| OLD | NEW |
| 1 /* | 1 /* |
| 2 * Copyright 2015 Google Inc. | 2 * Copyright 2015 Google Inc. |
| 3 * | 3 * |
| 4 * Use of this source code is governed by a BSD-style license that can be | 4 * Use of this source code is governed by a BSD-style license that can be |
| 5 * found in the LICENSE file. | 5 * found in the LICENSE file. |
| 6 */ | 6 */ |
| 7 | 7 |
| 8 #include "SkFloatingPoint.h" | |
| 9 #include "SkOpts.h" | 8 #include "SkOpts.h" |
| 9 |
| 10 #define SK_OPTS_NS neon | 10 #define SK_OPTS_NS neon |
| 11 #include "SkBlurImageFilter_opts.h" | 11 #include "SkBlurImageFilter_opts.h" |
| 12 #include "SkFloatingPoint_opts.h" |
| 13 #include "SkUtils_opts.h" |
| 12 #include "SkXfermode_opts.h" | 14 #include "SkXfermode_opts.h" |
| 13 | 15 |
| 14 namespace neon { // This helps identify methods from this file when debugging /
profiling. | |
| 15 | |
| 16 static float rsqrt(float x) { | |
| 17 return sk_float_rsqrt(x); // This sk_float_rsqrt copy will take the NEON co
mpile-time path. | |
| 18 } | |
| 19 | |
| 20 static void memset16(uint16_t* dst, uint16_t value, int n) { | |
| 21 uint16x8_t v8 = vdupq_n_u16(value); | |
| 22 uint16x8x4_t v32 = {{ v8, v8, v8, v8 }}; | |
| 23 | |
| 24 while (n >= 32) { | |
| 25 vst4q_u16(dst, v32); // This swizzles, but we don't care: all lanes are
the same, value. | |
| 26 dst += 32; | |
| 27 n -= 32; | |
| 28 } | |
| 29 switch (n / 8) { | |
| 30 case 3: vst1q_u16(dst, v8); dst += 8; | |
| 31 case 2: vst1q_u16(dst, v8); dst += 8; | |
| 32 case 1: vst1q_u16(dst, v8); dst += 8; | |
| 33 } | |
| 34 if (n & 4) { | |
| 35 vst1_u16(dst, vget_low_u16(v8)); | |
| 36 dst += 4; | |
| 37 } | |
| 38 switch (n & 3) { | |
| 39 case 3: *dst++ = value; | |
| 40 case 2: *dst++ = value; | |
| 41 case 1: *dst = value; | |
| 42 } | |
| 43 } | |
| 44 | |
| 45 static void memset32(uint32_t* dst, uint32_t value, int n) { | |
| 46 uint32x4_t v4 = vdupq_n_u32(value); | |
| 47 uint32x4x4_t v16 = {{ v4, v4, v4, v4 }}; | |
| 48 | |
| 49 while (n >= 16) { | |
| 50 vst4q_u32(dst, v16); // This swizzles, but we don't care: all lanes are
the same, value. | |
| 51 dst += 16; | |
| 52 n -= 16; | |
| 53 } | |
| 54 switch (n / 4) { | |
| 55 case 3: vst1q_u32(dst, v4); dst += 4; | |
| 56 case 2: vst1q_u32(dst, v4); dst += 4; | |
| 57 case 1: vst1q_u32(dst, v4); dst += 4; | |
| 58 } | |
| 59 if (n & 2) { | |
| 60 vst1_u32(dst, vget_low_u32(v4)); | |
| 61 dst += 2; | |
| 62 } | |
| 63 if (n & 1) { | |
| 64 *dst = value; | |
| 65 } | |
| 66 } | |
| 67 | |
| 68 } // namespace neon | |
| 69 | |
| 70 namespace SkOpts { | 16 namespace SkOpts { |
| 71 void Init_neon() { | 17 void Init_neon() { |
| 72 rsqrt = neon::rsqrt; | 18 rsqrt = neon::rsqrt; |
| 73 memset16 = neon::memset16; | 19 memset16 = neon::memset16; |
| 74 memset32 = neon::memset32; | 20 memset32 = neon::memset32; |
| 75 create_xfermode = SkCreate4pxXfermode; | 21 create_xfermode = SkCreate4pxXfermode; |
| 76 | 22 |
| 77 static const auto x = neon::kX, y = neon::kY; | 23 static const auto x = neon::kX, y = neon::kY; |
| 78 box_blur_xx = neon::box_blur<x,x>; | 24 box_blur_xx = neon::box_blur<x,x>; |
| 79 box_blur_xy = neon::box_blur<x,y>; | 25 box_blur_xy = neon::box_blur<x,y>; |
| 80 box_blur_yx = neon::box_blur<y,x>; | 26 box_blur_yx = neon::box_blur<y,x>; |
| 81 } | 27 } |
| 82 } | 28 } |
| OLD | NEW |