• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-igemm/wasmsimd-s4.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/igemm.h>
15 
16 
xnn_f32_igemm_minmax_ukernel_3x8s4__wasmsimd_arm(size_t mr,size_t nc,size_t kc,size_t ks,const float ** restrict a,const float * restrict w,float * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const float * zero,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_igemm_minmax_ukernel_3x8s4__wasmsimd_arm(
18     size_t mr,
19     size_t nc,
20     size_t kc,
21     size_t ks,
22     const float**restrict a,
23     const float*restrict w,
24     float*restrict c,
25     size_t cm_stride,
26     size_t cn_stride,
27     size_t a_offset,
28     const float* zero,
29     const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
30 {
31   assert(mr != 0);
32   assert(mr <= 3);
33   assert(nc != 0);
34   assert(kc != 0);
35   assert(kc % sizeof(float) == 0);
36   assert(ks != 0);
37   assert(ks % (3 * sizeof(void*)) == 0);
38   assert(a_offset % sizeof(float) == 0);
39   assert(a != NULL);
40   assert(w != NULL);
41   assert(c != NULL);
42 
43   float* c0 = c;
44   float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
45   if XNN_UNPREDICTABLE(mr < 2) {
46     c1 = c0;
47   }
48   float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
49   if XNN_UNPREDICTABLE(mr <= 2) {
50     c2 = c1;
51   }
52 
53   const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
54   const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
55   do {
56     v128_t vacc0x0123 = wasm_v128_load(w);
57     v128_t vacc0x4567 = wasm_v128_load(w + 4);
58     v128_t vacc1x0123 = vacc0x0123;
59     v128_t vacc1x4567 = vacc0x4567;
60     v128_t vacc2x0123 = vacc0x0123;
61     v128_t vacc2x4567 = vacc0x4567;
62     w += 8;
63 
64     size_t p = ks;
65     do {
66       const float* restrict a0 = a[0];
67       assert(a0 != NULL);
68       if XNN_UNPREDICTABLE(a0 != zero) {
69         a0 = (const float*) ((uintptr_t) a0 + a_offset);
70       }
71       const float* restrict a1 = a[1];
72       assert(a1 != NULL);
73       if XNN_UNPREDICTABLE(a1 != zero) {
74         a1 = (const float*) ((uintptr_t) a1 + a_offset);
75       }
76       const float* restrict a2 = a[2];
77       assert(a2 != NULL);
78       if XNN_UNPREDICTABLE(a2 != zero) {
79         a2 = (const float*) ((uintptr_t) a2 + a_offset);
80       }
81       a += 3;
82 
83       size_t k = kc;
84       while (k >= 4 * sizeof(float)) {
85         v128_t va0 = wasm_v128_load(a0);
86         a0 += 4;
87         v128_t va1 = wasm_v128_load(a1);
88         a1 += 4;
89         v128_t va2 = wasm_v128_load(a2);
90         a2 += 4;
91 
92 
93         const v128_t vb0123c0 = wasm_v128_load(w + 0);
94         const v128_t vb4567c0 = wasm_v128_load(w + 4);
95 
96         vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c0));
97         vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c0));
98         vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c0));
99         vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c0));
100         vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c0));
101         vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c0));
102 
103         va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
104         va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
105         va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
106 
107         const v128_t vb0123c1 = wasm_v128_load(w + 8);
108         const v128_t vb4567c1 = wasm_v128_load(w + 12);
109 
110         vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c1));
111         vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c1));
112         vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c1));
113         vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c1));
114         vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c1));
115         vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c1));
116 
117         va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
118         va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
119         va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
120 
121         const v128_t vb0123c2 = wasm_v128_load(w + 16);
122         const v128_t vb4567c2 = wasm_v128_load(w + 20);
123 
124         vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c2));
125         vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c2));
126         vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c2));
127         vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c2));
128         vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c2));
129         vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c2));
130 
131         va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
132         va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
133         va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
134 
135         const v128_t vb0123c3 = wasm_v128_load(w + 24);
136         const v128_t vb4567c3 = wasm_v128_load(w + 28);
137 
138         vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c3));
139         vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c3));
140         vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c3));
141         vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c3));
142         vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c3));
143         vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c3));
144 
145 
146         w += 32;
147         k -= 4 * sizeof(float);
148       }
149       if XNN_UNLIKELY(k != 0) {
150         do {
151           const v128_t vb0123 = wasm_v128_load(w);
152           const v128_t vb4567 = wasm_v128_load(w + 4);
153           w += 8;
154 
155           const v128_t va0 = wasm_v32x4_load_splat(a0);
156           a0 += 1;
157           const v128_t va1 = wasm_v32x4_load_splat(a1);
158           a1 += 1;
159           const v128_t va2 = wasm_v32x4_load_splat(a2);
160           a2 += 1;
161 
162           vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123));
163           vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567));
164           vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123));
165           vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567));
166           vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123));
167           vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567));
168           k -= sizeof(float);
169         } while (k != 0);
170       }
171       p -= 3 * sizeof(void*);
172     } while (p != 0);
173 
174     vacc0x0123 = wasm_f32x4_max(vacc0x0123, vmin);
175     vacc1x0123 = wasm_f32x4_max(vacc1x0123, vmin);
176     vacc2x0123 = wasm_f32x4_max(vacc2x0123, vmin);
177     vacc0x4567 = wasm_f32x4_max(vacc0x4567, vmin);
178     vacc1x4567 = wasm_f32x4_max(vacc1x4567, vmin);
179     vacc2x4567 = wasm_f32x4_max(vacc2x4567, vmin);
180 
181     vacc0x0123 = wasm_f32x4_min(vacc0x0123, vmax);
182     vacc1x0123 = wasm_f32x4_min(vacc1x0123, vmax);
183     vacc2x0123 = wasm_f32x4_min(vacc2x0123, vmax);
184     vacc0x4567 = wasm_f32x4_min(vacc0x4567, vmax);
185     vacc1x4567 = wasm_f32x4_min(vacc1x4567, vmax);
186     vacc2x4567 = wasm_f32x4_min(vacc2x4567, vmax);
187 
188     if XNN_LIKELY(nc >= 8) {
189       wasm_v128_store(c2, vacc2x0123);
190       wasm_v128_store(c2 + 4, vacc2x4567);
191       c2 = (float*) ((uintptr_t) c2 + cn_stride);
192       wasm_v128_store(c1, vacc1x0123);
193       wasm_v128_store(c1 + 4, vacc1x4567);
194       c1 = (float*) ((uintptr_t) c1 + cn_stride);
195       wasm_v128_store(c0, vacc0x0123);
196       wasm_v128_store(c0 + 4, vacc0x4567);
197       c0 = (float*) ((uintptr_t) c0 + cn_stride);
198 
199       a = (const float**restrict) ((uintptr_t) a - ks);
200       nc -= 8;
201     } else {
202       if (nc & 4) {
203         wasm_v128_store(c2, vacc2x0123);
204         wasm_v128_store(c1, vacc1x0123);
205         wasm_v128_store(c0, vacc0x0123);
206 
207         vacc2x0123 = vacc2x4567;
208         vacc1x0123 = vacc1x4567;
209         vacc0x0123 = vacc0x4567;
210 
211         c2 += 4;
212         c1 += 4;
213         c0 += 4;
214       }
215       if (nc & 2) {
216         *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
217         *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
218         *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
219 
220         vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
221         vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
222         vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
223 
224         c2 += 2;
225         c1 += 2;
226         c0 += 2;
227       }
228       if (nc & 1) {
229         *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
230         *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
231         *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
232       }
233 
234       nc = 0;
235     }
236   } while (nc != 0);
237 }
238