1 // Auto-generated file. Do not edit! 2 // Template: src/f32-vrelu/scalar.c.in 3 // Generator: tools/xngen 4 // 5 // Copyright 2020 Google LLC 6 // 7 // This source code is licensed under the BSD-style license found in the 8 // LICENSE file in the root directory of this source tree. 9 10 #include <assert.h> 11 12 #include <xnnpack/vunary.h> 13 #include <xnnpack/common.h> 14 #include <xnnpack/math.h> 15 xnn_f32_vrelu_ukernel__scalar_x8(size_t n,const float * x_ptr,float * y_ptr,const union xnn_f32_relu_params params[restrict XNN_MIN_ELEMENTS (1)])16void xnn_f32_vrelu_ukernel__scalar_x8( 17 size_t n, 18 const float* x_ptr, 19 float* y_ptr, 20 const union xnn_f32_relu_params params[restrict XNN_MIN_ELEMENTS(1)]) 21 { 22 assert(n != 0); 23 assert(n % sizeof(float) == 0); 24 assert(x_ptr != NULL); 25 assert(y_ptr != NULL); 26 27 const uint32_t* x = (const uint32_t*)x_ptr; 28 uint32_t* y = (uint32_t*)y_ptr; 29 30 for (; n >= 8 * sizeof(uint32_t); n -= 8 * sizeof(uint32_t)) { 31 uint32_t vacc0 = x[0]; 32 uint32_t vacc1 = x[1]; 33 uint32_t vacc2 = x[2]; 34 uint32_t vacc3 = x[3]; 35 uint32_t vacc4 = x[4]; 36 uint32_t vacc5 = x[5]; 37 uint32_t vacc6 = x[6]; 38 uint32_t vacc7 = x[7]; 39 x += 8; 40 41 vacc0 = ((vacc0 >> 31) - 1) & vacc0; 42 vacc1 = ((vacc1 >> 31) - 1) & vacc1; 43 vacc2 = ((vacc2 >> 31) - 1) & vacc2; 44 vacc3 = ((vacc3 >> 31) - 1) & vacc3; 45 vacc4 = ((vacc4 >> 31) - 1) & vacc4; 46 vacc5 = ((vacc5 >> 31) - 1) & vacc5; 47 vacc6 = ((vacc6 >> 31) - 1) & vacc6; 48 vacc7 = ((vacc7 >> 31) - 1) & vacc7; 49 50 y[0] = vacc0; 51 y[1] = vacc1; 52 y[2] = vacc2; 53 y[3] = vacc3; 54 y[4] = vacc4; 55 y[5] = vacc5; 56 y[6] = vacc6; 57 y[7] = vacc7; 58 y += 8; 59 } 60 if XNN_UNLIKELY(n != 0) { 61 do { 62 uint32_t vacc = *x++; 63 vacc = ((vacc >> 31) - 1) & vacc; 64 *y++ = vacc; 65 n -= sizeof(uint32_t); 66 } while (n != 0); 67 } 68 } 69