vpx/vp9/encoder/x86/vp9_quantize_ssse3.asm
Ronald S. Bultje c8defcfdee Update quantize SSSE3 SIMD to cover 32x32 transform case also.
Encode time of bus (speed 0) 50 frames @ 1500kbps goes from 2min14.4 to
2min10.1, i.e. a 2.3% overall speed increase.

Change-Id: I3699580e74ec26c7d24e03681bc47ba25ee1ee87
2013-07-01 11:36:33 -07:00

199 lines
9.2 KiB
NASM

;
; Copyright (c) 2010 The WebM project authors. All Rights Reserved.
;
; Use of this source code is governed by a BSD-style license
; that can be found in the LICENSE file in the root of the source
; tree. An additional intellectual property rights grant can be found
; in the file PATENTS. All contributing project authors may
; be found in the AUTHORS file in the root of the source tree.
;
%include "third_party/x86inc/x86inc.asm"
SECTION_RODATA
pw_1: times 8 dw 1
SECTION .text
%macro QUANTIZE_FN 1
cglobal quantize_%1, 0, 6, 15, coeff, ncoeff, skip, zbin, round, quant, \
shift, qcoeff, dqcoeff, dequant, zbin_oq, \
eob, scan, iscan
cmp dword skipm, 0
jne .blank
; actual quantize loop - setup pointers, rounders, etc.
movifnidn coeffq, coeffmp
movifnidn ncoeffq, ncoeffmp
mov r2, dequantmp
movifnidn zbinq, zbinmp
movifnidn roundq, roundmp
movifnidn quantq, quantmp
movd m4, dword zbin_oqm ; m4 = zbin_oq
mova m0, [zbinq] ; m0 = zbin
punpcklwd m4, m4
mova m1, [roundq] ; m1 = round
pshufd m4, m4, 0
mova m2, [quantq] ; m2 = quant
paddw m0, m4 ; m0 = zbin + zbin_oq
mova m3, [r2q] ; m3 = dequant
psubw m0, [pw_1]
mov r2, shiftmp
mov r3, qcoeffmp
mova m4, [r2] ; m4 = shift
mov r4, dqcoeffmp
mov r5, iscanmp
mov r2, eobmp
pxor m5, m5 ; m5 = dedicated zero
DEFINE_ARGS coeff, ncoeff, eob, qcoeff, dqcoeff, iscan
lea coeffq, [ coeffq+ncoeffq*2]
lea iscanq, [ iscanq+ncoeffq*2]
lea qcoeffq, [ qcoeffq+ncoeffq*2]
lea dqcoeffq, [dqcoeffq+ncoeffq*2]
neg ncoeffq
; get DC and first 15 AC coeffs
mova m9, [ coeffq+ncoeffq*2+ 0] ; m9 = c[i]
mova m10, [ coeffq+ncoeffq*2+16] ; m10 = c[i]
pabsw m6, m9 ; m6 = abs(m9)
pabsw m11, m10 ; m11 = abs(m10)
%ifidn %1, b_32x32
paddw m6, m6
paddw m11, m11
%endif
pcmpgtw m7, m6, m0 ; m7 = c[i] >= zbin
punpckhqdq m0, m0
pcmpgtw m12, m11, m0 ; m12 = c[i] >= zbin
paddw m6, m1 ; m6 += round
punpckhqdq m1, m1
paddw m11, m1 ; m11 += round
pmulhw m8, m6, m2 ; m8 = m6*q>>16
punpckhqdq m2, m2
pmulhw m13, m11, m2 ; m13 = m11*q>>16
paddw m8, m6 ; m8 += m6
paddw m13, m11 ; m13 += m11
pmulhw m8, m4 ; m8 = m8*qsh>>16
punpckhqdq m4, m4
pmulhw m13, m4 ; m13 = m13*qsh>>16
psignw m8, m9 ; m8 = reinsert sign
psignw m13, m10 ; m13 = reinsert sign
pand m8, m7
pand m13, m12
mova [qcoeffq+ncoeffq*2+ 0], m8
mova [qcoeffq+ncoeffq*2+16], m13
%ifidn %1, b_32x32
pabsw m8, m8
pabsw m13, m13
%endif
pmullw m8, m3 ; dqc[i] = qc[i] * q
punpckhqdq m3, m3
pmullw m13, m3 ; dqc[i] = qc[i] * q
%ifidn %1, b_32x32
psrlw m8, 1
psrlw m13, 1
psignw m8, m9
psignw m13, m10
%endif
mova [dqcoeffq+ncoeffq*2+ 0], m8
mova [dqcoeffq+ncoeffq*2+16], m13
pcmpeqw m8, m5 ; m8 = c[i] == 0
pcmpeqw m13, m5 ; m13 = c[i] == 0
mova m6, [ iscanq+ncoeffq*2+ 0] ; m6 = scan[i]
mova m11, [ iscanq+ncoeffq*2+16] ; m11 = scan[i]
psubw m6, m7 ; m6 = scan[i] + 1
psubw m11, m12 ; m11 = scan[i] + 1
pandn m8, m6 ; m8 = max(eob)
pandn m13, m11 ; m13 = max(eob)
pmaxsw m8, m13
add ncoeffq, mmsize
jz .accumulate_eob
.ac_only_loop:
mova m9, [ coeffq+ncoeffq*2+ 0] ; m9 = c[i]
mova m10, [ coeffq+ncoeffq*2+16] ; m10 = c[i]
pabsw m6, m9 ; m6 = abs(m9)
pabsw m11, m10 ; m11 = abs(m10)
%ifidn %1, b_32x32
paddw m6, m6
paddw m11, m11
%endif
pcmpgtw m7, m6, m0 ; m7 = c[i] >= zbin
pcmpgtw m12, m11, m0 ; m12 = c[i] >= zbin
paddw m6, m1 ; m6 += round
paddw m11, m1 ; m11 += round
pmulhw m14, m6, m2 ; m14 = m6*q>>16
pmulhw m13, m11, m2 ; m13 = m11*q>>16
paddw m14, m6 ; m14 += m6
paddw m13, m11 ; m13 += m11
pmulhw m14, m4 ; m14 = m14*qsh>>16
pmulhw m13, m4 ; m13 = m13*qsh>>16
psignw m14, m9 ; m14 = reinsert sign
psignw m13, m10 ; m13 = reinsert sign
pand m14, m7
pand m13, m12
mova [qcoeffq+ncoeffq*2+ 0], m14
mova [qcoeffq+ncoeffq*2+16], m13
%ifidn %1, b_32x32
pabsw m14, m14
pabsw m13, m13
%endif
pmullw m14, m3 ; dqc[i] = qc[i] * q
pmullw m13, m3 ; dqc[i] = qc[i] * q
%ifidn %1, b_32x32
psrlw m14, 1
psrlw m13, 1
psignw m14, m9
psignw m13, m10
%endif
mova [dqcoeffq+ncoeffq*2+ 0], m14
mova [dqcoeffq+ncoeffq*2+16], m13
pcmpeqw m14, m5 ; m14 = c[i] == 0
pcmpeqw m13, m5 ; m13 = c[i] == 0
mova m6, [ iscanq+ncoeffq*2+ 0] ; m6 = scan[i]
mova m11, [ iscanq+ncoeffq*2+16] ; m11 = scan[i]
psubw m6, m7 ; m6 = scan[i] + 1
psubw m11, m12 ; m11 = scan[i] + 1
pandn m14, m6 ; m14 = max(eob)
pandn m13, m11 ; m13 = max(eob)
pmaxsw m8, m14
pmaxsw m8, m13
add ncoeffq, mmsize
jl .ac_only_loop
.accumulate_eob:
; horizontally accumulate/max eobs and write into [eob] memory pointer
pshufd m7, m8, 0xe
pmaxsw m8, m7
pshuflw m7, m8, 0xe
pmaxsw m8, m7
pshuflw m7, m8, 0x1
pmaxsw m8, m7
pextrw [eobq], m8, 0
RET
; skip-block, i.e. just write all zeroes
.blank:
mov r0, dqcoeffmp
movifnidn ncoeffq, ncoeffmp
mov r2, qcoeffmp
mov r3, eobmp
DEFINE_ARGS dqcoeff, ncoeff, qcoeff, eob
lea dqcoeffq, [dqcoeffq+ncoeffq*2]
lea qcoeffq, [ qcoeffq+ncoeffq*2]
neg ncoeffq
pxor m7, m7
.blank_loop:
mova [dqcoeffq+ncoeffq*2+ 0], m7
mova [dqcoeffq+ncoeffq*2+16], m7
mova [qcoeffq+ncoeffq*2+ 0], m7
mova [qcoeffq+ncoeffq*2+16], m7
add ncoeffq, mmsize
jl .blank_loop
mov word [eobq], 0
RET
%endmacro
INIT_XMM ssse3
QUANTIZE_FN b
QUANTIZE_FN b_32x32