Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-2x8c8-minmax-wasmsdot-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_qd8_f32_qc8w_igemm_minmax_ukernel_2x8c8__wasmsdot_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 float* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const int8_t* zero_data,
31 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
32 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(mr != 0);
35 assert(mr <= 2);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (2 * sizeof(void*)) == 0);
40 assert(a_offset % sizeof(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 kc = round_up_po2(kc, 8 * sizeof(int8_t));
46 float* c0 = c;
47 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr != 2) {
49 c1 = c0;
50 }
51
52 const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
53 const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
54 XNN_FORCE_REALIZATION(vmin);
55 XNN_FORCE_REALIZATION(vmax);
56
57 do {
58 const v128_t vinput_zero_point = wasm_v128_load32_splat(&quantization_params->zero_point);
59 v128_t vksum0123 = wasm_v128_load(w); w = (const int32_t*) w + 4;
60 v128_t vsum0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
61 v128_t vacc0x01 = wasm_u64x2_extend_low_u32x4(vsum0x0123);
62 v128_t vacc0x23 = wasm_u64x2_extend_high_u32x4(vsum0x0123);
63 v128_t vsum1x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
64 v128_t vacc1x01 = wasm_u64x2_extend_low_u32x4(vsum1x0123);
65 v128_t vacc1x23 = wasm_u64x2_extend_high_u32x4(vsum1x0123);
66 v128_t vksum4567 = wasm_v128_load(w); w = (const int32_t*) w + 4;
67 v128_t vsum0x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point);
68 v128_t vacc0x45 = wasm_u64x2_extend_low_u32x4(vsum0x4567);
69 v128_t vacc0x67 = wasm_u64x2_extend_high_u32x4(vsum0x4567);
70 v128_t vsum1x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point);
71 v128_t vacc1x45 = wasm_u64x2_extend_low_u32x4(vsum1x4567);
72 v128_t vacc1x67 = wasm_u64x2_extend_high_u32x4(vsum1x4567);
73
74 size_t p = ks;
75 do {
76 const int8_t* restrict a0 = a[0];
77 if XNN_UNPREDICTABLE(a0 != zero) {
78 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
79 } else {
80 a0 = zero_data;
81 }
82 const int8_t* restrict a1 = a[1];
83 if XNN_UNPREDICTABLE(a1 != zero) {
84 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
85 } else {
86 a1 = zero_data;
87 }
88 a += 2;
89
90 size_t k = kc;
91 while (k >= 16 * sizeof(int8_t)) {
92 const v128_t va0x01 = wasm_v128_load64_splat(a0);
93 const v128_t va0x23 = wasm_v128_load64_splat((const int8_t*) a0 + 8);
94 a0 += 16;
95 const v128_t va1x01 = wasm_v128_load64_splat(a1);
96 const v128_t va1x23 = wasm_v128_load64_splat((const int8_t*) a1 + 8);
97 a1 += 16;
98
99 const v128_t vb01x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
100 const v128_t vb23x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
101 const v128_t vb45x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
102 const v128_t vb67x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
103 const v128_t vb01x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
104 const v128_t vb23x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
105 const v128_t vb45x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
106 const v128_t vb67x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
107
108 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va0x01, vacc0x01);
109 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va0x01, vacc0x23);
110 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va0x01, vacc0x45);
111 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va0x01, vacc0x67);
112 vacc1x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va1x01, vacc1x01);
113 vacc1x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va1x01, vacc1x23);
114 vacc1x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va1x01, vacc1x45);
115 vacc1x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va1x01, vacc1x67);
116 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x23, va0x23, vacc0x01);
117 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x23, va0x23, vacc0x23);
118 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x23, va0x23, vacc0x45);
119 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x23, va0x23, vacc0x67);
120 vacc1x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x23, va1x23, vacc1x01);
121 vacc1x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x23, va1x23, vacc1x23);
122 vacc1x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x23, va1x23, vacc1x45);
123 vacc1x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x23, va1x23, vacc1x67);
124
125 k -= 16 * sizeof(int8_t);
126 }
127
128 if (k != 0) {
129 const v128_t va0x01 = wasm_v128_load64_splat(a0);
130 a0 += 8;
131 const v128_t va1x01 = wasm_v128_load64_splat(a1);
132 a1 += 8;
133
134 const v128_t vb01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
135 const v128_t vb23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
136 const v128_t vb45 = wasm_v128_load(w); w = (const int8_t*) w + 16;
137 const v128_t vb67 = wasm_v128_load(w); w = (const int8_t*) w + 16;
138
139 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01, va0x01, vacc0x01);
140 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23, va0x01, vacc0x23);
141 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45, va0x01, vacc0x45);
142 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67, va0x01, vacc0x67);
143 vacc1x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01, va1x01, vacc1x01);
144 vacc1x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23, va1x01, vacc1x23);
145 vacc1x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45, va1x01, vacc1x45);
146 vacc1x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67, va1x01, vacc1x67);
147
148 k -= 8 * sizeof(int8_t);
149 }
150 p -= 2 * sizeof(void*);
151 } while (p != 0);
152
153 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x01, vacc0x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x01, vacc0x23, 1, 3, 5, 7));
154 v128_t vacc0x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x45, vacc0x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x45, vacc0x67, 1, 3, 5, 7));
155 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x01, vacc1x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc1x01, vacc1x23, 1, 3, 5, 7));
156 v128_t vacc1x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x45, vacc1x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc1x45, vacc1x67, 1, 3, 5, 7));
157
158 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
159 vacc0x4567 = wasm_f32x4_convert_i32x4(vacc0x4567);
160 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
161 vacc1x4567 = wasm_f32x4_convert_i32x4(vacc1x4567);
162
163 const v128_t vinput_scale = wasm_v128_load32_splat(&quantization_params->inv_scale);
164
165 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale);
166 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vinput_scale);
167 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale);
168 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vinput_scale);
169
170 const v128_t vfilter_output_scale0123 = wasm_v128_load(w); w = (const float*) w + 4;
171 const v128_t vfilter_output_scale4567 = wasm_v128_load(w); w = (const float*) w + 4;
172
173 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
174 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vfilter_output_scale4567);
175 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
176 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vfilter_output_scale4567);
177
178 const v128_t vbias0123 = wasm_v128_load(w); w = (const float*) w + 4;
179 const v128_t vbias4567 = wasm_v128_load(w); w = (const float*) w + 4;
180 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
181 vacc0x4567 = wasm_f32x4_add(vacc0x4567, vbias4567);
182 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
183 vacc1x4567 = wasm_f32x4_add(vacc1x4567, vbias4567);
184
185 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
186 vacc0x4567 = wasm_f32x4_pmax(vacc0x4567, vmin);
187 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
188 vacc1x4567 = wasm_f32x4_pmax(vacc1x4567, vmin);
189
190 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
191 vacc0x4567 = wasm_f32x4_pmin(vacc0x4567, vmax);
192 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
193 vacc1x4567 = wasm_f32x4_pmin(vacc1x4567, vmax);
194
195 if XNN_LIKELY(nc >= 8) {
196 wasm_v128_store(c1, vacc1x0123);
197 wasm_v128_store(c1 + 4, vacc1x4567);
198 wasm_v128_store(c0, vacc0x0123);
199 wasm_v128_store(c0 + 4, vacc0x4567);
200
201 c0 = (float*) ((uintptr_t) c0 + cn_stride);
202 c1 = (float*) ((uintptr_t) c1 + cn_stride);
203
204 a = (const int8_t**restrict) ((uintptr_t) a - ks);
205
206 nc -= 8;
207 } else {
208 if (nc & 4) {
209 wasm_v128_store(c1, vacc1x0123);
210 vacc1x0123 = vacc1x4567;
211 c1 += 4;
212 wasm_v128_store(c0, vacc0x0123);
213 vacc0x0123 = vacc0x4567;
214 c0 += 4;
215 }
216 if (nc & 2) {
217 wasm_v128_store64_lane(c1, vacc1x0123, 0);
218 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
219 c1 += 2;
220 wasm_v128_store64_lane(c0, vacc0x0123, 0);
221 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
222 c0 += 2;
223 }
224 if (nc & 1) {
225 wasm_v128_store32_lane(c1, vacc1x0123, 0);
226 wasm_v128_store32_lane(c0, vacc0x0123, 0);
227 }
228 nc = 0;
229 }
230 } while (nc != 0);
231 }