vpx/vp8/encoder/x86/quantize_sse2.asm
Johann 15f9bea73b update sse2 regular quantizer
about ~5% gain on 32bit. disabled for 64bit

unset executable bit on ssse3 version (cosmetic)

Change-Id: I1a5860839eb294ce4261f819caea2dcfa78e57ca
2011-01-14 14:26:10 -05:00

303 lines
8.2 KiB
NASM

;
; Copyright (c) 2010 The WebM project authors. All Rights Reserved.
;
; Use of this source code is governed by a BSD-style license and patent
; grant that can be found in the LICENSE file in the root of the source
; tree. All contributing project authors may be found in the AUTHORS
; file in the root of the source tree.
;
%include "vpx_ports/x86_abi_support.asm"
;int vp8_regular_quantize_b_impl_sse2(
; short *coeff_ptr,
; short *zbin_ptr,
; short *qcoeff_ptr,
; short *dequant_ptr,
; const int *default_zig_zag,
; short *round_ptr,
; short *quant_ptr,
; short *dqcoeff_ptr,
; unsigned short zbin_oq_value,
; short *zbin_boost_ptr,
; short *quant_shift);
;
global sym(vp8_regular_quantize_b_impl_sse2)
sym(vp8_regular_quantize_b_impl_sse2):
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 11
SAVE_XMM
push rsi
push rdi
push rbx
ALIGN_STACK 16, rax
%define abs_minus_zbin 0
%define temp_qcoeff 32
%define qcoeff 64
%define eob_tmp 96
%define stack_size 112
sub rsp, stack_size
; end prolog
mov rdx, arg(0) ; coeff_ptr
mov rcx, arg(1) ; zbin_ptr
movd xmm7, arg(8) ; zbin_oq_value
mov rdi, arg(5) ; round_ptr
mov rsi, arg(6) ; quant_ptr
; z
movdqa xmm0, OWORD PTR[rdx]
movdqa xmm4, OWORD PTR[rdx + 16]
pshuflw xmm7, xmm7, 0
punpcklwd xmm7, xmm7 ; duplicated zbin_oq_value
movdqa xmm1, xmm0
movdqa xmm5, xmm4
; sz
psraw xmm0, 15
psraw xmm4, 15
; (z ^ sz)
pxor xmm1, xmm0
pxor xmm5, xmm4
; x = abs(z)
psubw xmm1, xmm0
psubw xmm5, xmm4
movdqa xmm2, OWORD PTR[rcx]
movdqa xmm3, OWORD PTR[rcx + 16]
; *zbin_ptr + zbin_oq_value
paddw xmm2, xmm7
paddw xmm3, xmm7
; x - (*zbin_ptr + zbin_oq_value)
psubw xmm1, xmm2
psubw xmm5, xmm3
movdqa OWORD PTR[rsp + abs_minus_zbin], xmm1
movdqa OWORD PTR[rsp + abs_minus_zbin + 16], xmm5
; add (zbin_ptr + zbin_oq_value) back
paddw xmm1, xmm2
paddw xmm5, xmm3
movdqa xmm2, OWORD PTR[rdi]
movdqa xmm6, OWORD PTR[rdi + 16]
movdqa xmm3, OWORD PTR[rsi]
movdqa xmm7, OWORD PTR[rsi + 16]
; x + round
paddw xmm1, xmm2
paddw xmm5, xmm6
; y = x * quant_ptr >> 16
pmulhw xmm3, xmm1
pmulhw xmm7, xmm5
; y += x
paddw xmm1, xmm3
paddw xmm5, xmm7
movdqa OWORD PTR[rsp + temp_qcoeff], xmm1
movdqa OWORD PTR[rsp + temp_qcoeff + 16], xmm5
pxor xmm6, xmm6
; zero qcoeff
movdqa OWORD PTR[rsp + qcoeff], xmm6
movdqa OWORD PTR[rsp + qcoeff + 16], xmm6
mov [rsp + eob_tmp], DWORD -1 ; eob
mov rsi, arg(9) ; zbin_boost_ptr
mov rdi, arg(4) ; default_zig_zag
mov rax, arg(10) ; quant_shift_ptr
%macro ZIGZAG_LOOP 2
rq_zigzag_loop_%1:
movsxd rdx, DWORD PTR[rdi + (%1 * 4)] ; rc
movsx ebx, WORD PTR [rsi] ; *zbin_boost_ptr
lea rsi, [rsi + 2] ; zbin_boost_ptr++
; x
movsx ecx, WORD PTR[rsp + abs_minus_zbin + rdx *2]
; if (x >= zbin)
sub ecx, ebx ; x - zbin
jl rq_zigzag_loop_%2 ; x < zbin
movsx ebx, WORD PTR[rsp + temp_qcoeff + rdx *2]
; downshift by quant_shift[rdx]
movsx ecx, WORD PTR[rax + rdx*2] ; quant_shift_ptr[rc]
sar ebx, cl ; also sets Z bit
je rq_zigzag_loop_%2 ; !y
mov WORD PTR[rsp + qcoeff + rdx * 2], bx ;qcoeff_ptr[rc] = temp_qcoeff[rc]
mov rsi, arg(9) ; reset to b->zrun_zbin_boost
mov [rsp + eob_tmp], DWORD %1 ; eob = i
%endmacro
ZIGZAG_LOOP 0, 1
ZIGZAG_LOOP 1, 2
ZIGZAG_LOOP 2, 3
ZIGZAG_LOOP 3, 4
ZIGZAG_LOOP 4, 5
ZIGZAG_LOOP 5, 6
ZIGZAG_LOOP 6, 7
ZIGZAG_LOOP 7, 8
ZIGZAG_LOOP 8, 9
ZIGZAG_LOOP 9, 10
ZIGZAG_LOOP 10, 11
ZIGZAG_LOOP 11, 12
ZIGZAG_LOOP 12, 13
ZIGZAG_LOOP 13, 14
ZIGZAG_LOOP 14, 15
ZIGZAG_LOOP 15, end
rq_zigzag_loop_end:
mov rbx, arg(2) ; qcoeff_ptr
mov rcx, arg(3) ; dequant_ptr
mov rsi, arg(7) ; dqcoeff_ptr
mov rax, [rsp + eob_tmp] ; eob
movdqa xmm2, OWORD PTR[rsp + qcoeff]
movdqa xmm3, OWORD PTR[rsp + qcoeff + 16]
; y ^ sz
pxor xmm2, xmm0
pxor xmm3, xmm4
; x = (y ^ sz) - sz
psubw xmm2, xmm0
psubw xmm3, xmm4
movdqa xmm0, OWORD PTR[rcx]
movdqa xmm1, OWORD PTR[rcx + 16]
pmullw xmm0, xmm2
pmullw xmm1, xmm3
movdqa OWORD PTR[rbx], xmm2
movdqa OWORD PTR[rbx + 16], xmm3
movdqa OWORD PTR[rsi], xmm0 ; store dqcoeff
movdqa OWORD PTR[rsi + 16], xmm1 ; store dqcoeff
add rax, 1
; begin epilog
add rsp, stack_size
pop rsp
pop rbx
pop rdi
pop rsi
RESTORE_XMM
UNSHADOW_ARGS
pop rbp
ret
;int vp8_fast_quantize_b_impl_sse2(short *coeff_ptr,
; short *qcoeff_ptr,short *dequant_ptr,
; short *inv_scan_order, short *round_ptr,
; short *quant_ptr, short *dqcoeff_ptr);
global sym(vp8_fast_quantize_b_impl_sse2)
sym(vp8_fast_quantize_b_impl_sse2):
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 7
push rsi
push rdi
; end prolog
mov rdx, arg(0) ;coeff_ptr
mov rcx, arg(2) ;dequant_ptr
mov rdi, arg(4) ;round_ptr
mov rsi, arg(5) ;quant_ptr
movdqa xmm0, XMMWORD PTR[rdx]
movdqa xmm4, XMMWORD PTR[rdx + 16]
movdqa xmm2, XMMWORD PTR[rdi] ;round lo
movdqa xmm3, XMMWORD PTR[rdi + 16] ;round hi
movdqa xmm1, xmm0
movdqa xmm5, xmm4
psraw xmm0, 15 ;sign of z (aka sz)
psraw xmm4, 15 ;sign of z (aka sz)
pxor xmm1, xmm0
pxor xmm5, xmm4
psubw xmm1, xmm0 ;x = abs(z)
psubw xmm5, xmm4 ;x = abs(z)
paddw xmm1, xmm2
paddw xmm5, xmm3
pmulhw xmm1, XMMWORD PTR[rsi]
pmulhw xmm5, XMMWORD PTR[rsi + 16]
mov rdi, arg(1) ;qcoeff_ptr
mov rsi, arg(6) ;dqcoeff_ptr
movdqa xmm2, XMMWORD PTR[rcx]
movdqa xmm3, XMMWORD PTR[rcx + 16]
pxor xmm1, xmm0
pxor xmm5, xmm4
psubw xmm1, xmm0
psubw xmm5, xmm4
movdqa XMMWORD PTR[rdi], xmm1
movdqa XMMWORD PTR[rdi + 16], xmm5
pmullw xmm2, xmm1
pmullw xmm3, xmm5
mov rdi, arg(3) ;inv_scan_order
; Start with 16
pxor xmm4, xmm4 ;clear all bits
pcmpeqw xmm1, xmm4
pcmpeqw xmm5, xmm4
pcmpeqw xmm4, xmm4 ;set all bits
pxor xmm1, xmm4
pxor xmm5, xmm4
pand xmm1, XMMWORD PTR[rdi]
pand xmm5, XMMWORD PTR[rdi+16]
pmaxsw xmm1, xmm5
; now down to 8
pshufd xmm5, xmm1, 00001110b
pmaxsw xmm1, xmm5
; only 4 left
pshuflw xmm5, xmm1, 00001110b
pmaxsw xmm1, xmm5
; okay, just 2!
pshuflw xmm5, xmm1, 00000001b
pmaxsw xmm1, xmm5
movd rax, xmm1
and rax, 0xff
movdqa XMMWORD PTR[rsi], xmm2 ;store dqcoeff
movdqa XMMWORD PTR[rsi + 16], xmm3 ;store dqcoeff
; begin epilog
pop rdi
pop rsi
UNSHADOW_ARGS
pop rbp
ret