vpx/vpx_dsp/x86/quantize_avx_x86_64.asm
Geza Lore 9cfba09ac0 Optimize vpx_quantize_{b,b_32x32} assembler.
Added optimization of the 8 bit assembly quantizer routines. This makes
these functions up to 100% faster, depending on encoding parameters.

This patch maskes the encoder faster in both the high bitdepth and 8bit
configurations. In the high bitdepth configuration, it effects profile 0
only.

Based on my profiling using 1080p input the net gain is between 1-3% for
the 8 bit config, and around 2.5-4.5% for the high bitdepth config,
depending on target bitrate. The difference between the 8 bit and high
bitdepth configurations for the same encoder run is reduced by 1% in all
cases I have profiled.

Change-Id: I86714a6b7364da20cd468cd784247009663a5140
2015-10-20 10:11:19 +01:00

545 lines
21 KiB
NASM

;
; Copyright (c) 2015 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 .text
%macro QUANTIZE_FN 2
cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \
shift, qcoeff, dqcoeff, dequant, \
eob, scan, iscan
vzeroupper
; If we can skip this block, then just zero the output
cmp skipmp, 0
jne .blank
%ifnidn %1, b_32x32
; Special case for ncoeff == 16, as it is frequent and we can save on
; not setting up a loop.
cmp ncoeffmp, 16
jne .generic
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; Special case of ncoeff == 16
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
.single:
movifnidn coeffq, coeffmp
movifnidn zbinq, zbinmp
mova m0, [zbinq] ; m0 = zbin
; Get DC and first 15 AC coeffs - in this special case, that is all.
%if CONFIG_VP9_HIGHBITDEPTH
; coeff stored as 32bit numbers but we process them as 16 bit numbers
mova m9, [coeffq]
packssdw m9, [coeffq+16] ; m9 = c[i]
mova m10, [coeffq+32]
packssdw m10, [coeffq+48] ; m10 = c[i]
%else
mova m9, [coeffq] ; m9 = c[i]
mova m10, [coeffq+16] ; m10 = c[i]
%endif
mov r0, eobmp ; Output pointer
mov r1, qcoeffmp ; Output pointer
mov r2, dqcoeffmp ; Output pointer
pxor m5, m5 ; m5 = dedicated zero
pcmpeqw m4, m4 ; All word lanes -1
paddw m0, m4 ; m0 = zbin - 1
pabsw m6, m9 ; m6 = abs(m9)
pabsw m11, m10 ; m11 = abs(m10)
pcmpgtw m7, m6, m0 ; m7 = c[i] >= zbin
punpckhqdq m0, m0
pcmpgtw m12, m11, m0 ; m12 = c[i] >= zbin
; Check if all coeffs are less than zbin. If yes, we just write zeros
; to the outputs and we are done.
por m14, m7, m12
ptest m14, m14
jnz .single_nonzero
%if CONFIG_VP9_HIGHBITDEPTH
mova [r1 ], ymm5
mova [r1+32], ymm5
mova [r2 ], ymm5
mova [r2+32], ymm5
%else
mova [r1], ymm5
mova [r2], ymm5
%endif
mov [r0], word 0
vzeroupper
RET
.single_nonzero:
; Actual quantization of size 16 block - setup pointers, rounders, etc.
movifnidn r4, roundmp
movifnidn r5, quantmp
mov r3, dequantmp
mov r6, shiftmp
mova m1, [r4] ; m1 = round
mova m2, [r5] ; m2 = quant
mova m3, [r3] ; m3 = dequant
mova m4, [r6] ; m4 = shift
mov r3, iscanmp
DEFINE_ARGS eob, qcoeff, dqcoeff, iscan
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
paddsw m6, m1 ; m6 += round
punpckhqdq m1, m1
paddsw 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
%if CONFIG_VP9_HIGHBITDEPTH
; Store 16bit numbers as 32bit numbers in array pointed to by qcoeff
pcmpgtw m6, m5, m8
punpckhwd m6, m8, m6
pmovsxwd m11, m8
mova [qcoeffq ], m11
mova [qcoeffq+16], m6
pcmpgtw m6, m5, m13
punpckhwd m6, m13, m6
pmovsxwd m11, m13
mova [qcoeffq+32], m11
mova [qcoeffq+48], m6
%else
mova [qcoeffq ], m8
mova [qcoeffq+16], m13
%endif
pmullw m8, m3 ; dqc[i] = qc[i] * q
punpckhqdq m3, m3
pmullw m13, m3 ; dqc[i] = qc[i] * q
%if CONFIG_VP9_HIGHBITDEPTH
; Store 16bit numbers as 32bit numbers in array pointed to by qcoeff
pcmpgtw m6, m5, m8
punpckhwd m6, m8, m6
pmovsxwd m11, m8
mova [dqcoeffq ], m11
mova [dqcoeffq+16], m6
pcmpgtw m6, m5, m13
punpckhwd m6, m13, m6
pmovsxwd m11, m13
mova [dqcoeffq+32], m11
mova [dqcoeffq+48], m6
%else
mova [dqcoeffq ], m8
mova [dqcoeffq+16], m13
%endif
mova m6, [iscanq] ; m6 = scan[i]
mova m11, [iscanq+16] ; m11 = scan[i]
pcmpeqw m8, m8, m5 ; m8 = c[i] == 0
pcmpeqw m13, m13, m5 ; m13 = c[i] == 0
psubw m6, m6, m7 ; m6 = scan[i] + 1
psubw m11, m11, m12 ; m11 = scan[i] + 1
pandn m8, m8, m6 ; m8 = max(eob)
pandn m13, m13, m11 ; m13 = max(eob)
pmaxsw m8, m8, m13
; 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
movq rax, m8
mov [eobq], ax
vzeroupper
RET
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; Generic case of ncoeff != 16
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
.generic:
%endif ; %ifnidn %1, b_32x32
DEFINE_ARGS coeff, ncoeff, skip, zbin, round, quant, shift, \
qcoeff, dqcoeff, dequant, eob, scan, iscan
; Actual quantization loop - setup pointers, rounders, etc.
movifnidn coeffq, coeffmp
movifnidn ncoeffq, ncoeffmp
mov r2, dequantmp
movifnidn zbinq, zbinmp
movifnidn roundq, roundmp
movifnidn quantq, quantmp
mova m0, [zbinq] ; m0 = zbin
mova m1, [roundq] ; m1 = round
mova m2, [quantq] ; m2 = quant
mova m3, [r2] ; m3 = dequant
pcmpeqw m4, m4 ; All lanes -1
%ifidn %1, b_32x32
psubw m0, m4
psubw m1, m4
psrlw m0, 1 ; m0 = (m0 + 1) / 2
psrlw m1, 1 ; m1 = (m1 + 1) / 2
%endif
paddw m0, m4 ; m0 = m0 + 1
mov r2, shiftmp
mov r3, qcoeffmp
mova m4, [r2] ; m4 = shift
mov r4, dqcoeffmp
mov r5, iscanmp
%ifidn %1, b_32x32
psllw m4, 1
%endif
pxor m5, m5 ; m5 = dedicated zero
DEFINE_ARGS coeff, ncoeff, d1, qcoeff, dqcoeff, iscan, d2, d3, d4, d5, eob
%if CONFIG_VP9_HIGHBITDEPTH
lea coeffq, [ coeffq+ncoeffq*4]
lea qcoeffq, [ qcoeffq+ncoeffq*4]
lea dqcoeffq, [dqcoeffq+ncoeffq*4]
%else
lea coeffq, [ coeffq+ncoeffq*2]
lea qcoeffq, [ qcoeffq+ncoeffq*2]
lea dqcoeffq, [dqcoeffq+ncoeffq*2]
%endif
lea iscanq, [ iscanq+ncoeffq*2]
neg ncoeffq
; get DC and first 15 AC coeffs
%if CONFIG_VP9_HIGHBITDEPTH
; coeff stored as 32bit numbers & require 16bit numbers
mova m9, [coeffq+ncoeffq*4+ 0]
packssdw m9, [coeffq+ncoeffq*4+16]
mova m10, [coeffq+ncoeffq*4+32]
packssdw m10, [coeffq+ncoeffq*4+48]
%else
mova m9, [coeffq+ncoeffq*2+ 0] ; m9 = c[i]
mova m10, [coeffq+ncoeffq*2+16] ; m10 = c[i]
%endif
pabsw m6, m9 ; m6 = abs(m9)
pabsw m11, m10 ; m11 = abs(m10)
pcmpgtw m7, m6, m0 ; m7 = c[i] >= zbin
punpckhqdq m0, m0
pcmpgtw m12, m11, m0 ; m12 = c[i] >= zbin
; Check if all coeffs are less than zbin. If yes, skip forward quickly.
por m14, m7, m12
ptest m14, m14
jnz .first_nonzero
%if CONFIG_VP9_HIGHBITDEPTH
mova [qcoeffq+ncoeffq*4 ], ymm5
mova [qcoeffq+ncoeffq*4+32], ymm5
mova [dqcoeffq+ncoeffq*4 ], ymm5
mova [dqcoeffq+ncoeffq*4+32], ymm5
%else
mova [qcoeffq+ncoeffq*2], ymm5
mova [dqcoeffq+ncoeffq*2], ymm5
%endif
add ncoeffq, mmsize
punpckhqdq m1, m1
punpckhqdq m2, m2
punpckhqdq m3, m3
punpckhqdq m4, m4
pxor m8, m8
jmp .ac_only_loop
.first_nonzero:
paddsw m6, m1 ; m6 += round
punpckhqdq m1, m1
paddsw 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
%if CONFIG_VP9_HIGHBITDEPTH
; store 16bit numbers as 32bit numbers in array pointed to by qcoeff
pcmpgtw m6, m5, m8
punpckhwd m6, m8, m6
pmovsxwd m11, m8
mova [qcoeffq+ncoeffq*4+ 0], m11
mova [qcoeffq+ncoeffq*4+16], m6
pcmpgtw m6, m5, m13
punpckhwd m6, m13, m6
pmovsxwd m11, m13
mova [qcoeffq+ncoeffq*4+32], m11
mova [qcoeffq+ncoeffq*4+48], m6
%else
mova [qcoeffq+ncoeffq*2+ 0], m8
mova [qcoeffq+ncoeffq*2+16], m13
%endif
%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
%if CONFIG_VP9_HIGHBITDEPTH
; store 16bit numbers as 32bit numbers in array pointed to by qcoeff
pcmpgtw m6, m5, m8
punpckhwd m6, m8, m6
pmovsxwd m11, m8
mova [dqcoeffq+ncoeffq*4+ 0], m11
mova [dqcoeffq+ncoeffq*4+16], m6
pcmpgtw m6, m5, m13
punpckhwd m6, m13, m6
pmovsxwd m11, m13
mova [dqcoeffq+ncoeffq*4+32], m11
mova [dqcoeffq+ncoeffq*4+48], m6
%else
mova [dqcoeffq+ncoeffq*2+ 0], m8
mova [dqcoeffq+ncoeffq*2+16], m13
%endif
pcmpeqw m8, m5 ; m8 = c[i] == 0
pcmpeqw m13, m5 ; m13 = c[i] == 0
mova m6, [iscanq+ncoeffq*2] ; 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
.ac_only_loop:
%if CONFIG_VP9_HIGHBITDEPTH
; pack coeff from 32bit to 16bit array
mova m9, [coeffq+ncoeffq*4+ 0]
packssdw m9, [coeffq+ncoeffq*4+16]
mova m10, [coeffq+ncoeffq*4+32]
packssdw m10, [coeffq+ncoeffq*4+48]
%else
mova m9, [coeffq+ncoeffq*2+ 0] ; m9 = c[i]
mova m10, [coeffq+ncoeffq*2+16] ; m10 = c[i]
%endif
pabsw m6, m9 ; m6 = abs(m9)
pabsw m11, m10 ; m11 = abs(m10)
pcmpgtw m7, m6, m0 ; m7 = c[i] >= zbin
pcmpgtw m12, m11, m0 ; m12 = c[i] >= zbin
; Check if all coeffs are less than zbin. If yes, skip this itertion.
; And just write zeros as the result would be.
por m14, m7, m12
ptest m14, m14
jnz .rest_nonzero
%if CONFIG_VP9_HIGHBITDEPTH
mova [qcoeffq+ncoeffq*4+ 0], ymm5
mova [qcoeffq+ncoeffq*4+32], ymm5
mova [dqcoeffq+ncoeffq*4+ 0], ymm5
mova [dqcoeffq+ncoeffq*4+32], ymm5
%else
mova [qcoeffq+ncoeffq*2+ 0], ymm5
mova [dqcoeffq+ncoeffq*2+ 0], ymm5
%endif
add ncoeffq, mmsize
jnz .ac_only_loop
; Horizontally accumulate/max eobs and write into [eob] memory pointer
mov r2, eobmp
pshufd m7, m8, 0xe
pmaxsw m8, m7
pshuflw m7, m8, 0xe
pmaxsw m8, m7
pshuflw m7, m8, 0x1
pmaxsw m8, m7
movq rax, m8
mov [r2], ax
vzeroupper
RET
.rest_nonzero:
paddsw m6, m1 ; m6 += round
paddsw 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
%if CONFIG_VP9_HIGHBITDEPTH
; store 16bit numbers as 32bit numbers in array pointed to by qcoeff
pcmpgtw m6, m5, m14
punpckhwd m6, m14, m6
pmovsxwd m11, m14
mova [qcoeffq+ncoeffq*4+ 0], m11
mova [qcoeffq+ncoeffq*4+16], m6
pcmpgtw m6, m5, m13
punpckhwd m6, m13, m6
pmovsxwd m11, m13
mova [qcoeffq+ncoeffq*4+32], m11
mova [qcoeffq+ncoeffq*4+48], m6
%else
mova [qcoeffq+ncoeffq*2+ 0], m14
mova [qcoeffq+ncoeffq*2+16], m13
%endif
%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
%if CONFIG_VP9_HIGHBITDEPTH
; store 16bit numbers as 32bit numbers in array pointed to by qcoeff
pcmpgtw m6, m5, m14
punpckhwd m6, m14, m6
pmovsxwd m11, m14
mova [dqcoeffq+ncoeffq*4+ 0], m11
mova [dqcoeffq+ncoeffq*4+16], m6
pcmpgtw m6, m5, m13
punpckhwd m6, m13, m6
pmovsxwd m11, m13
mova [dqcoeffq+ncoeffq*4+32], m11
mova [dqcoeffq+ncoeffq*4+48], m6
%else
mova [dqcoeffq+ncoeffq*2+ 0], m14
mova [dqcoeffq+ncoeffq*2+16], m13
%endif
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
jnz .ac_only_loop
; Horizontally accumulate/max eobs and write into [eob] memory pointer
mov r2, eobmp
pshufd m7, m8, 0xe
pmaxsw m8, m7
pshuflw m7, m8, 0xe
pmaxsw m8, m7
pshuflw m7, m8, 0x1
pmaxsw m8, m7
movq rax, m8
mov [r2], ax
vzeroupper
RET
; Skip-block, i.e. just write all zeroes
.blank:
DEFINE_ARGS coeff, ncoeff, skip, zbin, round, quant, shift, \
qcoeff, dqcoeff, dequant, eob, scan, iscan
mov r0, dqcoeffmp
movifnidn ncoeffq, ncoeffmp
mov r2, qcoeffmp
mov r3, eobmp
DEFINE_ARGS dqcoeff, ncoeff, qcoeff, eob
%if CONFIG_VP9_HIGHBITDEPTH
lea dqcoeffq, [dqcoeffq+ncoeffq*4]
lea qcoeffq, [ qcoeffq+ncoeffq*4]
%else
lea dqcoeffq, [dqcoeffq+ncoeffq*2]
lea qcoeffq, [ qcoeffq+ncoeffq*2]
%endif
neg ncoeffq
pxor m7, m7
.blank_loop:
%if CONFIG_VP9_HIGHBITDEPTH
mova [dqcoeffq+ncoeffq*4+ 0], ymm7
mova [dqcoeffq+ncoeffq*4+32], ymm7
mova [qcoeffq+ncoeffq*4+ 0], ymm7
mova [qcoeffq+ncoeffq*4+32], ymm7
%else
mova [dqcoeffq+ncoeffq*2+ 0], ymm7
mova [qcoeffq+ncoeffq*2+ 0], ymm7
%endif
add ncoeffq, mmsize
jl .blank_loop
mov [eobq], word 0
vzeroupper
RET
%endmacro
INIT_XMM avx
QUANTIZE_FN b, 7
QUANTIZE_FN b_32x32, 7
END