1// Copyright 2021 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$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 9$assert OP in ["RNDNE", "RNDZ", "RNDU", "RNDD"] 10#include <assert.h> 11 12#include <wasm_simd128.h> 13 14#include <xnnpack/common.h> 15#include <xnnpack/math.h> 16#include <xnnpack/vunary.h> 17 18 19$WASM_F32X4_RND = { 20$ "RNDNE": "wasm_f32x4_nearest", 21$ "RNDZ": "wasm_f32x4_trunc", 22$ "RNDU": "wasm_f32x4_ceil", 23$ "RNDD": "wasm_f32x4_floor", 24$}[OP] 25void xnn_f32_v${OP.lower()}_ukernel__wasmsimd_x${BATCH_TILE}( 26 size_t n, 27 const float* x, 28 float* y, 29 const union xnn_f32_rnd_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS 30{ 31 assert(n != 0); 32 assert(n % sizeof(float) == 0); 33 34 for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) { 35 $for N in range(0, BATCH_TILE, 4): 36 const v128_t vx${ABC[N:N+4]} = wasm_v128_load(x); x += 4; 37 38 $for N in range(0, BATCH_TILE, 4): 39 const v128_t vy${ABC[N:N+4]} = ${WASM_F32X4_RND}(vx${ABC[N:N+4]}); 40 41 $for N in range(0, BATCH_TILE, 4): 42 wasm_v128_store(y, vy${ABC[N:N+4]}); y += 4; 43 } 44 $if BATCH_TILE > 4: 45 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) { 46 const v128_t vx = wasm_v128_load(x); x += 4; 47 const v128_t vy = ${WASM_F32X4_RND}(vx); 48 wasm_v128_store(y, vy); y += 4; 49 } 50 if XNN_UNLIKELY(n != 0) { 51 const v128_t vx = wasm_v128_load(x); 52 v128_t vy = ${WASM_F32X4_RND}(vx); 53 54 if (n & (2 * sizeof(float))) { 55 *((double*) y) = wasm_f64x2_extract_lane(vy, 0); 56 vy = wasm_v32x4_shuffle(vy, vy, 2, 3, 2, 3); 57 y += 2; 58 } 59 if (n & (1 * sizeof(float))) { 60 *y = wasm_f32x4_extract_lane(vy, 0); 61 } 62 } 63} 64