2 ; Copyright (c) 2010 The WebM project authors. All Rights Reserved.
4 ; Use of this source code is governed by a BSD-style license and patent
5 ; grant that can be found in the LICENSE file in the root of the source
6 ; tree. All contributing project authors may be found in the AUTHORS
7 ; file in the root of the source tree.
11 %include "vpx_ports/x86_abi_support.asm"
12 %include "vp9_asm_enc_offsets.asm"
15 ; void vp9_fast_quantize_b_ssse3 | arg
20 global sym(vp9_fast_quantize_b_ssse3)
21 sym(vp9_fast_quantize_b_ssse3):
30 %ifidn __OUTPUT_FORMAT__,x64
38 mov rdi, arg(0) ; BLOCK *b
39 mov rsi, arg(1) ; BLOCKD *d
41 %ifidn __OUTPUT_FORMAT__,x64
42 mov rdi, rcx ; BLOCK *b
43 mov rsi, rdx ; BLOCKD *d
45 ;mov rdi, rdi ; BLOCK *b
46 ;mov rsi, rsi ; BLOCKD *d
50 mov rax, [rdi + vp9_block_coeff]
51 mov rcx, [rdi + vp9_block_round]
52 mov rdx, [rdi + vp9_block_quant_fast]
56 movdqa xmm4, [rax + 16]
60 movdqa xmm3, [rcx + 16]
77 pmulhw xmm5, [rdx + 16]
79 mov rax, [rsi + vp9_blockd_qcoeff]
80 mov rdi, [rsi + vp9_blockd_dequant]
81 mov rcx, [rsi + vp9_blockd_dqcoeff]
89 movdqa [rax + 16], xmm5
92 movdqa xmm3, [rdi + 16]
98 pcmpeqw xmm1, xmm4 ;non zero mask
99 pcmpeqw xmm5, xmm4 ;non zero mask
101 pshufb xmm1, [GLOBAL(zz_shuf)]
107 xor dx, ax ;flip the bits for bsr
110 movdqa [rcx], xmm2 ;store dqcoeff
111 movdqa [rcx + 16], xmm3 ;store dqcoeff
113 sub edi, edx ;check for all zeros in bit mask
116 and eax, edi ;if the bit mask was all zero,
118 mov [rsi + vp9_blockd_eob], eax
125 %ifidn __OUTPUT_FORMAT__,x64
138 db 0, 1, 4, 8, 5, 2, 3, 6, 9, 12, 13, 10, 7, 11, 14, 15