Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-3x4c16-minmax-wasmsdot.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_3x4c16__wasmsdot(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 const int8_t* restrict a,
22 size_t a_stride,
23 const void* restrict w,
24 float* restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
28 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(mr != 0);
31 assert(mr <= 3);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(int8_t) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38
39 kc = round_up_po2(kc, 16 * sizeof(int8_t));
40 const int8_t* a0 = a;
41 float* c0 = c;
42 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
43 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr < 2) {
45 a1 = a0;
46 c1 = c0;
47 }
48 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
49 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 a2 = a1;
52 c2 = c1;
53 }
54
55 const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
56 const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
57 XNN_FORCE_REALIZATION(vmin);
58 XNN_FORCE_REALIZATION(vmax);
59
60 do {
61 v128_t vksum0 = wasm_v128_load32_zero(w);
62 v128_t vksum1 = wasm_v128_load32_zero((const int32_t*) w + 1);
63 v128_t vksum2 = wasm_v128_load32_zero((const int32_t*) w + 2);
64 v128_t vksum3 = wasm_v128_load32_zero((const int32_t*) w + 3);
65 const v128_t vinput_zero_point0 = wasm_v128_load32_splat(&quantization_params[0].zero_point);
66 const v128_t vinput_zero_point1 = wasm_v128_load32_splat(&quantization_params[1].zero_point);
67 const v128_t vinput_zero_point2 = wasm_v128_load32_splat(&quantization_params[2].zero_point);
68 v128_t vacc0x0 = wasm_i32x4_mul(vksum0, vinput_zero_point0);
69 v128_t vacc0x1 = wasm_i32x4_mul(vksum1, vinput_zero_point0);
70 v128_t vacc0x2 = wasm_i32x4_mul(vksum2, vinput_zero_point0);
71 v128_t vacc0x3 = wasm_i32x4_mul(vksum3, vinput_zero_point0);
72 v128_t vacc1x0 = wasm_i32x4_mul(vksum0, vinput_zero_point1);
73 v128_t vacc1x1 = wasm_i32x4_mul(vksum1, vinput_zero_point1);
74 v128_t vacc1x2 = wasm_i32x4_mul(vksum2, vinput_zero_point1);
75 v128_t vacc1x3 = wasm_i32x4_mul(vksum3, vinput_zero_point1);
76 v128_t vacc2x0 = wasm_i32x4_mul(vksum0, vinput_zero_point2);
77 v128_t vacc2x1 = wasm_i32x4_mul(vksum1, vinput_zero_point2);
78 v128_t vacc2x2 = wasm_i32x4_mul(vksum2, vinput_zero_point2);
79 v128_t vacc2x3 = wasm_i32x4_mul(vksum3, vinput_zero_point2);
80 w = (const int32_t*) w + 4;
81
82 size_t k = kc;
83 do {
84 const v128_t va0 = wasm_v128_load(a0);
85 a0 += 16;
86 const v128_t va1 = wasm_v128_load(a1);
87 a1 += 16;
88 const v128_t va2 = wasm_v128_load(a2);
89 a2 += 16;
90
91 const v128_t vb0 = wasm_v128_load(w);
92
93 vacc0x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va0, vacc0x0);
94 vacc1x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va1, vacc1x0);
95 vacc2x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va2, vacc2x0);
96 const v128_t vb1 = wasm_v128_load((const int8_t*) w + 16);
97
98 vacc0x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va0, vacc0x1);
99 vacc1x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va1, vacc1x1);
100 vacc2x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va2, vacc2x1);
101 const v128_t vb2 = wasm_v128_load((const int8_t*) w + 32);
102
103 vacc0x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va0, vacc0x2);
104 vacc1x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va1, vacc1x2);
105 vacc2x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va2, vacc2x2);
106 const v128_t vb3 = wasm_v128_load((const int8_t*) w + 48);
107
108 vacc0x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va0, vacc0x3);
109 vacc1x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va1, vacc1x3);
110 vacc2x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va2, vacc2x3);
111
112 w = (const int8_t*) w + 64;
113 k -= 16 * sizeof(int8_t);
114 } while (k != 0);
115
116 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));
117 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));
118 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));
119 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));
120 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));
121 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));
122
123 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
124 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
125 v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x02, vacc2x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x02, vacc2x13, 2, 6, 3, 7));
126
127 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
128 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
129 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
130
131 const v128_t vinput_scale0 = wasm_v128_load32_splat(&quantization_params[0].inv_scale);
132 const v128_t vinput_scale1 = wasm_v128_load32_splat(&quantization_params[1].inv_scale);
133 const v128_t vinput_scale2 = wasm_v128_load32_splat(&quantization_params[2].inv_scale);
134
135 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale0);
136 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale1);
137 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vinput_scale2);
138
139 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
140 w = (const float*) w + 4;
141 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
142 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
143 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vfilter_output_scale0123);
144
145 const v128_t vbias0123 = wasm_v128_load(w);
146 w = (const float*) w + 4;
147 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
148 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
149 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vbias0123);
150
151 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
152 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
153 vacc2x0123 = wasm_f32x4_pmax(vacc2x0123, vmin);
154
155 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
156 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
157 vacc2x0123 = wasm_f32x4_pmin(vacc2x0123, vmax);
158
159 if XNN_LIKELY(nc >= 4) {
160 wasm_v128_store(c0, vacc0x0123);
161 wasm_v128_store(c1, vacc1x0123);
162 wasm_v128_store(c2, vacc2x0123);
163
164 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
165 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
166 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
167
168 c0 = (float*) ((uintptr_t) c0 + cn_stride);
169 c1 = (float*) ((uintptr_t) c1 + cn_stride);
170 c2 = (float*) ((uintptr_t) c2 + cn_stride);
171
172 nc -= 4;
173 } else {
174 if (nc & 2) {
175 wasm_v128_store64_lane(c0, vacc0x0123, 0);
176 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
177 c0 += 2;
178 wasm_v128_store64_lane(c1, vacc1x0123, 0);
179 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
180 c1 += 2;
181 wasm_v128_store64_lane(c2, vacc2x0123, 0);
182 vacc2x0123 = wasm_v64x2_shuffle(vacc2x0123, vacc2x0123, 1, 1);
183 c2 += 2;
184 }
185 if (nc & 1) {
186 wasm_v128_store32_lane(c0, vacc0x0123, 0);
187 wasm_v128_store32_lane(c1, vacc1x0123, 0);
188 wasm_v128_store32_lane(c2, vacc2x0123, 0);
189 }
190 nc = 0;
191 }
192 } while (nc != 0);
193 }