1 // Auto-generated file. Do not edit!
2 // Template: src/f32-spmm/wasmsimd.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/spmm.h>
15
16
xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86_x4(size_t mc,size_t nc,const float * restrict input,const float * restrict weights,const int32_t * restrict widx_dmap,const uint32_t * restrict nidx_nnzmap,float * restrict output,size_t output_stride,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86_x4(
18 size_t mc,
19 size_t nc,
20 const float*restrict input,
21 const float*restrict weights,
22 const int32_t*restrict widx_dmap,
23 const uint32_t*restrict nidx_nnzmap,
24 float*restrict output,
25 size_t output_stride,
26 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
27 {
28 assert(mc != 0);
29 assert(mc % sizeof(float) == 0);
30 assert(nc != 0);
31
32 const v128_t vmin = wasm_v32x4_load_splat(¶ms->scalar.min);
33 const v128_t vmax = wasm_v32x4_load_splat(¶ms->scalar.max);
34 const v128_t vzero = wasm_f64x2_splat(0.0);
35 size_t output_decrement = output_stride * nc - 16 * sizeof(float);
36 while XNN_LIKELY(mc >= 16 * sizeof(float)) {
37 const float*restrict w = weights;
38 const int32_t* dmap = widx_dmap;
39 const uint32_t* nnzmap = nidx_nnzmap;
40 size_t n = nc;
41 do {
42 uint32_t nnz = *nnzmap++;
43 v128_t vacc0123x0 = wasm_v32x4_load_splat(w);
44 w += 1;
45 v128_t vacc0123x1 = vzero;
46 v128_t vacc0123x2 = vzero;
47 v128_t vacc0123x3 = vzero;
48 v128_t vacc4567x0 = vacc0123x0;
49 v128_t vacc4567x1 = vzero;
50 v128_t vacc4567x2 = vzero;
51 v128_t vacc4567x3 = vzero;
52 v128_t vacc89ABx0 = vacc0123x0;
53 v128_t vacc89ABx1 = vzero;
54 v128_t vacc89ABx2 = vzero;
55 v128_t vacc89ABx3 = vzero;
56 v128_t vaccCDEFx0 = vacc0123x0;
57 v128_t vaccCDEFx1 = vzero;
58 v128_t vaccCDEFx2 = vzero;
59 v128_t vaccCDEFx3 = vzero;
60 for (; nnz >= 4; nnz -= 4) {
61 const intptr_t diff0 = dmap[0];
62 const intptr_t diff1 = dmap[1];
63 const intptr_t diff2 = dmap[2];
64 const intptr_t diff3 = dmap[3];
65 dmap += 4;
66 const v128_t vi0123x0 = wasm_v128_load(input);
67 const v128_t vi4567x0 = wasm_v128_load(input + 4);
68 const v128_t vi89ABx0 = wasm_v128_load(input + 8);
69 const v128_t viCDEFx0 = wasm_v128_load(input + 12);
70 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff0);
71 const v128_t vw0 = wasm_v32x4_load_splat(w);
72 w += 1;
73 vacc0123x0 = wasm_f32x4_add(vacc0123x0, wasm_f32x4_mul(vi0123x0, vw0));
74 vacc4567x0 = wasm_f32x4_add(vacc4567x0, wasm_f32x4_mul(vi4567x0, vw0));
75 vacc89ABx0 = wasm_f32x4_add(vacc89ABx0, wasm_f32x4_mul(vi89ABx0, vw0));
76 vaccCDEFx0 = wasm_f32x4_add(vaccCDEFx0, wasm_f32x4_mul(viCDEFx0, vw0));
77 const v128_t vi0123x1 = wasm_v128_load(input);
78 const v128_t vi4567x1 = wasm_v128_load(input + 4);
79 const v128_t vi89ABx1 = wasm_v128_load(input + 8);
80 const v128_t viCDEFx1 = wasm_v128_load(input + 12);
81 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff1);
82 const v128_t vw1 = wasm_v32x4_load_splat(w);
83 w += 1;
84 vacc0123x1 = wasm_f32x4_add(vacc0123x1, wasm_f32x4_mul(vi0123x1, vw1));
85 vacc4567x1 = wasm_f32x4_add(vacc4567x1, wasm_f32x4_mul(vi4567x1, vw1));
86 vacc89ABx1 = wasm_f32x4_add(vacc89ABx1, wasm_f32x4_mul(vi89ABx1, vw1));
87 vaccCDEFx1 = wasm_f32x4_add(vaccCDEFx1, wasm_f32x4_mul(viCDEFx1, vw1));
88 const v128_t vi0123x2 = wasm_v128_load(input);
89 const v128_t vi4567x2 = wasm_v128_load(input + 4);
90 const v128_t vi89ABx2 = wasm_v128_load(input + 8);
91 const v128_t viCDEFx2 = wasm_v128_load(input + 12);
92 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff2);
93 const v128_t vw2 = wasm_v32x4_load_splat(w);
94 w += 1;
95 vacc0123x2 = wasm_f32x4_add(vacc0123x2, wasm_f32x4_mul(vi0123x2, vw2));
96 vacc4567x2 = wasm_f32x4_add(vacc4567x2, wasm_f32x4_mul(vi4567x2, vw2));
97 vacc89ABx2 = wasm_f32x4_add(vacc89ABx2, wasm_f32x4_mul(vi89ABx2, vw2));
98 vaccCDEFx2 = wasm_f32x4_add(vaccCDEFx2, wasm_f32x4_mul(viCDEFx2, vw2));
99 const v128_t vi0123x3 = wasm_v128_load(input);
100 const v128_t vi4567x3 = wasm_v128_load(input + 4);
101 const v128_t vi89ABx3 = wasm_v128_load(input + 8);
102 const v128_t viCDEFx3 = wasm_v128_load(input + 12);
103 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff3);
104 const v128_t vw3 = wasm_v32x4_load_splat(w);
105 w += 1;
106 vacc0123x3 = wasm_f32x4_add(vacc0123x3, wasm_f32x4_mul(vi0123x3, vw3));
107 vacc4567x3 = wasm_f32x4_add(vacc4567x3, wasm_f32x4_mul(vi4567x3, vw3));
108 vacc89ABx3 = wasm_f32x4_add(vacc89ABx3, wasm_f32x4_mul(vi89ABx3, vw3));
109 vaccCDEFx3 = wasm_f32x4_add(vaccCDEFx3, wasm_f32x4_mul(viCDEFx3, vw3));
110 }
111 v128_t vacc0123 = vacc0123x0;
112 v128_t vacc4567 = vacc4567x0;
113 v128_t vacc89AB = vacc89ABx0;
114 v128_t vaccCDEF = vaccCDEFx0;
115 vacc0123 = wasm_f32x4_add(vacc0123, vacc0123x1);
116 vacc4567 = wasm_f32x4_add(vacc4567, vacc4567x1);
117 vacc89AB = wasm_f32x4_add(vacc89AB, vacc89ABx1);
118 vaccCDEF = wasm_f32x4_add(vaccCDEF, vaccCDEFx1);
119 vacc0123 = wasm_f32x4_add(vacc0123, vacc0123x2);
120 vacc4567 = wasm_f32x4_add(vacc4567, vacc4567x2);
121 vacc89AB = wasm_f32x4_add(vacc89AB, vacc89ABx2);
122 vaccCDEF = wasm_f32x4_add(vaccCDEF, vaccCDEFx2);
123 vacc0123 = wasm_f32x4_add(vacc0123, vacc0123x3);
124 vacc4567 = wasm_f32x4_add(vacc4567, vacc4567x3);
125 vacc89AB = wasm_f32x4_add(vacc89AB, vacc89ABx3);
126 vaccCDEF = wasm_f32x4_add(vaccCDEF, vaccCDEFx3);
127 if XNN_LIKELY(nnz != 0) {
128 do {
129 const intptr_t diff = *dmap++;
130 const v128_t vi0123 = wasm_v128_load(input);
131 const v128_t vi4567 = wasm_v128_load(input + 4);
132 const v128_t vi89AB = wasm_v128_load(input + 8);
133 const v128_t viCDEF = wasm_v128_load(input + 12);
134 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
135 const v128_t vw = wasm_v32x4_load_splat(w); w += 1;
136 vacc0123 = wasm_f32x4_add(vacc0123, wasm_f32x4_mul(vi0123, vw));
137 vacc4567 = wasm_f32x4_add(vacc4567, wasm_f32x4_mul(vi4567, vw));
138 vacc89AB = wasm_f32x4_add(vacc89AB, wasm_f32x4_mul(vi89AB, vw));
139 vaccCDEF = wasm_f32x4_add(vaccCDEF, wasm_f32x4_mul(viCDEF, vw));
140 } while (--nnz != 0);
141 }
142 v128_t vout0123 = wasm_v128_bitselect(vacc0123, vmax, wasm_f32x4_le(vacc0123, vmax));
143 v128_t vout4567 = wasm_v128_bitselect(vacc4567, vmax, wasm_f32x4_le(vacc4567, vmax));
144 v128_t vout89AB = wasm_v128_bitselect(vacc89AB, vmax, wasm_f32x4_le(vacc89AB, vmax));
145 v128_t voutCDEF = wasm_v128_bitselect(vaccCDEF, vmax, wasm_f32x4_le(vaccCDEF, vmax));
146 vout0123 = wasm_v128_bitselect(vmin, vout0123, wasm_f32x4_lt(vout0123, vmin));
147 vout4567 = wasm_v128_bitselect(vmin, vout4567, wasm_f32x4_lt(vout4567, vmin));
148 vout89AB = wasm_v128_bitselect(vmin, vout89AB, wasm_f32x4_lt(vout89AB, vmin));
149 voutCDEF = wasm_v128_bitselect(vmin, voutCDEF, wasm_f32x4_lt(voutCDEF, vmin));
150 wasm_v128_store(output, vout0123);
151 wasm_v128_store(output + 4, vout4567);
152 wasm_v128_store(output + 8, vout89AB);
153 wasm_v128_store(output + 12, voutCDEF);
154 output = (float*restrict) ((uintptr_t) output + output_stride);
155 } while (--n != 0);
156 output = (float*restrict) ((uintptr_t) output - output_decrement);
157 input += 16;
158 mc -= 16 * sizeof(float);
159 }
160 if XNN_UNLIKELY(mc != 0) {
161 output_decrement += 8 * sizeof(float);
162 if (mc & (8 * sizeof(float))) {
163 const float*restrict w = weights;
164 const int32_t* dmap = widx_dmap;
165 const uint32_t* nnzmap = nidx_nnzmap;
166 size_t n = nc;
167 do {
168 uint32_t nnz = *nnzmap++;
169 v128_t vacc0123 = wasm_v32x4_load_splat(w); w += 1;
170 v128_t vacc4567 = vacc0123;
171 if XNN_LIKELY(nnz != 0) {
172 do {
173 const intptr_t diff = *dmap++;
174 const v128_t vi0123 = wasm_v128_load(input);
175 const v128_t vi4567 = wasm_v128_load(input + 4);
176 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
177 const v128_t vw = wasm_v32x4_load_splat(w); w += 1;
178 vacc0123 = wasm_f32x4_add(vacc0123, wasm_f32x4_mul(vi0123, vw));
179 vacc4567 = wasm_f32x4_add(vacc4567, wasm_f32x4_mul(vi4567, vw));
180 } while (--nnz != 0);
181 }
182 v128_t vout0123 = wasm_v128_bitselect(vacc0123, vmax, wasm_f32x4_le(vacc0123, vmax));
183 v128_t vout4567 = wasm_v128_bitselect(vacc4567, vmax, wasm_f32x4_le(vacc4567, vmax));
184 vout0123 = wasm_v128_bitselect(vmin, vout0123, wasm_f32x4_lt(vout0123, vmin));
185 vout4567 = wasm_v128_bitselect(vmin, vout4567, wasm_f32x4_lt(vout4567, vmin));
186 wasm_v128_store(output, vout0123);
187
188 wasm_v128_store(output + 4, vout4567);
189 output = (float*restrict) ((uintptr_t) output + output_stride);
190 } while (--n != 0);
191 output = (float*restrict) ((uintptr_t) output - output_decrement);
192 input += 8;
193 }
194 output_decrement += 4 * sizeof(float);
195 if (mc & (4 * sizeof(float))) {
196 const float*restrict w = weights;
197 const int32_t* dmap = widx_dmap;
198 const uint32_t* nnzmap = nidx_nnzmap;
199 size_t n = nc;
200 do {
201 uint32_t nnz = *nnzmap++;
202 v128_t vacc0123 = wasm_v32x4_load_splat(w); w += 1;
203 if XNN_LIKELY(nnz != 0) {
204 do {
205 const intptr_t diff = *dmap++;
206 const v128_t vi0123 = wasm_v128_load(input);
207 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
208 const v128_t vw = wasm_v32x4_load_splat(w); w += 1;
209 vacc0123 = wasm_f32x4_add(vacc0123, wasm_f32x4_mul(vi0123, vw));
210 } while (--nnz != 0);
211 }
212 v128_t vout0123 = wasm_v128_bitselect(vacc0123, vmax, wasm_f32x4_le(vacc0123, vmax));
213 vout0123 = wasm_v128_bitselect(vmin, vout0123, wasm_f32x4_lt(vout0123, vmin));
214 wasm_v128_store(output, vout0123);
215
216 output = (float*restrict) ((uintptr_t) output + output_stride);
217 } while (--n != 0);
218 output = (float*restrict) ((uintptr_t) output - output_decrement);
219 input += 4;
220 }
221 output_decrement += 2 * sizeof(float);
222 if (mc & (2 * sizeof(float))) {
223 const float*restrict w = weights;
224 const int32_t* dmap = widx_dmap;
225 const uint32_t* nnzmap = nidx_nnzmap;
226 size_t n = nc;
227 do {
228 uint32_t nnz = *nnzmap++;
229 v128_t vacc01 = wasm_v32x4_load_splat(w); w += 1;
230 if XNN_LIKELY(nnz != 0) {
231 do {
232 const intptr_t diff = *dmap++;
233 const v128_t vi01 = wasm_v64x2_load_splat(input);
234 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
235 const v128_t vw = wasm_v32x4_load_splat(w); w += 1;
236 vacc01 = wasm_f32x4_add(vacc01, wasm_f32x4_mul(vi01, vw));
237 } while (--nnz != 0);
238 }
239 v128_t vout01 = wasm_v128_bitselect(vacc01, vmax, wasm_f32x4_le(vacc01, vmax));
240 vout01 = wasm_v128_bitselect(vmin, vout01, wasm_f32x4_lt(vout01, vmin));
241 *((double*) output) = wasm_f64x2_extract_lane(vout01, 0);
242
243 output = (float*restrict) ((uintptr_t) output + output_stride);
244 } while (--n != 0);
245 output = (float*restrict) ((uintptr_t) output - output_decrement);
246 input += 2;
247 }
248 output_decrement += 1 * sizeof(float);
249 if (mc & (1 * sizeof(float))) {
250 const float*restrict w = weights;
251 const int32_t* dmap = widx_dmap;
252 const uint32_t* nnzmap = nidx_nnzmap;
253 size_t n = nc;
254 do {
255 uint32_t nnz = *nnzmap++;
256 v128_t vacc0 = wasm_v32x4_load_splat(w); w += 1;
257 if XNN_LIKELY(nnz != 0) {
258 do {
259 const intptr_t diff = *dmap++;
260 const v128_t vi0 = wasm_v32x4_load_splat(input);
261 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
262 const v128_t vw = wasm_v32x4_load_splat(w); w += 1;
263 vacc0 = wasm_f32x4_add(vacc0, wasm_f32x4_mul(vi0, vw));
264 } while (--nnz != 0);
265 }
266 v128_t vout0 = wasm_v128_bitselect(vacc0, vmax, wasm_f32x4_le(vacc0, vmax));
267 vout0 = wasm_v128_bitselect(vmin, vout0, wasm_f32x4_lt(vout0, vmin));
268 *output = wasm_f32x4_extract_lane(vout0, 0);
269
270 output = (float*restrict) ((uintptr_t) output + output_stride);
271 } while (--n != 0);
272 output = (float*restrict) ((uintptr_t) output - output_decrement);
273 input += 1;
274 }
275 }
276 }
277