1*4bdc9457SAndroid Build Coastguard Worker// Copyright 2020 Google LLC 2*4bdc9457SAndroid Build Coastguard Worker// 3*4bdc9457SAndroid Build Coastguard Worker// This source code is licensed under the BSD-style license found in the 4*4bdc9457SAndroid Build Coastguard Worker// LICENSE file in the root directory of this source tree. 5*4bdc9457SAndroid Build Coastguard Worker 6*4bdc9457SAndroid Build Coastguard Worker$assert BATCH_TILE >= 1 7*4bdc9457SAndroid Build Coastguard Worker$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 8*4bdc9457SAndroid Build Coastguard Worker#include <assert.h> 9*4bdc9457SAndroid Build Coastguard Worker#include <math.h> 10*4bdc9457SAndroid Build Coastguard Worker 11*4bdc9457SAndroid Build Coastguard Worker#include <xnnpack/common.h> 12*4bdc9457SAndroid Build Coastguard Worker#include <xnnpack/vunary.h> 13*4bdc9457SAndroid Build Coastguard Worker 14*4bdc9457SAndroid Build Coastguard Worker 15*4bdc9457SAndroid Build Coastguard Workervoid xnn_f32_vsqrt_ukernel__scalar_sqrt_x${BATCH_TILE}( 16*4bdc9457SAndroid Build Coastguard Worker size_t n, 17*4bdc9457SAndroid Build Coastguard Worker const float* x, 18*4bdc9457SAndroid Build Coastguard Worker float* y, 19*4bdc9457SAndroid Build Coastguard Worker const union xnn_f32_sqrt_params params[restrict XNN_MIN_ELEMENTS(1)]) 20*4bdc9457SAndroid Build Coastguard Worker{ 21*4bdc9457SAndroid Build Coastguard Worker assert(n != 0); 22*4bdc9457SAndroid Build Coastguard Worker assert(n % sizeof(float) == 0); 23*4bdc9457SAndroid Build Coastguard Worker 24*4bdc9457SAndroid Build Coastguard Worker $if BATCH_TILE > 1: 25*4bdc9457SAndroid Build Coastguard Worker for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) { 26*4bdc9457SAndroid Build Coastguard Worker $for N in range(BATCH_TILE): 27*4bdc9457SAndroid Build Coastguard Worker const float vx${ABC[N]} = x[${N}]; 28*4bdc9457SAndroid Build Coastguard Worker x += ${BATCH_TILE}; 29*4bdc9457SAndroid Build Coastguard Worker 30*4bdc9457SAndroid Build Coastguard Worker $for N in range(BATCH_TILE): 31*4bdc9457SAndroid Build Coastguard Worker const float vy${ABC[N]} = sqrtf(vx${ABC[N]}); 32*4bdc9457SAndroid Build Coastguard Worker 33*4bdc9457SAndroid Build Coastguard Worker $for N in range(BATCH_TILE): 34*4bdc9457SAndroid Build Coastguard Worker y[${N}] = vy${ABC[N]}; 35*4bdc9457SAndroid Build Coastguard Worker y += ${BATCH_TILE}; 36*4bdc9457SAndroid Build Coastguard Worker } 37*4bdc9457SAndroid Build Coastguard Worker if XNN_UNLIKELY(n != 0) { 38*4bdc9457SAndroid Build Coastguard Worker $if BATCH_TILE > 2: 39*4bdc9457SAndroid Build Coastguard Worker do { 40*4bdc9457SAndroid Build Coastguard Worker const float vx = *x++; 41*4bdc9457SAndroid Build Coastguard Worker const float vy = sqrtf(vx); 42*4bdc9457SAndroid Build Coastguard Worker *y++ = vy; 43*4bdc9457SAndroid Build Coastguard Worker n -= sizeof(float); 44*4bdc9457SAndroid Build Coastguard Worker } while (n != 0); 45*4bdc9457SAndroid Build Coastguard Worker $else: 46*4bdc9457SAndroid Build Coastguard Worker const float vx = *x; 47*4bdc9457SAndroid Build Coastguard Worker const float vy = sqrtf(vx); 48*4bdc9457SAndroid Build Coastguard Worker *y = vy; 49*4bdc9457SAndroid Build Coastguard Worker } 50*4bdc9457SAndroid Build Coastguard Worker $else: 51*4bdc9457SAndroid Build Coastguard Worker for (; n >= sizeof(float); n -= sizeof(float)) { 52*4bdc9457SAndroid Build Coastguard Worker const float vx = *x++; 53*4bdc9457SAndroid Build Coastguard Worker const float vy = sqrtf(vx); 54*4bdc9457SAndroid Build Coastguard Worker *y++ = vy; 55*4bdc9457SAndroid Build Coastguard Worker } 56*4bdc9457SAndroid Build Coastguard Worker} 57