gfx/ycbcr/yuv_row_arm.s@97036ab72558
gfx/ycbcr/yuv_row_arm.s
Tue, 06 Jan 2015 21:39:09 +0100
- author
- Michael Schloh von Bennewitz <michael@schloh.com>
- date
- Tue, 06 Jan 2015 21:39:09 +0100
- branch
- TOR_BUG_9701
- changeset 8
- 97036ab72558
- permissions
- -rw-r--r--
Conditionally force memory storage according to privacy.thirdparty.isolate;
This solves Tor bug #9701, complying with disk avoidance documented in
https://www.torproject.org/projects/torbrowser/design/#disk-avoidance.
1 /* This Source Code Form is subject to the terms of the Mozilla Public
2 * License, v. 2.0. If a copy of the MPL was not distributed with this
3 * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
5 .arch armv7-a
6 .fpu neon
7 /* Allow to build on targets not supporting neon, and force the object file
8 * target to avoid bumping the final binary target */
9 .object_arch armv4t
10 .text
11 .align
13 .balign 64
14 YCbCr42xToRGB565_DITHER03_CONSTS_NEON:
15 .short -14240
16 .short -14240+384
17 .short 8672
18 .short 8672+192
19 .short -17696
20 .short -17696+384
21 .byte 102
22 .byte 25
23 .byte 52
24 .byte 129
25 YCbCr42xToRGB565_DITHER12_CONSTS_NEON:
26 .short -14240+128
27 .short -14240+256
28 .short 8672+64
29 .short 8672+128
30 .short -17696+128
31 .short -17696+256
32 .byte 102
33 .byte 25
34 .byte 52
35 .byte 129
36 YCbCr42xToRGB565_DITHER21_CONSTS_NEON:
37 .short -14240+256
38 .short -14240+128
39 .short 8672+128
40 .short 8672+64
41 .short -17696+256
42 .short -17696+128
43 .byte 102
44 .byte 25
45 .byte 52
46 .byte 129
47 YCbCr42xToRGB565_DITHER30_CONSTS_NEON:
48 .short -14240+384
49 .short -14240
50 .short 8672+192
51 .short 8672
52 .short -17696+384
53 .short -17696
54 .byte 102
55 .byte 25
56 .byte 52
57 .byte 129
59 @ void ScaleYCbCr42xToRGB565_BilinearY_Row_NEON(
60 @ yuv2rgb565_row_scale_bilinear_ctx *ctx, int dither);
61 @
62 @ ctx = {
63 @ uint16_t *rgb_row; /*r0*/
64 @ const uint8_t *y_row; /*r1*/
65 @ const uint8_t *u_row; /*r2*/
66 @ const uint8_t *v_row; /*r3*/
67 @ int y_yweight; /*r4*/
68 @ int y_pitch; /*r5*/
69 @ int width; /*r6*/
70 @ int source_x0_q16; /*r7*/
71 @ int source_dx_q16; /*r8*/
72 @ int source_uv_xoffs_q16; /*r9*/
73 @ };
74 .global ScaleYCbCr42xToRGB565_BilinearY_Row_NEON
75 .type ScaleYCbCr42xToRGB565_BilinearY_Row_NEON, %function
76 .balign 64
77 .fnstart
78 ScaleYCbCr42xToRGB565_BilinearY_Row_NEON:
79 STMFD r13!,{r4-r9,r14} @ 8 words.
80 ADR r14,YCbCr42xToRGB565_DITHER03_CONSTS_NEON
81 VPUSH {Q4-Q7} @ 16 words.
82 ADD r14,r14,r1, LSL #4 @ Select the dither table to use
83 LDMIA r0, {r0-r9}
84 @ Set up image index registers.
85 ADD r12,r8, r8
86 VMOV.I32 D16,#0 @ Q8 = < 2| 2| 0| 0>*source_dx_q16
87 VDUP.32 D17,r12
88 ADD r12,r12,r12
89 VTRN.32 D16,D17 @ Q2 = < 2| 0| 2| 0>*source_dx_q16
90 VDUP.32 D19,r12 @ Q9 = < 4| 4| ?| ?>*source_dx_q16
91 ADD r12,r12,r12
92 VDUP.32 Q0, r7 @ Q0 = < 1| 1| 1| 1>*source_x0_q16
93 VADD.I32 D17,D17,D19 @ Q8 = < 6| 4| 2| 0>*source_dx_q16
94 CMP r8, #0 @ If source_dx_q16 is negative...
95 VDUP.32 Q9, r12 @ Q9 = < 8| 8| 8| 8>*source_dx_q16
96 ADDLT r7, r7, r8, LSL #4 @ Make r7 point to the end of the block
97 VADD.I32 Q0, Q0, Q8 @ Q0 = < 6| 4| 2| 0>*source_dx_q16+source_x0_q16
98 SUBLT r7, r7, r8 @ (i.e., the lowest address we'll use)
99 VADD.I32 Q1, Q0, Q9 @ Q1 = <14|12|10| 8>*source_dx_q16+source_x0_q16
100 VDUP.I32 Q9, r8 @ Q8 = < 1| 1| 1| 1>*source_dx_q16
101 VADD.I32 Q2, Q0, Q9 @ Q2 = < 7| 5| 3| 1>*source_dx_q16+source_x0_q16
102 VADD.I32 Q3, Q1, Q9 @ Q3 = <15|13|11| 9>*source_dx_q16+source_x0_q16
103 VLD1.64 {D30,D31},[r14,:128] @ Load some constants
104 VMOV.I8 D28,#52
105 VMOV.I8 D29,#129
106 @ The basic idea here is to do aligned loads of a block of data and then
107 @ index into it using VTBL to extract the data from the source X
108 @ coordinate corresponding to each destination pixel.
109 @ This is significantly less code and significantly fewer cycles than doing
110 @ a series of single-lane loads, but it means that the X step between
111 @ pixels must be limited to 2.0 or less, otherwise we couldn't guarantee
112 @ that we could read 8 pixels from a single aligned 32-byte block of data.
113 @ Q0...Q3 contain the 16.16 fixed-point X coordinates of each pixel,
114 @ separated into even pixels and odd pixels to make extracting offsets and
115 @ weights easier.
116 @ We then pull out two bytes from the middle of each coordinate: the top
117 @ byte corresponds to the integer part of the X coordinate, and the bottom
118 @ byte corresponds to the weight to use for bilinear blending.
119 @ These are separated out into different registers with VTRN.
120 @ Then by subtracting the integer X coordinate of the first pixel in the
121 @ data block we loaded, we produce an index register suitable for use by
122 @ VTBL.
123 s42xbily_neon_loop:
124 @ Load the Y' data.
125 MOV r12,r7, ASR #16
126 VRSHRN.S32 D16,Q0, #8
127 AND r12,r12,#~15 @ Read 16-byte aligned blocks
128 VDUP.I8 D20,r12
129 ADD r12,r1, r12 @ r12 = y_row+(source_x&~7)
130 VRSHRN.S32 D17,Q1, #8
131 PLD [r12,#64]
132 VLD1.64 {D8, D9, D10,D11},[r12,:128],r5 @ Load Y' top row
133 ADD r14,r7, r8, LSL #3
134 VRSHRN.S32 D18,Q2, #8
135 MOV r14,r14,ASR #16
136 VRSHRN.S32 D19,Q3, #8
137 AND r14,r14,#~15 @ Read 16-byte aligned blocks
138 VLD1.64 {D12,D13,D14,D15},[r12,:128] @ Load Y' bottom row
139 PLD [r12,#64]
140 VDUP.I8 D21,r14
141 ADD r14,r1, r14 @ r14 = y_row+(source_x&~7)
142 VMOV.I8 Q13,#1
143 PLD [r14,#64]
144 VTRN.8 Q8, Q9 @ Q8 = <wFwEwDwCwBwAw9w8w7w6w5w4w3w2w1w0>
145 @ Q9 = <xFxExDxCxBxAx9x8x7x6x5x4x3x2x1x0>
146 VSUB.S8 Q9, Q9, Q10 @ Make offsets relative to the data we loaded.
147 @ First 8 Y' pixels
148 VTBL.8 D20,{D8, D9, D10,D11},D18 @ Index top row at source_x
149 VTBL.8 D24,{D12,D13,D14,D15},D18 @ Index bottom row at source_x
150 VADD.S8 Q13,Q9, Q13 @ Add 1 to source_x
151 VTBL.8 D22,{D8, D9, D10,D11},D26 @ Index top row at source_x+1
152 VTBL.8 D26,{D12,D13,D14,D15},D26 @ Index bottom row at source_x+1
153 @ Next 8 Y' pixels
154 VLD1.64 {D8, D9, D10,D11},[r14,:128],r5 @ Load Y' top row
155 VLD1.64 {D12,D13,D14,D15},[r14,:128] @ Load Y' bottom row
156 PLD [r14,#64]
157 VTBL.8 D21,{D8, D9, D10,D11},D19 @ Index top row at source_x
158 VTBL.8 D25,{D12,D13,D14,D15},D19 @ Index bottom row at source_x
159 VTBL.8 D23,{D8, D9, D10,D11},D27 @ Index top row at source_x+1
160 VTBL.8 D27,{D12,D13,D14,D15},D27 @ Index bottom row at source_x+1
161 @ Blend Y'.
162 VDUP.I16 Q9, r4 @ Load the y weights.
163 VSUBL.U8 Q4, D24,D20 @ Q5:Q4 = c-a
164 VSUBL.U8 Q5, D25,D21
165 VSUBL.U8 Q6, D26,D22 @ Q7:Q6 = d-b
166 VSUBL.U8 Q7, D27,D23
167 VMUL.S16 Q4, Q4, Q9 @ Q5:Q4 = (c-a)*yweight
168 VMUL.S16 Q5, Q5, Q9
169 VMUL.S16 Q6, Q6, Q9 @ Q7:Q6 = (d-b)*yweight
170 VMUL.S16 Q7, Q7, Q9
171 VMOVL.U8 Q12,D16 @ Promote the x weights to 16 bits.
172 VMOVL.U8 Q13,D17 @ Sadly, there's no VMULW.
173 VRSHRN.S16 D8, Q4, #8 @ Q4 = (c-a)*yweight+128>>8
174 VRSHRN.S16 D9, Q5, #8
175 VRSHRN.S16 D12,Q6, #8 @ Q6 = (d-b)*yweight+128>>8
176 VRSHRN.S16 D13,Q7, #8
177 VADD.I8 Q10,Q10,Q4 @ Q10 = a+((c-a)*yweight+128>>8)
178 VADD.I8 Q11,Q11,Q6 @ Q11 = b+((d-b)*yweight+128>>8)
179 VSUBL.U8 Q4, D22,D20 @ Q5:Q4 = b-a
180 VSUBL.U8 Q5, D23,D21
181 VMUL.S16 Q4, Q4, Q12 @ Q5:Q4 = (b-a)*xweight
182 VMUL.S16 Q5, Q5, Q13
183 VRSHRN.S16 D8, Q4, #8 @ Q4 = (b-a)*xweight+128>>8
184 ADD r12,r7, r9
185 VRSHRN.S16 D9, Q5, #8
186 MOV r12,r12,ASR #17
187 VADD.I8 Q8, Q10,Q4 @ Q8 = a+((b-a)*xweight+128>>8)
188 @ Start extracting the chroma x coordinates, and load Cb and Cr.
189 AND r12,r12,#~15 @ Read 16-byte aligned blocks
190 VDUP.I32 Q9, r9 @ Q9 = source_uv_xoffs_q16 x 4
191 ADD r14,r2, r12
192 VADD.I32 Q10,Q0, Q9
193 VLD1.64 {D8, D9, D10,D11},[r14,:128] @ Load Cb
194 PLD [r14,#64]
195 VADD.I32 Q11,Q1, Q9
196 ADD r14,r3, r12
197 VADD.I32 Q12,Q2, Q9
198 VLD1.64 {D12,D13,D14,D15},[r14,:128] @ Load Cr
199 PLD [r14,#64]
200 VADD.I32 Q13,Q3, Q9
201 VRSHRN.S32 D20,Q10,#9 @ Q10 = <xEwExCwCxAwAx8w8x6w6x4w4x2w2x0w0>
202 VRSHRN.S32 D21,Q11,#9
203 VDUP.I8 Q9, r12
204 VRSHRN.S32 D22,Q12,#9 @ Q11 = <xFwFxDwDxBwBx9w9x7w7x5w5x3w3x1w1>
205 VRSHRN.S32 D23,Q13,#9
206 @ We don't actually need the x weights, but we get them for free.
207 @ Free ALU slot
208 VTRN.8 Q10,Q11 @ Q10 = <wFwEwDwCwBwAw9w8w7w6w5w4w3w2w1w0>
209 @ Free ALU slot @ Q11 = <xFxExDxCxBxAx9x8x7x6x5x4x3x2x1x0>
210 VSUB.S8 Q11,Q11,Q9 @ Make offsets relative to the data we loaded.
211 VTBL.8 D18,{D8, D9, D10,D11},D22 @ Index Cb at source_x
212 VMOV.I8 D24,#74
213 VTBL.8 D19,{D8, D9, D10,D11},D23
214 VMOV.I8 D26,#102
215 VTBL.8 D20,{D12,D13,D14,D15},D22 @ Index Cr at source_x
216 VMOV.I8 D27,#25
217 VTBL.8 D21,{D12,D13,D14,D15},D23
218 @ We now have Y' in Q8, Cb in Q9, and Cr in Q10
219 @ We use VDUP to expand constants, because it's a permute instruction, so
220 @ it can dual issue on the A8.
221 SUBS r6, r6, #16 @ width -= 16
222 VMULL.U8 Q4, D16,D24 @ Q5:Q4 = Y'*74
223 VDUP.32 Q6, D30[1] @ Q7:Q6 = bias_G
224 VMULL.U8 Q5, D17,D24
225 VDUP.32 Q7, D30[1]
226 VMLSL.U8 Q6, D18,D27 @ Q7:Q6 = -25*Cb+bias_G
227 VDUP.32 Q11,D30[0] @ Q12:Q11 = bias_R
228 VMLSL.U8 Q7, D19,D27
229 VDUP.32 Q12,D30[0]
230 VMLAL.U8 Q11,D20,D26 @ Q12:Q11 = 102*Cr+bias_R
231 VDUP.32 Q8, D31[0] @ Q13:Q8 = bias_B
232 VMLAL.U8 Q12,D21,D26
233 VDUP.32 Q13,D31[0]
234 VMLAL.U8 Q8, D18,D29 @ Q13:Q8 = 129*Cb+bias_B
235 VMLAL.U8 Q13,D19,D29
236 VMLSL.U8 Q6, D20,D28 @ Q7:Q6 = -25*Cb-52*Cr+bias_G
237 VMLSL.U8 Q7, D21,D28
238 VADD.S16 Q11,Q4, Q11 @ Q12:Q11 = 74*Y'+102*Cr+bias_R
239 VADD.S16 Q12,Q5, Q12
240 VQADD.S16 Q8, Q4, Q8 @ Q13:Q8 = 74*Y'+129*Cr+bias_B
241 VQADD.S16 Q13,Q5, Q13
242 VADD.S16 Q6, Q4, Q6 @ Q7:Q6 = 74*Y'-25*Cb-52*Cr+bias_G
243 VADD.S16 Q7, Q5, Q7
244 @ Push each value to the top of its word and saturate it.
245 VQSHLU.S16 Q11,Q11,#2
246 VQSHLU.S16 Q12,Q12,#2
247 VQSHLU.S16 Q6, Q6, #2
248 VQSHLU.S16 Q7, Q7, #2
249 VQSHLU.S16 Q8, Q8, #2
250 VQSHLU.S16 Q13,Q13,#2
251 @ Merge G and B into R.
252 VSRI.U16 Q11,Q6, #5
253 VSRI.U16 Q12,Q7, #5
254 VSRI.U16 Q11,Q8, #11
255 MOV r14,r8, LSL #4
256 VSRI.U16 Q12,Q13,#11
257 BLT s42xbily_neon_tail
258 VDUP.I32 Q13,r14
259 @ Store the result.
260 VST1.16 {D22,D23,D24,D25},[r0]!
261 BEQ s42xbily_neon_done
262 @ Advance the x coordinates.
263 VADD.I32 Q0, Q0, Q13
264 VADD.I32 Q1, Q1, Q13
265 ADD r7, r14
266 VADD.I32 Q2, Q2, Q13
267 VADD.I32 Q3, Q3, Q13
268 B s42xbily_neon_loop
269 s42xbily_neon_tail:
270 @ We have between 1 and 15 pixels left to write.
271 @ -r6 == the number of pixels we need to skip writing.
272 @ Adjust r0 to point to the last one we need to write, because we're going
273 @ to write them in reverse order.
274 ADD r0, r0, r6, LSL #1
275 MOV r14,#-2
276 ADD r0, r0, #30
277 @ Skip past the ones we don't need to write.
278 SUB PC, PC, r6, LSL #2
279 ORR r0, r0, r0
280 VST1.16 {D25[3]},[r0,:16],r14
281 VST1.16 {D25[2]},[r0,:16],r14
282 VST1.16 {D25[1]},[r0,:16],r14
283 VST1.16 {D25[0]},[r0,:16],r14
284 VST1.16 {D24[3]},[r0,:16],r14
285 VST1.16 {D24[2]},[r0,:16],r14
286 VST1.16 {D24[1]},[r0,:16],r14
287 VST1.16 {D24[0]},[r0,:16],r14
288 VST1.16 {D23[3]},[r0,:16],r14
289 VST1.16 {D23[2]},[r0,:16],r14
290 VST1.16 {D23[1]},[r0,:16],r14
291 VST1.16 {D23[0]},[r0,:16],r14
292 VST1.16 {D22[3]},[r0,:16],r14
293 VST1.16 {D22[2]},[r0,:16],r14
294 VST1.16 {D22[1]},[r0,:16],r14
295 VST1.16 {D22[0]},[r0,:16]
296 s42xbily_neon_done:
297 VPOP {Q4-Q7} @ 16 words.
298 LDMFD r13!,{r4-r9,PC} @ 8 words.
299 .fnend
300 .size ScaleYCbCr42xToRGB565_BilinearY_Row_NEON, .-ScaleYCbCr42xToRGB565_BilinearY_Row_NEON
302 #if defined(__ELF__)&&defined(__linux__)
303 .section .note.GNU-stack,"",%progbits
304 #endif
