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_16x1__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 union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f16_spmm_minmax_ukernel_16x1__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 union xnn_f16_minmax_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 vmax = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neon.max));
36 const float16x8_t vmin = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neon.min));
37
38 size_t output_decrement = output_stride * nc - 16 * sizeof(__fp16);
39 while XNN_LIKELY(mc >= 16 * sizeof(__fp16)) {
40 const __fp16*restrict w = (const __fp16*) weights;
41 const int32_t* dmap = widx_dmap;
42 const uint32_t* nnzmap = nidx_nnzmap;
43 size_t n = nc;
44 do {
45 uint32_t nnz = *nnzmap++;
46 float16x8_t vacc01234567 = vld1q_dup_f16(w); w += 1;
47 float16x8_t vacc89ABCDEF = vacc01234567;
48 if XNN_LIKELY(nnz != 0) {
49 do {
50 const intptr_t diff = *dmap++;
51 const float16x8_t va01234567 = vld1q_f16(i);
52 const float16x8_t va89ABCDEF = vld1q_f16(i + 8);
53 i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff);
54 const float16x8_t vb = vld1q_dup_f16(w); w += 1;
55 vacc01234567 = vfmaq_f16(vacc01234567, va01234567, vb);
56 vacc89ABCDEF = vfmaq_f16(vacc89ABCDEF, va89ABCDEF, vb);
57 } while (--nnz != 0);
58 }
59 float16x8_t vout01234567 = vminq_f16(vacc01234567, vmax);
60 float16x8_t vout89ABCDEF = vminq_f16(vacc89ABCDEF, vmax);
61 vout01234567 = vmaxq_f16(vout01234567, vmin);
62 vout89ABCDEF = vmaxq_f16(vout89ABCDEF, vmin);
63 vst1q_f16(o, vout01234567);
64 vst1q_f16(o + 8, vout89ABCDEF);
65 o = (__fp16*restrict) ((uintptr_t) o + output_stride);
66 } while (--n != 0);
67 o = (__fp16*restrict) ((uintptr_t) o - output_decrement);
68 i += 16;
69 mc -= 16 * sizeof(__fp16);
70 }
71 if XNN_UNLIKELY(mc != 0) {
72 output_decrement += 8 * sizeof(__fp16);
73 if (mc & (8 * sizeof(__fp16))) {
74 const __fp16*restrict w = (const __fp16*) weights;
75 const int32_t* dmap = widx_dmap;
76 const uint32_t* nnzmap = nidx_nnzmap;
77 size_t n = nc;
78 do {
79 uint32_t nnz = *nnzmap++;
80 float16x8_t vacc01234567 = vld1q_dup_f16(w); w += 1;
81 if XNN_LIKELY(nnz != 0) {
82 do {
83 const intptr_t diff = *dmap++;
84 const float16x8_t va01234567 = vld1q_f16(i);
85 i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff);
86 const float16x8_t vb = vld1q_dup_f16(w); w += 1;
87 vacc01234567 = vfmaq_f16(vacc01234567, va01234567, vb);
88 } while (--nnz != 0);
89 }
90 float16x8_t vout01234567 = vminq_f16(vacc01234567, vmax);
91 vout01234567 = vmaxq_f16(vout01234567, vmin);
92 vst1q_f16(o, vout01234567);
93 o = (__fp16*restrict) ((uintptr_t) o + output_stride);
94 } while (--n != 0);
95 o = (__fp16*restrict) ((uintptr_t) o - output_decrement);
96 i += 8;
97 }
98 output_decrement += 4 * sizeof(__fp16);
99 if (mc & (4 * sizeof(__fp16))) {
100 const __fp16*restrict w = (const __fp16*) weights;
101 const int32_t* dmap = widx_dmap;
102 const uint32_t* nnzmap = nidx_nnzmap;
103 size_t n = nc;
104 do {
105 uint32_t nnz = *nnzmap++;
106 float16x4_t vacc0123 = vld1_dup_f16(w); w += 1;
107 if XNN_LIKELY(nnz != 0) {
108 do {
109 const intptr_t diff = *dmap++;
110 const float16x4_t va0123 = vld1_f16(i);
111 i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff);
112 const float16x4_t vb = vld1_dup_f16(w); w += 1;
113 vacc0123 = vfma_f16(vacc0123, va0123, vb);
114 } while (--nnz != 0);
115 }
116 float16x4_t vout0123 = vmin_f16(vacc0123, vget_low_f16(vmax));
117 vout0123 = vmax_f16(vout0123, vget_low_f16(vmin));
118 vst1_f16(o, vout0123);
119 o = (__fp16*restrict) ((uintptr_t) o + output_stride);
120 } while (--n != 0);
121 o = (__fp16*restrict) ((uintptr_t) o - output_decrement);
122 i += 4;
123 }
124 output_decrement += 2 * sizeof(__fp16);
125 if (mc & (2 * sizeof(__fp16))) {
126 const __fp16*restrict w = (const __fp16*) weights;
127 const int32_t* dmap = widx_dmap;
128 const uint32_t* nnzmap = nidx_nnzmap;
129 size_t n = nc;
130 do {
131 uint32_t nnz = *nnzmap++;
132 float16x4_t vacc01 = vld1_dup_f16(w); w += 1;
133 if XNN_LIKELY(nnz != 0) {
134 do {
135 const intptr_t diff = *dmap++;
136 const float16x4_t va01 = vreinterpret_f16_f32(vld1_dup_f32((const void*) i));
137 i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff);
138 const float16x4_t vb = vld1_dup_f16(w); w += 1;
139 vacc01 = vfma_f16(vacc01, va01, vb);
140 } while (--nnz != 0);
141 }
142 float16x4_t vout01 = vmin_f16(vacc01, vget_low_f16(vmax));
143 vout01 = vmax_f16(vout01, vget_low_f16(vmin));
144 vst1_lane_f32((void*) o, vreinterpret_f32_f16(vout01), 0);
145 o = (__fp16*restrict) ((uintptr_t) o + output_stride);
146 } while (--n != 0);
147 o = (__fp16*restrict) ((uintptr_t) o - output_decrement);
148 i += 2;
149 }
150 output_decrement += 1 * sizeof(__fp16);
151 if (mc & (1 * sizeof(__fp16))) {
152 const __fp16*restrict w = (const __fp16*) weights;
153 const int32_t* dmap = widx_dmap;
154 const uint32_t* nnzmap = nidx_nnzmap;
155 size_t n = nc;
156 do {
157 uint32_t nnz = *nnzmap++;
158 float16x4_t vacc0 = vld1_dup_f16(w); w += 1;
159 if XNN_LIKELY(nnz != 0) {
160 do {
161 const intptr_t diff = *dmap++;
162 const float16x4_t va0 = vld1_dup_f16(i);
163 i = (const __fp16*restrict) ((uintptr_t) i + (uintptr_t) diff);
164 const float16x4_t vb = vld1_dup_f16(w); w += 1;
165 vacc0 = vfma_f16(vacc0, va0, vb);
166 } while (--nnz != 0);
167 }
168 float16x4_t vout0 = vmin_f16(vacc0, vget_low_f16(vmax));
169 vout0 = vmax_f16(vout0, vget_low_f16(vmin));
170 vst1_lane_f16(o, vout0, 0);
171 o = (__fp16*restrict) ((uintptr_t) o + output_stride);
172 } while (--n != 0);
173 o = (__fp16*restrict) ((uintptr_t) o - output_decrement);
174 i += 1;
175 }
176 }
177 }
178