1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-igemm/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/igemm.h>
15 
16 
xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat(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_6x8__wasmsimd_x86_loadsplat(
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 <= 6);
33   assert(nc != 0);
34   assert(kc != 0);
35   assert(kc % sizeof(float) == 0);
36   assert(ks != 0);
37   assert(ks % (6 * 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   float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
53   if XNN_UNPREDICTABLE(mr < 4) {
54     c3 = c2;
55   }
56   float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
57   if XNN_UNPREDICTABLE(mr <= 4) {
58     c4 = c3;
59   }
60   float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
61   if XNN_UNPREDICTABLE(mr != 6) {
62     c5 = c4;
63   }
64 
65   do {
66     v128_t vacc0x0123 = wasm_v128_load(w);
67     v128_t vacc0x4567 = wasm_v128_load(w + 4);
68     v128_t vacc1x0123 = vacc0x0123;
69     v128_t vacc1x4567 = vacc0x4567;
70     v128_t vacc2x0123 = vacc0x0123;
71     v128_t vacc2x4567 = vacc0x4567;
72     v128_t vacc3x0123 = vacc0x0123;
73     v128_t vacc3x4567 = vacc0x4567;
74     v128_t vacc4x0123 = vacc0x0123;
75     v128_t vacc4x4567 = vacc0x4567;
76     v128_t vacc5x0123 = vacc0x0123;
77     v128_t vacc5x4567 = vacc0x4567;
78     w += 8;
79 
80     size_t p = ks;
81     do {
82       const float* restrict a0 = a[0];
83       assert(a0 != NULL);
84       if XNN_UNPREDICTABLE(a0 != zero) {
85         a0 = (const float*) ((uintptr_t) a0 + a_offset);
86       }
87       const float* restrict a1 = a[1];
88       assert(a1 != NULL);
89       if XNN_UNPREDICTABLE(a1 != zero) {
90         a1 = (const float*) ((uintptr_t) a1 + a_offset);
91       }
92       const float* restrict a2 = a[2];
93       assert(a2 != NULL);
94       if XNN_UNPREDICTABLE(a2 != zero) {
95         a2 = (const float*) ((uintptr_t) a2 + a_offset);
96       }
97       const float* restrict a3 = a[3];
98       assert(a3 != NULL);
99       if XNN_UNPREDICTABLE(a3 != zero) {
100         a3 = (const float*) ((uintptr_t) a3 + a_offset);
101       }
102       const float* restrict a4 = a[4];
103       assert(a4 != NULL);
104       if XNN_UNPREDICTABLE(a4 != zero) {
105         a4 = (const float*) ((uintptr_t) a4 + a_offset);
106       }
107       const float* restrict a5 = a[5];
108       assert(a5 != NULL);
109       if XNN_UNPREDICTABLE(a5 != zero) {
110         a5 = (const float*) ((uintptr_t) a5 + a_offset);
111       }
112       a += 6;
113 
114       size_t k = kc;
115       do {
116         const v128_t vb0123 = wasm_v128_load(w);
117         const v128_t vb4567 = wasm_v128_load(w + 4);
118         w += 8;
119 
120         const v128_t va0 = wasm_v32x4_load_splat(a0);
121         a0 += 1;
122         const v128_t va1 = wasm_v32x4_load_splat(a1);
123         a1 += 1;
124         const v128_t va2 = wasm_v32x4_load_splat(a2);
125         a2 += 1;
126         const v128_t va3 = wasm_v32x4_load_splat(a3);
127         a3 += 1;
128         const v128_t va4 = wasm_v32x4_load_splat(a4);
129         a4 += 1;
130         const v128_t va5 = wasm_v32x4_load_splat(a5);
131         a5 += 1;
132 
133         vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123));
134         vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567));
135         vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123));
136         vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567));
137         vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123));
138         vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567));
139         vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123));
140         vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567));
141         vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123));
142         vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567));
143         vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(va5, vb0123));
144         vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(va5, vb4567));
145         k -= sizeof(float);
146       } while (k != 0);
147       p -= 6 * sizeof(void*);
148     } while (p != 0);
149 
150     const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
151     vacc0x0123 = wasm_v128_bitselect(vmin, vacc0x0123, wasm_f32x4_lt(vacc0x0123, vmin));
152     vacc1x0123 = wasm_v128_bitselect(vmin, vacc1x0123, wasm_f32x4_lt(vacc1x0123, vmin));
153     vacc2x0123 = wasm_v128_bitselect(vmin, vacc2x0123, wasm_f32x4_lt(vacc2x0123, vmin));
154     vacc3x0123 = wasm_v128_bitselect(vmin, vacc3x0123, wasm_f32x4_lt(vacc3x0123, vmin));
155     vacc4x0123 = wasm_v128_bitselect(vmin, vacc4x0123, wasm_f32x4_lt(vacc4x0123, vmin));
156     vacc5x0123 = wasm_v128_bitselect(vmin, vacc5x0123, wasm_f32x4_lt(vacc5x0123, vmin));
157     vacc0x4567 = wasm_v128_bitselect(vmin, vacc0x4567, wasm_f32x4_lt(vacc0x4567, vmin));
158     vacc1x4567 = wasm_v128_bitselect(vmin, vacc1x4567, wasm_f32x4_lt(vacc1x4567, vmin));
159     vacc2x4567 = wasm_v128_bitselect(vmin, vacc2x4567, wasm_f32x4_lt(vacc2x4567, vmin));
160     vacc3x4567 = wasm_v128_bitselect(vmin, vacc3x4567, wasm_f32x4_lt(vacc3x4567, vmin));
161     vacc4x4567 = wasm_v128_bitselect(vmin, vacc4x4567, wasm_f32x4_lt(vacc4x4567, vmin));
162     vacc5x4567 = wasm_v128_bitselect(vmin, vacc5x4567, wasm_f32x4_lt(vacc5x4567, vmin));
163 
164     const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
165     vacc0x0123 = wasm_v128_bitselect(vacc0x0123, vmax, wasm_f32x4_le(vacc0x0123, vmax));
166     vacc1x0123 = wasm_v128_bitselect(vacc1x0123, vmax, wasm_f32x4_le(vacc1x0123, vmax));
167     vacc2x0123 = wasm_v128_bitselect(vacc2x0123, vmax, wasm_f32x4_le(vacc2x0123, vmax));
168     vacc3x0123 = wasm_v128_bitselect(vacc3x0123, vmax, wasm_f32x4_le(vacc3x0123, vmax));
169     vacc4x0123 = wasm_v128_bitselect(vacc4x0123, vmax, wasm_f32x4_le(vacc4x0123, vmax));
170     vacc5x0123 = wasm_v128_bitselect(vacc5x0123, vmax, wasm_f32x4_le(vacc5x0123, vmax));
171     vacc0x4567 = wasm_v128_bitselect(vacc0x4567, vmax, wasm_f32x4_le(vacc0x4567, vmax));
172     vacc1x4567 = wasm_v128_bitselect(vacc1x4567, vmax, wasm_f32x4_le(vacc1x4567, vmax));
173     vacc2x4567 = wasm_v128_bitselect(vacc2x4567, vmax, wasm_f32x4_le(vacc2x4567, vmax));
174     vacc3x4567 = wasm_v128_bitselect(vacc3x4567, vmax, wasm_f32x4_le(vacc3x4567, vmax));
175     vacc4x4567 = wasm_v128_bitselect(vacc4x4567, vmax, wasm_f32x4_le(vacc4x4567, vmax));
176     vacc5x4567 = wasm_v128_bitselect(vacc5x4567, vmax, wasm_f32x4_le(vacc5x4567, vmax));
177 
178     if XNN_LIKELY(nc >= 8) {
179       wasm_v128_store(c5, vacc5x0123);
180       wasm_v128_store(c5 + 4, vacc5x4567);
181       c5 = (float*) ((uintptr_t) c5 + cn_stride);
182       wasm_v128_store(c4, vacc4x0123);
183       wasm_v128_store(c4 + 4, vacc4x4567);
184       c4 = (float*) ((uintptr_t) c4 + cn_stride);
185       wasm_v128_store(c3, vacc3x0123);
186       wasm_v128_store(c3 + 4, vacc3x4567);
187       c3 = (float*) ((uintptr_t) c3 + cn_stride);
188       wasm_v128_store(c2, vacc2x0123);
189       wasm_v128_store(c2 + 4, vacc2x4567);
190       c2 = (float*) ((uintptr_t) c2 + cn_stride);
191       wasm_v128_store(c1, vacc1x0123);
192       wasm_v128_store(c1 + 4, vacc1x4567);
193       c1 = (float*) ((uintptr_t) c1 + cn_stride);
194       wasm_v128_store(c0, vacc0x0123);
195       wasm_v128_store(c0 + 4, vacc0x4567);
196       c0 = (float*) ((uintptr_t) c0 + cn_stride);
197 
198       a = (const float**restrict) ((uintptr_t) a - ks);
199       nc -= 8;
200     } else {
201       if (nc & 4) {
202         wasm_v128_store(c5, vacc5x0123);
203         wasm_v128_store(c4, vacc4x0123);
204         wasm_v128_store(c3, vacc3x0123);
205         wasm_v128_store(c2, vacc2x0123);
206         wasm_v128_store(c1, vacc1x0123);
207         wasm_v128_store(c0, vacc0x0123);
208 
209         vacc5x0123 = vacc5x4567;
210         vacc4x0123 = vacc4x4567;
211         vacc3x0123 = vacc3x4567;
212         vacc2x0123 = vacc2x4567;
213         vacc1x0123 = vacc1x4567;
214         vacc0x0123 = vacc0x4567;
215 
216         c5 += 4;
217         c4 += 4;
218         c3 += 4;
219         c2 += 4;
220         c1 += 4;
221         c0 += 4;
222       }
223       if (nc & 2) {
224         *((double*) c5) = wasm_f64x2_extract_lane(vacc5x0123, 0);
225         *((double*) c4) = wasm_f64x2_extract_lane(vacc4x0123, 0);
226         *((double*) c3) = wasm_f64x2_extract_lane(vacc3x0123, 0);
227         *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
228         *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
229         *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
230 
231         vacc5x0123 = wasm_v32x4_shuffle(vacc5x0123, vacc5x0123, 2, 3, 2, 3);
232         vacc4x0123 = wasm_v32x4_shuffle(vacc4x0123, vacc4x0123, 2, 3, 2, 3);
233         vacc3x0123 = wasm_v32x4_shuffle(vacc3x0123, vacc3x0123, 2, 3, 2, 3);
234         vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
235         vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
236         vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
237 
238         c5 += 2;
239         c4 += 2;
240         c3 += 2;
241         c2 += 2;
242         c1 += 2;
243         c0 += 2;
244       }
245       if (nc & 1) {
246         *c5 = wasm_f32x4_extract_lane(vacc5x0123, 0);
247         *c4 = wasm_f32x4_extract_lane(vacc4x0123, 0);
248         *c3 = wasm_f32x4_extract_lane(vacc3x0123, 0);
249         *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
250         *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
251         *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
252       }
253 
254       nc = 0;
255     }
256   } while (nc != 0);
257 }
258