Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-2x4c2-minmax-wasmsimd-dot16x2-ld64.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx4c2-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_2x4c2__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 <= 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, 2 * 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
56 do {
57 const v128_t vksum0123 = wasm_v128_load(w);
58 const v128_t vinput_zero_point0 = wasm_v128_load32_splat(&quantization_params[0].zero_point);
59 const v128_t vinput_zero_point1 = wasm_v128_load32_splat(&quantization_params[1].zero_point);
60 v128_t vacc0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point0);
61 v128_t vacc1x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point1);
62 w = (const int32_t*) w + 4;
63
64 size_t k = kc;
65 while (k >= 8 * sizeof(int8_t)) {
66 const v128_t vxa0 = wasm_i16x8_load8x8((const v128_t*) a0);
67 a0 += 8;
68 const v128_t vxa1 = wasm_i16x8_load8x8((const v128_t*) a1);
69 a1 += 8;
70
71 const v128_t vxb0 = wasm_i16x8_load8x8(w);
72
73 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
74 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
75 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
76 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
77 const v128_t vxb1 = wasm_i16x8_load8x8((const int8_t*) w + 8);
78
79 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
80 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
81 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
82 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
83 const v128_t vxb2 = wasm_i16x8_load8x8((const int8_t*) w + 16);
84
85 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
86 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
87 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
88 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
89 const v128_t vxb3 = wasm_i16x8_load8x8((const int8_t*) w + 24);
90
91 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
92 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 3, 3, 3, 3), vxb3));
93 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
94 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 3, 3, 3, 3), vxb3));
95
96 w = (const int8_t*) w + 32;
97 k -= 8 * sizeof(int8_t);
98 }
99 if (k != 0) {
100 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
101 a0 = (const int8_t*) ((uintptr_t) a0 + k);
102 const v128_t vxa1 = wasm_i16x8_load8x8(a1);
103 a1 = (const int8_t*) ((uintptr_t) a1 + k);
104
105 const v128_t vxb0 = wasm_i16x8_load8x8(w);
106 w = (const int8_t*) w + 8;
107
108 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
109 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
110 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
111 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
112
113 if (k > 2 * sizeof(int8_t)) {
114 const v128_t vxb1 = wasm_i16x8_load8x8(w);
115 w = (const int8_t*) w + 8;
116
117 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
118 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
119 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
120 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
121
122 if (k > 4 * sizeof(int8_t)) {
123 const v128_t vxb2 = wasm_i16x8_load8x8(w);
124 w = (const int8_t*) w + 8;
125
126 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
127 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
128 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
129 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
130 }
131 }
132 }
133
134 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
135 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
136
137 const v128_t vinput_scale0 = wasm_v128_load32_splat(&quantization_params[0].inv_scale);
138 const v128_t vinput_scale1 = wasm_v128_load32_splat(&quantization_params[1].inv_scale);
139
140 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale0);
141 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale1);
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 w = (const float*) w + 4;
147
148 const v128_t vbias0123 = wasm_v128_load(w);
149 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
150 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
151 w = (const float*) w + 4;
152
153 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
154 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
155
156 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
157 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
158
159 if XNN_LIKELY(nc >= 4) {
160 wasm_v128_store(c0, vacc0x0123);
161 wasm_v128_store(c1, vacc1x0123);
162
163 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
164 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
165
166 c0 = (float*) ((uintptr_t) c0 + cn_stride);
167 c1 = (float*) ((uintptr_t) c1 + cn_stride);
168
169 nc -= 4;
170 } else {
171 if (nc & 2) {
172 wasm_v128_store64_lane(c0, vacc0x0123, 0);
173 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
174 c0 += 2;
175 wasm_v128_store64_lane(c1, vacc1x0123, 0);
176 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
177 c1 += 2;
178 }
179 if (nc & 1) {
180 wasm_v128_store32_lane(c0, vacc0x0123, 0);
181 wasm_v128_store32_lane(c1, vacc1x0123, 0);
182 }
183 nc = 0;
184 }
185 } while (nc != 0);
186 }