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$assert BATCH_TILE % 4 == 0 7$assert BATCH_TILE >= 4 8$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 9#include <assert.h> 10 11#include <wasm_simd128.h> 12 13#include <xnnpack/common.h> 14#include <xnnpack/hswish.h> 15 16 17void xnn_f32_hswish_ukernel__wasmsimd_x${BATCH_TILE}( 18 size_t n, 19 const float* x, 20 float* y, 21 const union xnn_f32_hswish_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN 22{ 23 assert(n != 0); 24 assert(n % sizeof(float) == 0); 25 26 const v128_t vsixth = wasm_v32x4_load_splat(¶ms->scalar.sixth); 27 const v128_t vthree = wasm_v32x4_load_splat(¶ms->scalar.three); 28 const v128_t vsix = wasm_v32x4_load_splat(¶ms->scalar.six); 29 const v128_t vzero = wasm_f32x4_splat(0.0f); 30 31 $if BATCH_TILE > 4: 32 for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) { 33 v128_t vx${ABC[0:4]} = wasm_v128_load(x); 34 $for N in range(4, BATCH_TILE, 4): 35 v128_t vx${ABC[N:N+4]} = wasm_v128_load(x + ${N}); 36 x += ${BATCH_TILE}; 37 38 $for N in range(0, BATCH_TILE, 4): 39 v128_t vacc${ABC[N:N+4]} = wasm_f32x4_add(vx${ABC[N:N+4]}, vthree); 40 vx${ABC[N:N+4]} = wasm_f32x4_mul(vx${ABC[N:N+4]}, vsixth); 41 42 $for N in range(0, BATCH_TILE, 4): 43 vacc${ABC[N:N+4]} = wasm_i32x4_max(vacc${ABC[N:N+4]}, vzero); 44 45 $for N in range(0, BATCH_TILE, 4): 46 vacc${ABC[N:N+4]} = wasm_i32x4_min(vacc${ABC[N:N+4]}, vsix); 47 48 $for N in range(0, BATCH_TILE, 4): 49 vacc${ABC[N:N+4]} = wasm_f32x4_mul(vacc${ABC[N:N+4]}, vx${ABC[N:N+4]}); 50 51 wasm_v128_store(y, vacc${ABC[0:4]}); 52 $for N in range(4, BATCH_TILE, 4): 53 wasm_v128_store(y + ${N}, vacc${ABC[N:N+4]}); 54 y += ${BATCH_TILE}; 55 } 56 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) { 57 v128_t vx = wasm_v128_load(x); 58 x += 4; 59 60 v128_t vacc = wasm_f32x4_add(vx, vthree); 61 vx = wasm_f32x4_mul(vx, vsixth); 62 vacc = wasm_i32x4_max(vacc, vzero); 63 vacc = wasm_i32x4_min(vacc, vsix); 64 vacc = wasm_f32x4_mul(vacc, vx); 65 66 wasm_v128_store(y, vacc); 67 y += 4; 68 } 69 if XNN_UNLIKELY(n != 0) { 70 v128_t vx = wasm_v128_load(x); 71 72 v128_t vacc = wasm_f32x4_add(vx, vthree); 73 vx = wasm_f32x4_mul(vx, vsixth); 74 vacc = wasm_i32x4_max(vacc, vzero); 75 vacc = wasm_i32x4_min(vacc, vsix); 76 vacc = wasm_f32x4_mul(vacc, vx); 77 78 if (n & (2 * sizeof(float))) { 79 *((double*) y) = wasm_f64x2_extract_lane(vacc, 0); 80 vacc = wasm_v32x4_shuffle(vacc, vacc, 2, 3, 2, 3); 81 y += 2; 82 } 83 if (n & (1 * sizeof(float))) { 84 *y = wasm_f32x4_extract_lane(vacc, 0); 85 } 86 } 87} 88