Marat Dukhan | 8c41796 | 2020-07-08 12:27:50 -0700 | [diff] [blame] | 1 | // Copyright 2020 Google LLC |
| 2 | // |
| 3 | // This source code is licensed under the BSD-style license found in the |
| 4 | // LICENSE file in the root directory of this source tree. |
| 5 | |
| 6 | #include <assert.h> |
| 7 | |
| 8 | #include <wasm_simd128.h> |
| 9 | |
| 10 | #include <xnnpack/rmax.h> |
| 11 | |
| 12 | |
| 13 | void xnn_f32_rmax_ukernel__wasmsimd_arm( |
| 14 | size_t n, |
| 15 | const float* x, |
| 16 | float* y) |
| 17 | { |
| 18 | assert(n != 0); |
| 19 | assert(n % sizeof(float) == 0); |
| 20 | |
| 21 | v128_t vmax0 = wasm_v32x4_load_splat(x); |
| 22 | v128_t vmax1 = vmax0; |
| 23 | v128_t vmax2 = vmax0; |
| 24 | v128_t vmax3 = vmax0; |
| 25 | for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) { |
| 26 | const v128_t vx0 = wasm_v128_load(x); |
| 27 | const v128_t vx1 = wasm_v128_load(x + 4); |
| 28 | const v128_t vx2 = wasm_v128_load(x + 8); |
| 29 | const v128_t vx3 = wasm_v128_load(x + 12); |
| 30 | x += 16; |
| 31 | |
| 32 | vmax0 = wasm_f32x4_max(vmax0, vx0); |
| 33 | vmax1 = wasm_f32x4_max(vmax1, vx1); |
| 34 | vmax2 = wasm_f32x4_max(vmax2, vx2); |
| 35 | vmax3 = wasm_f32x4_max(vmax3, vx3); |
| 36 | } |
| 37 | v128_t vmax0123 = wasm_f32x4_max(wasm_f32x4_max(vmax0, vmax1), wasm_f32x4_max(vmax2, vmax3)); |
| 38 | for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) { |
| 39 | const v128_t vx = wasm_v128_load(x); |
| 40 | vmax0123 = wasm_f32x4_max(vmax0123, vx); |
| 41 | x += 4; |
| 42 | } |
| 43 | vmax0123 = wasm_f32x4_max(vmax0123, wasm_v32x4_shuffle(vmax0123, vmax0123, 2, 3, 0, 1)); |
| 44 | float vmax = __builtin_wasm_max_f32(wasm_f32x4_extract_lane(vmax0123, 0), wasm_f32x4_extract_lane(vmax0123, 1)); |
| 45 | if XNN_UNLIKELY(n != 0) { |
| 46 | do { |
| 47 | const float vx = *x++; |
| 48 | vmax = __builtin_wasm_max_f32(vx, vmax); |
| 49 | n -= 4; |
| 50 | } while (n != 0); |
| 51 | } |
| 52 | *y = vmax; |
| 53 | } |