generic-library/vpx
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Fix ssse3 quantize_fp functions while skip=1

Browse Source 
In ssse3 functions, DEFINE_ARGS macro hard codes qcoeff and dqcoeff
to r3 and r4. If skip is 1, qcoeff and dqcoeff need to be loaded
from the stack, which doesn't work because of the above definitions.
Currently, skip=1 case is not used in the encoder. This patch fixed
the issue, so it can be turned on later.

Change-Id: I998d696b1a7a85dca2b3bcee790b21c21e039147
This commit is contained in:
Yunqing Wang 2015-02-24 10:37:05 -08:00
parent 44adb8e283
commit 58e0159c80
1 changed files with 34 additions and 33 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

65
vp9/encoder/x86/vp9_quantize_ssse3_x86_64.asm
View File

@ -15,6 +15,7 @@ pw_1: times 8 dw 1
SECTION .text SECTION .text
; TODO(yunqingwang)fix quantize_b code for skip=1 case.
%macro QUANTIZE_FN 2 %macro QUANTIZE_FN 2
cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \ cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \
shift, qcoeff, dqcoeff, dequant, \ shift, qcoeff, dqcoeff, dequant, \
@ -244,11 +245,11 @@ cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \
psllw m2, 1 psllw m2, 1
%endif %endif
pxor m5, m5 ; m5 = dedicated zero pxor m5, m5 ; m5 = dedicated zero
DEFINE_ARGS coeff, ncoeff, d1, qcoeff, dqcoeff, iscan, d2, d3, d4, d5, eob
lea coeffq, [ coeffq+ncoeffq*2] lea coeffq, [ coeffq+ncoeffq*2]
lea iscanq, [ iscanq+ncoeffq*2] lea r5q, [ r5q+ncoeffq*2]
lea qcoeffq, [ qcoeffq+ncoeffq*2] lea r3q, [ r3q+ncoeffq*2]
lea dqcoeffq, [dqcoeffq+ncoeffq*2] lea r4q, [r4q+ncoeffq*2]
neg ncoeffq neg ncoeffq
; get DC and first 15 AC coeffs ; get DC and first 15 AC coeffs
@ -266,15 +267,15 @@ cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \
pmulhw m13, m11, m2 ; m13 = m11*q>>16 pmulhw m13, m11, m2 ; m13 = m11*q>>16
psignw m8, m9 ; m8 = reinsert sign psignw m8, m9 ; m8 = reinsert sign
psignw m13, m10 ; m13 = reinsert sign psignw m13, m10 ; m13 = reinsert sign
mova [qcoeffq+ncoeffq*2+ 0], m8 mova [r3q+ncoeffq*2+ 0], m8
mova [qcoeffq+ncoeffq*2+16], m13 mova [r3q+ncoeffq*2+16], m13
%ifidn %1, fp_32x32 %ifidn %1, fp_32x32
pabsw m8, m8 pabsw m8, m8
pabsw m13, m13 pabsw m13, m13
%endif %endif
pmullw m8, m3 ; dqc[i] = qc[i] * q pmullw m8, m3 ; r4[i] = r3[i] * q
punpckhqdq m3, m3 punpckhqdq m3, m3
pmullw m13, m3 ; dqc[i] = qc[i] * q pmullw m13, m3 ; r4[i] = r3[i] * q
%ifidn %1, fp_32x32 %ifidn %1, fp_32x32
psrlw m8, 1 psrlw m8, 1
psrlw m13, 1 psrlw m13, 1
@ -282,12 +283,12 @@ cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \
psignw m13, m10 psignw m13, m10
psrlw m0, m3, 2 psrlw m0, m3, 2
%endif %endif
mova [dqcoeffq+ncoeffq*2+ 0], m8 mova [r4q+ncoeffq*2+ 0], m8
mova [dqcoeffq+ncoeffq*2+16], m13 mova [r4q+ncoeffq*2+16], m13
pcmpeqw m8, m5 ; m8 = c[i] == 0 pcmpeqw m8, m5 ; m8 = c[i] == 0
pcmpeqw m13, m5 ; m13 = c[i] == 0 pcmpeqw m13, m5 ; m13 = c[i] == 0
mova m6, [ iscanq+ncoeffq*2+ 0] ; m6 = scan[i] mova m6, [ r5q+ncoeffq*2+ 0] ; m6 = scan[i]
mova m11, [ iscanq+ncoeffq*2+16] ; m11 = scan[i] mova m11, [ r5q+ncoeffq*2+16] ; m11 = scan[i]
psubw m6, m7 ; m6 = scan[i] + 1 psubw m6, m7 ; m6 = scan[i] + 1
psubw m11, m7 ; m11 = scan[i] + 1 psubw m11, m7 ; m11 = scan[i] + 1
pandn m8, m6 ; m8 = max(eob) pandn m8, m6 ; m8 = max(eob)
@ -318,26 +319,26 @@ cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \
pmulhw m13, m11, m2 ; m13 = m11*q>>16 pmulhw m13, m11, m2 ; m13 = m11*q>>16
psignw m14, m9 ; m14 = reinsert sign psignw m14, m9 ; m14 = reinsert sign
psignw m13, m10 ; m13 = reinsert sign psignw m13, m10 ; m13 = reinsert sign
mova [qcoeffq+ncoeffq*2+ 0], m14 mova [r3q+ncoeffq*2+ 0], m14
mova [qcoeffq+ncoeffq*2+16], m13 mova [r3q+ncoeffq*2+16], m13
%ifidn %1, fp_32x32 %ifidn %1, fp_32x32
pabsw m14, m14 pabsw m14, m14
pabsw m13, m13 pabsw m13, m13
%endif %endif
pmullw m14, m3 ; dqc[i] = qc[i] * q pmullw m14, m3 ; r4[i] = r3[i] * q
pmullw m13, m3 ; dqc[i] = qc[i] * q pmullw m13, m3 ; r4[i] = r3[i] * q
%ifidn %1, fp_32x32 %ifidn %1, fp_32x32
psrlw m14, 1 psrlw m14, 1
psrlw m13, 1 psrlw m13, 1
psignw m14, m9 psignw m14, m9
psignw m13, m10 psignw m13, m10
%endif %endif
mova [dqcoeffq+ncoeffq*2+ 0], m14 mova [r4q+ncoeffq*2+ 0], m14
mova [dqcoeffq+ncoeffq*2+16], m13 mova [r4q+ncoeffq*2+16], m13
pcmpeqw m14, m5 ; m14 = c[i] == 0 pcmpeqw m14, m5 ; m14 = c[i] == 0
pcmpeqw m13, m5 ; m13 = c[i] == 0 pcmpeqw m13, m5 ; m13 = c[i] == 0
mova m6, [ iscanq+ncoeffq*2+ 0] ; m6 = scan[i] mova m6, [ r5q+ncoeffq*2+ 0] ; m6 = scan[i]
mova m11, [ iscanq+ncoeffq*2+16] ; m11 = scan[i] mova m11, [ r5q+ncoeffq*2+16] ; m11 = scan[i]
psubw m6, m7 ; m6 = scan[i] + 1 psubw m6, m7 ; m6 = scan[i] + 1
psubw m11, m7 ; m11 = scan[i] + 1 psubw m11, m7 ; m11 = scan[i] + 1
pandn m14, m6 ; m14 = max(eob) pandn m14, m6 ; m14 = max(eob)
@ -350,10 +351,10 @@ cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \
%ifidn %1, fp_32x32 %ifidn %1, fp_32x32
jmp .accumulate_eob jmp .accumulate_eob
.skip_iter: .skip_iter:
mova [qcoeffq+ncoeffq*2+ 0], m5 mova [r3q+ncoeffq*2+ 0], m5
mova [qcoeffq+ncoeffq*2+16], m5 mova [r3q+ncoeffq*2+16], m5
mova [dqcoeffq+ncoeffq*2+ 0], m5 mova [r4q+ncoeffq*2+ 0], m5
mova [dqcoeffq+ncoeffq*2+16], m5 mova [r4q+ncoeffq*2+16], m5
add ncoeffq, mmsize add ncoeffq, mmsize
jl .ac_only_loop jl .ac_only_loop
%endif %endif
@ -377,19 +378,19 @@ cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \
movifnidn ncoeffq, ncoeffmp movifnidn ncoeffq, ncoeffmp
mov r2, qcoeffmp mov r2, qcoeffmp
mov r3, eobmp mov r3, eobmp
DEFINE_ARGS dqcoeff, ncoeff, qcoeff, eob
lea dqcoeffq, [dqcoeffq+ncoeffq*2] lea r0q, [r0q+ncoeffq*2]
lea qcoeffq, [ qcoeffq+ncoeffq*2] lea r2q, [r2q+ncoeffq*2]
neg ncoeffq neg ncoeffq
pxor m7, m7 pxor m7, m7
.blank_loop: .blank_loop:
mova [dqcoeffq+ncoeffq*2+ 0], m7 mova [r0q+ncoeffq*2+ 0], m7
mova [dqcoeffq+ncoeffq*2+16], m7 mova [r0q+ncoeffq*2+16], m7
mova [qcoeffq+ncoeffq*2+ 0], m7 mova [r2q+ncoeffq*2+ 0], m7
mova [qcoeffq+ncoeffq*2+16], m7 mova [r2q+ncoeffq*2+16], m7
add ncoeffq, mmsize add ncoeffq, mmsize
jl .blank_loop jl .blank_loop
mov word [eobq], 0 mov word [r3q], 0
RET RET
%endmacro %endmacro