1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-gemm/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/gemm.h>
15 
16 
xnn_f32_gemm_minmax_ukernel_4x8s4__wasmsimd_x86(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_4x8s4__wasmsimd_x86(
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 <= 4);
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 
59   do {
60     v128_t vacc0x0123 = wasm_v128_load(w + 0);
61     v128_t vacc0x4567 = wasm_v128_load(w + 4);
62     v128_t vacc1x0123 = vacc0x0123;
63     v128_t vacc1x4567 = vacc0x4567;
64     v128_t vacc2x0123 = vacc0x0123;
65     v128_t vacc2x4567 = vacc0x4567;
66     v128_t vacc3x0123 = vacc0x0123;
67     v128_t vacc3x4567 = vacc0x4567;
68     w += 8;
69 
70     size_t k = kc;
71     while (k >= 4 * sizeof(float)) {
72       v128_t va0 = wasm_v128_load(a0);
73       a0 += 4;
74       v128_t va1 = wasm_v128_load(a1);
75       a1 += 4;
76       v128_t va2 = wasm_v128_load(a2);
77       a2 += 4;
78       v128_t va3 = wasm_v128_load(a3);
79       a3 += 4;
80 
81 
82       const v128_t vb0123c0 = wasm_v128_load(w + 0);
83       const v128_t vb4567c0 = wasm_v128_load(w + 4);
84 
85       vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c0));
86       vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c0));
87       vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c0));
88       vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c0));
89       vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c0));
90       vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c0));
91       vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c0));
92       vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c0));
93 
94       va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
95       va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
96       va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
97       va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
98 
99       const v128_t vb0123c1 = wasm_v128_load(w + 8);
100       const v128_t vb4567c1 = wasm_v128_load(w + 12);
101 
102       vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c1));
103       vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c1));
104       vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c1));
105       vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c1));
106       vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c1));
107       vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c1));
108       vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c1));
109       vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c1));
110 
111       va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
112       va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
113       va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
114       va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
115 
116       const v128_t vb0123c2 = wasm_v128_load(w + 16);
117       const v128_t vb4567c2 = wasm_v128_load(w + 20);
118 
119       vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c2));
120       vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c2));
121       vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c2));
122       vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c2));
123       vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c2));
124       vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c2));
125       vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c2));
126       vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c2));
127 
128       va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
129       va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
130       va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
131       va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
132 
133       const v128_t vb0123c3 = wasm_v128_load(w + 24);
134       const v128_t vb4567c3 = wasm_v128_load(w + 28);
135 
136       vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c3));
137       vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c3));
138       vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c3));
139       vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c3));
140       vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c3));
141       vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c3));
142       vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c3));
143       vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c3));
144 
145 
146       w += 32;
147       k -= 4 * sizeof(float);
148     }
149     if XNN_UNLIKELY(k != 0) {
150       do {
151         const v128_t va0 = wasm_v32x4_load_splat(a0);
152         a0 += 1;
153         const v128_t va1 = wasm_v32x4_load_splat(a1);
154         a1 += 1;
155         const v128_t va2 = wasm_v32x4_load_splat(a2);
156         a2 += 1;
157         const v128_t va3 = wasm_v32x4_load_splat(a3);
158         a3 += 1;
159 
160         const v128_t vb0123 = wasm_v128_load(w);
161         const v128_t vb4567 = wasm_v128_load(w + 4);
162         w += 8;
163 
164         vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123));
165         vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123));
166         vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123));
167         vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123));
168         vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567));
169         vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567));
170         vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567));
171         vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567));
172 
173         k -= sizeof(float);
174       } while (k != 0);
175     }
176 
177     const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
178     vacc0x0123 = wasm_v128_bitselect(vmin, vacc0x0123, wasm_f32x4_lt(vacc0x0123, vmin));
179     vacc1x0123 = wasm_v128_bitselect(vmin, vacc1x0123, wasm_f32x4_lt(vacc1x0123, vmin));
180     vacc2x0123 = wasm_v128_bitselect(vmin, vacc2x0123, wasm_f32x4_lt(vacc2x0123, vmin));
181     vacc3x0123 = wasm_v128_bitselect(vmin, vacc3x0123, wasm_f32x4_lt(vacc3x0123, vmin));
182     vacc0x4567 = wasm_v128_bitselect(vmin, vacc0x4567, wasm_f32x4_lt(vacc0x4567, vmin));
183     vacc1x4567 = wasm_v128_bitselect(vmin, vacc1x4567, wasm_f32x4_lt(vacc1x4567, vmin));
184     vacc2x4567 = wasm_v128_bitselect(vmin, vacc2x4567, wasm_f32x4_lt(vacc2x4567, vmin));
185     vacc3x4567 = wasm_v128_bitselect(vmin, vacc3x4567, wasm_f32x4_lt(vacc3x4567, vmin));
186 
187     const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
188     vacc0x0123 = wasm_v128_bitselect(vacc0x0123, vmax, wasm_f32x4_le(vacc0x0123, vmax));
189     vacc1x0123 = wasm_v128_bitselect(vacc1x0123, vmax, wasm_f32x4_le(vacc1x0123, vmax));
190     vacc2x0123 = wasm_v128_bitselect(vacc2x0123, vmax, wasm_f32x4_le(vacc2x0123, vmax));
191     vacc3x0123 = wasm_v128_bitselect(vacc3x0123, vmax, wasm_f32x4_le(vacc3x0123, vmax));
192     vacc0x4567 = wasm_v128_bitselect(vacc0x4567, vmax, wasm_f32x4_le(vacc0x4567, vmax));
193     vacc1x4567 = wasm_v128_bitselect(vacc1x4567, vmax, wasm_f32x4_le(vacc1x4567, vmax));
194     vacc2x4567 = wasm_v128_bitselect(vacc2x4567, vmax, wasm_f32x4_le(vacc2x4567, vmax));
195     vacc3x4567 = wasm_v128_bitselect(vacc3x4567, vmax, wasm_f32x4_le(vacc3x4567, vmax));
196 
197     if XNN_LIKELY(nc >= 8) {
198       wasm_v128_store(c3, vacc3x0123);
199       wasm_v128_store(c3 + 4, vacc3x4567);
200       c3 = (float*) ((uintptr_t) c3 + cn_stride);
201       wasm_v128_store(c2, vacc2x0123);
202       wasm_v128_store(c2 + 4, vacc2x4567);
203       c2 = (float*) ((uintptr_t) c2 + cn_stride);
204       wasm_v128_store(c1, vacc1x0123);
205       wasm_v128_store(c1 + 4, vacc1x4567);
206       c1 = (float*) ((uintptr_t) c1 + cn_stride);
207       wasm_v128_store(c0, vacc0x0123);
208       wasm_v128_store(c0 + 4, vacc0x4567);
209       c0 = (float*) ((uintptr_t) c0 + cn_stride);
210 
211       a3 = (const float*) ((uintptr_t) a3 - kc);
212       a2 = (const float*) ((uintptr_t) a2 - kc);
213       a1 = (const float*) ((uintptr_t) a1 - kc);
214       a0 = (const float*) ((uintptr_t) a0 - kc);
215 
216       nc -= 8;
217     } else {
218       if (nc & 4) {
219         wasm_v128_store(c3, vacc3x0123);
220         wasm_v128_store(c2, vacc2x0123);
221         wasm_v128_store(c1, vacc1x0123);
222         wasm_v128_store(c0, vacc0x0123);
223 
224         vacc3x0123 = vacc3x4567;
225         vacc2x0123 = vacc2x4567;
226         vacc1x0123 = vacc1x4567;
227         vacc0x0123 = vacc0x4567;
228 
229         c3 += 4;
230         c2 += 4;
231         c1 += 4;
232         c0 += 4;
233       }
234       if (nc & 2) {
235         *((double*) c3) = wasm_f64x2_extract_lane(vacc3x0123, 0);
236         *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
237         *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
238         *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
239 
240         vacc3x0123 = wasm_v32x4_shuffle(vacc3x0123, vacc3x0123, 2, 3, 2, 3);
241         vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
242         vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
243         vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
244 
245         c3 += 2;
246         c2 += 2;
247         c1 += 2;
248         c0 += 2;
249       }
250       if (nc & 1) {
251         *c3 = wasm_f32x4_extract_lane(vacc3x0123, 0);
252         *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
253         *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
254         *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
255       }
256 
257       nc = 0;
258     }
259   } while (nc != 0);
260 }
261