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_32x1__wasmsimd_x86(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_32x1__wasmsimd_x86(
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_v128_load64_splat(params->wasmsimd.min);
33 const v128_t vmax = wasm_v128_load64_splat(params->wasmsimd.max);
34 size_t output_decrement = output_stride * nc - 32 * sizeof(float);
35 while XNN_LIKELY(mc >= 32 * sizeof(float)) {
36 const float*restrict w = weights;
37 const int32_t* dmap = widx_dmap;
38 const uint32_t* nnzmap = nidx_nnzmap;
39 size_t n = nc;
40 do {
41 uint32_t nnz = *nnzmap++;
42 v128_t vacc0123 = wasm_v128_load32_splat(w); w += 1;
43 v128_t vacc4567 = vacc0123;
44 v128_t vacc89AB = vacc0123;
45 v128_t vaccCDEF = vacc0123;
46 v128_t vaccGHIJ = vacc0123;
47 v128_t vaccKLMN = vacc0123;
48 v128_t vaccOPQR = vacc0123;
49 v128_t vaccSTUV = vacc0123;
50 if XNN_LIKELY(nnz != 0) {
51 do {
52 const intptr_t diff = *dmap++;
53 const v128_t vi0123 = wasm_v128_load(input);
54 const v128_t vi4567 = wasm_v128_load(input + 4);
55 const v128_t vi89AB = wasm_v128_load(input + 8);
56 const v128_t viCDEF = wasm_v128_load(input + 12);
57 const v128_t viGHIJ = wasm_v128_load(input + 16);
58 const v128_t viKLMN = wasm_v128_load(input + 20);
59 const v128_t viOPQR = wasm_v128_load(input + 24);
60 const v128_t viSTUV = wasm_v128_load(input + 28);
61 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
62 const v128_t vw = wasm_v128_load32_splat(w); w += 1;
63 vacc0123 = wasm_f32x4_add(vacc0123, wasm_f32x4_mul(vi0123, vw));
64 vacc4567 = wasm_f32x4_add(vacc4567, wasm_f32x4_mul(vi4567, vw));
65 vacc89AB = wasm_f32x4_add(vacc89AB, wasm_f32x4_mul(vi89AB, vw));
66 vaccCDEF = wasm_f32x4_add(vaccCDEF, wasm_f32x4_mul(viCDEF, vw));
67 vaccGHIJ = wasm_f32x4_add(vaccGHIJ, wasm_f32x4_mul(viGHIJ, vw));
68 vaccKLMN = wasm_f32x4_add(vaccKLMN, wasm_f32x4_mul(viKLMN, vw));
69 vaccOPQR = wasm_f32x4_add(vaccOPQR, wasm_f32x4_mul(viOPQR, vw));
70 vaccSTUV = wasm_f32x4_add(vaccSTUV, wasm_f32x4_mul(viSTUV, vw));
71 } while (--nnz != 0);
72 }
73 v128_t vout0123 = wasm_f32x4_pmin(vmax, vacc0123);
74 v128_t vout4567 = wasm_f32x4_pmin(vmax, vacc4567);
75 v128_t vout89AB = wasm_f32x4_pmin(vmax, vacc89AB);
76 v128_t voutCDEF = wasm_f32x4_pmin(vmax, vaccCDEF);
77 v128_t voutGHIJ = wasm_f32x4_pmin(vmax, vaccGHIJ);
78 v128_t voutKLMN = wasm_f32x4_pmin(vmax, vaccKLMN);
79 v128_t voutOPQR = wasm_f32x4_pmin(vmax, vaccOPQR);
80 v128_t voutSTUV = wasm_f32x4_pmin(vmax, vaccSTUV);
81 vout0123 = wasm_f32x4_pmax(vmin, vout0123);
82 vout4567 = wasm_f32x4_pmax(vmin, vout4567);
83 vout89AB = wasm_f32x4_pmax(vmin, vout89AB);
84 voutCDEF = wasm_f32x4_pmax(vmin, voutCDEF);
85 voutGHIJ = wasm_f32x4_pmax(vmin, voutGHIJ);
86 voutKLMN = wasm_f32x4_pmax(vmin, voutKLMN);
87 voutOPQR = wasm_f32x4_pmax(vmin, voutOPQR);
88 voutSTUV = wasm_f32x4_pmax(vmin, voutSTUV);
89 wasm_v128_store(output, vout0123);
90 wasm_v128_store(output + 4, vout4567);
91 wasm_v128_store(output + 8, vout89AB);
92 wasm_v128_store(output + 12, voutCDEF);
93 wasm_v128_store(output + 16, voutGHIJ);
94 wasm_v128_store(output + 20, voutKLMN);
95 wasm_v128_store(output + 24, voutOPQR);
96 wasm_v128_store(output + 28, voutSTUV);
97 output = (float*restrict) ((uintptr_t) output + output_stride);
98 } while (--n != 0);
99 output = (float*restrict) ((uintptr_t) output - output_decrement);
100 input += 32;
101 mc -= 32 * sizeof(float);
102 }
103 if XNN_UNLIKELY(mc != 0) {
104 output_decrement += 16 * sizeof(float);
105 if (mc & (16 * sizeof(float))) {
106 const float*restrict w = weights;
107 const int32_t* dmap = widx_dmap;
108 const uint32_t* nnzmap = nidx_nnzmap;
109 size_t n = nc;
110 do {
111 uint32_t nnz = *nnzmap++;
112 v128_t vacc0123 = wasm_v128_load32_splat(w); w += 1;
113 v128_t vacc4567 = vacc0123;
114 v128_t vacc89AB = vacc0123;
115 v128_t vaccCDEF = vacc0123;
116 if XNN_LIKELY(nnz != 0) {
117 do {
118 const intptr_t diff = *dmap++;
119 const v128_t vi0123 = wasm_v128_load(input);
120 const v128_t vi4567 = wasm_v128_load(input + 4);
121 const v128_t vi89AB = wasm_v128_load(input + 8);
122 const v128_t viCDEF = wasm_v128_load(input + 12);
123 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
124 const v128_t vw = wasm_v128_load32_splat(w); w += 1;
125 vacc0123 = wasm_f32x4_add(vacc0123, wasm_f32x4_mul(vi0123, vw));
126 vacc4567 = wasm_f32x4_add(vacc4567, wasm_f32x4_mul(vi4567, vw));
127 vacc89AB = wasm_f32x4_add(vacc89AB, wasm_f32x4_mul(vi89AB, vw));
128 vaccCDEF = wasm_f32x4_add(vaccCDEF, wasm_f32x4_mul(viCDEF, vw));
129 } while (--nnz != 0);
130 }
131 v128_t vout0123 = wasm_f32x4_pmin(vmax, vacc0123);
132 v128_t vout4567 = wasm_f32x4_pmin(vmax, vacc4567);
133 v128_t vout89AB = wasm_f32x4_pmin(vmax, vacc89AB);
134 v128_t voutCDEF = wasm_f32x4_pmin(vmax, vaccCDEF);
135 vout0123 = wasm_f32x4_pmax(vmin, vout0123);
136 vout4567 = wasm_f32x4_pmax(vmin, vout4567);
137 vout89AB = wasm_f32x4_pmax(vmin, vout89AB);
138 voutCDEF = wasm_f32x4_pmax(vmin, voutCDEF);
139 wasm_v128_store(output, vout0123);
140
141 wasm_v128_store(output + 4, vout4567);
142 wasm_v128_store(output + 8, vout89AB);
143 wasm_v128_store(output + 12, voutCDEF);
144 output = (float*restrict) ((uintptr_t) output + output_stride);
145 } while (--n != 0);
146 output = (float*restrict) ((uintptr_t) output - output_decrement);
147 input += 16;
148 }
149 output_decrement += 8 * sizeof(float);
150 if (mc & (8 * sizeof(float))) {
151 const float*restrict w = weights;
152 const int32_t* dmap = widx_dmap;
153 const uint32_t* nnzmap = nidx_nnzmap;
154 size_t n = nc;
155 do {
156 uint32_t nnz = *nnzmap++;
157 v128_t vacc0123 = wasm_v128_load32_splat(w); w += 1;
158 v128_t vacc4567 = vacc0123;
159 if XNN_LIKELY(nnz != 0) {
160 do {
161 const intptr_t diff = *dmap++;
162 const v128_t vi0123 = wasm_v128_load(input);
163 const v128_t vi4567 = wasm_v128_load(input + 4);
164 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
165 const v128_t vw = wasm_v128_load32_splat(w); w += 1;
166 vacc0123 = wasm_f32x4_add(vacc0123, wasm_f32x4_mul(vi0123, vw));
167 vacc4567 = wasm_f32x4_add(vacc4567, wasm_f32x4_mul(vi4567, vw));
168 } while (--nnz != 0);
169 }
170 v128_t vout0123 = wasm_f32x4_pmin(vmax, vacc0123);
171 v128_t vout4567 = wasm_f32x4_pmin(vmax, vacc4567);
172 vout0123 = wasm_f32x4_pmax(vmin, vout0123);
173 vout4567 = wasm_f32x4_pmax(vmin, vout4567);
174 wasm_v128_store(output, vout0123);
175
176 wasm_v128_store(output + 4, vout4567);
177 output = (float*restrict) ((uintptr_t) output + output_stride);
178 } while (--n != 0);
179 output = (float*restrict) ((uintptr_t) output - output_decrement);
180 input += 8;
181 }
182 output_decrement += 4 * sizeof(float);
183 if (mc & (4 * sizeof(float))) {
184 const float*restrict w = weights;
185 const int32_t* dmap = widx_dmap;
186 const uint32_t* nnzmap = nidx_nnzmap;
187 size_t n = nc;
188 do {
189 uint32_t nnz = *nnzmap++;
190 v128_t vacc0123 = wasm_v128_load32_splat(w); w += 1;
191 if XNN_LIKELY(nnz != 0) {
192 do {
193 const intptr_t diff = *dmap++;
194 const v128_t vi0123 = wasm_v128_load(input);
195 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
196 const v128_t vw = wasm_v128_load32_splat(w); w += 1;
197 vacc0123 = wasm_f32x4_add(vacc0123, wasm_f32x4_mul(vi0123, vw));
198 } while (--nnz != 0);
199 }
200 v128_t vout0123 = wasm_f32x4_pmin(vmax, vacc0123);
201 vout0123 = wasm_f32x4_pmax(vmin, vout0123);
202 wasm_v128_store(output, vout0123);
203
204 output = (float*restrict) ((uintptr_t) output + output_stride);
205 } while (--n != 0);
206 output = (float*restrict) ((uintptr_t) output - output_decrement);
207 input += 4;
208 }
209 output_decrement += 2 * sizeof(float);
210 if (mc & (2 * sizeof(float))) {
211 const float*restrict w = weights;
212 const int32_t* dmap = widx_dmap;
213 const uint32_t* nnzmap = nidx_nnzmap;
214 size_t n = nc;
215 do {
216 uint32_t nnz = *nnzmap++;
217 v128_t vacc01 = wasm_v128_load32_splat(w); w += 1;
218 if XNN_LIKELY(nnz != 0) {
219 do {
220 const intptr_t diff = *dmap++;
221 const v128_t vi01 = wasm_v128_load64_splat(input);
222 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
223 const v128_t vw = wasm_v128_load32_splat(w); w += 1;
224 vacc01 = wasm_f32x4_add(vacc01, wasm_f32x4_mul(vi01, vw));
225 } while (--nnz != 0);
226 }
227 v128_t vout01 = wasm_f32x4_pmin(vmax, vacc01);
228 vout01 = wasm_f32x4_pmax(vmin, vout01);
229 *((double*) output) = wasm_f64x2_extract_lane(vout01, 0);
230
231 output = (float*restrict) ((uintptr_t) output + output_stride);
232 } while (--n != 0);
233 output = (float*restrict) ((uintptr_t) output - output_decrement);
234 input += 2;
235 }
236 output_decrement += 1 * sizeof(float);
237 if (mc & (1 * sizeof(float))) {
238 const float*restrict w = weights;
239 const int32_t* dmap = widx_dmap;
240 const uint32_t* nnzmap = nidx_nnzmap;
241 size_t n = nc;
242 do {
243 uint32_t nnz = *nnzmap++;
244 v128_t vacc0 = wasm_v128_load32_splat(w); w += 1;
245 if XNN_LIKELY(nnz != 0) {
246 do {
247 const intptr_t diff = *dmap++;
248 const v128_t vi0 = wasm_v128_load32_splat(input);
249 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
250 const v128_t vw = wasm_v128_load32_splat(w); w += 1;
251 vacc0 = wasm_f32x4_add(vacc0, wasm_f32x4_mul(vi0, vw));
252 } while (--nnz != 0);
253 }
254 v128_t vout0 = wasm_f32x4_pmin(vmax, vacc0);
255 vout0 = wasm_f32x4_pmax(vmin, vout0);
256 *output = wasm_f32x4_extract_lane(vout0, 0);
257
258 output = (float*restrict) ((uintptr_t) output + output_stride);
259 } while (--n != 0);
260 output = (float*restrict) ((uintptr_t) output - output_decrement);
261 input += 1;
262 }
263 }
264 }
265