Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-2x8c8-minmax-fp32-wasmusdot-u2.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c8-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_2x8c8__wasmusdot_u2(
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, 8 * 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 vacc0x01 = wasm_u64x2_load32x2(w); w = (const int32_t*) w + 2;
64 v128_t vacc0x23 = wasm_u64x2_load32x2(w); w = (const int32_t*) w + 2;
65 v128_t vacc0x45 = wasm_u64x2_load32x2(w); w = (const int32_t*) w + 2;
66 v128_t vacc0x67 = wasm_u64x2_load32x2(w); w = (const int32_t*) w + 2;
67 v128_t vacc1x01 = vacc0x01;
68 v128_t vacc1x23 = vacc0x23;
69 v128_t vacc1x45 = vacc0x45;
70 v128_t vacc1x67 = vacc0x67;
71
72 size_t p = ks;
73 do {
74 const int8_t* restrict a0 = a[0];
75 if XNN_UNPREDICTABLE(a0 != zero) {
76 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
77 }
78 const int8_t* restrict a1 = a[1];
79 if XNN_UNPREDICTABLE(a1 != zero) {
80 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
81 }
82 a += 2;
83
84 size_t k = kc;
85 while (k >= 16 * sizeof(int8_t)) {
86 const v128_t va0x01 = wasm_v128_xor(wasm_v128_load64_splat(a0), vsign_mask);
87 const v128_t va0x23 = wasm_v128_xor(wasm_v128_load64_splat((const int8_t*) a0 + 8), vsign_mask);
88 a0 += 16;
89 const v128_t va1x01 = wasm_v128_xor(wasm_v128_load64_splat(a1), vsign_mask);
90 const v128_t va1x23 = wasm_v128_xor(wasm_v128_load64_splat((const int8_t*) a1 + 8), vsign_mask);
91 a1 += 16;
92
93 const v128_t vb01x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
94 const v128_t vb23x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
95 const v128_t vb45x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
96 const v128_t vb67x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
97 const v128_t vb01x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
98 const v128_t vb23x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
99 const v128_t vb45x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
100 const v128_t vb67x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
101
102 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va0x01, vacc0x01);
103 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va0x01, vacc0x23);
104 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va0x01, vacc0x45);
105 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va0x01, vacc0x67);
106 vacc1x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va1x01, vacc1x01);
107 vacc1x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va1x01, vacc1x23);
108 vacc1x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va1x01, vacc1x45);
109 vacc1x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va1x01, vacc1x67);
110 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x23, va0x23, vacc0x01);
111 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x23, va0x23, vacc0x23);
112 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x23, va0x23, vacc0x45);
113 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x23, va0x23, vacc0x67);
114 vacc1x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x23, va1x23, vacc1x01);
115 vacc1x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x23, va1x23, vacc1x23);
116 vacc1x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x23, va1x23, vacc1x45);
117 vacc1x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x23, va1x23, vacc1x67);
118
119 k -= 16 * sizeof(int8_t);
120 }
121
122 if (k != 0) {
123 const v128_t va0x01 = wasm_v128_xor(wasm_v128_load64_splat(a0), vsign_mask);
124 a0 += 8;
125 const v128_t va1x01 = wasm_v128_xor(wasm_v128_load64_splat(a1), vsign_mask);
126 a1 += 8;
127
128 const v128_t vb01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
129 const v128_t vb23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
130 const v128_t vb45 = wasm_v128_load(w); w = (const int8_t*) w + 16;
131 const v128_t vb67 = wasm_v128_load(w); w = (const int8_t*) w + 16;
132
133 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01, va0x01, vacc0x01);
134 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23, va0x01, vacc0x23);
135 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45, va0x01, vacc0x45);
136 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67, va0x01, vacc0x67);
137 vacc1x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01, va1x01, vacc1x01);
138 vacc1x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23, va1x01, vacc1x23);
139 vacc1x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45, va1x01, vacc1x45);
140 vacc1x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67, va1x01, vacc1x67);
141
142 k -= 8 * sizeof(int8_t);
143 }
144 p -= 2 * sizeof(void*);
145 } while (p != 0);
146
147 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x01, vacc0x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x01, vacc0x23, 1, 3, 5, 7));
148 v128_t vacc0x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x45, vacc0x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x45, vacc0x67, 1, 3, 5, 7));
149 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x01, vacc1x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc1x01, vacc1x23, 1, 3, 5, 7));
150 v128_t vacc1x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x45, vacc1x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc1x45, vacc1x67, 1, 3, 5, 7));
151
152 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
153 vacc0x4567 = wasm_f32x4_convert_i32x4(vacc0x4567);
154 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
155 vacc1x4567 = wasm_f32x4_convert_i32x4(vacc1x4567);
156
157 const v128_t vscale0123 = wasm_v128_load(w);
158 w = (const float*) w + 4;
159 const v128_t vscale4567 = wasm_v128_load(w);
160 w = (const float*) w + 4;
161 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale0123);
162 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vscale4567);
163 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale0123);
164 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vscale4567);
165
166 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
167 vacc0x4567 = wasm_f32x4_add(vacc0x4567, vmagic_bias);
168 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
169 vacc1x4567 = wasm_f32x4_add(vacc1x4567, vmagic_bias);
170
171 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
172 vacc0x4567 = wasm_i32x4_max(vacc0x4567, vmagic_min);
173 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
174 vacc1x4567 = wasm_i32x4_max(vacc1x4567, vmagic_min);
175
176 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
177 vacc0x4567 = wasm_i32x4_sub(vacc0x4567, vmagic_bias_less_output_zero_point);
178 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
179 vacc1x4567 = wasm_i32x4_sub(vacc1x4567, vmagic_bias_less_output_zero_point);
180
181 v128_t vacc0x01234567 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc0x4567);
182 v128_t vacc1x01234567 = wasm_i16x8_narrow_i32x4(vacc1x0123, vacc1x4567);
183
184 v128_t vout0x01234567_1x01234567 = wasm_i8x16_narrow_i16x8(vacc0x01234567, vacc1x01234567);
185
186 vout0x01234567_1x01234567 = wasm_i8x16_min(vout0x01234567_1x01234567, voutput_max);
187
188 if (nc >= 8) {
189 wasm_v128_store64_lane(c1, vout0x01234567_1x01234567, 1);
190 wasm_v128_store64_lane(c0, vout0x01234567_1x01234567, 0);
191
192 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
193 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
194
195 a = (const int8_t**restrict) ((uintptr_t) a - ks);
196
197 nc -= 8;
198 } else {
199 if (nc & 4) {
200 wasm_v128_store32_lane(c1, vout0x01234567_1x01234567, 2);
201 c1 += 4;
202 wasm_v128_store32_lane(c0, vout0x01234567_1x01234567, 0);
203 c0 += 4;
204
205 vout0x01234567_1x01234567 = wasm_u64x2_shr(vout0x01234567_1x01234567, 32);
206 }
207 if (nc & 2) {
208 wasm_v128_store16_lane(c1, vout0x01234567_1x01234567, 4);
209 c1 += 2;
210 wasm_v128_store16_lane(c0, vout0x01234567_1x01234567, 0);
211 c0 += 2;
212
213 vout0x01234567_1x01234567 = wasm_u32x4_shr(vout0x01234567_1x01234567, 16);
214 }
215 if (nc & 1) {
216 wasm_v128_store8_lane(c1, vout0x01234567_1x01234567, 8);
217 wasm_v128_store8_lane(c0, vout0x01234567_1x01234567, 0);
218 }
219
220 nc = 0;
221 }
222 } while (nc != 0);
223 }