xref: /aosp_15_r20/external/XNNPACK/src/f32-spmm/gen/8x4-minmax-neonfma.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-spmm/neon-blocked.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2019 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 <arm_neon.h>
13 
14 #include <xnnpack/spmm.h>
15 
16 
xnn_f32_spmm_minmax_ukernel_8x4__neonfma(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_8x4__neonfma(
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 float32x4_t vmin = vld1q_dup_f32(&params->scalar.min);
33   const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max);
34   size_t output_decrement = output_stride * nc - 8 * sizeof(float);
35   while XNN_LIKELY(mc >= 8 * 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     while (n >= 4) {
41       uint32_t nnz = *nnzmap++;
42       float32x4_t vacc0123n0 = vld1q_dup_f32(w); w += 1;
43       float32x4_t vacc4567n0 = vacc0123n0;
44       float32x4_t vacc0123n1 = vld1q_dup_f32(w); w += 1;
45       float32x4_t vacc4567n1 = vacc0123n1;
46       float32x4_t vacc0123n2 = vld1q_dup_f32(w); w += 1;
47       float32x4_t vacc4567n2 = vacc0123n2;
48       float32x4_t vacc0123n3 = vld1q_dup_f32(w); w += 1;
49       float32x4_t vacc4567n3 = vacc0123n3;
50       if XNN_LIKELY(nnz != 0) {
51         do {
52           const intptr_t diff = *dmap++;
53           const float32x4_t vi0123 = vld1q_f32(input);
54           const float32x4_t vi4567 = vld1q_f32(input + 4);
55           input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
56           __builtin_prefetch(input + 16);
57           const float32x4_t vw = vld1q_f32(w); w += 4;
58           __builtin_prefetch(w + 32);
59           vacc0123n0 = vfmaq_laneq_f32(vacc0123n0, vi0123, vw, 0);
60           vacc4567n0 = vfmaq_laneq_f32(vacc4567n0, vi4567, vw, 0);
61           vacc0123n1 = vfmaq_laneq_f32(vacc0123n1, vi0123, vw, 1);
62           vacc4567n1 = vfmaq_laneq_f32(vacc4567n1, vi4567, vw, 1);
63           vacc0123n2 = vfmaq_laneq_f32(vacc0123n2, vi0123, vw, 2);
64           vacc4567n2 = vfmaq_laneq_f32(vacc4567n2, vi4567, vw, 2);
65           vacc0123n3 = vfmaq_laneq_f32(vacc0123n3, vi0123, vw, 3);
66           vacc4567n3 = vfmaq_laneq_f32(vacc4567n3, vi4567, vw, 3);
67         } while (--nnz != 0);
68       }
69       float32x4_t vout0123n0 = vminq_f32(vacc0123n0, vmax);
70       float32x4_t vout4567n0 = vminq_f32(vacc4567n0, vmax);
71       float32x4_t vout0123n1 = vminq_f32(vacc0123n1, vmax);
72       float32x4_t vout4567n1 = vminq_f32(vacc4567n1, vmax);
73       float32x4_t vout0123n2 = vminq_f32(vacc0123n2, vmax);
74       float32x4_t vout4567n2 = vminq_f32(vacc4567n2, vmax);
75       float32x4_t vout0123n3 = vminq_f32(vacc0123n3, vmax);
76       float32x4_t vout4567n3 = vminq_f32(vacc4567n3, vmax);
77 
78       vout0123n0 = vmaxq_f32(vout0123n0, vmin);
79       vout4567n0 = vmaxq_f32(vout4567n0, vmin);
80       vout0123n1 = vmaxq_f32(vout0123n1, vmin);
81       vout4567n1 = vmaxq_f32(vout4567n1, vmin);
82       vout0123n2 = vmaxq_f32(vout0123n2, vmin);
83       vout4567n2 = vmaxq_f32(vout4567n2, vmin);
84       vout0123n3 = vmaxq_f32(vout0123n3, vmin);
85       vout4567n3 = vmaxq_f32(vout4567n3, vmin);
86 
87       vst1q_f32(output + 0, vout0123n0);
88       vst1q_f32(output + 4, vout4567n0);
89       output = (float*restrict) ((uintptr_t) output + output_stride);
90       vst1q_f32(output + 0, vout0123n1);
91       vst1q_f32(output + 4, vout4567n1);
92       output = (float*restrict) ((uintptr_t) output + output_stride);
93       vst1q_f32(output + 0, vout0123n2);
94       vst1q_f32(output + 4, vout4567n2);
95       output = (float*restrict) ((uintptr_t) output + output_stride);
96       vst1q_f32(output + 0, vout0123n3);
97       vst1q_f32(output + 4, vout4567n3);
98       output = (float*restrict) ((uintptr_t) output + output_stride);
99       n -= 4;
100     }
101 
102     // clean up loop, fall back to nr=1
103     if XNN_UNLIKELY(n != 0) {
104       do {
105         uint32_t nnz = *nnzmap++;
106         float32x4_t vacc0123 = vld1q_dup_f32(w); w += 1;
107         float32x4_t vacc4567 = vacc0123;
108         if XNN_LIKELY(nnz != 0) {
109           do {
110             const intptr_t diff = *dmap++;
111             const float32x4_t vi0123 = vld1q_f32(input);
112             const float32x4_t vi4567 = vld1q_f32(input + 4);
113             input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
114             __builtin_prefetch(input + 16);
115             const float32x4_t vw = vld1q_dup_f32(w); w += 1;
116             __builtin_prefetch(w + 32);
117             vacc0123 = vfmaq_f32(vacc0123, vi0123, vw);
118             vacc4567 = vfmaq_f32(vacc4567, vi4567, vw);
119           } while (--nnz != 0);
120         }
121         float32x4_t vout0123 = vminq_f32(vacc0123, vmax);
122         float32x4_t vout4567 = vminq_f32(vacc4567, vmax);
123 
124         vout0123 = vmaxq_f32(vout0123, vmin);
125         vout4567 = vmaxq_f32(vout4567, vmin);
126 
127         vst1q_f32(output + 0, vout0123);
128         vst1q_f32(output + 4, vout4567);
129         output = (float*restrict) ((uintptr_t) output + output_stride);
130         n -= 1;
131       } while (n != 0);
132     }
133     output = (float*restrict) ((uintptr_t) output - output_decrement);
134     input += 8;
135     mc -= 8 * sizeof(float);
136   }
137   if XNN_UNLIKELY(mc != 0) {
138     output_decrement += 4 * sizeof(float);
139     if (mc & (4 * sizeof(float))) {
140       const float*restrict w = weights;
141       const int32_t* dmap = widx_dmap;
142       const uint32_t* nnzmap = nidx_nnzmap;
143       size_t n = nc;
144       while (n >= 4) {
145         uint32_t nnz = *nnzmap++;
146         float32x4_t vacc0123n0 = vld1q_dup_f32(w); w += 1;
147         float32x4_t vacc0123n1 = vld1q_dup_f32(w); w += 1;
148         float32x4_t vacc0123n2 = vld1q_dup_f32(w); w += 1;
149         float32x4_t vacc0123n3 = vld1q_dup_f32(w); w += 1;
150         if XNN_LIKELY(nnz != 0) {
151           do {
152             const intptr_t diff = *dmap++;
153             const float32x4_t vi0123 = vld1q_f32(input);
154             input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
155             const float32x4_t vw = vld1q_f32(w); w += 4;
156 
157             vacc0123n0 = vfmaq_laneq_f32(vacc0123n0, vi0123, vw, 0);
158             vacc0123n1 = vfmaq_laneq_f32(vacc0123n1, vi0123, vw, 1);
159             vacc0123n2 = vfmaq_laneq_f32(vacc0123n2, vi0123, vw, 2);
160             vacc0123n3 = vfmaq_laneq_f32(vacc0123n3, vi0123, vw, 3);
161           } while (--nnz != 0);
162         }
163         float32x4_t vout0123n0 = vminq_f32(vacc0123n0, vmax);
164         float32x4_t vout0123n1 = vminq_f32(vacc0123n1, vmax);
165         float32x4_t vout0123n2 = vminq_f32(vacc0123n2, vmax);
166         float32x4_t vout0123n3 = vminq_f32(vacc0123n3, vmax);
167 
168         vout0123n0 = vmaxq_f32(vout0123n0, vmin);
169         vout0123n1 = vmaxq_f32(vout0123n1, vmin);
170         vout0123n2 = vmaxq_f32(vout0123n2, vmin);
171         vout0123n3 = vmaxq_f32(vout0123n3, vmin);
172 
173         vst1q_f32(output + 0, vout0123n0);
174         output = (float*restrict) ((uintptr_t) output + output_stride);
175         vst1q_f32(output + 0, vout0123n1);
176         output = (float*restrict) ((uintptr_t) output + output_stride);
177         vst1q_f32(output + 0, vout0123n2);
178         output = (float*restrict) ((uintptr_t) output + output_stride);
179         vst1q_f32(output + 0, vout0123n3);
180         output = (float*restrict) ((uintptr_t) output + output_stride);
181         n -= 4;
182       }
183 
184       // clean up loop, fall back to nr=1
185       if XNN_UNLIKELY(n != 0) {
186         do {
187           uint32_t nnz = *nnzmap++;
188           float32x4_t vacc0123 = vld1q_dup_f32(w); w += 1;
189           if XNN_LIKELY(nnz != 0) {
190             do {
191               const intptr_t diff = *dmap++;
192               const float32x4_t vi0123 = vld1q_f32(input);
193               input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
194               const float32x4_t vw = vld1q_dup_f32(w); w += 1;
195               vacc0123 = vfmaq_f32(vacc0123, vi0123, vw);
196             } while (--nnz != 0);
197           }
198           float32x4_t vout0123 = vminq_f32(vacc0123, vmax);
199 
200           vout0123 = vmaxq_f32(vout0123, vmin);
201 
202           vst1q_f32(output + 0, vout0123);
203           output = (float*restrict) ((uintptr_t) output + output_stride);
204           n -= 1;
205         } while (n != 0);
206       }
207       output = (float*restrict) ((uintptr_t) output - output_decrement);
208       input += 4;
209     }
210     output_decrement += 2 * sizeof(float);
211     if (mc & (2 * sizeof(float))) {
212       const float*restrict w = weights;
213       const int32_t* dmap = widx_dmap;
214       const uint32_t* nnzmap = nidx_nnzmap;
215       size_t n = nc;
216       while (n >= 4) {
217         uint32_t nnz = *nnzmap++;
218         float32x2_t vacc01n0 = vld1_dup_f32(w); w += 1;
219         float32x2_t vacc01n1 = vld1_dup_f32(w); w += 1;
220         float32x2_t vacc01n2 = vld1_dup_f32(w); w += 1;
221         float32x2_t vacc01n3 = vld1_dup_f32(w); w += 1;
222         if XNN_LIKELY(nnz != 0) {
223           do {
224             const intptr_t diff = *dmap++;
225             const float32x2_t vi01 = vld1_f32(input);
226             input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
227             const float32x4_t vw = vld1q_f32(w); w += 4;
228 
229             vacc01n0 = vfma_laneq_f32(vacc01n0, vi01, vw, 0);
230             vacc01n1 = vfma_laneq_f32(vacc01n1, vi01, vw, 1);
231             vacc01n2 = vfma_laneq_f32(vacc01n2, vi01, vw, 2);
232             vacc01n3 = vfma_laneq_f32(vacc01n3, vi01, vw, 3);
233           } while (--nnz != 0);
234         }
235         float32x2_t vout01n0 = vmin_f32(vacc01n0, vget_low_f32(vmax));
236         float32x2_t vout01n1 = vmin_f32(vacc01n1, vget_low_f32(vmax));
237         float32x2_t vout01n2 = vmin_f32(vacc01n2, vget_low_f32(vmax));
238         float32x2_t vout01n3 = vmin_f32(vacc01n3, vget_low_f32(vmax));
239 
240         vout01n0 = vmax_f32(vout01n0, vget_low_f32(vmin));
241         vout01n1 = vmax_f32(vout01n1, vget_low_f32(vmin));
242         vout01n2 = vmax_f32(vout01n2, vget_low_f32(vmin));
243         vout01n3 = vmax_f32(vout01n3, vget_low_f32(vmin));
244 
245         vst1_f32(output + 0, vout01n0);
246         output = (float*restrict) ((uintptr_t) output + output_stride);
247         vst1_f32(output + 0, vout01n1);
248         output = (float*restrict) ((uintptr_t) output + output_stride);
249         vst1_f32(output + 0, vout01n2);
250         output = (float*restrict) ((uintptr_t) output + output_stride);
251         vst1_f32(output + 0, vout01n3);
252         output = (float*restrict) ((uintptr_t) output + output_stride);
253         n -= 4;
254       }
255 
256       // clean up loop, fall back to nr=1
257       if XNN_UNLIKELY(n != 0) {
258         do {
259           uint32_t nnz = *nnzmap++;
260           float32x2_t vacc01 = vld1_dup_f32(w); w += 1;
261           if XNN_LIKELY(nnz != 0) {
262             do {
263               const intptr_t diff = *dmap++;
264               const float32x2_t vi01 = vld1_f32(input);
265               input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
266               const float32x2_t vw = vld1_dup_f32(w); w += 1;
267               vacc01 = vfma_f32(vacc01, vi01, vw);
268             } while (--nnz != 0);
269           }
270           float32x2_t vout01 = vmin_f32(vacc01, vget_low_f32(vmax));
271           vout01 = vmax_f32(vout01, vget_low_f32(vmin));
272 
273           vst1_f32(output, vout01);
274           output = (float*restrict) ((uintptr_t) output + output_stride);
275           n -= 1;
276         } while (n != 0);
277       }
278       output = (float*restrict) ((uintptr_t) output - output_decrement);
279       input += 2;
280     }
281     output_decrement += 1 * sizeof(float);
282     if (mc & (1 * sizeof(float))) {
283       const float*restrict w = weights;
284       const int32_t* dmap = widx_dmap;
285       const uint32_t* nnzmap = nidx_nnzmap;
286       size_t n = nc;
287       while (n >= 4) {
288         uint32_t nnz = *nnzmap++;
289         float32x2_t vacc0n0 = vld1_dup_f32(w); w += 1;
290         float32x2_t vacc0n1 = vld1_dup_f32(w); w += 1;
291         float32x2_t vacc0n2 = vld1_dup_f32(w); w += 1;
292         float32x2_t vacc0n3 = vld1_dup_f32(w); w += 1;
293         if XNN_LIKELY(nnz != 0) {
294           do {
295             const intptr_t diff = *dmap++;
296             const float32x2_t vi0 = vld1_dup_f32(input);
297             input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
298             const float32x4_t vw = vld1q_f32(w); w += 4;
299 
300             vacc0n0 = vfma_laneq_f32(vacc0n0, vi0, vw, 0);
301             vacc0n1 = vfma_laneq_f32(vacc0n1, vi0, vw, 1);
302             vacc0n2 = vfma_laneq_f32(vacc0n2, vi0, vw, 2);
303             vacc0n3 = vfma_laneq_f32(vacc0n3, vi0, vw, 3);
304           } while (--nnz != 0);
305         }
306         float32x2_t vout0n0 = vmin_f32(vacc0n0, vget_low_f32(vmax));
307         float32x2_t vout0n1 = vmin_f32(vacc0n1, vget_low_f32(vmax));
308         float32x2_t vout0n2 = vmin_f32(vacc0n2, vget_low_f32(vmax));
309         float32x2_t vout0n3 = vmin_f32(vacc0n3, vget_low_f32(vmax));
310 
311         vout0n0 = vmax_f32(vout0n0, vget_low_f32(vmin));
312         vout0n1 = vmax_f32(vout0n1, vget_low_f32(vmin));
313         vout0n2 = vmax_f32(vout0n2, vget_low_f32(vmin));
314         vout0n3 = vmax_f32(vout0n3, vget_low_f32(vmin));
315 
316         vst1_lane_f32(output + 0, vout0n0, 0);
317         output = (float*restrict) ((uintptr_t) output + output_stride);
318         vst1_lane_f32(output + 0, vout0n1, 0);
319         output = (float*restrict) ((uintptr_t) output + output_stride);
320         vst1_lane_f32(output + 0, vout0n2, 0);
321         output = (float*restrict) ((uintptr_t) output + output_stride);
322         vst1_lane_f32(output + 0, vout0n3, 0);
323         output = (float*restrict) ((uintptr_t) output + output_stride);
324         n -= 4;
325       }
326 
327       // clean up loop, fall back to nr=1
328       if XNN_UNLIKELY(n != 0) {
329         do {
330           uint32_t nnz = *nnzmap++;
331           float32x2_t vacc0 = vld1_dup_f32(w); w += 1;
332           if XNN_LIKELY(nnz != 0) {
333             do {
334               const intptr_t diff = *dmap++;
335               const float32x2_t vi0 = vld1_dup_f32(input);
336               input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
337               const float32x2_t vw = vld1_dup_f32(w); w += 1;
338               vacc0 = vfma_f32(vacc0, vi0, vw);
339             } while (--nnz != 0);
340           }
341           float32x2_t vout0 = vmin_f32(vacc0, vget_low_f32(vmax));
342           vout0 = vmax_f32(vout0, vget_low_f32(vmin));
343 
344           vst1_lane_f32(output, vout0, 1);
345           output = (float*restrict) ((uintptr_t) output + output_stride);
346           n -= 1;
347         } while (n != 0);
348       }
349       output = (float*restrict) ((uintptr_t) output - output_decrement);
350       input += 1;
351     }
352     }
353 }
354