1 ; Copyright (c) 2011 The Chromium Authors. All rights reserved.
2 ; Use of this source code is governed by a BSD-style license that can be
3 ; found in the LICENSE file.
5 %include "media/base/simd/media_export.asm"
6 %include "third_party/x86inc/x86inc.asm"
9 ; This file uses MMX instructions.
14 ;void LinearScaleYUVToRGB32Row_MMX_X64(const uint8* y_buf,
19 ; ptrdiff_t source_dx);
20 %define SYMBOL LinearScaleYUVToRGB32Row_MMX_X64
25 %assign stack_offset 0
26 extern mangle(kCoefficientsRgbY)
28 ; Parameters are in the following order:
35 ; 7. Conversion lookup table
37 PROLOGUE 7, 7, 3, Y, U, V, ARGB, WIDTH, SOURCE_DX, R1
65 imul WIDTHq, SOURCE_DXq ; source_width = width * source_dx
67 cmp SOURCE_DXq, 0x20000
69 mov Xq, 0x8000 ; x = 0.5 for 1/2 or less
76 movzx COMPLd, BYTE [Uq + INDEXq]
77 movzx COMPRd, BYTE [Uq + INDEXq + 1]
79 and FRACTIONq, 0x1fffe
80 imul COMPRq, FRACTIONq
81 xor FRACTIONq, 0x1fffe
82 imul COMPLq, FRACTIONq
85 movq mm0, [TABLEq + 2048 + 8 * COMPLq]
88 movzx COMPLd, BYTE [Vq + INDEXq]
89 movzx COMPRd, BYTE [Vq + INDEXq + 1]
90 ; Trick here to imul COMPL first then COMPR.
91 ; Saves two instruction. :)
92 imul COMPLq, FRACTIONq
93 xor FRACTIONq, 0x1fffe
94 imul COMPRq, FRACTIONq
97 paddsw mm0, [TABLEq + 4096 + 8 * COMPLq]
99 ; Interpolate first Y1.
100 lea INDEXq, [Xq + SOURCE_DXq] ; INDEXq now points to next pixel.
101 ; Xq points to current pixel.
104 movzx COMPLd, BYTE [Yq + Xq]
105 movzx COMPRd, BYTE [Yq + Xq + 1]
106 and FRACTIONq, 0xffff
107 imul COMPRq, FRACTIONq
108 xor FRACTIONq, 0xffff
109 imul COMPLq, FRACTIONq
112 movq mm1, [TABLEq + 8 * COMPLq]
114 ; Interpolate Y2 if available.
118 lea Xq, [INDEXq + SOURCE_DXq] ; Xq points to next pixel.
119 ; INDEXq points to current pixel.
120 mov FRACTIONq, INDEXq
122 movzx COMPLd, BYTE [Yq + INDEXq]
123 movzx COMPRd, BYTE [Yq + INDEXq + 1]
124 and FRACTIONq, 0xffff
125 imul COMPRq, FRACTIONq
126 xor FRACTIONq, 0xffff
127 imul COMPLq, FRACTIONq
130 movq mm2, [TABLEq + 8 * COMPLq]