Chromium Code Reviews
chromiumcodereview-hr@appspot.gserviceaccount.com (chromiumcodereview-hr) | Please choose your nickname with Settings | Help | Chromium Project | Gerrit Changes | Sign out
(15)

Issues Search
    My Issues | Starred     Open | Closed | All

Side by Side Diff: src/opts/SkBitmapProcState_filter_neon.h

Issue 21915004: ARM Skia NEON patches - 19 - Intrinsics version of the Filter32 routines (Closed) Base URL: https://skia.googlecode.com/svn/trunk
Patch Set: Created 7 years, 4 months ago
Use n/p to move between diff chunks; N/P to move between comments. Draft comments are only viewable by you.
Jump to:
View unified diff | Download patch | Annotate | Revision Log
« no previous file with comments | « no previous file | no next file » | no next file with comments »
Toggle Intra-line Diffs ('i') | Expand Comments ('e') | Collapse Comments ('c') | Show Comments Hide Comments ('s')
OLDNEW
1 1
2 /* 2 /*
3 * Copyright 2012 The Android Open Source Project 3 * Copyright 2012 The Android Open Source Project
4 * 4 *
5 * Use of this source code is governed by a BSD-style license that can be 5 * Use of this source code is governed by a BSD-style license that can be
6 * found in the LICENSE file. 6 * found in the LICENSE file.
7 */ 7 */
8 8
9 9
10 #include <arm_neon.h>
10 #include "SkColorPriv.h" 11 #include "SkColorPriv.h"
11 12
12 /* 13 /*
13 Filter_32_opaque 14 * Filter_32_opaque
14 15 *
15 There is no hard-n-fast rule that the filtering must produce 16 * There is no hard-n-fast rule that the filtering must produce
16 exact results for the color components, but if the 4 incoming colors are 17 * exact results for the color components, but if the 4 incoming colors are
17 all opaque, then the output color must also be opaque. Subsequent parts of 18 * all opaque, then the output color must also be opaque. Subsequent parts of
18 the drawing pipeline may rely on this (e.g. which blitrow proc to use). 19 * the drawing pipeline may rely on this (e.g. which blitrow proc to use).
19 */ 20 */
20 21
21 static inline void Filter_32_opaque_neon(unsigned x, unsigned y, 22 static inline void Filter_32_opaque_neon(unsigned x, unsigned y,
22 SkPMColor a00, SkPMColor a01, 23 SkPMColor a00, SkPMColor a01,
23 SkPMColor a10, SkPMColor a11, 24 SkPMColor a10, SkPMColor a11,
24 SkPMColor *dst) { 25 SkPMColor *dst) {
25 asm volatile( 26 uint8x8_t vy, vconst16_8, v16_y, vres;
26 "vdup.8 d0, %[y] \n\t" // duplicate y into d0 27 uint16x4_t vx, vconst16_16, v16_x, tmp;
27 "vmov.u8 d16, #16 \n\t" // set up const ant in d16 28 uint32x2_t va0, va1;
28 "vsub.u8 d1, d16, d0 \n\t" // d1 = 16-y 29 uint16x8_t tmp1, tmp2;
29 30
30 "vdup.32 d4, %[a00] \n\t" // duplicate a0 0 into d4 31 vy = vdup_n_u8(y); // duplicate y into vy
31 "vdup.32 d5, %[a10] \n\t" // duplicate a1 0 into d5 32 vconst16_8 = vmov_n_u8(16); // set up constant in vconst16_8
32 "vmov.32 d4[1], %[a01] \n\t" // set top of d 4 to a01 33 v16_y = vsub_u8(vconst16_8, vy); // v16_y = 16-y
33 "vmov.32 d5[1], %[a11] \n\t" // set top of d 5 to a11
34 34
35 "vmull.u8 q3, d4, d1 \n\t" // q3 = [a01|a0 0] * (16-y) 35 va0 = vdup_n_u32(a00); // duplicate a00
36 "vmull.u8 q0, d5, d0 \n\t" // q0 = [a11|a1 0] * y 36 va1 = vdup_n_u32(a10); // duplicate a10
37 va0 = vset_lane_u32(a01, va0, 1); // set top to a01
38 va1 = vset_lane_u32(a11, va1, 1); // set top to a11
37 39
38 "vdup.16 d5, %[x] \n\t" // duplicate x into d5 40 tmp1 = vmull_u8(vreinterpret_u8_u32(va0), v16_y); // tmp1 = [a01|a00] * (16- y)
39 "vmov.u16 d16, #16 \n\t" // set up const ant in d16 41 tmp2 = vmull_u8(vreinterpret_u8_u32(va1), vy); // tmp2 = [a11|a10] * y
40 "vsub.u16 d3, d16, d5 \n\t" // d3 = 16-x
41 42
42 "vmul.i16 d4, d7, d5 \n\t" // d4 = a01 * x 43 vx = vdup_n_u16(x); // duplicate x into vx
43 "vmla.i16 d4, d1, d5 \n\t" // d4 += a11 * x 44 vconst16_16 = vmov_n_u16(16); // set up constant in vconst16_16
44 "vmla.i16 d4, d6, d3 \n\t" // d4 += a00 * (16-x) 45 v16_x = vsub_u16(vconst16_16, vx); // v16_x = 16-x
45 "vmla.i16 d4, d0, d3 \n\t" // d4 += a10 * (16-x) 46
46 "vshrn.i16 d0, q2, #8 \n\t" // shift down r esult by 8 47 tmp = vmul_u16(vget_high_u16(tmp1), vx); // tmp = a01 * x
47 "vst1.32 {d0[0]}, [%[dst]] \n\t" // store result 48 tmp = vmla_u16(tmp, vget_high_u16(tmp2), vx); // tmp += a11 * x
48 : 49 tmp = vmla_u16(tmp, vget_low_u16(tmp1), v16_x); // tmp += a00 * (16-x)
49 : [x] "r" (x), [y] "r" (y), [a00] "r" (a00), [a01] "r" (a01), [ a10] "r" (a10), [a11] "r" (a11), [dst] "r" (dst) 50 tmp = vmla_u16(tmp, vget_low_u16(tmp2), v16_x); // tmp += a10 * (16-x)
50 : "cc", "memory", "d0", "d1", "d3", "d4", "d5", "d6", "d7", "d1 6" 51
51 ); 52 vres = vshrn_n_u16(vcombine_u16(tmp, vcreate_u16(0)), 8); // shift down resu lt by 8
53 vst1_lane_u32(dst, vreinterpret_u32_u8(vres), 0); // store result
52 } 54 }
53 55
54 static inline void Filter_32_alpha_neon(unsigned x, unsigned y, 56 static inline void Filter_32_alpha_neon(unsigned x, unsigned y,
55 SkPMColor a00, SkPMColor a01, 57 SkPMColor a00, SkPMColor a01,
56 SkPMColor a10, SkPMColor a11, 58 SkPMColor a10, SkPMColor a11,
57 SkPMColor *dst, uint16_t scale) { 59 SkPMColor *dst, uint16_t scale) {
58 asm volatile( 60 uint8x8_t vy, vconst16_8, v16_y, vres;
59 "vdup.8 d0, %[y] \n\t" // duplicate y into d0 61 uint16x4_t vx, vconst16_16, v16_x, tmp, vscale;
60 "vmov.u8 d16, #16 \n\t" // set up const ant in d16 62 uint32x2_t va0, va1;
61 "vsub.u8 d1, d16, d0 \n\t" // d1 = 16-y 63 uint16x8_t tmp1, tmp2;
62 64
63 "vdup.32 d4, %[a00] \n\t" // duplicate a0 0 into d4 65 vy = vdup_n_u8(y); // duplicate y into vy
64 "vdup.32 d5, %[a10] \n\t" // duplicate a1 0 into d5 66 vconst16_8 = vmov_n_u8(16); // set up constant in vconst16_8
65 "vmov.32 d4[1], %[a01] \n\t" // set top of d 4 to a01 67 v16_y = vsub_u8(vconst16_8, vy); // v16_y = 16-y
66 "vmov.32 d5[1], %[a11] \n\t" // set top of d 5 to a11
67 68
68 "vmull.u8 q3, d4, d1 \n\t" // q3 = [a01|a0 0] * (16-y) 69 va0 = vdup_n_u32(a00); // duplicate a00
69 "vmull.u8 q0, d5, d0 \n\t" // q0 = [a11|a1 0] * y 70 va1 = vdup_n_u32(a10); // duplicate a10
71 va0 = vset_lane_u32(a01, va0, 1); // set top to a01
72 va1 = vset_lane_u32(a11, va1, 1); // set top to a11
70 73
71 "vdup.16 d5, %[x] \n\t" // duplicate x into d5 74 tmp1 = vmull_u8(vreinterpret_u8_u32(va0), v16_y); // tmp1 = [a01|a00] * (16- y)
72 "vmov.u16 d16, #16 \n\t" // set up const ant in d16 75 tmp2 = vmull_u8(vreinterpret_u8_u32(va1), vy); // tmp2 = [a11|a10] * y
73 "vsub.u16 d3, d16, d5 \n\t" // d3 = 16-x
74 76
75 "vmul.i16 d4, d7, d5 \n\t" // d4 = a01 * x 77 vx = vdup_n_u16(x); // duplicate x into vx
76 "vmla.i16 d4, d1, d5 \n\t" // d4 += a11 * x 78 vconst16_16 = vmov_n_u16(16); // set up constant in vconst16_16
77 "vmla.i16 d4, d6, d3 \n\t" // d4 += a00 * (16-x) 79 v16_x = vsub_u16(vconst16_16, vx); // v16_x = 16-x
78 "vmla.i16 d4, d0, d3 \n\t" // d4 += a10 * (16-x) 80
79 "vdup.16 d3, %[scale] \n\t" // duplicate sc ale into d3 81 tmp = vmul_u16(vget_high_u16(tmp1), vx); // tmp = a01 * x
80 "vshr.u16 d4, d4, #8 \n\t" // shift down r esult by 8 82 tmp = vmla_u16(tmp, vget_high_u16(tmp2), vx); // tmp += a11 * x
81 "vmul.i16 d4, d4, d3 \n\t" // multiply res ult by scale 83 tmp = vmla_u16(tmp, vget_low_u16(tmp1), v16_x); // tmp += a00 * (16-x)
82 "vshrn.i16 d0, q2, #8 \n\t" // shift down r esult by 8 84 tmp = vmla_u16(tmp, vget_low_u16(tmp2), v16_x); // tmp += a10 * (16-x)
83 "vst1.32 {d0[0]}, [%[dst]] \n\t" // store result 85
84 : 86 vscale = vdup_n_u16(scale); // duplicate scale
85 : [x] "r" (x), [y] "r" (y), [a00] "r" (a00), [a01] "r" (a01), [ a10] "r" (a10), [a11] "r" (a11), [dst] "r" (dst), [scale] "r" (scale) 87 tmp = vshr_n_u16(tmp, 8); // shift down result by 8
86 : "cc", "memory", "d0", "d1", "d3", "d4", "d5", "d6", "d7", "d1 6" 88 tmp = vmul_u16(tmp, vscale); // multiply result by scale
87 ); 89
90 vres = vshrn_n_u16(vcombine_u16(tmp, vcreate_u16(0)), 8); // shift down resu lt by 8
91 vst1_lane_u32(dst, vreinterpret_u32_u8(vres), 0); // store result
88 } 92 }
OLDNEW
« no previous file with comments | « no previous file | no next file » | no next file with comments »

Powered by Google App Engine
This is Rietveld 408576698