Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame^] | 1 | // Auto-generated file. Do not edit! |
| 2 | // Template: src/f32-clamp/wasmsimd.c.in |
| 3 | // Generator: tools/xngen |
| 4 | // |
| 5 | // Copyright 2020 Google LLC |
| 6 | // |
| 7 | // This source code is licensed under the BSD-style license found in the |
| 8 | // LICENSE file in the root directory of this source tree. |
| 9 | |
| 10 | #include <assert.h> |
| 11 | |
| 12 | #include <wasm_simd128.h> |
| 13 | |
| 14 | #include <xnnpack/clamp.h> |
| 15 | #include <xnnpack/common.h> |
| 16 | |
| 17 | |
| 18 | void xnn_f32_clamp_ukernel__wasmsimd_x86_x8( |
| 19 | size_t n, |
| 20 | const float* x, |
| 21 | float* y, |
| 22 | const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN |
| 23 | { |
| 24 | assert(n != 0); |
| 25 | assert(n % sizeof(float) == 0); |
| 26 | |
| 27 | const v128_t vy_min = wasm_v32x4_load_splat(¶ms->scalar.min); |
| 28 | const v128_t vy_max = wasm_v32x4_load_splat(¶ms->scalar.max); |
| 29 | |
| 30 | for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) { |
| 31 | v128_t vacc0123 = wasm_v128_load(x); |
| 32 | v128_t vacc4567 = wasm_v128_load(x + 4); |
| 33 | x += 8; |
| 34 | |
| 35 | const v128_t vmaskmin0123 = wasm_f32x4_lt(vacc0123, vy_min); |
| 36 | const v128_t vmaskmin4567 = wasm_f32x4_lt(vacc4567, vy_min); |
| 37 | |
| 38 | const v128_t vmaskmax0123 = wasm_f32x4_le(vy_max, vacc0123); |
| 39 | vacc0123 = wasm_v128_bitselect(vy_min, vacc0123, vmaskmin0123); |
| 40 | const v128_t vmaskmax4567 = wasm_f32x4_le(vy_max, vacc4567); |
| 41 | vacc4567 = wasm_v128_bitselect(vy_min, vacc4567, vmaskmin4567); |
| 42 | |
| 43 | vacc0123 = wasm_v128_bitselect(vy_max, vacc0123, vmaskmax0123); |
| 44 | vacc4567 = wasm_v128_bitselect(vy_max, vacc4567, vmaskmax4567); |
| 45 | |
| 46 | wasm_v128_store(y, vacc0123); |
| 47 | wasm_v128_store(y + 4, vacc4567); |
| 48 | y += 8; |
| 49 | } |
| 50 | for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) { |
| 51 | v128_t vacc = wasm_v128_load(x); |
| 52 | x += 4; |
| 53 | |
| 54 | const v128_t vmaskmin = wasm_f32x4_lt(vacc, vy_min); |
| 55 | const v128_t vmaskmax = wasm_f32x4_le(vy_max, vacc); |
| 56 | vacc = wasm_v128_bitselect(vy_min, vacc, vmaskmin); |
| 57 | vacc = wasm_v128_bitselect(vy_max, vacc, vmaskmax); |
| 58 | |
| 59 | wasm_v128_store(y, vacc); |
| 60 | y += 4; |
| 61 | } |
| 62 | if XNN_UNLIKELY(n != 0) { |
| 63 | v128_t vacc = wasm_v128_load(x); |
| 64 | |
| 65 | const v128_t vmaskmin = wasm_f32x4_lt(vacc, vy_min); |
| 66 | const v128_t vmaskmax = wasm_f32x4_le(vy_max, vacc); |
| 67 | vacc = wasm_v128_bitselect(vy_min, vacc, vmaskmin); |
| 68 | vacc = wasm_v128_bitselect(vy_max, vacc, vmaskmax); |
| 69 | |
| 70 | if (n & (2 * sizeof(float))) { |
| 71 | *((double*) y) = wasm_f64x2_extract_lane(vacc, 0); |
| 72 | vacc = wasm_v32x4_shuffle(vacc, vacc, 2, 3, 2, 3); |
| 73 | y += 2; |
| 74 | } |
| 75 | if (n & (1 * sizeof(float))) { |
| 76 | *y = wasm_f32x4_extract_lane(vacc, 0); |
| 77 | } |
| 78 | } |
| 79 | } |