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$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 9$assert OP in ["ABS", "NEG", "SQR"] 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$WASM_F32X4_OP = { 20$ "ABS": lambda x: "wasm_v128_and(%s, vnonsign_mask)" % x, 21$ "NEG": lambda x: "wasm_v128_xor(%s, vsign_mask)" % x, 22$ "SQR": lambda x: "wasm_f32x4_mul(%s, %s)" % (x, x), 23$}[OP] 24$PARAMS = { 25$ "ABS": "xnn_f32_abs_params", 26$ "NEG": "xnn_f32_neg_params", 27$ "SQR": "xnn_f32_default_params", 28$}[OP] 29void xnn_f32_v${OP.lower()}_ukernel__wasmsimd_x${BATCH_TILE}( 30 size_t n, 31 const float* x, 32 float* y, 33 const union ${PARAMS} params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS 34{ 35 assert(n != 0); 36 assert(n % sizeof(float) == 0); 37 assert(x != NULL); 38 assert(y != NULL); 39 40 $if OP == "ABS": 41 const v128_t vnonsign_mask = wasm_v128_load64_splat(¶ms->wasmsimd.nonsign_mask); 42 $elif OP == "NEG": 43 const v128_t vsign_mask = wasm_v128_load64_splat(¶ms->wasmsimd.sign_mask); 44 $if BATCH_TILE > 4: 45 for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) { 46 const v128_t vx${ABC[0:4]} = wasm_v128_load(x); 47 $for N in range(4, BATCH_TILE, 4): 48 const v128_t vx${ABC[N:N+4]} = wasm_v128_load(x + ${N}); 49 x += ${BATCH_TILE}; 50 51 $for N in range(0, BATCH_TILE, 4): 52 const v128_t vy${ABC[N:N+4]} = ${WASM_F32X4_OP("vx" + ABC[N:N+4])}; 53 54 wasm_v128_store(y, vy${ABC[0:4]}); 55 $for N in range(4, BATCH_TILE, 4): 56 wasm_v128_store(y + ${N}, vy${ABC[N:N+4]}); 57 y += ${BATCH_TILE}; 58 } 59 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) { 60 const v128_t vx = wasm_v128_load(x); 61 x += 4; 62 63 const v128_t vy = ${WASM_F32X4_OP("vx")}; 64 65 wasm_v128_store(y, vy); 66 y += 4; 67 } 68 if XNN_UNLIKELY(n != 0) { 69 const v128_t vx = wasm_v128_load(x); 70 v128_t vy = ${WASM_F32X4_OP("vx")}; 71 if (n & (2 * sizeof(float))) { 72 *((double*) y) = wasm_f64x2_extract_lane(vy, 0); 73 vy = wasm_v32x4_shuffle(vy, vy, 2, 3, 2, 3); 74 y += 2; 75 } 76 if (n & (1 * sizeof(float))) { 77 *y = wasm_f32x4_extract_lane(vy, 0); 78 } 79 } 80} 81