1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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 <xnnpack/math.h>
13 #include <xnnpack/gemm.h>
14 #include <xnnpack/unaligned.h>
15
16
xnn_qu8_igemm_minmax_fp32_ukernel_4x2__wasm_fmagic(size_t mr,size_t nc,size_t kc,size_t ks,const uint8_t ** restrict a,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qu8_igemm_minmax_fp32_ukernel_4x2__wasm_fmagic(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 size_t ks,
22 const uint8_t**restrict a,
23 const void*restrict w,
24 uint8_t*restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 size_t a_offset,
28 const uint8_t* zero,
29 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
30 {
31 assert(mr != 0);
32 assert(mr <= 4);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(ks != 0);
36 assert(ks % (4 * sizeof(void*)) == 0);
37 assert(a != NULL);
38 assert(w != NULL);
39 assert(c != NULL);
40
41 uint8_t* c0 = c;
42 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
43 if XNN_UNPREDICTABLE(mr < 2) {
44 c1 = c0;
45 }
46 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
47 if XNN_UNPREDICTABLE(mr <= 2) {
48 c2 = c1;
49 }
50 uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
51 if XNN_UNPREDICTABLE(mr != 4) {
52 c3 = c2;
53 }
54
55 const int32_t vb_zero_point = params->fp32_scalar_fmagic.kernel_zero_point;
56 do {
57 int32_t vacc0x0 = unaligned_indexed_load_s32(w, 0);
58 int32_t vacc0x1 = unaligned_indexed_load_s32(w, 1);
59 int32_t vacc1x0 = vacc0x0;
60 int32_t vacc1x1 = vacc0x1;
61 int32_t vacc2x0 = vacc0x0;
62 int32_t vacc2x1 = vacc0x1;
63 int32_t vacc3x0 = vacc0x0;
64 int32_t vacc3x1 = vacc0x1;
65 w = (const void*) ((const int32_t*) w + 2);
66
67 size_t p = ks;
68 do {
69 const uint8_t* restrict a0 = a[0];
70 assert(a0 != NULL);
71 if XNN_UNPREDICTABLE(a0 != zero) {
72 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
73 }
74 const uint8_t* restrict a1 = a[1];
75 assert(a1 != NULL);
76 if XNN_UNPREDICTABLE(a1 != zero) {
77 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
78 }
79 const uint8_t* restrict a2 = a[2];
80 assert(a2 != NULL);
81 if XNN_UNPREDICTABLE(a2 != zero) {
82 a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
83 }
84 const uint8_t* restrict a3 = a[3];
85 assert(a3 != NULL);
86 if XNN_UNPREDICTABLE(a3 != zero) {
87 a3 = (const uint8_t*) ((uintptr_t) a3 + a_offset);
88 }
89 a += 4;
90
91 size_t k = kc;
92 do {
93 const int32_t va0 = (int32_t) (uint32_t) *a0++;
94 const int32_t va1 = (int32_t) (uint32_t) *a1++;
95 const int32_t va2 = (int32_t) (uint32_t) *a2++;
96 const int32_t va3 = (int32_t) (uint32_t) *a3++;
97
98 const int32_t vb0 = (int32_t) (uint32_t) ((const uint8_t*) w)[0] - vb_zero_point;
99 const int32_t vb1 = (int32_t) (uint32_t) ((const uint8_t*) w)[1] - vb_zero_point;
100 w = (const void*) ((const uint8_t*) w + 2);
101
102 vacc0x0 += va0 * vb0;
103 vacc0x1 += va0 * vb1;
104 vacc1x0 += va1 * vb0;
105 vacc1x1 += va1 * vb1;
106 vacc2x0 += va2 * vb0;
107 vacc2x1 += va2 * vb1;
108 vacc3x0 += va3 * vb0;
109 vacc3x1 += va3 * vb1;
110
111 k -= sizeof(uint8_t);
112 } while (k != 0);
113 p -= 4 * sizeof(void*);
114 } while (p != 0);
115
116 float vfpacc0x0 = (float) vacc0x0;
117 float vfpacc0x1 = (float) vacc0x1;
118 float vfpacc1x0 = (float) vacc1x0;
119 float vfpacc1x1 = (float) vacc1x1;
120 float vfpacc2x0 = (float) vacc2x0;
121 float vfpacc2x1 = (float) vacc2x1;
122 float vfpacc3x0 = (float) vacc3x0;
123 float vfpacc3x1 = (float) vacc3x1;
124
125 const float vscale = params->fp32_scalar_fmagic.scale;
126 vfpacc0x0 *= vscale;
127 vfpacc0x1 *= vscale;
128 vfpacc1x0 *= vscale;
129 vfpacc1x1 *= vscale;
130 vfpacc2x0 *= vscale;
131 vfpacc2x1 *= vscale;
132 vfpacc3x0 *= vscale;
133 vfpacc3x1 *= vscale;
134
135 const float voutput_min_less_zero_point = params->fp32_scalar_fmagic.output_min_less_zero_point;
136 vfpacc0x0 = __builtin_wasm_max_f32(vfpacc0x0, voutput_min_less_zero_point);
137 vfpacc0x1 = __builtin_wasm_max_f32(vfpacc0x1, voutput_min_less_zero_point);
138 vfpacc1x0 = __builtin_wasm_max_f32(vfpacc1x0, voutput_min_less_zero_point);
139 vfpacc1x1 = __builtin_wasm_max_f32(vfpacc1x1, voutput_min_less_zero_point);
140 vfpacc2x0 = __builtin_wasm_max_f32(vfpacc2x0, voutput_min_less_zero_point);
141 vfpacc2x1 = __builtin_wasm_max_f32(vfpacc2x1, voutput_min_less_zero_point);
142 vfpacc3x0 = __builtin_wasm_max_f32(vfpacc3x0, voutput_min_less_zero_point);
143 vfpacc3x1 = __builtin_wasm_max_f32(vfpacc3x1, voutput_min_less_zero_point);
144
145 const float voutput_max_less_zero_point = params->fp32_scalar_fmagic.output_max_less_zero_point;
146 vfpacc0x0 = __builtin_wasm_min_f32(vfpacc0x0, voutput_max_less_zero_point);
147 vfpacc0x1 = __builtin_wasm_min_f32(vfpacc0x1, voutput_max_less_zero_point);
148 vfpacc1x0 = __builtin_wasm_min_f32(vfpacc1x0, voutput_max_less_zero_point);
149 vfpacc1x1 = __builtin_wasm_min_f32(vfpacc1x1, voutput_max_less_zero_point);
150 vfpacc2x0 = __builtin_wasm_min_f32(vfpacc2x0, voutput_max_less_zero_point);
151 vfpacc2x1 = __builtin_wasm_min_f32(vfpacc2x1, voutput_max_less_zero_point);
152 vfpacc3x0 = __builtin_wasm_min_f32(vfpacc3x0, voutput_max_less_zero_point);
153 vfpacc3x1 = __builtin_wasm_min_f32(vfpacc3x1, voutput_max_less_zero_point);
154
155 const float vmagic_bias = params->fp32_scalar_fmagic.magic_bias;
156 vfpacc0x0 += vmagic_bias;
157 vfpacc0x1 += vmagic_bias;
158 vfpacc1x0 += vmagic_bias;
159 vfpacc1x1 += vmagic_bias;
160 vfpacc2x0 += vmagic_bias;
161 vfpacc2x1 += vmagic_bias;
162 vfpacc3x0 += vmagic_bias;
163 vfpacc3x1 += vmagic_bias;
164
165 const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar_fmagic.magic_bias_less_output_zero_point;
166 int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0) - vmagic_bias_less_output_zero_point;
167 int32_t vout0x1 = (int32_t) float_as_uint32(vfpacc0x1) - vmagic_bias_less_output_zero_point;
168 int32_t vout1x0 = (int32_t) float_as_uint32(vfpacc1x0) - vmagic_bias_less_output_zero_point;
169 int32_t vout1x1 = (int32_t) float_as_uint32(vfpacc1x1) - vmagic_bias_less_output_zero_point;
170 int32_t vout2x0 = (int32_t) float_as_uint32(vfpacc2x0) - vmagic_bias_less_output_zero_point;
171 int32_t vout2x1 = (int32_t) float_as_uint32(vfpacc2x1) - vmagic_bias_less_output_zero_point;
172 int32_t vout3x0 = (int32_t) float_as_uint32(vfpacc3x0) - vmagic_bias_less_output_zero_point;
173 int32_t vout3x1 = (int32_t) float_as_uint32(vfpacc3x1) - vmagic_bias_less_output_zero_point;
174
175 if XNN_LIKELY(nc >= 2) {
176 c3[0] = (uint8_t) vout3x0;
177 c3[1] = (uint8_t) vout3x1;
178 c2[0] = (uint8_t) vout2x0;
179 c2[1] = (uint8_t) vout2x1;
180 c1[0] = (uint8_t) vout1x0;
181 c1[1] = (uint8_t) vout1x1;
182 c0[0] = (uint8_t) vout0x0;
183 c0[1] = (uint8_t) vout0x1;
184
185 c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
186 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
187 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
188 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
189
190 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
191 nc -= 2;
192 } else {
193 if (nc & 1) {
194 c3[0] = (uint8_t) vout3x0;
195 c2[0] = (uint8_t) vout2x0;
196 c1[0] = (uint8_t) vout1x0;
197 c0[0] = (uint8_t) vout0x0;
198 }
199
200 nc = 0;
201 }
202 } while (nc != 0);
203 }
204