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 | |
| 12 | #include <fp16.h> |
| 13 | |
| 14 | #include <xnnpack/math.h> |
| 15 | #include <xnnpack/gemm.h> |
| 16 | |
| 17 | |
| 18 | void xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_magic( |
| 19 | size_t mr, |
| 20 | size_t nc, |
| 21 | size_t kc, |
| 22 | const int8_t* restrict a, |
| 23 | size_t a_stride, |
| 24 | const void* restrict w, |
| 25 | int8_t* restrict c, |
| 26 | size_t cm_stride, |
| 27 | size_t cn_stride, |
| 28 | const union xnn_qs8_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) |
| 29 | { |
| 30 | assert(mr != 0); |
| 31 | assert(mr <= 3); |
| 32 | assert(nc != 0); |
| 33 | assert(kc != 0); |
| 34 | |
| 35 | const int8_t* a0 = a; |
| 36 | int8_t* c0 = c; |
| 37 | const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride); |
| 38 | int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride); |
| 39 | if XNN_UNPREDICTABLE(mr < 2) { |
| 40 | a1 = a0; |
| 41 | c1 = c0; |
| 42 | } |
| 43 | const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride); |
| 44 | int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride); |
| 45 | if XNN_UNPREDICTABLE(mr <= 2) { |
| 46 | a2 = a1; |
| 47 | c2 = c1; |
| 48 | } |
| 49 | |
| 50 | do { |
| 51 | int32_t vacc0x0 = ((const int32_t*) w)[0]; |
| 52 | int32_t vacc0x1 = ((const int32_t*) w)[1]; |
| 53 | int32_t vacc0x2 = ((const int32_t*) w)[2]; |
| 54 | int32_t vacc0x3 = ((const int32_t*) w)[3]; |
| 55 | int32_t vacc1x0 = vacc0x0; |
| 56 | int32_t vacc1x1 = vacc0x1; |
| 57 | int32_t vacc1x2 = vacc0x2; |
| 58 | int32_t vacc1x3 = vacc0x3; |
| 59 | int32_t vacc2x0 = vacc0x0; |
| 60 | int32_t vacc2x1 = vacc0x1; |
| 61 | int32_t vacc2x2 = vacc0x2; |
| 62 | int32_t vacc2x3 = vacc0x3; |
Marat Dukhan | 927d474 | 2021-07-15 13:42:49 -0700 | [diff] [blame] | 63 | w = (const void*) ((const int32_t*) w + 4); |
Marat Dukhan | d602154 | 2021-06-30 09:04:20 -0700 | [diff] [blame] | 64 | |
| 65 | size_t k = kc; |
| 66 | do { |
| 67 | const int32_t va0 = (int32_t) *a0++; |
| 68 | const int32_t va1 = (int32_t) *a1++; |
| 69 | const int32_t va2 = (int32_t) *a2++; |
| 70 | |
Marat Dukhan | 927d474 | 2021-07-15 13:42:49 -0700 | [diff] [blame] | 71 | const int32_t vb0 = (int32_t) ((const int8_t*) w)[0]; |
| 72 | const int32_t vb1 = (int32_t) ((const int8_t*) w)[1]; |
| 73 | const int32_t vb2 = (int32_t) ((const int8_t*) w)[2]; |
| 74 | const int32_t vb3 = (int32_t) ((const int8_t*) w)[3]; |
| 75 | w = (const void*) ((const int8_t*) w + 4); |
Marat Dukhan | d602154 | 2021-06-30 09:04:20 -0700 | [diff] [blame] | 76 | |
| 77 | vacc0x0 += va0 * vb0; |
| 78 | vacc0x1 += va0 * vb1; |
| 79 | vacc0x2 += va0 * vb2; |
| 80 | vacc0x3 += va0 * vb3; |
| 81 | vacc1x0 += va1 * vb0; |
| 82 | vacc1x1 += va1 * vb1; |
| 83 | vacc1x2 += va1 * vb2; |
| 84 | vacc1x3 += va1 * vb3; |
| 85 | vacc2x0 += va2 * vb0; |
| 86 | vacc2x1 += va2 * vb1; |
| 87 | vacc2x2 += va2 * vb2; |
| 88 | vacc2x3 += va2 * vb3; |
| 89 | |
| 90 | k -= sizeof(int8_t); |
| 91 | } while (k != 0); |
| 92 | |
| 93 | float vfpacc0x0 = (float) vacc0x0; |
| 94 | float vfpacc0x1 = (float) vacc0x1; |
| 95 | float vfpacc0x2 = (float) vacc0x2; |
| 96 | float vfpacc0x3 = (float) vacc0x3; |
| 97 | float vfpacc1x0 = (float) vacc1x0; |
| 98 | float vfpacc1x1 = (float) vacc1x1; |
| 99 | float vfpacc1x2 = (float) vacc1x2; |
| 100 | float vfpacc1x3 = (float) vacc1x3; |
| 101 | float vfpacc2x0 = (float) vacc2x0; |
| 102 | float vfpacc2x1 = (float) vacc2x1; |
| 103 | float vfpacc2x2 = (float) vacc2x2; |
| 104 | float vfpacc2x3 = (float) vacc2x3; |
| 105 | |
| 106 | const float vscale0 = ((const float*) w)[0]; |
| 107 | vfpacc0x0 *= vscale0; |
| 108 | vfpacc1x0 *= vscale0; |
| 109 | vfpacc2x0 *= vscale0; |
| 110 | const float vscale1 = ((const float*) w)[1]; |
| 111 | vfpacc0x1 *= vscale1; |
| 112 | vfpacc1x1 *= vscale1; |
| 113 | vfpacc2x1 *= vscale1; |
| 114 | const float vscale2 = ((const float*) w)[2]; |
| 115 | vfpacc0x2 *= vscale2; |
| 116 | vfpacc1x2 *= vscale2; |
| 117 | vfpacc2x2 *= vscale2; |
| 118 | const float vscale3 = ((const float*) w)[3]; |
| 119 | vfpacc0x3 *= vscale3; |
| 120 | vfpacc1x3 *= vscale3; |
| 121 | vfpacc2x3 *= vscale3; |
| 122 | w = (const void*) ((const float*) w + 4); |
| 123 | |
| 124 | const float voutput_min_less_zero_point = params->scalar_magic.output_min_less_zero_point; |
| 125 | vfpacc0x0 = math_max_f32(vfpacc0x0, voutput_min_less_zero_point); |
| 126 | vfpacc0x1 = math_max_f32(vfpacc0x1, voutput_min_less_zero_point); |
| 127 | vfpacc0x2 = math_max_f32(vfpacc0x2, voutput_min_less_zero_point); |
| 128 | vfpacc0x3 = math_max_f32(vfpacc0x3, voutput_min_less_zero_point); |
| 129 | vfpacc1x0 = math_max_f32(vfpacc1x0, voutput_min_less_zero_point); |
| 130 | vfpacc1x1 = math_max_f32(vfpacc1x1, voutput_min_less_zero_point); |
| 131 | vfpacc1x2 = math_max_f32(vfpacc1x2, voutput_min_less_zero_point); |
| 132 | vfpacc1x3 = math_max_f32(vfpacc1x3, voutput_min_less_zero_point); |
| 133 | vfpacc2x0 = math_max_f32(vfpacc2x0, voutput_min_less_zero_point); |
| 134 | vfpacc2x1 = math_max_f32(vfpacc2x1, voutput_min_less_zero_point); |
| 135 | vfpacc2x2 = math_max_f32(vfpacc2x2, voutput_min_less_zero_point); |
| 136 | vfpacc2x3 = math_max_f32(vfpacc2x3, voutput_min_less_zero_point); |
| 137 | |
| 138 | const float voutput_max_less_zero_point = params->scalar_magic.output_max_less_zero_point; |
| 139 | vfpacc0x0 = math_min_f32(vfpacc0x0, voutput_max_less_zero_point); |
| 140 | vfpacc0x1 = math_min_f32(vfpacc0x1, voutput_max_less_zero_point); |
| 141 | vfpacc0x2 = math_min_f32(vfpacc0x2, voutput_max_less_zero_point); |
| 142 | vfpacc0x3 = math_min_f32(vfpacc0x3, voutput_max_less_zero_point); |
| 143 | vfpacc1x0 = math_min_f32(vfpacc1x0, voutput_max_less_zero_point); |
| 144 | vfpacc1x1 = math_min_f32(vfpacc1x1, voutput_max_less_zero_point); |
| 145 | vfpacc1x2 = math_min_f32(vfpacc1x2, voutput_max_less_zero_point); |
| 146 | vfpacc1x3 = math_min_f32(vfpacc1x3, voutput_max_less_zero_point); |
| 147 | vfpacc2x0 = math_min_f32(vfpacc2x0, voutput_max_less_zero_point); |
| 148 | vfpacc2x1 = math_min_f32(vfpacc2x1, voutput_max_less_zero_point); |
| 149 | vfpacc2x2 = math_min_f32(vfpacc2x2, voutput_max_less_zero_point); |
| 150 | vfpacc2x3 = math_min_f32(vfpacc2x3, voutput_max_less_zero_point); |
| 151 | |
| 152 | const float vmagic_bias = params->scalar_magic.magic_bias; |
| 153 | vfpacc0x0 += vmagic_bias; |
| 154 | vfpacc0x1 += vmagic_bias; |
| 155 | vfpacc0x2 += vmagic_bias; |
| 156 | vfpacc0x3 += vmagic_bias; |
| 157 | vfpacc1x0 += vmagic_bias; |
| 158 | vfpacc1x1 += vmagic_bias; |
| 159 | vfpacc1x2 += vmagic_bias; |
| 160 | vfpacc1x3 += vmagic_bias; |
| 161 | vfpacc2x0 += vmagic_bias; |
| 162 | vfpacc2x1 += vmagic_bias; |
| 163 | vfpacc2x2 += vmagic_bias; |
| 164 | vfpacc2x3 += vmagic_bias; |
| 165 | |
| 166 | const int32_t vmagic_bias_less_output_zero_point = params->scalar_magic.magic_bias_less_output_zero_point; |
| 167 | int32_t vout0x0 = (int32_t) fp32_to_bits(vfpacc0x0) - vmagic_bias_less_output_zero_point; |
| 168 | int32_t vout0x1 = (int32_t) fp32_to_bits(vfpacc0x1) - vmagic_bias_less_output_zero_point; |
| 169 | int32_t vout0x2 = (int32_t) fp32_to_bits(vfpacc0x2) - vmagic_bias_less_output_zero_point; |
| 170 | int32_t vout0x3 = (int32_t) fp32_to_bits(vfpacc0x3) - vmagic_bias_less_output_zero_point; |
| 171 | int32_t vout1x0 = (int32_t) fp32_to_bits(vfpacc1x0) - vmagic_bias_less_output_zero_point; |
| 172 | int32_t vout1x1 = (int32_t) fp32_to_bits(vfpacc1x1) - vmagic_bias_less_output_zero_point; |
| 173 | int32_t vout1x2 = (int32_t) fp32_to_bits(vfpacc1x2) - vmagic_bias_less_output_zero_point; |
| 174 | int32_t vout1x3 = (int32_t) fp32_to_bits(vfpacc1x3) - vmagic_bias_less_output_zero_point; |
| 175 | int32_t vout2x0 = (int32_t) fp32_to_bits(vfpacc2x0) - vmagic_bias_less_output_zero_point; |
| 176 | int32_t vout2x1 = (int32_t) fp32_to_bits(vfpacc2x1) - vmagic_bias_less_output_zero_point; |
| 177 | int32_t vout2x2 = (int32_t) fp32_to_bits(vfpacc2x2) - vmagic_bias_less_output_zero_point; |
| 178 | int32_t vout2x3 = (int32_t) fp32_to_bits(vfpacc2x3) - vmagic_bias_less_output_zero_point; |
| 179 | |
| 180 | if XNN_LIKELY(nc >= 4) { |
| 181 | c0[0] = (int8_t) vout0x0; |
| 182 | c0[1] = (int8_t) vout0x1; |
| 183 | c0[2] = (int8_t) vout0x2; |
| 184 | c0[3] = (int8_t) vout0x3; |
| 185 | c1[0] = (int8_t) vout1x0; |
| 186 | c1[1] = (int8_t) vout1x1; |
| 187 | c1[2] = (int8_t) vout1x2; |
| 188 | c1[3] = (int8_t) vout1x3; |
| 189 | c2[0] = (int8_t) vout2x0; |
| 190 | c2[1] = (int8_t) vout2x1; |
| 191 | c2[2] = (int8_t) vout2x2; |
| 192 | c2[3] = (int8_t) vout2x3; |
| 193 | |
| 194 | a0 = (const int8_t*) ((uintptr_t) a0 - kc); |
| 195 | a1 = (const int8_t*) ((uintptr_t) a1 - kc); |
| 196 | a2 = (const int8_t*) ((uintptr_t) a2 - kc); |
| 197 | |
| 198 | c0 = (int8_t*) ((uintptr_t) c0 + cn_stride); |
| 199 | c1 = (int8_t*) ((uintptr_t) c1 + cn_stride); |
| 200 | c2 = (int8_t*) ((uintptr_t) c2 + cn_stride); |
| 201 | |
| 202 | nc -= 4; |
| 203 | } else { |
| 204 | if (nc & 2) { |
| 205 | c0[0] = (int8_t) vout0x0; |
| 206 | c0[1] = (int8_t) vout0x1; |
| 207 | vout0x0 = vout0x2; |
| 208 | c0 += 2; |
| 209 | c1[0] = (int8_t) vout1x0; |
| 210 | c1[1] = (int8_t) vout1x1; |
| 211 | vout1x0 = vout1x2; |
| 212 | c1 += 2; |
| 213 | c2[0] = (int8_t) vout2x0; |
| 214 | c2[1] = (int8_t) vout2x1; |
| 215 | vout2x0 = vout2x2; |
| 216 | c2 += 2; |
| 217 | } |
| 218 | if (nc & 1) { |
| 219 | c0[0] = (int8_t) vout0x0; |
| 220 | c1[0] = (int8_t) vout1x0; |
| 221 | c2[0] = (int8_t) vout2x0; |
| 222 | } |
| 223 | |
| 224 | nc = 0; |
| 225 | } |
| 226 | } while (nc != 0); |
| 227 | } |