Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-2x8c8-minmax-wasmusdot-u2.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_2x8c8__wasmusdot_u2(
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 <= 2);
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
50 const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
51 const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
52 XNN_FORCE_REALIZATION(vmin);
53 XNN_FORCE_REALIZATION(vmax);
54
55 const v128_t vsign_mask = wasm_u8x16_const_splat(UINT8_C(0x80));
56 do {
57 const v128_t vinput_zero_point0 = wasm_i32x4_splat((int32_t) quantization_params[0].zero_point + 128);
58 const v128_t vinput_zero_point1 = wasm_i32x4_splat((int32_t) quantization_params[1].zero_point + 128);
59 v128_t vksum0123 = wasm_v128_load(w); w = (const int32_t*) w + 4;
60 v128_t vsum0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point0);
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_point1);
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_point0);
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_point1);
71 v128_t vacc1x45 = wasm_u64x2_extend_low_u32x4(vsum1x4567);
72 v128_t vacc1x67 = wasm_u64x2_extend_high_u32x4(vsum1x4567);
73
74 size_t k = kc;
75 while (k >= 16 * sizeof(int8_t)) {
76 const v128_t va0x01 = wasm_v128_xor(wasm_v128_load64_splat(a0), vsign_mask);
77 const v128_t va0x23 = wasm_v128_xor(wasm_v128_load64_splat((const int8_t*) a0 + 8), vsign_mask);
78 a0 += 16;
79 const v128_t va1x01 = wasm_v128_xor(wasm_v128_load64_splat(a1), vsign_mask);
80 const v128_t va1x23 = wasm_v128_xor(wasm_v128_load64_splat((const int8_t*) a1 + 8), vsign_mask);
81 a1 += 16;
82
83 const v128_t vb01x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
84 const v128_t vb23x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
85 const v128_t vb45x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
86 const v128_t vb67x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
87 const v128_t vb01x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
88 const v128_t vb23x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
89 const v128_t vb45x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
90 const v128_t vb67x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
91
92 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va0x01, vacc0x01);
93 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va0x01, vacc0x23);
94 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va0x01, vacc0x45);
95 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va0x01, vacc0x67);
96 vacc1x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va1x01, vacc1x01);
97 vacc1x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va1x01, vacc1x23);
98 vacc1x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va1x01, vacc1x45);
99 vacc1x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va1x01, vacc1x67);
100 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x23, va0x23, vacc0x01);
101 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x23, va0x23, vacc0x23);
102 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x23, va0x23, vacc0x45);
103 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x23, va0x23, vacc0x67);
104 vacc1x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x23, va1x23, vacc1x01);
105 vacc1x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x23, va1x23, vacc1x23);
106 vacc1x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x23, va1x23, vacc1x45);
107 vacc1x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x23, va1x23, vacc1x67);
108
109 k -= 16 * sizeof(int8_t);
110 }
111
112 if (k != 0) {
113 const v128_t va0x01 = wasm_v128_xor(wasm_v128_load64_splat(a0), vsign_mask);
114 a0 += 8;
115 const v128_t va1x01 = wasm_v128_xor(wasm_v128_load64_splat(a1), vsign_mask);
116 a1 += 8;
117
118 const v128_t vb01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
119 const v128_t vb23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
120 const v128_t vb45 = wasm_v128_load(w); w = (const int8_t*) w + 16;
121 const v128_t vb67 = wasm_v128_load(w); w = (const int8_t*) w + 16;
122
123 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01, va0x01, vacc0x01);
124 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23, va0x01, vacc0x23);
125 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45, va0x01, vacc0x45);
126 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67, va0x01, vacc0x67);
127 vacc1x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01, va1x01, vacc1x01);
128 vacc1x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23, va1x01, vacc1x23);
129 vacc1x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45, va1x01, vacc1x45);
130 vacc1x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67, va1x01, vacc1x67);
131
132 k -= 8 * sizeof(int8_t);
133 }
134
135 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x01, vacc0x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x01, vacc0x23, 1, 3, 5, 7));
136 v128_t vacc0x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x45, vacc0x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x45, vacc0x67, 1, 3, 5, 7));
137 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x01, vacc1x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc1x01, vacc1x23, 1, 3, 5, 7));
138 v128_t vacc1x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x45, vacc1x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc1x45, vacc1x67, 1, 3, 5, 7));
139
140 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
141 vacc0x4567 = wasm_f32x4_convert_i32x4(vacc0x4567);
142 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
143 vacc1x4567 = wasm_f32x4_convert_i32x4(vacc1x4567);
144
145 const v128_t vinput_scale0 = wasm_v128_load32_splat(&quantization_params[0].inv_scale);
146 const v128_t vinput_scale1 = wasm_v128_load32_splat(&quantization_params[1].inv_scale);
147
148 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale0);
149 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vinput_scale0);
150 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale1);
151 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vinput_scale1);
152
153 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
154 w = (const float*) w + 4;
155 const v128_t vfilter_output_scale4567 = wasm_v128_load(w);
156 w = (const float*) w + 4;
157 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
158 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vfilter_output_scale4567);
159 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
160 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vfilter_output_scale4567);
161
162 const v128_t vbias0123 = wasm_v128_load(w);
163 w = (const float*) w + 4;
164 const v128_t vbias4567 = wasm_v128_load(w);
165 w = (const float*) w + 4;
166 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
167 vacc0x4567 = wasm_f32x4_add(vacc0x4567, vbias4567);
168 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
169 vacc1x4567 = wasm_f32x4_add(vacc1x4567, vbias4567);
170
171 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
172 vacc0x4567 = wasm_f32x4_pmax(vacc0x4567, vmin);
173 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
174 vacc1x4567 = wasm_f32x4_pmax(vacc1x4567, vmin);
175
176 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
177 vacc0x4567 = wasm_f32x4_pmin(vacc0x4567, vmax);
178 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
179 vacc1x4567 = wasm_f32x4_pmin(vacc1x4567, vmax);
180
181 if XNN_LIKELY(nc >= 8) {
182 wasm_v128_store(c0, vacc0x0123);
183 wasm_v128_store(c0 + 4, vacc0x4567);
184 wasm_v128_store(c1, vacc1x0123);
185 wasm_v128_store(c1 + 4, vacc1x4567);
186
187 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
188 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
189
190 c0 = (float*) ((uintptr_t) c0 + cn_stride);
191 c1 = (float*) ((uintptr_t) c1 + cn_stride);
192
193 nc -= 8;
194 } else {
195 if (nc & 4) {
196 wasm_v128_store(c0, vacc0x0123);
197 vacc0x0123 = vacc0x4567;
198 c0 += 4;
199 wasm_v128_store(c1, vacc1x0123);
200 vacc1x0123 = vacc1x4567;
201 c1 += 4;
202 }
203 if (nc & 2) {
204 wasm_v128_store64_lane(c0, vacc0x0123, 0);
205 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
206 c0 += 2;
207 wasm_v128_store64_lane(c1, vacc1x0123, 0);
208 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
209 c1 += 2;
210 }
211 if (nc & 1) {
212 wasm_v128_store32_lane(c0, vacc0x0123, 0);
213 wasm_v128_store32_lane(c1, vacc1x0123, 0);
214 }
215 nc = 0;
216 }
217 } while (nc != 0);
218 }