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 #ifndef SkNx_neon_DEFINED | 8 #ifndef SkNx_neon_DEFINED |
9 #define SkNx_neon_DEFINED | 9 #define SkNx_neon_DEFINED |
10 | 10 |
11 #define SKNX_IS_FAST | 11 #define SKNX_IS_FAST |
12 | 12 |
13 // ARMv8 has vrndmq_f32 to floor 4 floats. Here we emulate it: | 13 // ARMv8 has vrndmq_f32 to floor 4 floats. Here we emulate it: |
14 // - round by adding (1<<23) with our sign, then subtracting it; | 14 // - roundtrip through integers via truncation |
15 // - if that rounded value is bigger than our input, subtract 1. | 15 // - subtract 1 if that's too big (possible for negative values). |
| 16 // This restricts the domain of our inputs to a maximum somehwere around 2^31.
Seems plenty big. |
16 static inline float32x4_t armv7_vrndmq_f32(float32x4_t v) { | 17 static inline float32x4_t armv7_vrndmq_f32(float32x4_t v) { |
17 auto sign = vandq_u32((uint32x4_t)v, vdupq_n_u32(1<<31)); | 18 auto roundtrip = vcvtq_f32_s32(vcvtq_s32_f32(v)); |
18 auto bias = (float32x4_t)(vorrq_u32((uint32x4_t)vdupq_n_f32(1<<23), sign)); | 19 auto too_big = vcgtq_f32(roundtrip, v); |
19 auto rounded = vsubq_f32(vaddq_f32(v, bias), bias); | 20 return vsubq_f32(roundtrip, (float32x4_t)vandq_u32(too_big, (uint32x4_t)vdup
q_n_f32(1))); |
20 auto too_big = vcgtq_f32(rounded, v); | |
21 return vsubq_f32(rounded, (float32x4_t)vandq_u32(too_big, (uint32x4_t)vdupq_
n_f32(1))); | |
22 } | 21 } |
23 | 22 |
24 // Well, this is absurd. The shifts require compile-time constant arguments. | 23 // Well, this is absurd. The shifts require compile-time constant arguments. |
25 | 24 |
26 #define SHIFT8(op, v, bits) switch(bits) { \ | 25 #define SHIFT8(op, v, bits) switch(bits) { \ |
27 case 1: return op(v, 1); case 2: return op(v, 2); case 3: return op(v
, 3); \ | 26 case 1: return op(v, 1); case 2: return op(v, 2); case 3: return op(v
, 3); \ |
28 case 4: return op(v, 4); case 5: return op(v, 5); case 6: return op(v
, 6); \ | 27 case 4: return op(v, 4); case 5: return op(v, 5); case 6: return op(v
, 6); \ |
29 case 7: return op(v, 7); \ | 28 case 7: return op(v, 7); \ |
30 } return fVec | 29 } return fVec |
31 | 30 |
(...skipping 356 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
388 | 387 |
389 template<> inline Sk4h SkNx_cast<uint16_t, uint8_t>(const Sk4b& src) { | 388 template<> inline Sk4h SkNx_cast<uint16_t, uint8_t>(const Sk4b& src) { |
390 return vget_low_u16(vmovl_u8(src.fVec)); | 389 return vget_low_u16(vmovl_u8(src.fVec)); |
391 } | 390 } |
392 | 391 |
393 template<> inline Sk4b SkNx_cast<uint8_t, uint16_t>(const Sk4h& src) { | 392 template<> inline Sk4b SkNx_cast<uint8_t, uint16_t>(const Sk4h& src) { |
394 return vmovn_u16(vcombine_u16(src.fVec, src.fVec)); | 393 return vmovn_u16(vcombine_u16(src.fVec, src.fVec)); |
395 } | 394 } |
396 | 395 |
397 #endif//SkNx_neon_DEFINED | 396 #endif//SkNx_neon_DEFINED |
OLD | NEW |