1 // Auto-generated file. Do not edit!
2 // Template: src/f16-spmm/neonfp16arith.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_f16_spmm_minmax_ukernel_24x1__neonfp16arith(size_t mc,size_t nc,const void * restrict input,const void * restrict weights,const int32_t * restrict widx_dmap,const uint32_t * restrict nidx_nnzmap,void * restrict output,size_t output_stride,const struct xnn_f16_scaleminmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f16_spmm_minmax_ukernel_24x1__neonfp16arith(
18 size_t mc,
19 size_t nc,
20 const void*restrict input,
21 const void*restrict weights,
22 const int32_t*restrict widx_dmap,
23 const uint32_t*restrict nidx_nnzmap,
24 void*restrict output,
25 size_t output_stride,
26 const struct xnn_f16_scaleminmax_params params[restrict XNN_MIN_ELEMENTS(1)])
27 {
28 assert(mc != 0);
29 assert(mc % sizeof(__fp16) == 0);
30 assert(nc != 0);
31
32 const __fp16*restrict i = (const __fp16*) input;
33 __fp16*restrict o = (__fp16*) output;
34
35 const float16x8_t vscale = vld1q_dup_f16((const __fp16*) ¶ms->scale);
36 const float16x8_t vmax = vld1q_dup_f16((const __fp16*) ¶ms->max);
37 const float16x8_t vmin = vld1q_dup_f16((const __fp16*) ¶ms->min);
38
39 size_t output_decrement = output_stride * nc - 24 * sizeof(__fp16);
40 while XNN_LIKELY(mc >= 24 * sizeof(__fp16)) {
41 const __fp16*restrict w = (const __fp16*) weights;
42 const int32_t* dmap = widx_dmap;
43 const uint32_t* nnzmap = nidx_nnzmap;
44 size_t n = nc;
45 do {
46 uint32_t nnz = *nnzmap++;
47 float16x8_t vacc01234567 = vld1q_dup_f16(w); w += 1;
48 float16x8_t vacc89ABCDEF = vacc01234567;
49 float16x8_t vaccGHIJKLMN = vacc01234567;
50 if XNN_LIKELY(nnz != 0) {
51 do {
52 const intptr_t diff = *dmap++;
53 const float16x8_t va01234567 = vld1q_f16(i);
54 const float16x8_t va89ABCDEF = vld1q_f16(i + 8);
55 const float16x8_t vaGHIJKLMN = vld1q_f16(i + 16);
56 i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff);
57 const float16x8_t vb = vld1q_dup_f16(w); w += 1;
58 vacc01234567 = vfmaq_f16(vacc01234567, va01234567, vb);
59 vacc89ABCDEF = vfmaq_f16(vacc89ABCDEF, va89ABCDEF, vb);
60 vaccGHIJKLMN = vfmaq_f16(vaccGHIJKLMN, vaGHIJKLMN, vb);
61 } while (--nnz != 0);
62 }
63 float16x8_t vout01234567 = vmulq_f16(vacc01234567, vscale);
64 float16x8_t vout89ABCDEF = vmulq_f16(vacc89ABCDEF, vscale);
65 float16x8_t voutGHIJKLMN = vmulq_f16(vaccGHIJKLMN, vscale);
66 vout01234567 = vminq_f16(vout01234567, vmax);
67 vout89ABCDEF = vminq_f16(vout89ABCDEF, vmax);
68 voutGHIJKLMN = vminq_f16(voutGHIJKLMN, vmax);
69 vout01234567 = vmaxq_f16(vout01234567, vmin);
70 vout89ABCDEF = vmaxq_f16(vout89ABCDEF, vmin);
71 voutGHIJKLMN = vmaxq_f16(voutGHIJKLMN, vmin);
72 vst1q_f16(o, vout01234567);
73 vst1q_f16(o + 8, vout89ABCDEF);
74 vst1q_f16(o + 16, voutGHIJKLMN);
75 o = (__fp16*restrict) ((uintptr_t) o + output_stride);
76 } while (--n != 0);
77 o = (__fp16*restrict) ((uintptr_t) o - output_decrement);
78 i += 24;
79 mc -= 24 * sizeof(__fp16);
80 }
81 if XNN_UNLIKELY(mc != 0) {
82 output_decrement += 8 * sizeof(__fp16);
83 if (mc & (16 * sizeof(__fp16))) {
84 const __fp16*restrict w = (const __fp16*) weights;
85 const int32_t* dmap = widx_dmap;
86 const uint32_t* nnzmap = nidx_nnzmap;
87 size_t n = nc;
88 do {
89 uint32_t nnz = *nnzmap++;
90 float16x8_t vacc01234567 = vld1q_dup_f16(w); w += 1;
91 float16x8_t vacc89ABCDEF = vacc01234567;
92 if XNN_LIKELY(nnz != 0) {
93 do {
94 const intptr_t diff = *dmap++;
95 const float16x8_t va01234567 = vld1q_f16(i);
96 const float16x8_t va89ABCDEF = vld1q_f16(i + 8);
97 i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff);
98 const float16x8_t vb = vld1q_dup_f16(w); w += 1;
99 vacc01234567 = vfmaq_f16(vacc01234567, va01234567, vb);
100 vacc89ABCDEF = vfmaq_f16(vacc89ABCDEF, va89ABCDEF, vb);
101 } while (--nnz != 0);
102 }
103 float16x8_t vout01234567 = vminq_f16(vacc01234567, vmax);
104 float16x8_t vout89ABCDEF = vminq_f16(vacc89ABCDEF, vmax);
105 vout01234567 = vmaxq_f16(vout01234567, vmin);
106 vout89ABCDEF = vmaxq_f16(vout89ABCDEF, vmin);
107 vst1q_f16(o, vout01234567);
108 vst1q_f16(o + 8, vout89ABCDEF);
109 o = (__fp16*restrict) ((uintptr_t) o + output_stride);
110 } while (--n != 0);
111 o = (__fp16*restrict) ((uintptr_t) o - output_decrement);
112 i += 16;
113 }
114 output_decrement += 8 * sizeof(__fp16);
115 if (mc & (8 * sizeof(__fp16))) {
116 const __fp16*restrict w = (const __fp16*) weights;
117 const int32_t* dmap = widx_dmap;
118 const uint32_t* nnzmap = nidx_nnzmap;
119 size_t n = nc;
120 do {
121 uint32_t nnz = *nnzmap++;
122 float16x8_t vacc01234567 = vld1q_dup_f16(w); w += 1;
123 if XNN_LIKELY(nnz != 0) {
124 do {
125 const intptr_t diff = *dmap++;
126 const float16x8_t va01234567 = vld1q_f16(i);
127 i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff);
128 const float16x8_t vb = vld1q_dup_f16(w); w += 1;
129 vacc01234567 = vfmaq_f16(vacc01234567, va01234567, vb);
130 } while (--nnz != 0);
131 }
132 float16x8_t vout01234567 = vminq_f16(vacc01234567, vmax);
133 vout01234567 = vmaxq_f16(vout01234567, vmin);
134 vst1q_f16(o, vout01234567);
135 o = (__fp16*restrict) ((uintptr_t) o + output_stride);
136 } while (--n != 0);
137 o = (__fp16*restrict) ((uintptr_t) o - output_decrement);
138 i += 8;
139 }
140 output_decrement += 4 * sizeof(__fp16);
141 if (mc & (4 * sizeof(__fp16))) {
142 const __fp16*restrict w = (const __fp16*) weights;
143 const int32_t* dmap = widx_dmap;
144 const uint32_t* nnzmap = nidx_nnzmap;
145 size_t n = nc;
146 do {
147 uint32_t nnz = *nnzmap++;
148 float16x4_t vacc0123 = vld1_dup_f16(w); w += 1;
149 if XNN_LIKELY(nnz != 0) {
150 do {
151 const intptr_t diff = *dmap++;
152 const float16x4_t va0123 = vld1_f16(i);
153 i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff);
154 const float16x4_t vb = vld1_dup_f16(w); w += 1;
155 vacc0123 = vfma_f16(vacc0123, va0123, vb);
156 } while (--nnz != 0);
157 }
158 float16x4_t vout0123 = vmin_f16(vacc0123, vget_low_f16(vmax));
159 vout0123 = vmax_f16(vout0123, vget_low_f16(vmin));
160 vst1_f16(o, vout0123);
161 o = (__fp16*restrict) ((uintptr_t) o + output_stride);
162 } while (--n != 0);
163 o = (__fp16*restrict) ((uintptr_t) o - output_decrement);
164 i += 4;
165 }
166 output_decrement += 2 * sizeof(__fp16);
167 if (mc & (2 * sizeof(__fp16))) {
168 const __fp16*restrict w = (const __fp16*) weights;
169 const int32_t* dmap = widx_dmap;
170 const uint32_t* nnzmap = nidx_nnzmap;
171 size_t n = nc;
172 do {
173 uint32_t nnz = *nnzmap++;
174 float16x4_t vacc01 = vld1_dup_f16(w); w += 1;
175 if XNN_LIKELY(nnz != 0) {
176 do {
177 const intptr_t diff = *dmap++;
178 const float16x4_t va01 = vreinterpret_f16_f32(vld1_dup_f32(__builtin_assume_aligned(i, 1)));
179 i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff);
180 const float16x4_t vb = vld1_dup_f16(w); w += 1;
181 vacc01 = vfma_f16(vacc01, va01, vb);
182 } while (--nnz != 0);
183 }
184 float16x4_t vout01 = vmin_f16(vacc01, vget_low_f16(vmax));
185 vout01 = vmax_f16(vout01, vget_low_f16(vmin));
186 vst1_lane_f32(__builtin_assume_aligned(o, 1), vreinterpret_f32_f16(vout01), 0);
187 o = (__fp16*restrict) ((uintptr_t) o + output_stride);
188 } while (--n != 0);
189 o = (__fp16*restrict) ((uintptr_t) o - output_decrement);
190 i += 2;
191 }
192 output_decrement += 1 * sizeof(__fp16);
193 if (mc & (1 * sizeof(__fp16))) {
194 const __fp16*restrict w = (const __fp16*) weights;
195 const int32_t* dmap = widx_dmap;
196 const uint32_t* nnzmap = nidx_nnzmap;
197 size_t n = nc;
198 do {
199 uint32_t nnz = *nnzmap++;
200 float16x4_t vacc0 = vld1_dup_f16(w); w += 1;
201 if XNN_LIKELY(nnz != 0) {
202 do {
203 const intptr_t diff = *dmap++;
204 const float16x4_t va0 = vld1_dup_f16(i);
205 i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff);
206 const float16x4_t vb = vld1_dup_f16(w); w += 1;
207 vacc0 = vfma_f16(vacc0, va0, vb);
208 } while (--nnz != 0);
209 }
210 float16x4_t vout0 = vmin_f16(vacc0, vget_low_f16(vmax));
211 vout0 = vmax_f16(vout0, vget_low_f16(vmin));
212 vst1_lane_f16(o, vout0, 0);
213 o = (__fp16*restrict) ((uintptr_t) o + output_stride);
214 } while (--n != 0);
215 o = (__fp16*restrict) ((uintptr_t) o - output_decrement);
216 i += 1;
217 }
218 }
219 }
220