1 // Auto-generated file. Do not edit!
2 // Template: src/f32-gemm/wasmsimd-loadsplat.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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
16
xnn_f32_gemminc_minmax_ukernel_6x8__wasmrelaxedsimd_loadsplat(size_t mr,size_t nc,size_t kc,const float * restrict a,size_t a_stride,const float * restrict w,float * restrict c,size_t cm_stride,size_t cn_stride,const float * restrict acc,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_gemminc_minmax_ukernel_6x8__wasmrelaxedsimd_loadsplat(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 const float*restrict a,
22 size_t a_stride,
23 const float*restrict w,
24 float*restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 const float*restrict acc,
28 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
29 {
30 assert(mr != 0);
31 assert(mr <= 6);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(float) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38 assert(acc != NULL);
39
40 const float* a0 = a;
41 float* c0 = c;
42 const float* a1 = (const float*) ((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 float* a2 = (const float*) ((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 const float* a3 = (const float*) ((uintptr_t) a2 + a_stride);
55 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr < 4) {
57 a3 = a2;
58 c3 = c2;
59 }
60 const float* a4 = (const float*) ((uintptr_t) a3 + a_stride);
61 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
62 if XNN_UNPREDICTABLE(mr <= 4) {
63 a4 = a3;
64 c4 = c3;
65 }
66 const float* a5 = (const float*) ((uintptr_t) a4 + a_stride);
67 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
68 if XNN_UNPREDICTABLE(mr != 6) {
69 a5 = a4;
70 c5 = c4;
71 }
72
73 const v128_t vmin = wasm_v128_load64_splat(params->wasmsimd.min);
74 const v128_t vmax = wasm_v128_load64_splat(params->wasmsimd.max);
75 do {
76 v128_t vacc0x0123 = wasm_v128_load(acc + 0);
77 v128_t vacc0x4567 = wasm_v128_load(acc + 4);
78 v128_t vacc1x0123 = wasm_v128_load(acc + 8);
79 v128_t vacc1x4567 = wasm_v128_load(acc + 12);
80 v128_t vacc2x0123 = wasm_v128_load(acc + 16);
81 v128_t vacc2x4567 = wasm_v128_load(acc + 20);
82 v128_t vacc3x0123 = wasm_v128_load(acc + 24);
83 v128_t vacc3x4567 = wasm_v128_load(acc + 28);
84 v128_t vacc4x0123 = wasm_v128_load(acc + 32);
85 v128_t vacc4x4567 = wasm_v128_load(acc + 36);
86 v128_t vacc5x0123 = wasm_v128_load(acc + 40);
87 v128_t vacc5x4567 = wasm_v128_load(acc + 44);
88 acc += 48;
89
90 size_t k = kc;
91 do {
92 const v128_t va0 = wasm_v128_load32_splat(a0);
93 a0 += 1;
94 const v128_t va1 = wasm_v128_load32_splat(a1);
95 a1 += 1;
96 const v128_t va2 = wasm_v128_load32_splat(a2);
97 a2 += 1;
98 const v128_t va3 = wasm_v128_load32_splat(a3);
99 a3 += 1;
100 const v128_t va4 = wasm_v128_load32_splat(a4);
101 a4 += 1;
102 const v128_t va5 = wasm_v128_load32_splat(a5);
103 a5 += 1;
104
105 const v128_t vb0123 = wasm_v128_load(w);
106 const v128_t vb4567 = wasm_v128_load(w + 4);
107 w += 8;
108
109 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123));
110 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123));
111 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123));
112 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123));
113 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123));
114 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(va5, vb0123));
115 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567));
116 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567));
117 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567));
118 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567));
119 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567));
120 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(va5, vb4567));
121
122 k -= sizeof(float);
123 } while (k != 0);
124
125 vacc0x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc0x0123);
126 vacc1x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc1x0123);
127 vacc2x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc2x0123);
128 vacc3x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc3x0123);
129 vacc4x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc4x0123);
130 vacc5x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc5x0123);
131 vacc0x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc0x4567);
132 vacc1x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc1x4567);
133 vacc2x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc2x4567);
134 vacc3x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc3x4567);
135 vacc4x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc4x4567);
136 vacc5x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc5x4567);
137
138 vacc0x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc0x0123);
139 vacc1x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc1x0123);
140 vacc2x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc2x0123);
141 vacc3x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc3x0123);
142 vacc4x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc4x0123);
143 vacc5x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc5x0123);
144 vacc0x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc0x4567);
145 vacc1x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc1x4567);
146 vacc2x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc2x4567);
147 vacc3x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc3x4567);
148 vacc4x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc4x4567);
149 vacc5x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc5x4567);
150
151 if XNN_LIKELY(nc >= 8) {
152 wasm_v128_store(c5, vacc5x0123);
153 wasm_v128_store(c5 + 4, vacc5x4567);
154 c5 = (float*) ((uintptr_t) c5 + cn_stride);
155 wasm_v128_store(c4, vacc4x0123);
156 wasm_v128_store(c4 + 4, vacc4x4567);
157 c4 = (float*) ((uintptr_t) c4 + cn_stride);
158 wasm_v128_store(c3, vacc3x0123);
159 wasm_v128_store(c3 + 4, vacc3x4567);
160 c3 = (float*) ((uintptr_t) c3 + cn_stride);
161 wasm_v128_store(c2, vacc2x0123);
162 wasm_v128_store(c2 + 4, vacc2x4567);
163 c2 = (float*) ((uintptr_t) c2 + cn_stride);
164 wasm_v128_store(c1, vacc1x0123);
165 wasm_v128_store(c1 + 4, vacc1x4567);
166 c1 = (float*) ((uintptr_t) c1 + cn_stride);
167 wasm_v128_store(c0, vacc0x0123);
168 wasm_v128_store(c0 + 4, vacc0x4567);
169 c0 = (float*) ((uintptr_t) c0 + cn_stride);
170
171 a5 = (const float*) ((uintptr_t) a5 - kc);
172 a4 = (const float*) ((uintptr_t) a4 - kc);
173 a3 = (const float*) ((uintptr_t) a3 - kc);
174 a2 = (const float*) ((uintptr_t) a2 - kc);
175 a1 = (const float*) ((uintptr_t) a1 - kc);
176 a0 = (const float*) ((uintptr_t) a0 - kc);
177
178 nc -= 8;
179 } else {
180 if (nc & 4) {
181 wasm_v128_store(c5, vacc5x0123);
182 wasm_v128_store(c4, vacc4x0123);
183 wasm_v128_store(c3, vacc3x0123);
184 wasm_v128_store(c2, vacc2x0123);
185 wasm_v128_store(c1, vacc1x0123);
186 wasm_v128_store(c0, vacc0x0123);
187
188 vacc5x0123 = vacc5x4567;
189 vacc4x0123 = vacc4x4567;
190 vacc3x0123 = vacc3x4567;
191 vacc2x0123 = vacc2x4567;
192 vacc1x0123 = vacc1x4567;
193 vacc0x0123 = vacc0x4567;
194
195 c5 += 4;
196 c4 += 4;
197 c3 += 4;
198 c2 += 4;
199 c1 += 4;
200 c0 += 4;
201 }
202 if (nc & 2) {
203 *((double*) c5) = wasm_f64x2_extract_lane(vacc5x0123, 0);
204 *((double*) c4) = wasm_f64x2_extract_lane(vacc4x0123, 0);
205 *((double*) c3) = wasm_f64x2_extract_lane(vacc3x0123, 0);
206 *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
207 *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
208 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
209
210 vacc5x0123 = wasm_v32x4_shuffle(vacc5x0123, vacc5x0123, 2, 3, 2, 3);
211 vacc4x0123 = wasm_v32x4_shuffle(vacc4x0123, vacc4x0123, 2, 3, 2, 3);
212 vacc3x0123 = wasm_v32x4_shuffle(vacc3x0123, vacc3x0123, 2, 3, 2, 3);
213 vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
214 vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
215 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
216
217 c5 += 2;
218 c4 += 2;
219 c3 += 2;
220 c2 += 2;
221 c1 += 2;
222 c0 += 2;
223 }
224 if (nc & 1) {
225 *c5 = wasm_f32x4_extract_lane(vacc5x0123, 0);
226 *c4 = wasm_f32x4_extract_lane(vacc4x0123, 0);
227 *c3 = wasm_f32x4_extract_lane(vacc3x0123, 0);
228 *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
229 *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
230 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
231 }
232
233 nc = 0;
234 }
235 } while (nc != 0);
236 }
237