Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-2x4c8-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_qc4w_gemm_minmax_ukernel_2x4c8__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 struct xnn_f32_qc4w_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 vmask = wasm_i8x16_const_splat(0xF0);
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 const v128_t vinit0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point0);
61 const v128_t vinit1x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point1);
62
63 const v128_t vzero = wasm_i32x4_const_splat(0);
64 v128_t vacc0x0 = wasm_v32x4_shuffle(vinit0x0123, vzero, 0, 5, 6, 7);
65 v128_t vacc0x1 = wasm_v32x4_shuffle(vinit0x0123, vzero, 4, 1, 6, 7);
66 v128_t vacc0x2 = wasm_v32x4_shuffle(vinit0x0123, vzero, 4, 5, 2, 7);
67 v128_t vacc0x3 = wasm_v32x4_shuffle(vinit0x0123, vzero, 4, 5, 6, 3);
68 v128_t vacc1x0 = wasm_v32x4_shuffle(vinit1x0123, vzero, 0, 5, 6, 7);
69 v128_t vacc1x1 = wasm_v32x4_shuffle(vinit1x0123, vzero, 4, 1, 6, 7);
70 v128_t vacc1x2 = wasm_v32x4_shuffle(vinit1x0123, vzero, 4, 5, 2, 7);
71 v128_t vacc1x3 = wasm_v32x4_shuffle(vinit1x0123, vzero, 4, 5, 6, 3);
72 w = (const int32_t*) w + 4;
73
74 size_t k = kc;
75 while (k >= 16 * sizeof(int8_t)) {
76 v128_t vxa0;
77 v128_t vxa1;
78 vxa0 = wasm_i16x8_load8x8(a0);
79 a0 += 8;
80 vxa1 = wasm_i16x8_load8x8(a1);
81 a1 += 8;
82 const v128_t vb0 = wasm_v128_load(w);
83 const v128_t vb2 = wasm_v128_load((const int8_t*) w + 16);
84 const v128_t vbm0 = wasm_i8x16_shl(vb0, 4);
85 const v128_t vbm2 = wasm_i8x16_shl(vb2, 4);
86 const v128_t vxb0 = wasm_i16x8_extend_low_i8x16(vbm0);
87 const v128_t vxb1 = wasm_i16x8_extend_high_i8x16(vbm0);
88 const v128_t vxb2 = wasm_i16x8_extend_low_i8x16(vbm2);
89 const v128_t vxb3 = wasm_i16x8_extend_high_i8x16(vbm2);
90 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_dot_i16x8(vxa0, vxb0));
91 vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_dot_i16x8(vxa1, vxb0));
92 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_dot_i16x8(vxa0, vxb1));
93 vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_dot_i16x8(vxa1, vxb1));
94 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_dot_i16x8(vxa0, vxb2));
95 vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_dot_i16x8(vxa1, vxb2));
96 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_dot_i16x8(vxa0, vxb3));
97 vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_dot_i16x8(vxa1, vxb3));
98 vxa0 = wasm_i16x8_load8x8(a0);
99 a0 += 8;
100 vxa1 = wasm_i16x8_load8x8(a1);
101 a1 += 8;
102 const v128_t vbm4 = wasm_v128_and(vb0, vmask);
103 const v128_t vbm6 = wasm_v128_and(vb2, vmask);
104 const v128_t vxb4 = wasm_i16x8_extend_low_i8x16(vbm4);
105 const v128_t vxb5 = wasm_i16x8_extend_high_i8x16(vbm4);
106 const v128_t vxb6 = wasm_i16x8_extend_low_i8x16(vbm6);
107 const v128_t vxb7 = wasm_i16x8_extend_high_i8x16(vbm6);
108 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_dot_i16x8(vxa0, vxb4));
109 vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_dot_i16x8(vxa1, vxb4));
110 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_dot_i16x8(vxa0, vxb5));
111 vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_dot_i16x8(vxa1, vxb5));
112 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_dot_i16x8(vxa0, vxb6));
113 vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_dot_i16x8(vxa1, vxb6));
114 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_dot_i16x8(vxa0, vxb7));
115 vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_dot_i16x8(vxa1, vxb7));
116 w = (const int8_t*) w + 32;
117 k -= 16 * sizeof(int8_t);
118 }
119
120 while (k >= 8 * sizeof(int8_t)) {
121 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
122 a0 += 8;
123 const v128_t vxa1 = wasm_i16x8_load8x8(a1);
124 a1 += 8;
125
126 const v128_t vb0 = wasm_v128_load(w);
127 const v128_t vb2 = wasm_v128_load((const int8_t*) w + 16);
128 const v128_t vbm0 = wasm_i8x16_shl(vb0, 4);
129 const v128_t vbm2 = wasm_i8x16_shl(vb2, 4);
130 const v128_t vxb0 = wasm_i16x8_extend_low_i8x16(vbm0);
131 const v128_t vxb1 = wasm_i16x8_extend_high_i8x16(vbm0);
132 const v128_t vxb2 = wasm_i16x8_extend_low_i8x16(vbm2);
133 const v128_t vxb3 = wasm_i16x8_extend_high_i8x16(vbm2);
134
135 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_dot_i16x8(vxa0, vxb0));
136 vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_dot_i16x8(vxa1, vxb0));
137
138 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_dot_i16x8(vxa0, vxb1));
139 vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_dot_i16x8(vxa1, vxb1));
140
141 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_dot_i16x8(vxa0, vxb2));
142 vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_dot_i16x8(vxa1, vxb2));
143
144 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_dot_i16x8(vxa0, vxb3));
145 vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_dot_i16x8(vxa1, vxb3));
146
147 w = (const int8_t*) w + 32; // only low nibble used
148 k -= 8 * sizeof(int8_t);
149 };
150
151 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));
152 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));
153 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));
154 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));
155
156 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
157 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
158
159 vacc0x0123 = wasm_i32x4_shr(vacc0x0123, 4);
160 vacc1x0123 = wasm_i32x4_shr(vacc1x0123, 4);
161 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
162 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
163
164 const v128_t vinput_scale0 = wasm_v128_load32_splat(&quantization_params[0].inv_scale);
165 const v128_t vinput_scale1 = wasm_v128_load32_splat(&quantization_params[1].inv_scale);
166
167 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale0);
168 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale1);
169
170 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
171 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
172 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
173 w = (const float*) w + 4;
174
175 const v128_t vbias0123 = wasm_v128_load(w);
176 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
177 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
178 w = (const float*) w + 4;
179
180 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
181 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
182
183 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
184 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
185
186 if XNN_LIKELY(nc >= 4) {
187 wasm_v128_store(c0, vacc0x0123);
188 wasm_v128_store(c1, vacc1x0123);
189
190 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
191 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
192
193 c0 = (float*) ((uintptr_t) c0 + cn_stride);
194 c1 = (float*) ((uintptr_t) c1 + cn_stride);
195
196 nc -= 4;
197 } else {
198 if (nc & 2) {
199 wasm_v128_store64_lane(c0, vacc0x0123, 0);
200 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
201 c0 += 2;
202 wasm_v128_store64_lane(c1, vacc1x0123, 0);
203 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
204 c1 += 2;
205 }
206 if (nc & 1) {
207 wasm_v128_store32_lane(c0, vacc0x0123, 0);
208 wasm_v128_store32_lane(c1, vacc1x0123, 0);
209 }
210 nc = 0;
211 }
212 } while (nc != 0);
213 }