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$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 9#include <assert.h> 10 11#include <wasm_simd128.h> 12 13#include <xnnpack/common.h> 14#include <xnnpack/vunary.h> 15 16 17void xnn_f32_vlrelu_ukernel__wasmsimd_bitselect_x${BATCH_TILE}( 18 size_t n, 19 const float* x, 20 float* y, 21 const union xnn_f32_lrelu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS 22{ 23 assert(n != 0); 24 assert(n % sizeof(float) == 0); 25 26 const v128_t vslope = wasm_v128_load64_splat(params->wasmsimd.slope); 27 $if BATCH_TILE > 4: 28 for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) { 29 const v128_t vx${ABC[0:4]} = wasm_v128_load(x); 30 $for N in range(4, BATCH_TILE, 4): 31 const v128_t vx${ABC[N:N+4]} = wasm_v128_load(x + ${N}); 32 x += ${BATCH_TILE}; 33 34 $for N in range(0, BATCH_TILE, 4): 35 v128_t vacc${ABC[N:N+4]} = wasm_f32x4_mul(vx${ABC[N:N+4]}, vslope); 36 const v128_t vmask${ABC[N:N+4]} = wasm_i32x4_shr(vx${ABC[N:N+4]}, 31); 37 38 $for N in range(0, BATCH_TILE, 4): 39 vacc${ABC[N:N+4]} = wasm_v128_bitselect(vacc${ABC[N:N+4]}, vx${ABC[N:N+4]}, vmask${ABC[N:N+4]}); 40 41 wasm_v128_store(y, vacc${ABC[0:4]}); 42 $for N in range(4, BATCH_TILE, 4): 43 wasm_v128_store(y + ${N}, vacc${ABC[N:N+4]}); 44 y += ${BATCH_TILE}; 45 } 46 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) { 47 const v128_t vx = wasm_v128_load(x); 48 x += 4; 49 v128_t vacc = wasm_f32x4_mul(vx, vslope); 50 const v128_t vmask = wasm_i32x4_shr(vx, 31); 51 vacc = wasm_v128_bitselect(vacc, vx, vmask); 52 wasm_v128_store(y, vacc); 53 y += 4; 54 } 55 if XNN_UNLIKELY(n != 0) { 56 const v128_t vx = wasm_v128_load(x); 57 v128_t vacc = wasm_f32x4_mul(vx, vslope); 58 const v128_t vmask = wasm_i32x4_shr(vx, 31); 59 vacc = wasm_v128_bitselect(vacc, vx, vmask); 60 61 if (n & (2 * sizeof(float))) { 62 *((double*) y) = wasm_f64x2_extract_lane(vacc, 0); 63 vacc = wasm_v32x4_shuffle(vacc, vacc, 2, 3, 2, 3); 64 y += 2; 65 } 66 if (n & (1 * sizeof(float))) { 67 *y = wasm_f32x4_extract_lane(vacc, 0); 68 } 69 } 70} 71