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 #include <assert.h> 7*4bdc9457SAndroid Build Coastguard Worker #include <stddef.h> 8*4bdc9457SAndroid Build Coastguard Worker 9*4bdc9457SAndroid Build Coastguard Worker #include <smmintrin.h> 10*4bdc9457SAndroid Build Coastguard Worker 11*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/math-stubs.h> 12*4bdc9457SAndroid Build Coastguard Worker 13*4bdc9457SAndroid Build Coastguard Worker xnn_math_f32_roundd__sse41(size_t n,const float * input,float * output)14*4bdc9457SAndroid Build Coastguard Workervoid xnn_math_f32_roundd__sse41( 15*4bdc9457SAndroid Build Coastguard Worker size_t n, 16*4bdc9457SAndroid Build Coastguard Worker const float* input, 17*4bdc9457SAndroid Build Coastguard Worker float* output) 18*4bdc9457SAndroid Build Coastguard Worker { 19*4bdc9457SAndroid Build Coastguard Worker assert(n % (4 * sizeof(float)) == 0); 20*4bdc9457SAndroid Build Coastguard Worker 21*4bdc9457SAndroid Build Coastguard Worker for (; n != 0; n -= 4 * sizeof(float)) { 22*4bdc9457SAndroid Build Coastguard Worker const __m128 vx = _mm_load_ps(input); 23*4bdc9457SAndroid Build Coastguard Worker input += 4; 24*4bdc9457SAndroid Build Coastguard Worker 25*4bdc9457SAndroid Build Coastguard Worker const __m128 vy = _mm_round_ps(vx, _MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC); 26*4bdc9457SAndroid Build Coastguard Worker 27*4bdc9457SAndroid Build Coastguard Worker _mm_store_ps(output, vy); 28*4bdc9457SAndroid Build Coastguard Worker output += 4; 29*4bdc9457SAndroid Build Coastguard Worker } 30*4bdc9457SAndroid Build Coastguard Worker } 31