Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-3x8c8-minmax-wasmusdot.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/c8-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_qd8_f32_qc8w_gemm_minmax_ukernel_3x8c8__wasmusdot(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const int8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 float* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
29 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 3);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(int8_t) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 kc = round_up_po2(kc, 8 * sizeof(int8_t));
41 const int8_t* a0 = a;
42 float* c0 = c;
43 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
44 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 a1 = a0;
47 c1 = c0;
48 }
49 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
50 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 a2 = a1;
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 const v128_t vsign_mask = wasm_u8x16_const_splat(UINT8_C(0x80));
62 do {
63 const v128_t vinput_zero_point0 = wasm_i32x4_splat((int32_t) quantization_params[0].zero_point + 128);
64 const v128_t vinput_zero_point1 = wasm_i32x4_splat((int32_t) quantization_params[1].zero_point + 128);
65 const v128_t vinput_zero_point2 = wasm_i32x4_splat((int32_t) quantization_params[2].zero_point + 128);
66 v128_t vksum0123 = wasm_v128_load(w); w = (const int32_t*) w + 4;
67 v128_t vsum0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point0);
68 v128_t vacc0x01 = wasm_u64x2_extend_low_u32x4(vsum0x0123);
69 v128_t vacc0x23 = wasm_u64x2_extend_high_u32x4(vsum0x0123);
70 v128_t vsum1x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point1);
71 v128_t vacc1x01 = wasm_u64x2_extend_low_u32x4(vsum1x0123);
72 v128_t vacc1x23 = wasm_u64x2_extend_high_u32x4(vsum1x0123);
73 v128_t vsum2x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point2);
74 v128_t vacc2x01 = wasm_u64x2_extend_low_u32x4(vsum2x0123);
75 v128_t vacc2x23 = wasm_u64x2_extend_high_u32x4(vsum2x0123);
76 v128_t vksum4567 = wasm_v128_load(w); w = (const int32_t*) w + 4;
77 v128_t vsum0x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point0);
78 v128_t vacc0x45 = wasm_u64x2_extend_low_u32x4(vsum0x4567);
79 v128_t vacc0x67 = wasm_u64x2_extend_high_u32x4(vsum0x4567);
80 v128_t vsum1x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point1);
81 v128_t vacc1x45 = wasm_u64x2_extend_low_u32x4(vsum1x4567);
82 v128_t vacc1x67 = wasm_u64x2_extend_high_u32x4(vsum1x4567);
83 v128_t vsum2x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point2);
84 v128_t vacc2x45 = wasm_u64x2_extend_low_u32x4(vsum2x4567);
85 v128_t vacc2x67 = wasm_u64x2_extend_high_u32x4(vsum2x4567);
86
87 size_t k = kc;
88 while (k != 0) {
89 const v128_t va0x01 = wasm_v128_xor(wasm_v128_load64_splat(a0), vsign_mask);
90 a0 += 8;
91 const v128_t va1x01 = wasm_v128_xor(wasm_v128_load64_splat(a1), vsign_mask);
92 a1 += 8;
93 const v128_t va2x01 = wasm_v128_xor(wasm_v128_load64_splat(a2), vsign_mask);
94 a2 += 8;
95
96 const v128_t vb01x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
97 const v128_t vb23x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
98 const v128_t vb45x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
99 const v128_t vb67x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
100
101 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va0x01, vacc0x01);
102 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va0x01, vacc0x23);
103 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va0x01, vacc0x45);
104 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va0x01, vacc0x67);
105 vacc1x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va1x01, vacc1x01);
106 vacc1x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va1x01, vacc1x23);
107 vacc1x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va1x01, vacc1x45);
108 vacc1x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va1x01, vacc1x67);
109 vacc2x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va2x01, vacc2x01);
110 vacc2x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va2x01, vacc2x23);
111 vacc2x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va2x01, vacc2x45);
112 vacc2x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va2x01, vacc2x67);
113
114 k -= 8 * sizeof(int8_t);
115 }
116
117
118 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x01, vacc0x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x01, vacc0x23, 1, 3, 5, 7));
119 v128_t vacc0x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x45, vacc0x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x45, vacc0x67, 1, 3, 5, 7));
120 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x01, vacc1x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc1x01, vacc1x23, 1, 3, 5, 7));
121 v128_t vacc1x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x45, vacc1x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc1x45, vacc1x67, 1, 3, 5, 7));
122 v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x01, vacc2x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc2x01, vacc2x23, 1, 3, 5, 7));
123 v128_t vacc2x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x45, vacc2x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc2x45, vacc2x67, 1, 3, 5, 7));
124
125 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
126 vacc0x4567 = wasm_f32x4_convert_i32x4(vacc0x4567);
127 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
128 vacc1x4567 = wasm_f32x4_convert_i32x4(vacc1x4567);
129 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
130 vacc2x4567 = wasm_f32x4_convert_i32x4(vacc2x4567);
131
132 const v128_t vinput_scale0 = wasm_v128_load32_splat(&quantization_params[0].inv_scale);
133 const v128_t vinput_scale1 = wasm_v128_load32_splat(&quantization_params[1].inv_scale);
134 const v128_t vinput_scale2 = wasm_v128_load32_splat(&quantization_params[2].inv_scale);
135
136 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale0);
137 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vinput_scale0);
138 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale1);
139 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vinput_scale1);
140 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vinput_scale2);
141 vacc2x4567 = wasm_f32x4_mul(vacc2x4567, vinput_scale2);
142
143 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
144 w = (const float*) w + 4;
145 const v128_t vfilter_output_scale4567 = wasm_v128_load(w);
146 w = (const float*) w + 4;
147 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
148 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vfilter_output_scale4567);
149 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
150 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vfilter_output_scale4567);
151 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vfilter_output_scale0123);
152 vacc2x4567 = wasm_f32x4_mul(vacc2x4567, vfilter_output_scale4567);
153
154 const v128_t vbias0123 = wasm_v128_load(w);
155 w = (const float*) w + 4;
156 const v128_t vbias4567 = wasm_v128_load(w);
157 w = (const float*) w + 4;
158 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
159 vacc0x4567 = wasm_f32x4_add(vacc0x4567, vbias4567);
160 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
161 vacc1x4567 = wasm_f32x4_add(vacc1x4567, vbias4567);
162 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vbias0123);
163 vacc2x4567 = wasm_f32x4_add(vacc2x4567, vbias4567);
164
165 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
166 vacc0x4567 = wasm_f32x4_pmax(vacc0x4567, vmin);
167 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
168 vacc1x4567 = wasm_f32x4_pmax(vacc1x4567, vmin);
169 vacc2x0123 = wasm_f32x4_pmax(vacc2x0123, vmin);
170 vacc2x4567 = wasm_f32x4_pmax(vacc2x4567, vmin);
171
172 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
173 vacc0x4567 = wasm_f32x4_pmin(vacc0x4567, vmax);
174 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
175 vacc1x4567 = wasm_f32x4_pmin(vacc1x4567, vmax);
176 vacc2x0123 = wasm_f32x4_pmin(vacc2x0123, vmax);
177 vacc2x4567 = wasm_f32x4_pmin(vacc2x4567, vmax);
178
179 if XNN_LIKELY(nc >= 8) {
180 wasm_v128_store(c0, vacc0x0123);
181 wasm_v128_store(c0 + 4, vacc0x4567);
182 wasm_v128_store(c1, vacc1x0123);
183 wasm_v128_store(c1 + 4, vacc1x4567);
184 wasm_v128_store(c2, vacc2x0123);
185 wasm_v128_store(c2 + 4, vacc2x4567);
186
187 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
188 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
189 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
190
191 c0 = (float*) ((uintptr_t) c0 + cn_stride);
192 c1 = (float*) ((uintptr_t) c1 + cn_stride);
193 c2 = (float*) ((uintptr_t) c2 + cn_stride);
194
195 nc -= 8;
196 } else {
197 if (nc & 4) {
198 wasm_v128_store(c0, vacc0x0123);
199 vacc0x0123 = vacc0x4567;
200 c0 += 4;
201 wasm_v128_store(c1, vacc1x0123);
202 vacc1x0123 = vacc1x4567;
203 c1 += 4;
204 wasm_v128_store(c2, vacc2x0123);
205 vacc2x0123 = vacc2x4567;
206 c2 += 4;
207 }
208 if (nc & 2) {
209 wasm_v128_store64_lane(c0, vacc0x0123, 0);
210 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
211 c0 += 2;
212 wasm_v128_store64_lane(c1, vacc1x0123, 0);
213 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
214 c1 += 2;
215 wasm_v128_store64_lane(c2, vacc2x0123, 0);
216 vacc2x0123 = wasm_v64x2_shuffle(vacc2x0123, vacc2x0123, 1, 1);
217 c2 += 2;
218 }
219 if (nc & 1) {
220 wasm_v128_store32_lane(c0, vacc0x0123, 0);
221 wasm_v128_store32_lane(c1, vacc1x0123, 0);
222 wasm_v128_store32_lane(c2, vacc2x0123, 0);
223 }
224 nc = 0;
225 }
226 } while (nc != 0);
227 }