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$assert BATCH_TILE % 4 == 0 7*4bdc9457SAndroid Build Coastguard Worker$assert BATCH_TILE >= 4 8*4bdc9457SAndroid Build Coastguard Worker$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 9*4bdc9457SAndroid Build Coastguard Worker#include <assert.h> 10*4bdc9457SAndroid Build Coastguard Worker 11*4bdc9457SAndroid Build Coastguard Worker#include <xmmintrin.h> 12*4bdc9457SAndroid Build Coastguard Worker 13*4bdc9457SAndroid Build Coastguard Worker#include <xnnpack/vunary.h> 14*4bdc9457SAndroid Build Coastguard Worker#include <xnnpack/common.h> 15*4bdc9457SAndroid Build Coastguard Worker 16*4bdc9457SAndroid Build Coastguard Worker 17*4bdc9457SAndroid Build Coastguard Workervoid xnn_f32_vrelu_ukernel__sse_x${BATCH_TILE}( 18*4bdc9457SAndroid Build Coastguard Worker size_t n, 19*4bdc9457SAndroid Build Coastguard Worker const float* x, 20*4bdc9457SAndroid Build Coastguard Worker float* y, 21*4bdc9457SAndroid Build Coastguard Worker const union xnn_f32_relu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS 22*4bdc9457SAndroid Build Coastguard Worker{ 23*4bdc9457SAndroid Build Coastguard Worker assert(n != 0); 24*4bdc9457SAndroid Build Coastguard Worker assert(n % sizeof(float) == 0); 25*4bdc9457SAndroid Build Coastguard Worker assert(x != NULL); 26*4bdc9457SAndroid Build Coastguard Worker assert(y != NULL); 27*4bdc9457SAndroid Build Coastguard Worker 28*4bdc9457SAndroid Build Coastguard Worker const __m128 vzero = _mm_setzero_ps(); 29*4bdc9457SAndroid Build Coastguard Worker 30*4bdc9457SAndroid Build Coastguard Worker for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) { 31*4bdc9457SAndroid Build Coastguard Worker __m128 vacc${ABC[0:4]} = _mm_loadu_ps(x); 32*4bdc9457SAndroid Build Coastguard Worker $for N in range(4, BATCH_TILE, 4): 33*4bdc9457SAndroid Build Coastguard Worker __m128 vacc${ABC[N:N+4]} = _mm_loadu_ps(x + ${N}); 34*4bdc9457SAndroid Build Coastguard Worker x += ${BATCH_TILE}; 35*4bdc9457SAndroid Build Coastguard Worker 36*4bdc9457SAndroid Build Coastguard Worker $for N in range(0, BATCH_TILE, 4): 37*4bdc9457SAndroid Build Coastguard Worker vacc${ABC[N:N+4]} = _mm_max_ps(vacc${ABC[N:N+4]}, vzero); 38*4bdc9457SAndroid Build Coastguard Worker 39*4bdc9457SAndroid Build Coastguard Worker _mm_storeu_ps(y, vacc${ABC[0:4]}); 40*4bdc9457SAndroid Build Coastguard Worker $for N in range(4, BATCH_TILE, 4): 41*4bdc9457SAndroid Build Coastguard Worker _mm_storeu_ps(y + ${N}, vacc${ABC[N:N+4]}); 42*4bdc9457SAndroid Build Coastguard Worker y += ${BATCH_TILE}; 43*4bdc9457SAndroid Build Coastguard Worker } 44*4bdc9457SAndroid Build Coastguard Worker $if BATCH_TILE > 4: 45*4bdc9457SAndroid Build Coastguard Worker for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) { 46*4bdc9457SAndroid Build Coastguard Worker __m128 vacc = _mm_loadu_ps(x); 47*4bdc9457SAndroid Build Coastguard Worker x += 4; 48*4bdc9457SAndroid Build Coastguard Worker 49*4bdc9457SAndroid Build Coastguard Worker vacc = _mm_max_ps(vacc, vzero); 50*4bdc9457SAndroid Build Coastguard Worker 51*4bdc9457SAndroid Build Coastguard Worker _mm_storeu_ps(y, vacc); 52*4bdc9457SAndroid Build Coastguard Worker y += 4; 53*4bdc9457SAndroid Build Coastguard Worker } 54*4bdc9457SAndroid Build Coastguard Worker if XNN_UNLIKELY(n != 0) { 55*4bdc9457SAndroid Build Coastguard Worker __m128 vacc = _mm_loadu_ps(x); 56*4bdc9457SAndroid Build Coastguard Worker vacc = _mm_max_ps(vacc, vzero); 57*4bdc9457SAndroid Build Coastguard Worker 58*4bdc9457SAndroid Build Coastguard Worker if (n & (2 * sizeof(float))) { 59*4bdc9457SAndroid Build Coastguard Worker _mm_storel_pi((__m64*) y, vacc); 60*4bdc9457SAndroid Build Coastguard Worker vacc = _mm_movehl_ps(vacc, vacc); 61*4bdc9457SAndroid Build Coastguard Worker y += 2; 62*4bdc9457SAndroid Build Coastguard Worker } 63*4bdc9457SAndroid Build Coastguard Worker if (n & (1 * sizeof(float))) { 64*4bdc9457SAndroid Build Coastguard Worker _mm_store_ss(y, vacc); 65*4bdc9457SAndroid Build Coastguard Worker } 66*4bdc9457SAndroid Build Coastguard Worker } 67*4bdc9457SAndroid Build Coastguard Worker} 68