Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-2x8c16-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_2x8c16__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 <= 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 const v128_t vsign_mask = wasm_u8x16_const_splat(UINT8_C(0x80));
61 XNN_FORCE_REALIZATION(vsign_mask);
62 do {
63 v128_t vacc0x0 = wasm_v128_load32_zero(w);
64 v128_t vacc0x1 = wasm_v128_load32_zero((const int32_t*) w + 1);
65 v128_t vacc0x2 = wasm_v128_load32_zero((const int32_t*) w + 2);
66 v128_t vacc0x3 = wasm_v128_load32_zero((const int32_t*) w + 3);
67 v128_t vacc0x4 = wasm_v128_load32_zero((const int32_t*) w + 4);
68 v128_t vacc0x5 = wasm_v128_load32_zero((const int32_t*) w + 5);
69 v128_t vacc0x6 = wasm_v128_load32_zero((const int32_t*) w + 6);
70 v128_t vacc0x7 = wasm_v128_load32_zero((const int32_t*) w + 7);
71 v128_t vacc1x0 = vacc0x0;
72 v128_t vacc1x1 = vacc0x1;
73 v128_t vacc1x2 = vacc0x2;
74 v128_t vacc1x3 = vacc0x3;
75 v128_t vacc1x4 = vacc0x4;
76 v128_t vacc1x5 = vacc0x5;
77 v128_t vacc1x6 = vacc0x6;
78 v128_t vacc1x7 = vacc0x7;
79 w = (const int32_t*) w + 8;
80
81 size_t p = ks;
82 do {
83 const int8_t* restrict a0 = a[0];
84 if XNN_UNPREDICTABLE(a0 != zero) {
85 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
86 }
87 const int8_t* restrict a1 = a[1];
88 if XNN_UNPREDICTABLE(a1 != zero) {
89 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
90 }
91 a += 2;
92
93 size_t k = kc;
94 do {
95 const v128_t va0 = wasm_v128_xor(wasm_v128_load(a0), vsign_mask);
96 a0 += 16;
97 const v128_t va1 = wasm_v128_xor(wasm_v128_load(a1), vsign_mask);
98 a1 += 16;
99
100 const v128_t vb0 = wasm_v128_load(w);
101 vacc0x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va0, vacc0x0);
102 vacc1x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va1, vacc1x0);
103 const v128_t vb1 = wasm_v128_load((const int8_t*) w + 16);
104 vacc0x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va0, vacc0x1);
105 vacc1x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va1, vacc1x1);
106 const v128_t vb2 = wasm_v128_load((const int8_t*) w + 32);
107 vacc0x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va0, vacc0x2);
108 vacc1x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va1, vacc1x2);
109 const v128_t vb3 = wasm_v128_load((const int8_t*) w + 48);
110 vacc0x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va0, vacc0x3);
111 vacc1x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va1, vacc1x3);
112 const v128_t vb4 = wasm_v128_load((const int8_t*) w + 64);
113 vacc0x4 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4, va0, vacc0x4);
114 vacc1x4 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4, va1, vacc1x4);
115 const v128_t vb5 = wasm_v128_load((const int8_t*) w + 80);
116 vacc0x5 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb5, va0, vacc0x5);
117 vacc1x5 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb5, va1, vacc1x5);
118 const v128_t vb6 = wasm_v128_load((const int8_t*) w + 96);
119 vacc0x6 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb6, va0, vacc0x6);
120 vacc1x6 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb6, va1, vacc1x6);
121 const v128_t vb7 = wasm_v128_load((const int8_t*) w + 112);
122 vacc0x7 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb7, va0, vacc0x7);
123 vacc1x7 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb7, va1, vacc1x7);
124
125 w = (const int8_t*) w + 128;
126 k -= 16 * sizeof(int8_t);
127 } while (k != 0);
128 p -= 2 * sizeof(void*);
129 } while (p != 0);
130
131 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));
132 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));
133 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));
134 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));
135 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));
136 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));
137 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));
138 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));
139
140 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
141 v128_t vacc0x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x46, vacc0x57, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x46, vacc0x57, 2, 6, 3, 7));
142 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
143 v128_t vacc1x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x46, vacc1x57, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x46, vacc1x57, 2, 6, 3, 7));
144
145 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
146 vacc0x4567 = wasm_f32x4_convert_i32x4(vacc0x4567);
147 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
148 vacc1x4567 = wasm_f32x4_convert_i32x4(vacc1x4567);
149
150 const v128_t vscale0123 = wasm_v128_load(w);
151 w = (const float*) w + 4;
152 const v128_t vscale4567 = wasm_v128_load(w);
153 w = (const float*) w + 4;
154 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale0123);
155 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vscale4567);
156 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale0123);
157 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vscale4567);
158
159 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
160 vacc0x4567 = wasm_f32x4_add(vacc0x4567, vmagic_bias);
161 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
162 vacc1x4567 = wasm_f32x4_add(vacc1x4567, vmagic_bias);
163
164 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
165 vacc0x4567 = wasm_i32x4_max(vacc0x4567, vmagic_min);
166 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
167 vacc1x4567 = wasm_i32x4_max(vacc1x4567, vmagic_min);
168
169 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
170 vacc0x4567 = wasm_i32x4_sub(vacc0x4567, vmagic_bias_less_output_zero_point);
171 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
172 vacc1x4567 = wasm_i32x4_sub(vacc1x4567, vmagic_bias_less_output_zero_point);
173
174 v128_t vacc0x01234567 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc0x4567);
175 v128_t vacc1x01234567 = wasm_i16x8_narrow_i32x4(vacc1x0123, vacc1x4567);
176
177 v128_t vout0x01234567_1x01234567 = wasm_i8x16_narrow_i16x8(vacc0x01234567, vacc1x01234567);
178
179 vout0x01234567_1x01234567 = wasm_i8x16_min(vout0x01234567_1x01234567, voutput_max);
180
181 if (nc >= 8) {
182 wasm_v128_store64_lane(c1, vout0x01234567_1x01234567, 1);
183 wasm_v128_store64_lane(c0, vout0x01234567_1x01234567, 0);
184
185 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
186 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
187
188 a = (const int8_t**restrict) ((uintptr_t) a - ks);
189
190 nc -= 8;
191 } else {
192 if (nc & 4) {
193 wasm_v128_store32_lane(c1, vout0x01234567_1x01234567, 2);
194 c1 += 4;
195 wasm_v128_store32_lane(c0, vout0x01234567_1x01234567, 0);
196 c0 += 4;
197
198 vout0x01234567_1x01234567 = wasm_u64x2_shr(vout0x01234567_1x01234567, 32);
199 }
200 if (nc & 2) {
201 wasm_v128_store16_lane(c1, vout0x01234567_1x01234567, 4);
202 c1 += 2;
203 wasm_v128_store16_lane(c0, vout0x01234567_1x01234567, 0);
204 c0 += 2;
205
206 vout0x01234567_1x01234567 = wasm_u32x4_shr(vout0x01234567_1x01234567, 16);
207 }
208 if (nc & 1) {
209 wasm_v128_store8_lane(c1, vout0x01234567_1x01234567, 8);
210 wasm_v128_store8_lane(c0, vout0x01234567_1x01234567, 0);
211 }
212
213 nc = 0;
214 }
215 } while (nc != 0);
216 }