Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-1x16c4-minmax-fp32-wasmusdot-u2-acc2.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx16c4-wasmdot.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2024 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_1x16c4__wasmusdot_u2_acc2(
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 <= 1);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (1 * 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, 4 * sizeof(int8_t));
44 int8_t* c0 = c;
45
46 const v128_t vmagic_bias = wasm_f32x4_const_splat(12582912.0f);
47 const int32_t output_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point;
48 const v128_t vmagic_min = wasm_i32x4_splat((int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point));
49 const v128_t vmagic_bias_less_output_zero_point = wasm_i32x4_splat(INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point);
50 const v128_t voutput_max = wasm_i8x16_splat(params->fp32_scalar.output_max);
51 //XNN_FORCE_REALIZATION(vmagic_bias);
52 //XNN_FORCE_REALIZATION(vmagic_min);
53 //XNN_FORCE_REALIZATION(vmagic_bias_less_output_zero_point);
54 //XNN_FORCE_REALIZATION(voutput_max);
55
56 const v128_t vsign_mask = wasm_u8x16_const_splat(UINT8_C(0x80));
57 XNN_FORCE_REALIZATION(vsign_mask);
58 do {
59 v128_t vacc0x0x0123 = wasm_v128_load(w);
60 v128_t vacc0x0x4567 = wasm_v128_load((const int32_t*) w + 4);
61 v128_t vacc0x0x89AB = wasm_v128_load((const int32_t*) w + 8);
62 v128_t vacc0x0xCDEF = wasm_v128_load((const int32_t*) w + 12);
63 v128_t vacc0x1x0123 = wasm_u32x4_const(0, 0, 0, 0);
64 v128_t vacc0x1x4567 = wasm_u32x4_const(0, 0, 0, 0);
65 v128_t vacc0x1x89AB = wasm_u32x4_const(0, 0, 0, 0);
66 v128_t vacc0x1xCDEF = wasm_u32x4_const(0, 0, 0, 0);
67 w = (const int32_t*) w + 16;
68
69 size_t p = ks;
70 do {
71 const int8_t* restrict a0 = a[0];
72 if XNN_UNPREDICTABLE(a0 != zero) {
73 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
74 }
75 a += 1;
76
77 size_t k = kc;
78 while (k >= 8 * sizeof(int8_t)) {
79 v128_t va0x0x0123 = wasm_v128_load32_splat(a0);
80 v128_t va0x1x0123 = wasm_v128_load32_splat(a0 + 4);
81 a0 += 8;
82
83 va0x0x0123 = wasm_v128_xor(va0x0x0123, vsign_mask);
84 va0x1x0123 = wasm_v128_xor(va0x1x0123, vsign_mask);
85
86 const v128_t vb0x0123 = wasm_v128_load((const int8_t*) w);
87 const v128_t vb0x4567 = wasm_v128_load((const int8_t*) w + 16);
88 const v128_t vb0x89AB = wasm_v128_load((const int8_t*) w + 32);
89 const v128_t vb0xCDEF = wasm_v128_load((const int8_t*) w + 48);
90 const v128_t vb1x0123 = wasm_v128_load((const int8_t*) w + 64);
91 const v128_t vb1x4567 = wasm_v128_load((const int8_t*) w + 80);
92 const v128_t vb1x89AB = wasm_v128_load((const int8_t*) w + 96);
93 const v128_t vb1xCDEF = wasm_v128_load((const int8_t*) w + 112);
94
95 vacc0x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x0123, va0x0x0123, vacc0x0x0123);
96 vacc0x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x4567, va0x0x0123, vacc0x0x4567);
97 vacc0x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x89AB, va0x0x0123, vacc0x0x89AB);
98 vacc0x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0xCDEF, va0x0x0123, vacc0x0xCDEF);
99 vacc0x1x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x0123, va0x1x0123, vacc0x1x0123);
100 vacc0x1x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x4567, va0x1x0123, vacc0x1x4567);
101 vacc0x1x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x89AB, va0x1x0123, vacc0x1x89AB);
102 vacc0x1xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1xCDEF, va0x1x0123, vacc0x1xCDEF);
103
104 w = (const int8_t*) w + 128;
105 k -= 8 * sizeof(int8_t);
106 }
107 vacc0x0x0123 = wasm_i32x4_add(vacc0x0x0123, vacc0x1x0123);
108 vacc0x0x4567 = wasm_i32x4_add(vacc0x0x4567, vacc0x1x4567);
109 vacc0x0x89AB = wasm_i32x4_add(vacc0x0x89AB, vacc0x1x89AB);
110 vacc0x0xCDEF = wasm_i32x4_add(vacc0x0xCDEF, vacc0x1xCDEF);
111
112 while (k != 0) {
113 v128_t va0x0123 = wasm_v128_load32_splat(a0);
114 a0 += 4;
115
116 va0x0123 = wasm_v128_xor(va0x0123, vsign_mask);
117
118 const v128_t vb0123 = wasm_v128_load((const int8_t*) w);
119 const v128_t vb4567 = wasm_v128_load((const int8_t*) w + 16);
120 const v128_t vb89AB = wasm_v128_load((const int8_t*) w + 32);
121 const v128_t vbCDEF = wasm_v128_load((const int8_t*) w + 48);
122
123 vacc0x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va0x0123, vacc0x0x0123);
124 vacc0x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va0x0123, vacc0x0x4567);
125 vacc0x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va0x0123, vacc0x0x89AB);
126 vacc0x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va0x0123, vacc0x0xCDEF);
127
128 w = (const int8_t*) w + 64;
129 k -= 4 * sizeof(int8_t);
130 }
131
132 p -= 1 * sizeof(void*);
133 } while (p != 0);
134
135 vacc0x0x0123 = wasm_f32x4_convert_i32x4(vacc0x0x0123);
136 vacc0x0x4567 = wasm_f32x4_convert_i32x4(vacc0x0x4567);
137 vacc0x0x89AB = wasm_f32x4_convert_i32x4(vacc0x0x89AB);
138 vacc0x0xCDEF = wasm_f32x4_convert_i32x4(vacc0x0xCDEF);
139
140 const v128_t vscale0123 = wasm_v128_load(w);
141 w = (const float*) w + 4;
142 const v128_t vscale4567 = wasm_v128_load(w);
143 w = (const float*) w + 4;
144 const v128_t vscale89AB = wasm_v128_load(w);
145 w = (const float*) w + 4;
146 const v128_t vscaleCDEF = wasm_v128_load(w);
147 w = (const float*) w + 4;
148 vacc0x0x0123 = wasm_f32x4_mul(vacc0x0x0123, vscale0123);
149 vacc0x0x4567 = wasm_f32x4_mul(vacc0x0x4567, vscale4567);
150 vacc0x0x89AB = wasm_f32x4_mul(vacc0x0x89AB, vscale89AB);
151 vacc0x0xCDEF = wasm_f32x4_mul(vacc0x0xCDEF, vscaleCDEF);
152
153 vacc0x0x0123 = wasm_f32x4_add(vacc0x0x0123, vmagic_bias);
154 vacc0x0x4567 = wasm_f32x4_add(vacc0x0x4567, vmagic_bias);
155 vacc0x0x89AB = wasm_f32x4_add(vacc0x0x89AB, vmagic_bias);
156 vacc0x0xCDEF = wasm_f32x4_add(vacc0x0xCDEF, vmagic_bias);
157
158 vacc0x0x0123 = wasm_i32x4_max(vacc0x0x0123, vmagic_min);
159 vacc0x0x4567 = wasm_i32x4_max(vacc0x0x4567, vmagic_min);
160 vacc0x0x89AB = wasm_i32x4_max(vacc0x0x89AB, vmagic_min);
161 vacc0x0xCDEF = wasm_i32x4_max(vacc0x0xCDEF, vmagic_min);
162
163 vacc0x0x0123 = wasm_i32x4_sub(vacc0x0x0123, vmagic_bias_less_output_zero_point);
164 vacc0x0x4567 = wasm_i32x4_sub(vacc0x0x4567, vmagic_bias_less_output_zero_point);
165 vacc0x0x89AB = wasm_i32x4_sub(vacc0x0x89AB, vmagic_bias_less_output_zero_point);
166 vacc0x0xCDEF = wasm_i32x4_sub(vacc0x0xCDEF, vmagic_bias_less_output_zero_point);
167
168 v128_t vacc0x01234567 = wasm_i16x8_narrow_i32x4(vacc0x0x0123, vacc0x0x4567);
169 v128_t vacc0x89ABCDEF = wasm_i16x8_narrow_i32x4(vacc0x0x89AB, vacc0x0xCDEF);
170
171 vacc0x01234567 = wasm_i8x16_narrow_i16x8(vacc0x01234567, vacc0x01234567);
172 vacc0x89ABCDEF = wasm_i8x16_narrow_i16x8(vacc0x89ABCDEF, vacc0x89ABCDEF);
173
174 vacc0x01234567 = wasm_i8x16_min(vacc0x01234567, voutput_max);
175 vacc0x89ABCDEF = wasm_i8x16_min(vacc0x89ABCDEF, voutput_max);
176
177 if (nc >= 16) {
178 wasm_v128_store64_lane(c0, vacc0x01234567, 0);
179 wasm_v128_store64_lane(c0 + 8, vacc0x89ABCDEF, 0);
180
181 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
182
183 a = (const int8_t**restrict) ((uintptr_t) a - ks);
184
185 nc -= 16;
186 } else {
187 if (nc & 8) {
188 wasm_v128_store64_lane(c0, vacc0x01234567, 0);
189 c0 += 8;
190
191 vacc0x01234567 = vacc0x89ABCDEF;
192 }
193 if (nc & 4) {
194 wasm_v128_store32_lane(c0, vacc0x01234567, 0);
195 c0 += 4;
196
197 vacc0x01234567 = wasm_u64x2_shr(vacc0x01234567, 32);
198 }
199 if (nc & 2) {
200 wasm_v128_store16_lane(c0, vacc0x01234567, 0);
201 c0 += 2;
202
203 vacc0x01234567 = wasm_u32x4_shr(vacc0x01234567, 16);
204 }
205 if (nc & 1) {
206 wasm_v128_store8_lane(c0, vacc0x01234567, 0);
207 }
208
209 nc = 0;
210 }
211 } while (nc != 0);
212 }