Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-4x4c16-minmax-fp32-wasmsdot.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx4c16-wasmdot.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 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 <wasm_simd128.h>
13
14 #include "xnnpack/gemm.h"
15 #include "xnnpack/math.h"
16
17
18 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_4x4c16__wasmsdot(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 4);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (4 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 16 * sizeof(int8_t));
44 int8_t* c0 = c;
45 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr < 2) {
47 c1 = c0;
48 }
49 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 c2 = c1;
52 }
53 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
54 if XNN_UNPREDICTABLE(mr != 4) {
55 c3 = c2;
56 }
57
58 const v128_t vmagic_bias = wasm_f32x4_const_splat(12582912.0f);
59 const int32_t output_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point;
60 const v128_t vmagic_min = wasm_i32x4_splat((int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point));
61 const v128_t vmagic_bias_less_output_zero_point = wasm_i32x4_splat(INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point);
62 const v128_t voutput_max = wasm_i8x16_splat(params->fp32_scalar.output_max);
63 XNN_FORCE_REALIZATION(vmagic_bias);
64 XNN_FORCE_REALIZATION(vmagic_min);
65 XNN_FORCE_REALIZATION(vmagic_bias_less_output_zero_point);
66 XNN_FORCE_REALIZATION(voutput_max);
67
68 do {
69 v128_t vacc0x0 = wasm_v128_load32_zero(w);
70 v128_t vacc0x1 = wasm_v128_load32_zero((const int32_t*) w + 1);
71 v128_t vacc0x2 = wasm_v128_load32_zero((const int32_t*) w + 2);
72 v128_t vacc0x3 = wasm_v128_load32_zero((const int32_t*) w + 3);
73 v128_t vacc1x0 = vacc0x0;
74 v128_t vacc1x1 = vacc0x1;
75 v128_t vacc1x2 = vacc0x2;
76 v128_t vacc1x3 = vacc0x3;
77 v128_t vacc2x0 = vacc0x0;
78 v128_t vacc2x1 = vacc0x1;
79 v128_t vacc2x2 = vacc0x2;
80 v128_t vacc2x3 = vacc0x3;
81 v128_t vacc3x0 = vacc0x0;
82 v128_t vacc3x1 = vacc0x1;
83 v128_t vacc3x2 = vacc0x2;
84 v128_t vacc3x3 = vacc0x3;
85 w = (const int32_t*) w + 4;
86
87 size_t p = ks;
88 do {
89 const int8_t* restrict a0 = a[0];
90 if XNN_UNPREDICTABLE(a0 != zero) {
91 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
92 }
93 const int8_t* restrict a1 = a[1];
94 if XNN_UNPREDICTABLE(a1 != zero) {
95 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
96 }
97 const int8_t* restrict a2 = a[2];
98 if XNN_UNPREDICTABLE(a2 != zero) {
99 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
100 }
101 const int8_t* restrict a3 = a[3];
102 if XNN_UNPREDICTABLE(a3 != zero) {
103 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
104 }
105 a += 4;
106
107 size_t k = kc;
108 do {
109 const v128_t va0 = wasm_v128_load(a0);
110 a0 += 16;
111 const v128_t va1 = wasm_v128_load(a1);
112 a1 += 16;
113 const v128_t va2 = wasm_v128_load(a2);
114 a2 += 16;
115 const v128_t va3 = wasm_v128_load(a3);
116 a3 += 16;
117
118 const v128_t vb0 = wasm_v128_load(w);
119 vacc0x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va0, vacc0x0);
120 vacc1x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va1, vacc1x0);
121 vacc2x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va2, vacc2x0);
122 vacc3x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va3, vacc3x0);
123 const v128_t vb1 = wasm_v128_load((const int8_t*) w + 16);
124 vacc0x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va0, vacc0x1);
125 vacc1x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va1, vacc1x1);
126 vacc2x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va2, vacc2x1);
127 vacc3x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va3, vacc3x1);
128 const v128_t vb2 = wasm_v128_load((const int8_t*) w + 32);
129 vacc0x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va0, vacc0x2);
130 vacc1x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va1, vacc1x2);
131 vacc2x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va2, vacc2x2);
132 vacc3x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va3, vacc3x2);
133 const v128_t vb3 = wasm_v128_load((const int8_t*) w + 48);
134 vacc0x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va0, vacc0x3);
135 vacc1x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va1, vacc1x3);
136 vacc2x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va2, vacc2x3);
137 vacc3x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va3, vacc3x3);
138
139 w = (const int8_t*) w + 64;
140 k -= 16 * sizeof(int8_t);
141 } while (k != 0);
142 p -= 4 * sizeof(void*);
143 } while (p != 0);
144
145 const v128_t vacc0x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x0, vacc0x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x0, vacc0x2, 2, 6, 3, 7));
146 const v128_t vacc0x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x1, vacc0x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x1, vacc0x3, 2, 6, 3, 7));
147 const v128_t vacc1x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x0, vacc1x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x0, vacc1x2, 2, 6, 3, 7));
148 const v128_t vacc1x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x1, vacc1x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x1, vacc1x3, 2, 6, 3, 7));
149 const v128_t vacc2x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x0, vacc2x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x0, vacc2x2, 2, 6, 3, 7));
150 const v128_t vacc2x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x1, vacc2x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x1, vacc2x3, 2, 6, 3, 7));
151 const v128_t vacc3x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x0, vacc3x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc3x0, vacc3x2, 2, 6, 3, 7));
152 const v128_t vacc3x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x1, vacc3x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc3x1, vacc3x3, 2, 6, 3, 7));
153
154 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
155 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
156 v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x02, vacc2x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x02, vacc2x13, 2, 6, 3, 7));
157 v128_t vacc3x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x02, vacc3x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc3x02, vacc3x13, 2, 6, 3, 7));
158
159 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
160 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
161 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
162 vacc3x0123 = wasm_f32x4_convert_i32x4(vacc3x0123);
163
164 const v128_t vscale0123 = wasm_v128_load(w);
165 w = (const float*) w + 4;
166 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale0123);
167 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale0123);
168 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vscale0123);
169 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vscale0123);
170
171 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
172 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
173 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vmagic_bias);
174 vacc3x0123 = wasm_f32x4_add(vacc3x0123, vmagic_bias);
175
176 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
177 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
178 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vmagic_min);
179 vacc3x0123 = wasm_i32x4_max(vacc3x0123, vmagic_min);
180
181 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
182 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
183 vacc2x0123 = wasm_i32x4_sub(vacc2x0123, vmagic_bias_less_output_zero_point);
184 vacc3x0123 = wasm_i32x4_sub(vacc3x0123, vmagic_bias_less_output_zero_point);
185
186 v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123);
187 v128_t vacc23x0123 = wasm_i16x8_narrow_i32x4(vacc2x0123, vacc3x0123);
188
189 v128_t vout = wasm_i8x16_narrow_i16x8(vacc01x0123, vacc23x0123);
190
191 vout = wasm_i8x16_min(vout, voutput_max);
192
193 if (nc >= 4) {
194 wasm_v128_store32_lane(c3, vout, 3);
195 wasm_v128_store32_lane(c2, vout, 2);
196 wasm_v128_store32_lane(c1, vout, 1);
197 wasm_v128_store32_lane(c0, vout, 0);
198
199 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
200 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
201 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
202 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
203
204 a = (const int8_t**restrict) ((uintptr_t) a - ks);
205
206 nc -= 4;
207 } else {
208 if (nc & 2) {
209 wasm_v128_store16_lane(c3, vout, 6);
210 c3 += 2;
211 wasm_v128_store16_lane(c2, vout, 4);
212 c2 += 2;
213 wasm_v128_store16_lane(c1, vout, 2);
214 c1 += 2;
215 wasm_v128_store16_lane(c0, vout, 0);
216 c0 += 2;
217
218 vout = wasm_u32x4_shr(vout, 16);
219 }
220 if (nc & 1) {
221 wasm_v128_store8_lane(c3, vout, 12);
222 wasm_v128_store8_lane(c2, vout, 8);
223 wasm_v128_store8_lane(c1, vout, 4);
224 wasm_v128_store8_lane(c0, vout, 0);
225 }
226
227 nc = 0;
228 }
229 } while (nc != 0);
230 }