Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-4x4c16-minmax-fp32-wasmusdot.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__wasmusdot(
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 const v128_t vsign_mask = wasm_u8x16_const_splat(UINT8_C(0x80));
69 XNN_FORCE_REALIZATION(vsign_mask);
70 do {
71 v128_t vacc0x0 = wasm_v128_load32_zero(w);
72 v128_t vacc0x1 = wasm_v128_load32_zero((const int32_t*) w + 1);
73 v128_t vacc0x2 = wasm_v128_load32_zero((const int32_t*) w + 2);
74 v128_t vacc0x3 = wasm_v128_load32_zero((const int32_t*) w + 3);
75 v128_t vacc1x0 = vacc0x0;
76 v128_t vacc1x1 = vacc0x1;
77 v128_t vacc1x2 = vacc0x2;
78 v128_t vacc1x3 = vacc0x3;
79 v128_t vacc2x0 = vacc0x0;
80 v128_t vacc2x1 = vacc0x1;
81 v128_t vacc2x2 = vacc0x2;
82 v128_t vacc2x3 = vacc0x3;
83 v128_t vacc3x0 = vacc0x0;
84 v128_t vacc3x1 = vacc0x1;
85 v128_t vacc3x2 = vacc0x2;
86 v128_t vacc3x3 = vacc0x3;
87 w = (const int32_t*) w + 4;
88
89 size_t p = ks;
90 do {
91 const int8_t* restrict a0 = a[0];
92 if XNN_UNPREDICTABLE(a0 != zero) {
93 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
94 }
95 const int8_t* restrict a1 = a[1];
96 if XNN_UNPREDICTABLE(a1 != zero) {
97 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
98 }
99 const int8_t* restrict a2 = a[2];
100 if XNN_UNPREDICTABLE(a2 != zero) {
101 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
102 }
103 const int8_t* restrict a3 = a[3];
104 if XNN_UNPREDICTABLE(a3 != zero) {
105 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
106 }
107 a += 4;
108
109 size_t k = kc;
110 do {
111 const v128_t va0 = wasm_v128_xor(wasm_v128_load(a0), vsign_mask);
112 a0 += 16;
113 const v128_t va1 = wasm_v128_xor(wasm_v128_load(a1), vsign_mask);
114 a1 += 16;
115 const v128_t va2 = wasm_v128_xor(wasm_v128_load(a2), vsign_mask);
116 a2 += 16;
117 const v128_t va3 = wasm_v128_xor(wasm_v128_load(a3), vsign_mask);
118 a3 += 16;
119
120 const v128_t vb0 = wasm_v128_load(w);
121 vacc0x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va0, vacc0x0);
122 vacc1x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va1, vacc1x0);
123 vacc2x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va2, vacc2x0);
124 vacc3x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va3, vacc3x0);
125 const v128_t vb1 = wasm_v128_load((const int8_t*) w + 16);
126 vacc0x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va0, vacc0x1);
127 vacc1x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va1, vacc1x1);
128 vacc2x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va2, vacc2x1);
129 vacc3x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va3, vacc3x1);
130 const v128_t vb2 = wasm_v128_load((const int8_t*) w + 32);
131 vacc0x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va0, vacc0x2);
132 vacc1x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va1, vacc1x2);
133 vacc2x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va2, vacc2x2);
134 vacc3x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va3, vacc3x2);
135 const v128_t vb3 = wasm_v128_load((const int8_t*) w + 48);
136 vacc0x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va0, vacc0x3);
137 vacc1x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va1, vacc1x3);
138 vacc2x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va2, vacc2x3);
139 vacc3x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va3, vacc3x3);
140
141 w = (const int8_t*) w + 64;
142 k -= 16 * sizeof(int8_t);
143 } while (k != 0);
144 p -= 4 * sizeof(void*);
145 } while (p != 0);
146
147 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));
148 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));
149 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));
150 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));
151 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));
152 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));
153 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));
154 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));
155
156 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
157 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
158 v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x02, vacc2x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x02, vacc2x13, 2, 6, 3, 7));
159 v128_t vacc3x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x02, vacc3x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc3x02, vacc3x13, 2, 6, 3, 7));
160
161 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
162 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
163 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
164 vacc3x0123 = wasm_f32x4_convert_i32x4(vacc3x0123);
165
166 const v128_t vscale0123 = wasm_v128_load(w);
167 w = (const float*) w + 4;
168 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale0123);
169 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale0123);
170 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vscale0123);
171 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vscale0123);
172
173 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
174 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
175 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vmagic_bias);
176 vacc3x0123 = wasm_f32x4_add(vacc3x0123, vmagic_bias);
177
178 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
179 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
180 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vmagic_min);
181 vacc3x0123 = wasm_i32x4_max(vacc3x0123, vmagic_min);
182
183 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
184 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
185 vacc2x0123 = wasm_i32x4_sub(vacc2x0123, vmagic_bias_less_output_zero_point);
186 vacc3x0123 = wasm_i32x4_sub(vacc3x0123, vmagic_bias_less_output_zero_point);
187
188 v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123);
189 v128_t vacc23x0123 = wasm_i16x8_narrow_i32x4(vacc2x0123, vacc3x0123);
190
191 v128_t vout = wasm_i8x16_narrow_i16x8(vacc01x0123, vacc23x0123);
192
193 vout = wasm_i8x16_min(vout, voutput_max);
194
195 if (nc >= 4) {
196 wasm_v128_store32_lane(c3, vout, 3);
197 wasm_v128_store32_lane(c2, vout, 2);
198 wasm_v128_store32_lane(c1, vout, 1);
199 wasm_v128_store32_lane(c0, vout, 0);
200
201 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
202 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
203 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
204 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
205
206 a = (const int8_t**restrict) ((uintptr_t) a - ks);
207
208 nc -= 4;
209 } else {
210 if (nc & 2) {
211 wasm_v128_store16_lane(c3, vout, 6);
212 c3 += 2;
213 wasm_v128_store16_lane(c2, vout, 4);
214 c2 += 2;
215 wasm_v128_store16_lane(c1, vout, 2);
216 c1 += 2;
217 wasm_v128_store16_lane(c0, vout, 0);
218 c0 += 2;
219
220 vout = wasm_u32x4_shr(vout, 16);
221 }
222 if (nc & 1) {
223 wasm_v128_store8_lane(c3, vout, 12);
224 wasm_v128_store8_lane(c2, vout, 8);
225 wasm_v128_store8_lane(c1, vout, 4);
226 wasm_v128_store8_lane(c0, vout, 0);
227 }
228
229 nc = 0;
230 }
231 } while (nc != 0);
232 }