Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-3x4c8-minmax-wasmsimd-dot16x2-ld64.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx4c8-wasmsimd-dot16x2.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 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_3x4c8__wasmsimd_dot16x2_ld64(
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 do {
62 const v128_t vksum0123 = wasm_v128_load(w);
63 const v128_t vinput_zero_point0 = wasm_v128_load32_splat(&quantization_params[0].zero_point);
64 const v128_t vinput_zero_point1 = wasm_v128_load32_splat(&quantization_params[1].zero_point);
65 const v128_t vinput_zero_point2 = wasm_v128_load32_splat(&quantization_params[2].zero_point);
66 const v128_t vinit0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point0);
67 const v128_t vinit1x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point1);
68 const v128_t vinit2x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point2);
69
70 const v128_t vzero = wasm_i32x4_const_splat(0);
71 v128_t vacc0x0 = wasm_v32x4_shuffle(vinit0x0123, vzero, 0, 5, 6, 7);
72 v128_t vacc0x1 = wasm_v32x4_shuffle(vinit0x0123, vzero, 4, 1, 6, 7);
73 v128_t vacc0x2 = wasm_v32x4_shuffle(vinit0x0123, vzero, 4, 5, 2, 7);
74 v128_t vacc0x3 = wasm_v32x4_shuffle(vinit0x0123, vzero, 4, 5, 6, 3);
75 v128_t vacc1x0 = wasm_v32x4_shuffle(vinit1x0123, vzero, 0, 5, 6, 7);
76 v128_t vacc1x1 = wasm_v32x4_shuffle(vinit1x0123, vzero, 4, 1, 6, 7);
77 v128_t vacc1x2 = wasm_v32x4_shuffle(vinit1x0123, vzero, 4, 5, 2, 7);
78 v128_t vacc1x3 = wasm_v32x4_shuffle(vinit1x0123, vzero, 4, 5, 6, 3);
79 v128_t vacc2x0 = wasm_v32x4_shuffle(vinit2x0123, vzero, 0, 5, 6, 7);
80 v128_t vacc2x1 = wasm_v32x4_shuffle(vinit2x0123, vzero, 4, 1, 6, 7);
81 v128_t vacc2x2 = wasm_v32x4_shuffle(vinit2x0123, vzero, 4, 5, 2, 7);
82 v128_t vacc2x3 = wasm_v32x4_shuffle(vinit2x0123, vzero, 4, 5, 6, 3);
83 w = (const int32_t*) w + 4;
84
85 size_t k = kc;
86
87 while (k >= 8 * sizeof(int8_t)) {
88 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
89 a0 += 8;
90 const v128_t vxa1 = wasm_i16x8_load8x8(a1);
91 a1 += 8;
92 const v128_t vxa2 = wasm_i16x8_load8x8(a2);
93 a2 += 8;
94
95 const v128_t vxb0 = wasm_i16x8_load8x8(w);
96
97 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_dot_i16x8(vxa0, vxb0));
98 vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_dot_i16x8(vxa1, vxb0));
99 vacc2x0 = wasm_i32x4_add(vacc2x0, wasm_i32x4_dot_i16x8(vxa2, vxb0));
100 const v128_t vxb1 = wasm_i16x8_load8x8((const int8_t*) w + 8);
101
102 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_dot_i16x8(vxa0, vxb1));
103 vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_dot_i16x8(vxa1, vxb1));
104 vacc2x1 = wasm_i32x4_add(vacc2x1, wasm_i32x4_dot_i16x8(vxa2, vxb1));
105 const v128_t vxb2 = wasm_i16x8_load8x8((const int8_t*) w + 16);
106
107 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_dot_i16x8(vxa0, vxb2));
108 vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_dot_i16x8(vxa1, vxb2));
109 vacc2x2 = wasm_i32x4_add(vacc2x2, wasm_i32x4_dot_i16x8(vxa2, vxb2));
110 const v128_t vxb3 = wasm_i16x8_load8x8((const int8_t*) w + 24);
111
112 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_dot_i16x8(vxa0, vxb3));
113 vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_dot_i16x8(vxa1, vxb3));
114 vacc2x3 = wasm_i32x4_add(vacc2x3, wasm_i32x4_dot_i16x8(vxa2, vxb3));
115
116 w = (const int8_t*) w + 32;
117 k -= 8 * sizeof(int8_t);
118 };
119
120 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));
121 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));
122 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));
123 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));
124 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));
125 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));
126
127 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
128 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
129 v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x02, vacc2x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x02, vacc2x13, 2, 6, 3, 7));
130
131 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
132 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
133 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
134
135 const v128_t vinput_scale0 = wasm_v128_load32_splat(&quantization_params[0].inv_scale);
136 const v128_t vinput_scale1 = wasm_v128_load32_splat(&quantization_params[1].inv_scale);
137 const v128_t vinput_scale2 = wasm_v128_load32_splat(&quantization_params[2].inv_scale);
138
139 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale0);
140 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale1);
141 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vinput_scale2);
142
143 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
144 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
145 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
146 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vfilter_output_scale0123);
147 w = (const float*) w + 4;
148
149 const v128_t vbias0123 = wasm_v128_load(w);
150 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
151 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
152 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vbias0123);
153 w = (const float*) w + 4;
154
155 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
156 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
157 vacc2x0123 = wasm_f32x4_pmax(vacc2x0123, vmin);
158
159 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
160 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
161 vacc2x0123 = wasm_f32x4_pmin(vacc2x0123, vmax);
162
163 if XNN_LIKELY(nc >= 4) {
164 wasm_v128_store(c0, vacc0x0123);
165 wasm_v128_store(c1, vacc1x0123);
166 wasm_v128_store(c2, vacc2x0123);
167
168 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
169 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
170 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
171
172 c0 = (float*) ((uintptr_t) c0 + cn_stride);
173 c1 = (float*) ((uintptr_t) c1 + cn_stride);
174 c2 = (float*) ((uintptr_t) c2 + cn_stride);
175
176 nc -= 4;
177 } else {
178 if (nc & 2) {
179 wasm_v128_store64_lane(c0, vacc0x0123, 0);
180 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
181 c0 += 2;
182 wasm_v128_store64_lane(c1, vacc1x0123, 0);
183 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
184 c1 += 2;
185 wasm_v128_store64_lane(c2, vacc2x0123, 0);
186 vacc2x0123 = wasm_v64x2_shuffle(vacc2x0123, vacc2x0123, 1, 1);
187 c2 += 2;
188 }
189 if (nc & 1) {
190 wasm_v128_store32_lane(c0, vacc0x0123, 0);
191 wasm_v128_store32_lane(c1, vacc1x0123, 0);
192 wasm_v128_store32_lane(c2, vacc2x0123, 0);
193 }
194 nc = 0;
195 }
196 } while (nc != 0);
197 }