl.u8 q1, d5, d15
vsubl.u8 q5, d6, d16
vsubl.u8 q6, d7, d17
vpadal.s16 q13, q9 ;sum
vmlal.s16 q14, d18, d18 ;sse
vmlal.s16 q15, d19, d19
vpadal.s16 q13, q10 ;sum
vmlal.s16 q14, d20, d20 ;sse
vmlal.s16 q15, d21, d21
vpadal.s16 q13, q11 ;sum
vmlal.s16 q14, d22, d22 ;sse
vmlal.s16 q15, d23, d23
vpadal.s16 q13, q12 ;sum
vmlal.s16 q14, d24, d24 ;sse
vmlal.s16 q15, d25, d25
subs r12, r12, #1
vpadal.s16 q13, q0 ;sum
vmlal.s16 q14, d0, d0 ;sse
vmlal.s16 q15, d1, d1
vpadal.s16 q13, q1 ;sum
vmlal.s16 q14, d2, d2 ;sse
vmlal.s16 q15, d3, d3
vpadal.s16 q13, q5 ;sum
vmlal.s16 q14, d10, d10 ;sse
vmlal.s16 q15, d11, d11
vmov q0, q4
vpadal.s16 q13, q6 ;sum
vmlal.s16 q14, d12, d12 ;sse
vmlal.s16 q15, d13, d13
bne vp8_filt16x16s_4_4_loop_neon
vadd.u32 q15, q14, q15 ;accumulate sse
vpaddl.s32 q0, q13 ;accumulate sum
vpaddl.u32 q1, q15
vadd.s64 d0, d0, d1
vadd.u64 d1, d2, d3
vmull.s32 q5, d0, d0
vst1.32 {d1[0]}, [lr] ;store sse
Fix variance overflow In the variance calculations the difference is summed and later squared. When the sum exceeds sqrt(2^31) the value is treated as a negative when it is shifted which gives incorrect results. To fix this we cast the result of the multiplication as unsigned. The alternative fix is to shift sum down by 4 before multiplying. However that will reduce precision. For 16x16 blocks the maximum sum is 65280 and sqrt(2^31) is 46340 (and change). PPC change is untested. Change-Id: I1bad27ea0720067def6d71a6da5f789508cec265
2012-02-09 21:38:31 +01:00
vshr.u32 d10, d10, #8
vsub.u32 d0, d1, d10
Initial WebM release
2010-05-18 17:58:33 +02:00
vmov.32 r0, d0[0] ;return
pop {pc}
ENDP
;==============================
; r0 unsigned char *src_ptr,
; r1 int src_pixels_per_line,
; r2 int xoffset,
; r3 int yoffset,
; stack unsigned char *dst_ptr,
; stack int dst_pixels_per_line,
; stack unsigned int *sse
;note: in vp8_find_best_half_pixel_step()(called when 8<Speed<15), and first call of vp8_find_best_sub_pixel_step()
;(called when speed<=8). xoffset/yoffset can only be 4 or 0, which means either by pass the filter,
;or filter coeff is {64, 64}. This simplified program only works in this situation.
;note: It happens that both xoffset and yoffset are zero. This can be handled in c code later.
|vp8_sub_pixel_variance16x16s_neon| PROC
push {r4, lr}
ldr r4, [sp, #8] ;load *dst_ptr from stack
ldr r12, [sp, #12] ;load dst_pixels_per_line from stack
ldr lr, [sp, #16] ;load *sse from stack
cmp r2, #0 ;skip first_pass filter if xoffset=0
beq secondpass_bfilter16x16s_only
cmp r3, #0 ;skip second_pass filter if yoffset=0
beq firstpass_bfilter16x16s_only
vld1.u8 {d0, d1, d2, d3}, [r0], r1 ;load src data
sub sp, sp, #256 ;reserve space on stack for temporary storage
vext.8 q1, q0, q1, #1 ;construct src_ptr[1]
mov r3, sp
mov r2, #4 ;loop counter
vrhadd.u8 q0, q0, q1 ;(src_ptr[0]+src_ptr[1])/round/shift right 1
;First Pass: output_height lines x output_width columns (17x16)
vp8e_filt_blk2d_fp16x16s_loop_neon
vld1.u8 {d4, d5, d6, d7}, [r0], r1
vld1.u8 {d8, d9, d10, d11}, [r0], r1
vld1.u8 {d12, d13, d14, d15}, [r0], r1
vld1.u8 {d16, d17, d18, d19}, [r0], r1
;pld [r0]
;pld [r0, r1]
;pld [r0, r1, lsl #1]
vext.8 q3, q2, q3, #1 ;construct src_ptr[1]
vext.8 q5, q4, q5, #1
vext.8 q7, q6, q7, #1
vext.8 q9, q8, q9, #1
vrhadd.u8 q1, q2, q3 ;(src_ptr[0]+src_ptr[1])/round/shift right 1
vrhadd.u8 q2, q4, q5
vrhadd.u8 q3, q6, q7
vrhadd.u8 q4, q8, q9
vrhadd.u8 q0, q0, q1
vrhadd.u8 q1, q1, q2
vrhadd.u8 q2, q2, q3
vrhadd.u8 q3, q3, q4
subs r2, r2, #1
vst1.u8 {d0, d1 ,d2, d3}, [r3]! ;store result
vmov q0, q4
vst1.u8 {d4, d5, d6, d7}, [r3]!
bne vp8e_filt_blk2d_fp16x16s_loop_neon
b sub_pixel_variance16x16s_neon
;--------------------
firstpass_bfilter16x16s_only
mov r2, #2 ;loop counter
sub sp, sp, #256 ;reserve space on stack for temporary storage
mov r3, sp
;First Pass: output_height lines x output_width columns (16x16)
vp8e_filt_blk2d_fpo16x16s_loop_neon
vld1.u8 {d0, d1, d2, d3}, [r0], r1 ;load src data
vld1.u8 {d4, d5, d6, d7}, [r0], r1
vld1.u8 {d8, d9, d10, d11}, [r0], r1
vld1.u8 {d12, d13, d14, d15}, [r0], r1
;pld [r0]
;pld [r0, r1]
;pld [r0, r1, lsl #1]
vext.8 q1, q0, q1, #1 ;construct src_ptr[1]
vld1.u8 {d16, d17, d18, d19}, [r0], r1
vext.8 q3, q2, q3, #1
vld1.u8 {d20, d21, d22, d23}, [r0], r1
vext.8 q5, q4, q5, #1
vld1.u8 {d24, d25, d26, d27}, [r0], r1
vext.8 q7, q6, q7, #1
vld1.u8 {d28, d29, d30, d31}, [r0], r1
vext.8 q9, q8, q9, #1
vext.8 q11, q10, q11, #1
vext.8 q13, q12, q13, #1
vext.8 q15, q14, q15, #1
vrhadd.u8 q0, q0, q1 ;(src_ptr[0]+src_ptr[1])/round/shift right 1
vrhadd.u8 q1, q2, q3
vrhadd.u8 q2, q4, q5
vrhadd.u8 q3, q6, q7
vrhadd.u8 q4, q8, q9
vrhadd.u8 q5, q10, q11
vrhadd.u8 q6, q12, q13
vrhadd.u8 q7, q14, q15
subs r2, r2, #1
vst1.u8 {d0, d1, d2, d3}, [r3]! ;store result
vst1.u8 {d4, d5, d6, d7}, [r3]!
vst1.u8 {d8, d9, d10, d11}, [r3]!
vst1.u8 {d12, d13, d14, d15}, [r3]!
bne vp8e_filt_blk2d_fpo16x16s_loop_neon
b sub_pixel_variance16x16s_neon
;---------------------
secondpass_bfilter16x16s_only
sub sp, sp, #256 ;reserve space on stack for temporary storage
mov r2, #2 ;loop counter
vld1.u8 {d0, d1}, [r0], r1 ;load src data
mov r3, sp
vp8e_filt_blk2d_spo16x16s_loop_neon
vld1.u8 {d2, d3}, [r0], r1
vld1.u8 {d4, d5}, [r0], r1
vld1.u8 {d6, d7}, [r0], r1
vld1.u8 {d8, d9}, [r0], r1
vrhadd.u8 q0, q0, q1
vld1.u8 {d10, d11}, [r0], r1
vrhadd.u8 q1, q1, q2
vld1.u8 {d12, d13}, [r0], r1
vrhadd.u8 q2, q2, q3
vld1.u8 {d14, d15}, [r0], r1
vrhadd.u8 q3, q3, q4
vld1.u8 {d16, d17}, [r0], r1
vrhadd.u8 q4, q4, q5
vrhadd.u8 q5, q5, q6
vrhadd.u8 q6, q6, q7
vrhadd.u8 q7, q7, q8
subs r2, r2, #1
vst1.u8 {d0, d1, d2, d3}, [r3]! ;store result
vmov q0, q8
vst1.u8 {d4, d5, d6, d7}, [r3]!
vst1.u8 {d8, d9, d10, d11}, [r3]! ;store result
vst1.u8 {d12, d13, d14, d15}, [r3]!
bne vp8e_filt_blk2d_spo16x16s_loop_neon
b sub_pixel_variance16x16s_neon
;----------------------------
;variance16x16
sub_pixel_variance16x16s_neon
vmov.i8 q8, #0 ;q8 - sum
vmov.i8 q9, #0 ;q9, q10 - sse
vmov.i8 q10, #0
sub r3, r3, #256
mov r2, #4
sub_pixel_variance16x16s_neon_loop
vld1.8 {q0}, [r3]! ;Load up source and reference
vld1.8 {q1}, [r4], r12
vld1.8 {q2}, [r3]!
vld1.8 {q3}, [r4], r12
vld1.8 {q4}, [r3]!
vld1.8 {q5}, [r4], r12
vld1.8 {q6}, [r3]!
vld1.8 {q7}, [r4], r12
vsubl.u8 q11, d0, d2 ;diff
vsubl.u8 q12, d1, d3
vsubl.u8 q13, d4, d6
vsubl.u8 q14, d5, d7
vsubl.u8 q0, d8, d10
vsubl.u8 q1, d9, d11
vsubl.u8 q2, d12, d14
vsubl.u8 q3, d13, d15
vpadal.s16 q8, q11 ;sum
vmlal.s16 q9, d22, d22 ;sse
vmlal.s16 q10, d23, d23
subs r2, r2, #1
vpadal.s16 q8, q12
vmlal.s16 q9, d24, d24
vmlal.s16 q10, d25, d25
vpadal.s16 q8, q13
vmlal.s16 q9, d26, d26
vmlal.s16 q10, d27, d27
vpadal.s16 q8, q14
vmlal.s16 q9, d28, d28
vmlal.s16 q10, d29, d29
vpadal.s16 q8, q0 ;sum
vmlal.s16 q9, d0, d0 ;sse
vmlal.s16 q10, d1, d1
vpadal.s16 q8, q1
vmlal.s16 q9, d2, d2
vmlal.s16 q10, d3, d3
vpadal.s16 q8, q2
vmlal.s16 q9, d4, d4
vmlal.s16 q10, d5, d5
vpadal.s16 q8, q3
vmlal.s16 q9, d6, d6
vmlal.s16 q10, d7, d7
bne sub_pixel_variance16x16s_neon_loop
vadd.u32 q10, q9, q10 ;accumulate sse
vpaddl.s32 q0, q8 ;accumulate sum
vpaddl.u32 q1, q10
vadd.s64 d0, d0, d1
vadd.u64 d1, d2, d3
vmull.s32 q5, d0, d0
vst1.32 {d1[0]}, [lr] ;store sse
Fix variance overflow In the variance calculations the difference is summed and later squared. When the sum exceeds sqrt(2^31) the value is treated as a negative when it is shifted which gives incorrect results. To fix this we cast the result of the multiplication as unsigned. The alternative fix is to shift sum down by 4 before multiplying. However that will reduce precision. For 16x16 blocks the maximum sum is 65280 and sqrt(2^31) is 46340 (and change). PPC change is untested. Change-Id: I1bad27ea0720067def6d71a6da5f789508cec265
2012-02-09 21:38:31 +01:00
vshr.u32 d10, d10, #8
vsub.u32 d0, d1, d10
Initial WebM release
2010-05-18 17:58:33 +02:00
add sp, sp, #256
vmov.32 r0, d0[0] ;return
pop {r4, pc}
ENDP
END
Reference in New Issue Copy Permalink