Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-2x8c16-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_2x8c16__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 <= 2);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (2 * 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
50 const v128_t vmagic_bias = wasm_f32x4_const_splat(12582912.0f);
51 const int32_t output_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point;
52 const v128_t vmagic_min = wasm_i32x4_splat((int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point));
53 const v128_t vmagic_bias_less_output_zero_point = wasm_i32x4_splat(INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point);
54 const v128_t voutput_max = wasm_i8x16_splat(params->fp32_scalar.output_max);
55 XNN_FORCE_REALIZATION(vmagic_bias);
56 XNN_FORCE_REALIZATION(vmagic_min);
57 XNN_FORCE_REALIZATION(vmagic_bias_less_output_zero_point);
58 XNN_FORCE_REALIZATION(voutput_max);
59
60 do {
61 v128_t vacc0x0 = wasm_v128_load32_zero(w);
62 v128_t vacc0x1 = wasm_v128_load32_zero((const int32_t*) w + 1);
63 v128_t vacc0x2 = wasm_v128_load32_zero((const int32_t*) w + 2);
64 v128_t vacc0x3 = wasm_v128_load32_zero((const int32_t*) w + 3);
65 v128_t vacc0x4 = wasm_v128_load32_zero((const int32_t*) w + 4);
66 v128_t vacc0x5 = wasm_v128_load32_zero((const int32_t*) w + 5);
67 v128_t vacc0x6 = wasm_v128_load32_zero((const int32_t*) w + 6);
68 v128_t vacc0x7 = wasm_v128_load32_zero((const int32_t*) w + 7);
69 v128_t vacc1x0 = vacc0x0;
70 v128_t vacc1x1 = vacc0x1;
71 v128_t vacc1x2 = vacc0x2;
72 v128_t vacc1x3 = vacc0x3;
73 v128_t vacc1x4 = vacc0x4;
74 v128_t vacc1x5 = vacc0x5;
75 v128_t vacc1x6 = vacc0x6;
76 v128_t vacc1x7 = vacc0x7;
77 w = (const int32_t*) w + 8;
78
79 size_t p = ks;
80 do {
81 const int8_t* restrict a0 = a[0];
82 if XNN_UNPREDICTABLE(a0 != zero) {
83 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
84 }
85 const int8_t* restrict a1 = a[1];
86 if XNN_UNPREDICTABLE(a1 != zero) {
87 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
88 }
89 a += 2;
90
91 size_t k = kc;
92 do {
93 const v128_t va0 = wasm_v128_load(a0);
94 a0 += 16;
95 const v128_t va1 = wasm_v128_load(a1);
96 a1 += 16;
97
98 const v128_t vb0 = wasm_v128_load(w);
99 vacc0x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va0, vacc0x0);
100 vacc1x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va1, vacc1x0);
101 const v128_t vb1 = wasm_v128_load((const int8_t*) w + 16);
102 vacc0x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va0, vacc0x1);
103 vacc1x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va1, vacc1x1);
104 const v128_t vb2 = wasm_v128_load((const int8_t*) w + 32);
105 vacc0x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va0, vacc0x2);
106 vacc1x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va1, vacc1x2);
107 const v128_t vb3 = wasm_v128_load((const int8_t*) w + 48);
108 vacc0x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va0, vacc0x3);
109 vacc1x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va1, vacc1x3);
110 const v128_t vb4 = wasm_v128_load((const int8_t*) w + 64);
111 vacc0x4 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4, va0, vacc0x4);
112 vacc1x4 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4, va1, vacc1x4);
113 const v128_t vb5 = wasm_v128_load((const int8_t*) w + 80);
114 vacc0x5 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb5, va0, vacc0x5);
115 vacc1x5 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb5, va1, vacc1x5);
116 const v128_t vb6 = wasm_v128_load((const int8_t*) w + 96);
117 vacc0x6 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb6, va0, vacc0x6);
118 vacc1x6 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb6, va1, vacc1x6);
119 const v128_t vb7 = wasm_v128_load((const int8_t*) w + 112);
120 vacc0x7 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb7, va0, vacc0x7);
121 vacc1x7 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb7, va1, vacc1x7);
122
123 w = (const int8_t*) w + 128;
124 k -= 16 * sizeof(int8_t);
125 } while (k != 0);
126 p -= 2 * sizeof(void*);
127 } while (p != 0);
128
129 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));
130 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));
131 const v128_t vacc0x46 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x4, vacc0x6, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x4, vacc0x6, 2, 6, 3, 7));
132 const v128_t vacc0x57 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x5, vacc0x7, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x5, vacc0x7, 2, 6, 3, 7));
133 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));
134 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));
135 const v128_t vacc1x46 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x4, vacc1x6, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x4, vacc1x6, 2, 6, 3, 7));
136 const v128_t vacc1x57 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x5, vacc1x7, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x5, vacc1x7, 2, 6, 3, 7));
137
138 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
139 v128_t vacc0x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x46, vacc0x57, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x46, vacc0x57, 2, 6, 3, 7));
140 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
141 v128_t vacc1x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x46, vacc1x57, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x46, vacc1x57, 2, 6, 3, 7));
142
143 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
144 vacc0x4567 = wasm_f32x4_convert_i32x4(vacc0x4567);
145 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
146 vacc1x4567 = wasm_f32x4_convert_i32x4(vacc1x4567);
147
148 const v128_t vscale0123 = wasm_v128_load(w);
149 w = (const float*) w + 4;
150 const v128_t vscale4567 = wasm_v128_load(w);
151 w = (const float*) w + 4;
152 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale0123);
153 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vscale4567);
154 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale0123);
155 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vscale4567);
156
157 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
158 vacc0x4567 = wasm_f32x4_add(vacc0x4567, vmagic_bias);
159 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
160 vacc1x4567 = wasm_f32x4_add(vacc1x4567, vmagic_bias);
161
162 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
163 vacc0x4567 = wasm_i32x4_max(vacc0x4567, vmagic_min);
164 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
165 vacc1x4567 = wasm_i32x4_max(vacc1x4567, vmagic_min);
166
167 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
168 vacc0x4567 = wasm_i32x4_sub(vacc0x4567, vmagic_bias_less_output_zero_point);
169 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
170 vacc1x4567 = wasm_i32x4_sub(vacc1x4567, vmagic_bias_less_output_zero_point);
171
172 v128_t vacc0x01234567 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc0x4567);
173 v128_t vacc1x01234567 = wasm_i16x8_narrow_i32x4(vacc1x0123, vacc1x4567);
174
175 v128_t vout0x01234567_1x01234567 = wasm_i8x16_narrow_i16x8(vacc0x01234567, vacc1x01234567);
176
177 vout0x01234567_1x01234567 = wasm_i8x16_min(vout0x01234567_1x01234567, voutput_max);
178
179 if (nc >= 8) {
180 wasm_v128_store64_lane(c1, vout0x01234567_1x01234567, 1);
181 wasm_v128_store64_lane(c0, vout0x01234567_1x01234567, 0);
182
183 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
184 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
185
186 a = (const int8_t**restrict) ((uintptr_t) a - ks);
187
188 nc -= 8;
189 } else {
190 if (nc & 4) {
191 wasm_v128_store32_lane(c1, vout0x01234567_1x01234567, 2);
192 c1 += 4;
193 wasm_v128_store32_lane(c0, vout0x01234567_1x01234567, 0);
194 c0 += 4;
195
196 vout0x01234567_1x01234567 = wasm_u64x2_shr(vout0x01234567_1x01234567, 32);
197 }
198 if (nc & 2) {
199 wasm_v128_store16_lane(c1, vout0x01234567_1x01234567, 4);
200 c1 += 2;
201 wasm_v128_store16_lane(c0, vout0x01234567_1x01234567, 0);
202 c0 += 2;
203
204 vout0x01234567_1x01234567 = wasm_u32x4_shr(vout0x01234567_1x01234567, 16);
205 }
206 if (nc & 1) {
207 wasm_v128_store8_lane(c1, vout0x01234567_1x01234567, 8);
208 wasm_v128_store8_lane(c0, vout0x01234567_1x01234567, 0);
209 }
210
211 nc = 0;
212 }
213 } while (nc != 0);
214 }