generic-library/vpx
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity
b334bfc322
vpx/vp8/common/ppc/variance_altivec.asm

376 lines
6.9 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
;
.globl vp8_get8x8var_ppc
.globl vp8_get16x16var_ppc
.globl vp8_mse16x16_ppc
.globl vp8_variance16x16_ppc
.globl vp8_variance16x8_ppc
.globl vp8_variance8x16_ppc
.globl vp8_variance8x8_ppc
.globl vp8_variance4x4_ppc
.macro load_aligned_16 V R O
lvsl v3, 0, \R ;# permutate value for alignment
lvx v1, 0, \R
lvx v2, \O, \R
vperm \V, v1, v2, v3
.endm
.macro prologue
mfspr r11, 256 ;# get old VRSAVE
oris r12, r11, 0xffc0
mtspr 256, r12 ;# set VRSAVE
stwu r1, -32(r1) ;# create space on the stack
li r10, 16 ;# load offset and loop counter
vspltisw v7, 0 ;# zero for merging
vspltisw v8, 0 ;# zero out total to start
vspltisw v9, 0 ;# zero out total for dif^2
.endm
.macro epilogue
addi r1, r1, 32 ;# recover stack
mtspr 256, r11 ;# reset old VRSAVE
.endm
.macro compute_sum_sse
;# Compute sum first. Unpack to so signed subract
;# can be used. Only have a half word signed
;# subract. Do high, then low.
vmrghb v2, v7, v4
vmrghb v3, v7, v5
vsubshs v2, v2, v3
vsum4shs v8, v2, v8
vmrglb v2, v7, v4
vmrglb v3, v7, v5
vsubshs v2, v2, v3
vsum4shs v8, v2, v8
;# Now compute sse.
vsububs v2, v4, v5
vsububs v3, v5, v4
vor v2, v2, v3
vmsumubm v9, v2, v2, v9
.endm
.macro variance_16 DS loop_label store_sum
\loop_label:
;# only one of the inputs should need to be aligned.
load_aligned_16 v4, r3, r10
load_aligned_16 v5, r5, r10
;# move onto the next line
add r3, r3, r4
add r5, r5, r6
compute_sum_sse
bdnz \loop_label
vsumsws v8, v8, v7
vsumsws v9, v9, v7
stvx v8, 0, r1
lwz r3, 12(r1)
stvx v9, 0, r1
lwz r4, 12(r1)
.if \store_sum
stw r3, 0(r8) ;# sum
.endif
stw r4, 0(r7) ;# sse
mullw r3, r3, r3 ;# sum*sum
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
srlwi r3, r3, \DS ;# (sum*sum) >> DS
Initial WebM release
2010-05-18 17:58:33 +02:00
subf r3, r3, r4 ;# sse - ((sum*sum) >> DS)
.endm
.macro variance_8 DS loop_label store_sum
\loop_label:
;# only one of the inputs should need to be aligned.
load_aligned_16 v4, r3, r10
load_aligned_16 v5, r5, r10
;# move onto the next line
add r3, r3, r4
add r5, r5, r6
;# only one of the inputs should need to be aligned.
load_aligned_16 v6, r3, r10
load_aligned_16 v0, r5, r10
;# move onto the next line
add r3, r3, r4
add r5, r5, r6
vmrghb v4, v4, v6
vmrghb v5, v5, v0
compute_sum_sse
bdnz \loop_label
vsumsws v8, v8, v7
vsumsws v9, v9, v7
stvx v8, 0, r1
lwz r3, 12(r1)
stvx v9, 0, r1
lwz r4, 12(r1)
.if \store_sum
stw r3, 0(r8) ;# sum
.endif
stw r4, 0(r7) ;# sse
mullw r3, r3, r3 ;# sum*sum
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
srlwi r3, r3, \DS ;# (sum*sum) >> 8
Initial WebM release
2010-05-18 17:58:33 +02:00
subf r3, r3, r4 ;# sse - ((sum*sum) >> 8)
.endm
.align 2
;# r3 unsigned char *src_ptr
;# r4 int source_stride
;# r5 unsigned char *ref_ptr
;# r6 int recon_stride
;# r7 unsigned int *SSE
;# r8 int *Sum
;#
;# r3 return value
vp8_get8x8var_ppc:
prologue
li r9, 4
mtctr r9
variance_8 6, get8x8var_loop, 1
epilogue
blr
.align 2
;# r3 unsigned char *src_ptr
;# r4 int source_stride
;# r5 unsigned char *ref_ptr
;# r6 int recon_stride
;# r7 unsigned int *SSE
;# r8 int *Sum
;#
;# r3 return value
vp8_get16x16var_ppc:
prologue
mtctr r10
variance_16 8, get16x16var_loop, 1
epilogue
blr
.align 2
;# r3 unsigned char *src_ptr
;# r4 int source_stride
;# r5 unsigned char *ref_ptr
;# r6 int recon_stride
;# r7 unsigned int *sse
;#
;# r 3 return value
vp8_mse16x16_ppc:
prologue
mtctr r10
mse16x16_loop:
;# only one of the inputs should need to be aligned.
load_aligned_16 v4, r3, r10
load_aligned_16 v5, r5, r10
;# move onto the next line
add r3, r3, r4
add r5, r5, r6
;# Now compute sse.
vsububs v2, v4, v5
vsububs v3, v5, v4
vor v2, v2, v3
vmsumubm v9, v2, v2, v9
bdnz mse16x16_loop
vsumsws v9, v9, v7
stvx v9, 0, r1
lwz r3, 12(r1)
stvx v9, 0, r1
lwz r3, 12(r1)
stw r3, 0(r7) ;# sse
epilogue
blr
.align 2
;# r3 unsigned char *src_ptr
;# r4 int source_stride
;# r5 unsigned char *ref_ptr
;# r6 int recon_stride
;# r7 unsigned int *sse
;#
;# r3 return value
vp8_variance16x16_ppc:
prologue
mtctr r10
variance_16 8, variance16x16_loop, 0
epilogue
blr
.align 2
;# r3 unsigned char *src_ptr
;# r4 int source_stride
;# r5 unsigned char *ref_ptr
;# r6 int recon_stride
;# r7 unsigned int *sse
;#
;# r3 return value
vp8_variance16x8_ppc:
prologue
li r9, 8
mtctr r9
variance_16 7, variance16x8_loop, 0
epilogue
blr
.align 2
;# r3 unsigned char *src_ptr
;# r4 int source_stride
;# r5 unsigned char *ref_ptr
;# r6 int recon_stride
;# r7 unsigned int *sse
;#
;# r3 return value
vp8_variance8x16_ppc:
prologue
li r9, 8
mtctr r9
variance_8 7, variance8x16_loop, 0
epilogue
blr
.align 2
;# r3 unsigned char *src_ptr
;# r4 int source_stride
;# r5 unsigned char *ref_ptr
;# r6 int recon_stride
;# r7 unsigned int *sse
;#
;# r3 return value
vp8_variance8x8_ppc:
prologue
li r9, 4
mtctr r9
variance_8 6, variance8x8_loop, 0
epilogue
blr
.macro transfer_4x4 I P
lwz r0, 0(\I)
add \I, \I, \P
lwz r10,0(\I)
add \I, \I, \P
lwz r8, 0(\I)
add \I, \I, \P
lwz r9, 0(\I)
stw r0, 0(r1)
stw r10, 4(r1)
stw r8, 8(r1)
stw r9, 12(r1)
.endm
.align 2
;# r3 unsigned char *src_ptr
;# r4 int source_stride
;# r5 unsigned char *ref_ptr
;# r6 int recon_stride
;# r7 unsigned int *sse
;#
;# r3 return value
vp8_variance4x4_ppc:
prologue
transfer_4x4 r3, r4
lvx v4, 0, r1
transfer_4x4 r5, r6
lvx v5, 0, r1
compute_sum_sse
vsumsws v8, v8, v7
vsumsws v9, v9, v7
stvx v8, 0, r1
lwz r3, 12(r1)
stvx v9, 0, r1
lwz r4, 12(r1)
stw r4, 0(r7) ;# sse
mullw r3, r3, r3 ;# sum*sum
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
srlwi r3, r3, 4 ;# (sum*sum) >> 4
Initial WebM release
2010-05-18 17:58:33 +02:00
subf r3, r3, r4 ;# sse - ((sum*sum) >> 4)
epilogue
blr
Reference in New Issue Copy Permalink