Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-3x4c16-minmax-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_qd8_f32_qc8w_igemm_minmax_ukernel_3x4c16__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 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 <= 3);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (3 * 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, 16 * 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 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 c2 = c1;
54 }
55
56 const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
57 const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
58 XNN_FORCE_REALIZATION(vmin);
59 XNN_FORCE_REALIZATION(vmax);
60
61 do {
62 v128_t vksum0 = wasm_v128_load32_zero(w);
63 v128_t vksum1 = wasm_v128_load32_zero((const int32_t*) w + 1);
64 v128_t vksum2 = wasm_v128_load32_zero((const int32_t*) w + 2);
65 v128_t vksum3 = wasm_v128_load32_zero((const int32_t*) w + 3);
66 const v128_t vinput_zero_point = wasm_v128_load32_splat(&quantization_params->zero_point);
67 v128_t vacc0x0 = wasm_i32x4_mul(vksum0, vinput_zero_point);
68 v128_t vacc0x1 = wasm_i32x4_mul(vksum1, vinput_zero_point);
69 v128_t vacc0x2 = wasm_i32x4_mul(vksum2, vinput_zero_point);
70 v128_t vacc0x3 = wasm_i32x4_mul(vksum3, vinput_zero_point);
71 v128_t vacc1x0 = wasm_i32x4_mul(vksum0, vinput_zero_point);
72 v128_t vacc1x1 = wasm_i32x4_mul(vksum1, vinput_zero_point);
73 v128_t vacc1x2 = wasm_i32x4_mul(vksum2, vinput_zero_point);
74 v128_t vacc1x3 = wasm_i32x4_mul(vksum3, vinput_zero_point);
75 v128_t vacc2x0 = wasm_i32x4_mul(vksum0, vinput_zero_point);
76 v128_t vacc2x1 = wasm_i32x4_mul(vksum1, vinput_zero_point);
77 v128_t vacc2x2 = wasm_i32x4_mul(vksum2, vinput_zero_point);
78 v128_t vacc2x3 = wasm_i32x4_mul(vksum3, vinput_zero_point);
79 w = (const int32_t*) w + 4;
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 } else {
87 a0 = zero_data;
88 }
89 const int8_t* restrict a1 = a[1];
90 if XNN_UNPREDICTABLE(a1 != zero) {
91 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
92 } else {
93 a1 = zero_data;
94 }
95 const int8_t* restrict a2 = a[2];
96 if XNN_UNPREDICTABLE(a2 != zero) {
97 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
98 } else {
99 a2 = zero_data;
100 }
101 a += 3;
102
103 size_t k = kc;
104 do {
105 const v128_t va0 = wasm_v128_load(a0);
106 a0 += 16;
107 const v128_t va1 = wasm_v128_load(a1);
108 a1 += 16;
109 const v128_t va2 = wasm_v128_load(a2);
110 a2 += 16;
111
112 const v128_t vb0 = wasm_v128_load(w);
113 vacc0x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va0, vacc0x0);
114 vacc1x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va1, vacc1x0);
115 vacc2x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va2, vacc2x0);
116 const v128_t vb1 = wasm_v128_load((const int8_t*) w + 16);
117 vacc0x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va0, vacc0x1);
118 vacc1x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va1, vacc1x1);
119 vacc2x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va2, vacc2x1);
120 const v128_t vb2 = wasm_v128_load((const int8_t*) w + 32);
121 vacc0x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va0, vacc0x2);
122 vacc1x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va1, vacc1x2);
123 vacc2x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va2, vacc2x2);
124 const v128_t vb3 = wasm_v128_load((const int8_t*) w + 48);
125 vacc0x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va0, vacc0x3);
126 vacc1x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va1, vacc1x3);
127 vacc2x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va2, vacc2x3);
128
129 w = (const int8_t*) w + 64;
130 k -= 16 * sizeof(int8_t);
131 } while (k != 0);
132 p -= 3 * sizeof(void*);
133 } while (p != 0);
134
135 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));
136 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));
137 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));
138 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));
139 const v128_t vacc2x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x0, vacc2x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x0, vacc2x2, 2, 6, 3, 7));
140 const v128_t vacc2x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x1, vacc2x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x1, vacc2x3, 2, 6, 3, 7));
141
142 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
143 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
144 v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x02, vacc2x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x02, vacc2x13, 2, 6, 3, 7));
145
146 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
147 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
148 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
149
150 const v128_t vinput_scale = wasm_v128_load32_splat(&quantization_params->inv_scale);
151
152 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale);
153 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale);
154 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vinput_scale);
155
156 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
157 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
158 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
159 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vfilter_output_scale0123);
160 w = (const float*) w + 4;
161
162 const v128_t vbias0123 = wasm_v128_load(w);
163 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
164 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
165 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vbias0123);
166 w = (const float*) w + 4;
167
168 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
169 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
170 vacc2x0123 = wasm_f32x4_pmax(vacc2x0123, vmin);
171
172 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
173 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
174 vacc2x0123 = wasm_f32x4_pmin(vacc2x0123, vmax);
175
176 if XNN_LIKELY(nc >= 4) {
177 wasm_v128_store(c2, vacc2x0123);
178 wasm_v128_store(c1, vacc1x0123);
179 wasm_v128_store(c0, vacc0x0123);
180
181 c0 = (float*) ((uintptr_t) c0 + cn_stride);
182 c1 = (float*) ((uintptr_t) c1 + cn_stride);
183 c2 = (float*) ((uintptr_t) c2 + cn_stride);
184
185 a = (const int8_t**restrict) ((uintptr_t) a - ks);
186
187 nc -= 4;
188 } else {
189 if (nc & 2) {
190 wasm_v128_store64_lane(c2, vacc2x0123, 0);
191 vacc2x0123 = wasm_v64x2_shuffle(vacc2x0123, vacc2x0123, 1, 1);
192 c2 += 2;
193 wasm_v128_store64_lane(c1, vacc1x0123, 0);
194 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
195 c1 += 2;
196 wasm_v128_store64_lane(c0, vacc0x0123, 0);
197 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
198 c0 += 2;
199 }
200 if (nc & 1) {
201 wasm_v128_store32_lane(c2, vacc2x0123, 0);
202 wasm_v128_store32_lane(c1, vacc1x0123, 0);
203 wasm_v128_store32_lane(c0, vacc0x0123, 0);
204 }
205 nc = 0;
206 }
207 } while (nc != 0);
208 }