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_gemm_minmax_ukernel_6x8__wasmsimd_arm_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 union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_arm_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 union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29 assert(mr != 0);
30 assert(mr <= 6);
31 assert(nc != 0);
32 assert(kc != 0);
33 assert(kc % sizeof(float) == 0);
34 assert(a != NULL);
35 assert(w != NULL);
36 assert(c != NULL);
37
38 const float* a0 = a;
39 float* c0 = c;
40 const float* a1 = (const float*) ((uintptr_t) a0 + a_stride);
41 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
42 if XNN_UNPREDICTABLE(mr < 2) {
43 a1 = a0;
44 c1 = c0;
45 }
46 const float* a2 = (const float*) ((uintptr_t) a1 + a_stride);
47 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
48 if XNN_UNPREDICTABLE(mr <= 2) {
49 a2 = a1;
50 c2 = c1;
51 }
52 const float* a3 = (const float*) ((uintptr_t) a2 + a_stride);
53 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
54 if XNN_UNPREDICTABLE(mr < 4) {
55 a3 = a2;
56 c3 = c2;
57 }
58 const float* a4 = (const float*) ((uintptr_t) a3 + a_stride);
59 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
60 if XNN_UNPREDICTABLE(mr <= 4) {
61 a4 = a3;
62 c4 = c3;
63 }
64 const float* a5 = (const float*) ((uintptr_t) a4 + a_stride);
65 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
66 if XNN_UNPREDICTABLE(mr != 6) {
67 a5 = a4;
68 c5 = c4;
69 }
70
71 const v128_t vmin = wasm_v128_load64_splat(params->wasmsimd.min);
72 const v128_t vmax = wasm_v128_load64_splat(params->wasmsimd.max);
73 do {
74 v128_t vacc0x0123 = wasm_v128_load(w + 0);
75 v128_t vacc0x4567 = wasm_v128_load(w + 4);
76 v128_t vacc1x0123 = vacc0x0123;
77 v128_t vacc1x4567 = vacc0x4567;
78 v128_t vacc2x0123 = vacc0x0123;
79 v128_t vacc2x4567 = vacc0x4567;
80 v128_t vacc3x0123 = vacc0x0123;
81 v128_t vacc3x4567 = vacc0x4567;
82 v128_t vacc4x0123 = vacc0x0123;
83 v128_t vacc4x4567 = vacc0x4567;
84 v128_t vacc5x0123 = vacc0x0123;
85 v128_t vacc5x4567 = vacc0x4567;
86 w += 8;
87
88 size_t k = kc;
89 do {
90 const v128_t va0 = wasm_v128_load32_splat(a0);
91 a0 += 1;
92 const v128_t va1 = wasm_v128_load32_splat(a1);
93 a1 += 1;
94 const v128_t va2 = wasm_v128_load32_splat(a2);
95 a2 += 1;
96 const v128_t va3 = wasm_v128_load32_splat(a3);
97 a3 += 1;
98 const v128_t va4 = wasm_v128_load32_splat(a4);
99 a4 += 1;
100 const v128_t va5 = wasm_v128_load32_splat(a5);
101 a5 += 1;
102
103 const v128_t vb0123 = wasm_v128_load(w);
104 const v128_t vb4567 = wasm_v128_load(w + 4);
105 w += 8;
106
107 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123));
108 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123));
109 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123));
110 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123));
111 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123));
112 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(va5, vb0123));
113 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567));
114 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567));
115 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567));
116 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567));
117 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567));
118 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(va5, vb4567));
119
120 k -= sizeof(float);
121 } while (k != 0);
122
123 vacc0x0123 = wasm_f32x4_max(vmin, vacc0x0123);
124 vacc1x0123 = wasm_f32x4_max(vmin, vacc1x0123);
125 vacc2x0123 = wasm_f32x4_max(vmin, vacc2x0123);
126 vacc3x0123 = wasm_f32x4_max(vmin, vacc3x0123);
127 vacc4x0123 = wasm_f32x4_max(vmin, vacc4x0123);
128 vacc5x0123 = wasm_f32x4_max(vmin, vacc5x0123);
129 vacc0x4567 = wasm_f32x4_max(vmin, vacc0x4567);
130 vacc1x4567 = wasm_f32x4_max(vmin, vacc1x4567);
131 vacc2x4567 = wasm_f32x4_max(vmin, vacc2x4567);
132 vacc3x4567 = wasm_f32x4_max(vmin, vacc3x4567);
133 vacc4x4567 = wasm_f32x4_max(vmin, vacc4x4567);
134 vacc5x4567 = wasm_f32x4_max(vmin, vacc5x4567);
135
136 vacc0x0123 = wasm_f32x4_min(vmax, vacc0x0123);
137 vacc1x0123 = wasm_f32x4_min(vmax, vacc1x0123);
138 vacc2x0123 = wasm_f32x4_min(vmax, vacc2x0123);
139 vacc3x0123 = wasm_f32x4_min(vmax, vacc3x0123);
140 vacc4x0123 = wasm_f32x4_min(vmax, vacc4x0123);
141 vacc5x0123 = wasm_f32x4_min(vmax, vacc5x0123);
142 vacc0x4567 = wasm_f32x4_min(vmax, vacc0x4567);
143 vacc1x4567 = wasm_f32x4_min(vmax, vacc1x4567);
144 vacc2x4567 = wasm_f32x4_min(vmax, vacc2x4567);
145 vacc3x4567 = wasm_f32x4_min(vmax, vacc3x4567);
146 vacc4x4567 = wasm_f32x4_min(vmax, vacc4x4567);
147 vacc5x4567 = wasm_f32x4_min(vmax, vacc5x4567);
148
149 if XNN_LIKELY(nc >= 8) {
150 wasm_v128_store(c5, vacc5x0123);
151 wasm_v128_store(c5 + 4, vacc5x4567);
152 c5 = (float*) ((uintptr_t) c5 + cn_stride);
153 wasm_v128_store(c4, vacc4x0123);
154 wasm_v128_store(c4 + 4, vacc4x4567);
155 c4 = (float*) ((uintptr_t) c4 + cn_stride);
156 wasm_v128_store(c3, vacc3x0123);
157 wasm_v128_store(c3 + 4, vacc3x4567);
158 c3 = (float*) ((uintptr_t) c3 + cn_stride);
159 wasm_v128_store(c2, vacc2x0123);
160 wasm_v128_store(c2 + 4, vacc2x4567);
161 c2 = (float*) ((uintptr_t) c2 + cn_stride);
162 wasm_v128_store(c1, vacc1x0123);
163 wasm_v128_store(c1 + 4, vacc1x4567);
164 c1 = (float*) ((uintptr_t) c1 + cn_stride);
165 wasm_v128_store(c0, vacc0x0123);
166 wasm_v128_store(c0 + 4, vacc0x4567);
167 c0 = (float*) ((uintptr_t) c0 + cn_stride);
168
169 a5 = (const float*) ((uintptr_t) a5 - kc);
170 a4 = (const float*) ((uintptr_t) a4 - kc);
171 a3 = (const float*) ((uintptr_t) a3 - kc);
172 a2 = (const float*) ((uintptr_t) a2 - kc);
173 a1 = (const float*) ((uintptr_t) a1 - kc);
174 a0 = (const float*) ((uintptr_t) a0 - kc);
175
176 nc -= 8;
177 } else {
178 if (nc & 4) {
179 wasm_v128_store(c5, vacc5x0123);
180 wasm_v128_store(c4, vacc4x0123);
181 wasm_v128_store(c3, vacc3x0123);
182 wasm_v128_store(c2, vacc2x0123);
183 wasm_v128_store(c1, vacc1x0123);
184 wasm_v128_store(c0, vacc0x0123);
185
186 vacc5x0123 = vacc5x4567;
187 vacc4x0123 = vacc4x4567;
188 vacc3x0123 = vacc3x4567;
189 vacc2x0123 = vacc2x4567;
190 vacc1x0123 = vacc1x4567;
191 vacc0x0123 = vacc0x4567;
192
193 c5 += 4;
194 c4 += 4;
195 c3 += 4;
196 c2 += 4;
197 c1 += 4;
198 c0 += 4;
199 }
200 if (nc & 2) {
201 *((double*) c5) = wasm_f64x2_extract_lane(vacc5x0123, 0);
202 *((double*) c4) = wasm_f64x2_extract_lane(vacc4x0123, 0);
203 *((double*) c3) = wasm_f64x2_extract_lane(vacc3x0123, 0);
204 *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
205 *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
206 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
207
208 vacc5x0123 = wasm_v32x4_shuffle(vacc5x0123, vacc5x0123, 2, 3, 2, 3);
209 vacc4x0123 = wasm_v32x4_shuffle(vacc4x0123, vacc4x0123, 2, 3, 2, 3);
210 vacc3x0123 = wasm_v32x4_shuffle(vacc3x0123, vacc3x0123, 2, 3, 2, 3);
211 vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
212 vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
213 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
214
215 c5 += 2;
216 c4 += 2;
217 c3 += 2;
218 c2 += 2;
219 c1 += 2;
220 c0 += 2;
221 }
222 if (nc & 1) {
223 *c5 = wasm_f32x4_extract_lane(vacc5x0123, 0);
224 *c4 = wasm_f32x4_extract_lane(vacc4x0123, 0);
225 *c3 = wasm_f32x4_extract_lane(vacc3x0123, 0);
226 *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
227 *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
228 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
229 }
230
231 nc = 0;
232 }
233 } while (nc != 0);
234 }
235