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$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 9#include <assert.h> 10 11#include <wasm_simd128.h> 12 13#include <xnnpack/common.h> 14#include <xnnpack/math.h> 15#include <xnnpack/vunary.h> 16 17 18void xnn_f32_vrndd_ukernel__wasmsimd_cvt_x${BATCH_TILE}( 19 size_t n, 20 const float* x, 21 float* y, 22 const union xnn_f32_rnd_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS 23{ 24 assert(n != 0); 25 assert(n % sizeof(float) == 0); 26 27 const v128_t vsign_mask = wasm_v128_load64_splat(params->wasmsimd.sign_mask); 28 const v128_t vmagic_bias = wasm_v128_load64_splat(params->wasmsimd.magic_bias); 29 const v128_t vone = wasm_v128_load64_splat(params->wasmsimd.one); 30 $if BATCH_TILE > 4: 31 for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) { 32 const v128_t vx${ABC[0:4]} = wasm_v128_load(x); 33 $for N in range(4, BATCH_TILE, 4): 34 const v128_t vx${ABC[N:N+4]} = wasm_v128_load(x + ${N}); 35 x += ${BATCH_TILE}; 36 37 $for N in range(0, BATCH_TILE, 4): 38 const v128_t vintx${ABC[N:N+4]} = wasm_i32x4_trunc_sat_f32x4(vx${ABC[N:N+4]}); 39 const v128_t vabsx${ABC[N:N+4]} = wasm_f32x4_abs(vx${ABC[N:N+4]}); 40 41 $for N in range(0, BATCH_TILE, 4): 42 const v128_t vprerndx${ABC[N:N+4]} = wasm_f32x4_convert_i32x4(vintx${ABC[N:N+4]}); 43 const v128_t vrndmask${ABC[N:N+4]} = wasm_v128_andnot(wasm_f32x4_lt(vabsx${ABC[N:N+4]}, vmagic_bias), vsign_mask); 44 45 $for N in range(0, BATCH_TILE, 4): 46 const v128_t vrndx${ABC[N:N+4]} = wasm_v128_bitselect(vprerndx${ABC[N:N+4]}, vx${ABC[N:N+4]}, vrndmask${ABC[N:N+4]}); 47 48 $for N in range(0, BATCH_TILE, 4): 49 const v128_t vadj${ABC[N:N+4]} = wasm_v128_and(wasm_f32x4_lt(vx${ABC[N:N+4]}, vrndx${ABC[N:N+4]}), vone); 50 51 $for N in range(0, BATCH_TILE, 4): 52 const v128_t vy${ABC[N:N+4]} = wasm_f32x4_sub(vrndx${ABC[N:N+4]}, vadj${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 vintx = wasm_i32x4_trunc_sat_f32x4(vx); 64 const v128_t vabsx = wasm_f32x4_abs(vx); 65 const v128_t vprerndx = wasm_f32x4_convert_i32x4(vintx); 66 const v128_t vrndmask = wasm_v128_andnot(wasm_f32x4_lt(vabsx, vmagic_bias), vsign_mask); 67 const v128_t vrndx = wasm_v128_bitselect(vprerndx, vx, vrndmask); 68 const v128_t vadj = wasm_v128_and(wasm_f32x4_lt(vx, vrndx), vone); 69 const v128_t vy = wasm_f32x4_sub(vrndx, vadj); 70 71 wasm_v128_store(y, vy); 72 y += 4; 73 } 74 if XNN_UNLIKELY(n != 0) { 75 const v128_t vx = wasm_v128_load(x); 76 77 const v128_t vintx = wasm_i32x4_trunc_sat_f32x4(vx); 78 const v128_t vabsx = wasm_f32x4_abs(vx); 79 const v128_t vprerndx = wasm_f32x4_convert_i32x4(vintx); 80 const v128_t vrndmask = wasm_v128_andnot(wasm_f32x4_lt(vabsx, vmagic_bias), vsign_mask); 81 const v128_t vrndx = wasm_v128_bitselect(vprerndx, vx, vrndmask); 82 const v128_t vadj = wasm_v128_and(wasm_f32x4_lt(vx, vrndx), vone); 83 v128_t vy = wasm_f32x4_sub(vrndx, vadj); 84 85 if (n & (2 * sizeof(float))) { 86 *((double*) y) = wasm_f64x2_extract_lane(vy, 0); 87 vy = wasm_v32x4_shuffle(vy, vy, 2, 3, 2, 3); 88 y += 2; 89 } 90 if (n & (1 * sizeof(float))) { 91 *y = wasm_f32x4_extract_lane(vy, 0); 92 } 93 } 94} 95