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$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 9#include <assert.h> 10 11#include <wasm_simd128.h> 12 13#include <xnnpack/common.h> 14#include <xnnpack/vunary.h> 15 16 17void xnn_f32_vclamp_ukernel__wasmsimd_${"x86" if X86 else "arm"}_x${BATCH_TILE}( 18 size_t n, 19 const float* x, 20 float* y, 21 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS 22{ 23 assert(n != 0); 24 assert(n % sizeof(float) == 0); 25 assert(x != NULL); 26 assert(y != NULL); 27 28 const v128_t vy_min = wasm_v128_load64_splat(params->wasmsimd.min); 29 const v128_t vy_max = wasm_v128_load64_splat(params->wasmsimd.max); 30 31 $if BATCH_TILE > 4: 32 for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) { 33 v128_t vacc${ABC[0:4]} = wasm_v128_load(x); 34 $for N in range(4, BATCH_TILE, 4): 35 v128_t vacc${ABC[N:N+4]} = wasm_v128_load(x + ${N}); 36 x += ${BATCH_TILE}; 37 38 $if X86: 39 $for N in range(0, BATCH_TILE, 4): 40 vacc${ABC[N:N+4]} = wasm_f32x4_pmax(vy_min, vacc${ABC[N:N+4]}); 41 42 $for N in range(0, BATCH_TILE, 4): 43 vacc${ABC[N:N+4]} = wasm_f32x4_pmin(vy_max, vacc${ABC[N:N+4]}); 44 $else: 45 $for N in range(0, BATCH_TILE, 4): 46 vacc${ABC[N:N+4]} = wasm_f32x4_max(vacc${ABC[N:N+4]}, vy_min); 47 48 $for N in range(0, BATCH_TILE, 4): 49 vacc${ABC[N:N+4]} = wasm_f32x4_min(vacc${ABC[N:N+4]}, vy_max); 50 51 wasm_v128_store(y, vacc${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 vacc = wasm_v128_load(x); 58 x += 4; 59 60 $if X86: 61 vacc = wasm_f32x4_pmax(vy_min, vacc); 62 vacc = wasm_f32x4_pmin(vy_max, vacc); 63 $else: 64 vacc = wasm_f32x4_max(vacc, vy_min); 65 vacc = wasm_f32x4_min(vacc, vy_max); 66 67 wasm_v128_store(y, vacc); 68 y += 4; 69 } 70 if XNN_UNLIKELY(n != 0) { 71 v128_t vacc = wasm_v128_load(x); 72 73 $if X86: 74 vacc = wasm_f32x4_pmax(vy_min, vacc); 75 vacc = wasm_f32x4_pmin(vy_max, vacc); 76 $else: 77 vacc = wasm_f32x4_max(vacc, vy_min); 78 vacc = wasm_f32x4_min(vacc, vy_max); 79 80 if (n & (2 * sizeof(float))) { 81 *((double*) y) = wasm_f64x2_extract_lane(vacc, 0); 82 vacc = wasm_v32x4_shuffle(vacc, vacc, 2, 3, 2, 3); 83 y += 2; 84 } 85 if (n & (1 * sizeof(float))) { 86 *y = wasm_f32x4_extract_lane(vacc, 0); 87 } 88 } 89} 90