Skip to content

GitLab

Commit 0ced7014 authored by Fritz Koenig's avatar Fritz Koenig Committed by Code Review
Browse files
Options

Merge "vp8 fast quantizer sse2 optimizations for eob."

parents cb969895 e0cf330c
Hide whitespace changes
Inline Side-by-side
Showing with 40 additions and 64 deletions
vp8/common/entropy.c
View file @ 0ced7014
......@@ -36,6 +36,14 @@ DECLARE_ALIGNED(16, const int, vp8_default_zig_zag1d[16]) =
7, 11, 14, 15,
};
DECLARE_ALIGNED(16, const short, vp8_default_inv_zig_zag[16]) =
{
1, 2, 6, 7,
3, 5, 8, 13,
4, 9, 12, 14,
10, 11, 15, 16
};
DECLARE_ALIGNED(16, short, vp8_default_zig_zag_mask[16]);
const int vp8_mb_feature_data_bits[MB_LVL_MAX] = {7, 6};
......
vp8/common/entropy.h
View file @ 0ced7014
......@@ -95,6 +95,7 @@ struct VP8Common;
void vp8_default_coef_probs(struct VP8Common *);
extern DECLARE_ALIGNED(16, const int, vp8_default_zig_zag1d[16]);
extern DECLARE_ALIGNED(16, const short, vp8_default_inv_zig_zag[16]);
extern short vp8_default_zig_zag_mask[16];
extern const int vp8_mb_feature_data_bits[MB_LVL_MAX];
......
vp8/encoder/x86/quantize_sse2.asm
View file @ 0ced7014
......@@ -253,10 +253,9 @@ rq_zigzag_1c:
pop rbp
ret
;int vp8_fast_quantize_b_impl_sse2(short *coeff_ptr,
; short *qcoeff_ptr,short *dequant_ptr,
; short *scan_mask, short *round_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):
......@@ -265,32 +264,18 @@ sym(vp8_fast_quantize_b_impl_sse2):
SHADOW_ARGS_TO_STACK 7
push rsi
push rdi
push rbx
; end prolog
ALIGN_STACK 16, rax
%define save_xmm6 0
%define save_xmm7 16
%define vp8_fastquantizeb_stack_size save_xmm7 + 16
sub rsp, vp8_fastquantizeb_stack_size
movdqa XMMWORD PTR[rsp + save_xmm6], xmm6
movdqa XMMWORD PTR[rsp + save_xmm7], xmm7
mov rdx, arg(0) ;coeff_ptr
mov rcx, arg(2) ;dequant_ptr
mov rax, arg(3) ;scan_mask
mov rdi, arg(4) ;round_ptr
mov rsi, arg(5) ;quant_ptr
movdqa xmm0, XMMWORD PTR[rdx]
movdqa xmm4, XMMWORD PTR[rdx + 16]
movdqa xmm6, XMMWORD PTR[rdi] ;round lo
movdqa xmm7, XMMWORD PTR[rdi + 16] ;round hi
movdqa xmm2, XMMWORD PTR[rdi] ;round lo
movdqa xmm3, XMMWORD PTR[rdi + 16] ;round hi
movdqa xmm1, xmm0
movdqa xmm5, xmm4
......@@ -303,8 +288,8 @@ sym(vp8_fast_quantize_b_impl_sse2):
psubw xmm1, xmm0 ;x = abs(z)
psubw xmm5, xmm4 ;x = abs(z)
paddw xmm1, xmm6
paddw xmm5, xmm7
paddw xmm1, xmm2
paddw xmm5, xmm3
pmulhw xmm1, XMMWORD PTR[rsi]
pmulhw xmm5, XMMWORD PTR[rsi + 16]
......@@ -312,8 +297,8 @@ sym(vp8_fast_quantize_b_impl_sse2):
mov rdi, arg(1) ;qcoeff_ptr
mov rsi, arg(6) ;dqcoeff_ptr
movdqa xmm6, XMMWORD PTR[rcx]
movdqa xmm7, XMMWORD PTR[rcx + 16]
movdqa xmm2, XMMWORD PTR[rcx]
movdqa xmm3, XMMWORD PTR[rcx + 16]
pxor xmm1, xmm0
pxor xmm5, xmm4
......@@ -323,64 +308,47 @@ sym(vp8_fast_quantize_b_impl_sse2):
movdqa XMMWORD PTR[rdi], xmm1
movdqa XMMWORD PTR[rdi + 16], xmm5
pmullw xmm6, xmm1
pmullw xmm7, xmm5
pmullw xmm2, xmm1
pmullw xmm3, xmm5
movdqa xmm2, XMMWORD PTR[rax]
movdqa xmm3, XMMWORD PTR[rax+16];
mov rdi, arg(3) ;inv_scan_order
pxor xmm4, xmm4 ;clear all bits
; Start with 16
pxor xmm4, xmm4 ;clear all bits
pcmpeqw xmm1, xmm4
pcmpeqw xmm5, xmm4
pcmpeqw xmm4, xmm4 ;set all bits
pcmpeqw xmm4, xmm4 ;set all bits
pxor xmm1, xmm4
pxor xmm5, xmm4
psrlw xmm1, 15
psrlw xmm5, 15
pmaddwd xmm1, xmm2
pmaddwd xmm5, xmm3
movq xmm2, xmm1
movq xmm3, xmm5
psrldq xmm1, 8
psrldq xmm5, 8
pand xmm1, XMMWORD PTR[rdi]
pand xmm5, XMMWORD PTR[rdi+16]
paddd xmm1, xmm5
paddd xmm2, xmm3
pmaxsw xmm1, xmm5
paddd xmm1, xmm2
movq xmm5, xmm1
; now down to 8
pshufd xmm5, xmm1, 00001110b
psrldq xmm1, 4
paddd xmm5, xmm1
pmaxsw xmm1, xmm5
movq rcx, xmm5
and rcx, 0xffff
; only 4 left
pshuflw xmm5, xmm1, 00001110b
xor rdx, rdx
sub rdx, rcx
pmaxsw xmm1, xmm5
bsr rax, rcx
inc rax
; okay, just 2!
pshuflw xmm5, xmm1, 00000001b
sar rdx, 31
and rax, rdx
pmaxsw xmm1, xmm5
movdqa XMMWORD PTR[rsi], xmm6 ;store dqcoeff
movdqa XMMWORD PTR[rsi + 16], xmm7 ;store dqcoeff
movd rax, xmm1
and rax, 0xff
movdqa xmm6, XMMWORD PTR[rsp + save_xmm6]
movdqa xmm7, XMMWORD PTR[rsp + save_xmm7]
add rsp, vp8_fastquantizeb_stack_size
pop rsp
movdqa XMMWORD PTR[rsi], xmm2 ;store dqcoeff
movdqa XMMWORD PTR[rsi + 16], xmm3 ;store dqcoeff
; begin epilog
pop rbx
pop rdi
pop rsi
UNSHADOW_ARGS
......
vp8/encoder/x86/x86_csystemdependent.c
View file @ 0ced7014
......@@ -83,7 +83,7 @@ void vp8_subtract_b_mmx(BLOCK *be, BLOCKD *bd, int pitch)
#if HAVE_SSE2
int vp8_fast_quantize_b_impl_sse2(short *coeff_ptr,
short *qcoeff_ptr, short *dequant_ptr,
short *scan_mask, short *round_ptr,
const short *inv_scan_order, short *round_ptr,
short *quant_ptr, short *dqcoeff_ptr);
void vp8_fast_quantize_b_sse2(BLOCK *b, BLOCKD *d)
{
......@@ -99,8 +99,7 @@ void vp8_fast_quantize_b_sse2(BLOCK *b, BLOCKD *d)
coeff_ptr,
qcoeff_ptr,
dequant_ptr,
scan_mask,
vp8_default_inv_zig_zag,
round_ptr,
quant_ptr,
dqcoeff_ptr
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment