Marat Dukhan | d602154 | 2021-06-30 09:04:20 -0700 | [diff] [blame] | 1 | // Auto-generated file. Do not edit! |
| 2 | // Template: src/qs8-gemm/scalar.c.in |
| 3 | // Generator: tools/xngen |
| 4 | // |
| 5 | // Copyright 2021 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 | #include <math.h> |
| 12 | |
| 13 | #include <xnnpack/math.h> |
| 14 | #include <xnnpack/gemm.h> |
| 15 | |
| 16 | |
| 17 | void xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrint( |
| 18 | size_t mr, |
| 19 | size_t nc, |
| 20 | size_t kc, |
| 21 | const int8_t* restrict a, |
| 22 | size_t a_stride, |
| 23 | const void* restrict w, |
| 24 | int8_t* restrict c, |
| 25 | size_t cm_stride, |
| 26 | size_t cn_stride, |
| 27 | const union xnn_qs8_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) |
| 28 | { |
| 29 | assert(mr != 0); |
| 30 | assert(mr <= 1); |
| 31 | assert(nc != 0); |
| 32 | assert(kc != 0); |
| 33 | |
| 34 | const int8_t* a0 = a; |
| 35 | int8_t* c0 = c; |
| 36 | |
| 37 | do { |
| 38 | int32_t vacc0x0 = ((const int32_t*) w)[0]; |
| 39 | int32_t vacc0x1 = ((const int32_t*) w)[1]; |
| 40 | int32_t vacc0x2 = ((const int32_t*) w)[2]; |
| 41 | int32_t vacc0x3 = ((const int32_t*) w)[3]; |
Marat Dukhan | 927d474 | 2021-07-15 13:42:49 -0700 | [diff] [blame] | 42 | w = (const void*) ((const int32_t*) w + 4); |
Marat Dukhan | d602154 | 2021-06-30 09:04:20 -0700 | [diff] [blame] | 43 | |
| 44 | size_t k = kc; |
| 45 | do { |
| 46 | const int32_t va0 = (int32_t) *a0++; |
| 47 | |
Marat Dukhan | 927d474 | 2021-07-15 13:42:49 -0700 | [diff] [blame] | 48 | const int32_t vb0 = (int32_t) ((const int8_t*) w)[0]; |
| 49 | const int32_t vb1 = (int32_t) ((const int8_t*) w)[1]; |
| 50 | const int32_t vb2 = (int32_t) ((const int8_t*) w)[2]; |
| 51 | const int32_t vb3 = (int32_t) ((const int8_t*) w)[3]; |
| 52 | w = (const void*) ((const int8_t*) w + 4); |
Marat Dukhan | d602154 | 2021-06-30 09:04:20 -0700 | [diff] [blame] | 53 | |
| 54 | vacc0x0 += va0 * vb0; |
| 55 | vacc0x1 += va0 * vb1; |
| 56 | vacc0x2 += va0 * vb2; |
| 57 | vacc0x3 += va0 * vb3; |
| 58 | |
| 59 | k -= sizeof(int8_t); |
| 60 | } while (k != 0); |
| 61 | |
| 62 | float vfpacc0x0 = (float) vacc0x0; |
| 63 | float vfpacc0x1 = (float) vacc0x1; |
| 64 | float vfpacc0x2 = (float) vacc0x2; |
| 65 | float vfpacc0x3 = (float) vacc0x3; |
| 66 | |
| 67 | const float vscale0 = ((const float*) w)[0]; |
| 68 | vfpacc0x0 *= vscale0; |
| 69 | const float vscale1 = ((const float*) w)[1]; |
| 70 | vfpacc0x1 *= vscale1; |
| 71 | const float vscale2 = ((const float*) w)[2]; |
| 72 | vfpacc0x2 *= vscale2; |
| 73 | const float vscale3 = ((const float*) w)[3]; |
| 74 | vfpacc0x3 *= vscale3; |
| 75 | w = (const void*) ((const float*) w + 4); |
| 76 | |
| 77 | long vrndacc0x0 = lrintf(vfpacc0x0); |
| 78 | long vrndacc0x1 = lrintf(vfpacc0x1); |
| 79 | long vrndacc0x2 = lrintf(vfpacc0x2); |
| 80 | long vrndacc0x3 = lrintf(vfpacc0x3); |
| 81 | |
| 82 | const long voutput_min_less_zero_point = params->scalar_lrint.output_min_less_zero_point; |
| 83 | vrndacc0x0 = XNN_UNPREDICTABLE(vrndacc0x0 < voutput_min_less_zero_point) ? voutput_min_less_zero_point : vrndacc0x0; |
| 84 | vrndacc0x1 = XNN_UNPREDICTABLE(vrndacc0x1 < voutput_min_less_zero_point) ? voutput_min_less_zero_point : vrndacc0x1; |
| 85 | vrndacc0x2 = XNN_UNPREDICTABLE(vrndacc0x2 < voutput_min_less_zero_point) ? voutput_min_less_zero_point : vrndacc0x2; |
| 86 | vrndacc0x3 = XNN_UNPREDICTABLE(vrndacc0x3 < voutput_min_less_zero_point) ? voutput_min_less_zero_point : vrndacc0x3; |
| 87 | |
| 88 | const long voutput_max_less_zero_point = params->scalar_lrint.output_max_less_zero_point; |
| 89 | vrndacc0x0 = XNN_UNPREDICTABLE(vrndacc0x0 > voutput_max_less_zero_point) ? voutput_max_less_zero_point : vrndacc0x0; |
| 90 | vrndacc0x1 = XNN_UNPREDICTABLE(vrndacc0x1 > voutput_max_less_zero_point) ? voutput_max_less_zero_point : vrndacc0x1; |
| 91 | vrndacc0x2 = XNN_UNPREDICTABLE(vrndacc0x2 > voutput_max_less_zero_point) ? voutput_max_less_zero_point : vrndacc0x2; |
| 92 | vrndacc0x3 = XNN_UNPREDICTABLE(vrndacc0x3 > voutput_max_less_zero_point) ? voutput_max_less_zero_point : vrndacc0x3; |
| 93 | |
| 94 | const int32_t voutput_zero_point = params->scalar_lrint.output_zero_point; |
| 95 | int32_t vout0x0 = (int32_t) vrndacc0x0 + voutput_zero_point; |
| 96 | int32_t vout0x1 = (int32_t) vrndacc0x1 + voutput_zero_point; |
| 97 | int32_t vout0x2 = (int32_t) vrndacc0x2 + voutput_zero_point; |
| 98 | int32_t vout0x3 = (int32_t) vrndacc0x3 + voutput_zero_point; |
| 99 | |
| 100 | if XNN_LIKELY(nc >= 4) { |
| 101 | c0[0] = (int8_t) vout0x0; |
| 102 | c0[1] = (int8_t) vout0x1; |
| 103 | c0[2] = (int8_t) vout0x2; |
| 104 | c0[3] = (int8_t) vout0x3; |
| 105 | |
| 106 | a0 = (const int8_t*) ((uintptr_t) a0 - kc); |
| 107 | |
| 108 | c0 = (int8_t*) ((uintptr_t) c0 + cn_stride); |
| 109 | |
| 110 | nc -= 4; |
| 111 | } else { |
| 112 | if (nc & 2) { |
| 113 | c0[0] = (int8_t) vout0x0; |
| 114 | c0[1] = (int8_t) vout0x1; |
| 115 | vout0x0 = vout0x2; |
| 116 | c0 += 2; |
| 117 | } |
| 118 | if (nc & 1) { |
| 119 | c0[0] = (int8_t) vout0x0; |
| 120 | } |
| 121 | |
| 122 | nc = 0; |
| 123 | } |
| 124 | } while (nc != 0); |
| 125 | } |