1 // Auto-generated file. Do not edit!
2 // Template: src/bf16-gemm/c8-neon.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2022 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
11 #include <assert.h>
12
13 #include <arm_neon.h>
14
15 #include <xnnpack/gemm.h>
16
17
xnn_bf16_gemm_minmax_ukernel_2x4c8__neonfma_zip(size_t mr,size_t nc,size_t kc,const void * restrict a,size_t a_stride,const void * restrict w_ptr,void * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_bf16_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_bf16_gemm_minmax_ukernel_2x4c8__neonfma_zip(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const void* restrict a,
23 size_t a_stride,
24 const void* restrict w_ptr,
25 void* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_bf16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
29 {
30 assert(mr != 0);
31 assert(mr <= 2);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(uint16_t) == 0);
35 assert(a != NULL);
36 assert(w_ptr != NULL);
37 assert(c != NULL);
38
39 const uint16_t* a0 = (const uint16_t*) a;
40 uint16_t* c0 = (uint16_t*) c;
41 const uint16_t* a1 = (const uint16_t*) ((uintptr_t) a0 + a_stride);
42 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
43 if XNN_UNPREDICTABLE(mr != 2) {
44 a1 = a0;
45 c1 = c0;
46 }
47
48 const uint16_t* w = (const uint16_t*) w_ptr;
49 const uint16x8_t vzero = vmovq_n_u16(0);
50 do {
51 float32x4_t vacc0x0 = vreinterpretq_f32_u32(vshll_n_u16(vld1_lane_u16(w, vdup_n_u16(0), 0), 16)); w += 1;
52 float32x4_t vacc0x1 = vreinterpretq_f32_u32(vshll_n_u16(vld1_lane_u16(w, vdup_n_u16(0), 0), 16)); w += 1;
53 float32x4_t vacc0x2 = vreinterpretq_f32_u32(vshll_n_u16(vld1_lane_u16(w, vdup_n_u16(0), 0), 16)); w += 1;
54 float32x4_t vacc0x3 = vreinterpretq_f32_u32(vshll_n_u16(vld1_lane_u16(w, vdup_n_u16(0), 0), 16)); w += 1;
55 float32x4_t vacc1x0 = vacc0x0;
56 float32x4_t vacc1x1 = vacc0x1;
57 float32x4_t vacc1x2 = vacc0x2;
58 float32x4_t vacc1x3 = vacc0x3;
59
60 size_t k = kc;
61 for (; k >= 8 * sizeof(uint16_t); k -= 8 * sizeof(uint16_t)) {
62 const uint16x8_t va0 = vld1q_u16(a0); a0 += 8;
63 const uint16x8_t va1 = vld1q_u16(a1); a1 += 8;
64
65 const uint16x8_t vb0 = vld1q_u16(w); w += 8;
66 const uint16x8_t vb1 = vld1q_u16(w); w += 8;
67 const uint16x8_t vb2 = vld1q_u16(w); w += 8;
68 const uint16x8_t vb3 = vld1q_u16(w); w += 8;
69
70 const float32x4_t va0e = vreinterpretq_f32_u16(vzip1q_u16(vzero, va0));
71 const float32x4_t va1e = vreinterpretq_f32_u16(vzip1q_u16(vzero, va1));
72
73 const float32x4_t vb0e = vreinterpretq_f32_u16(vzip1q_u16(vzero, vb0));
74 const float32x4_t vb1e = vreinterpretq_f32_u16(vzip1q_u16(vzero, vb1));
75 const float32x4_t vb2e = vreinterpretq_f32_u16(vzip1q_u16(vzero, vb2));
76 const float32x4_t vb3e = vreinterpretq_f32_u16(vzip1q_u16(vzero, vb3));
77
78 vacc0x0 = vfmaq_f32(vacc0x0, va0e, vb0e);
79 vacc1x0 = vfmaq_f32(vacc1x0, va1e, vb0e);
80 vacc0x1 = vfmaq_f32(vacc0x1, va0e, vb1e);
81 vacc1x1 = vfmaq_f32(vacc1x1, va1e, vb1e);
82 vacc0x2 = vfmaq_f32(vacc0x2, va0e, vb2e);
83 vacc1x2 = vfmaq_f32(vacc1x2, va1e, vb2e);
84 vacc0x3 = vfmaq_f32(vacc0x3, va0e, vb3e);
85 vacc1x3 = vfmaq_f32(vacc1x3, va1e, vb3e);
86
87 const float32x4_t va0o = vreinterpretq_f32_u16(vzip2q_u16(vzero, va0));
88 const float32x4_t va1o = vreinterpretq_f32_u16(vzip2q_u16(vzero, va1));
89
90 const float32x4_t vb0o = vreinterpretq_f32_u16(vzip2q_u16(vzero, vb0));
91 const float32x4_t vb1o = vreinterpretq_f32_u16(vzip2q_u16(vzero, vb1));
92 const float32x4_t vb2o = vreinterpretq_f32_u16(vzip2q_u16(vzero, vb2));
93 const float32x4_t vb3o = vreinterpretq_f32_u16(vzip2q_u16(vzero, vb3));
94
95 vacc0x0 = vfmaq_f32(vacc0x0, va0o, vb0o);
96 vacc1x0 = vfmaq_f32(vacc1x0, va1o, vb0o);
97 vacc0x1 = vfmaq_f32(vacc0x1, va0o, vb1o);
98 vacc1x1 = vfmaq_f32(vacc1x1, va1o, vb1o);
99 vacc0x2 = vfmaq_f32(vacc0x2, va0o, vb2o);
100 vacc1x2 = vfmaq_f32(vacc1x2, va1o, vb2o);
101 vacc0x3 = vfmaq_f32(vacc0x3, va0o, vb3o);
102 vacc1x3 = vfmaq_f32(vacc1x3, va1o, vb3o);
103 }
104 if XNN_UNLIKELY(k != 0) {
105 const uint16x8_t va0 = vld1q_u16(a0); a0 = (const uint16_t*) ((uintptr_t) a0 + k);
106 const uint16x8_t va1 = vld1q_u16(a1); a1 = (const uint16_t*) ((uintptr_t) a1 + k);
107
108 const uint16x8_t vb0 = vld1q_u16(w); w += 8;
109 const uint16x8_t vb1 = vld1q_u16(w); w += 8;
110 const uint16x8_t vb2 = vld1q_u16(w); w += 8;
111 const uint16x8_t vb3 = vld1q_u16(w); w += 8;
112
113 const uint16x8_t vm0 = vceqq_u16(vb0, vmovq_n_u16(0));
114 const uint16x8_t vm1 = vceqq_u16(vb1, vmovq_n_u16(0));
115 const uint16x8_t vm2 = vceqq_u16(vb2, vmovq_n_u16(0));
116 const uint16x8_t vm3 = vceqq_u16(vb3, vmovq_n_u16(0));
117
118 const float32x4_t vb0e = vreinterpretq_f32_u16(vzip1q_u16(vzero, vb0));
119 const float32x4_t vb1e = vreinterpretq_f32_u16(vzip1q_u16(vzero, vb1));
120 const float32x4_t vb2e = vreinterpretq_f32_u16(vzip1q_u16(vzero, vb2));
121 const float32x4_t vb3e = vreinterpretq_f32_u16(vzip1q_u16(vzero, vb3));
122
123 const uint16x8_t va0x0 = vbicq_u16(va0, vm0);
124 const uint16x8_t va1x0 = vbicq_u16(va1, vm0);
125 const uint16x8_t va0x1 = vbicq_u16(va0, vm1);
126 const uint16x8_t va1x1 = vbicq_u16(va1, vm1);
127 const uint16x8_t va0x2 = vbicq_u16(va0, vm2);
128 const uint16x8_t va1x2 = vbicq_u16(va1, vm2);
129 const uint16x8_t va0x3 = vbicq_u16(va0, vm3);
130 const uint16x8_t va1x3 = vbicq_u16(va1, vm3);
131
132 const float32x4_t va0x0e = vreinterpretq_f32_u16(vzip1q_u16(vzero, va0x0));
133 const float32x4_t va1x0e = vreinterpretq_f32_u16(vzip1q_u16(vzero, va1x0));
134 const float32x4_t va0x1e = vreinterpretq_f32_u16(vzip1q_u16(vzero, va0x1));
135 const float32x4_t va1x1e = vreinterpretq_f32_u16(vzip1q_u16(vzero, va1x1));
136 const float32x4_t va0x2e = vreinterpretq_f32_u16(vzip1q_u16(vzero, va0x2));
137 const float32x4_t va1x2e = vreinterpretq_f32_u16(vzip1q_u16(vzero, va1x2));
138 const float32x4_t va0x3e = vreinterpretq_f32_u16(vzip1q_u16(vzero, va0x3));
139 const float32x4_t va1x3e = vreinterpretq_f32_u16(vzip1q_u16(vzero, va1x3));
140
141 vacc0x0 = vfmaq_f32(vacc0x0, va0x0e, vb0e);
142 vacc1x0 = vfmaq_f32(vacc1x0, va1x0e, vb0e);
143 vacc0x1 = vfmaq_f32(vacc0x1, va0x1e, vb1e);
144 vacc1x1 = vfmaq_f32(vacc1x1, va1x1e, vb1e);
145 vacc0x2 = vfmaq_f32(vacc0x2, va0x2e, vb2e);
146 vacc1x2 = vfmaq_f32(vacc1x2, va1x2e, vb2e);
147 vacc0x3 = vfmaq_f32(vacc0x3, va0x3e, vb3e);
148 vacc1x3 = vfmaq_f32(vacc1x3, va1x3e, vb3e);
149
150 const float32x4_t vb0o = vreinterpretq_f32_u16(vzip2q_u16(vzero, vb0));
151 const float32x4_t vb1o = vreinterpretq_f32_u16(vzip2q_u16(vzero, vb1));
152 const float32x4_t vb2o = vreinterpretq_f32_u16(vzip2q_u16(vzero, vb2));
153 const float32x4_t vb3o = vreinterpretq_f32_u16(vzip2q_u16(vzero, vb3));
154
155 const float32x4_t va0x0o = vreinterpretq_f32_u16(vzip2q_u16(vzero, va0x0));
156 const float32x4_t va1x0o = vreinterpretq_f32_u16(vzip2q_u16(vzero, va1x0));
157 const float32x4_t va0x1o = vreinterpretq_f32_u16(vzip2q_u16(vzero, va0x1));
158 const float32x4_t va1x1o = vreinterpretq_f32_u16(vzip2q_u16(vzero, va1x1));
159 const float32x4_t va0x2o = vreinterpretq_f32_u16(vzip2q_u16(vzero, va0x2));
160 const float32x4_t va1x2o = vreinterpretq_f32_u16(vzip2q_u16(vzero, va1x2));
161 const float32x4_t va0x3o = vreinterpretq_f32_u16(vzip2q_u16(vzero, va0x3));
162 const float32x4_t va1x3o = vreinterpretq_f32_u16(vzip2q_u16(vzero, va1x3));
163
164 vacc0x0 = vfmaq_f32(vacc0x0, va0x0o, vb0o);
165 vacc1x0 = vfmaq_f32(vacc1x0, va1x0o, vb0o);
166 vacc0x1 = vfmaq_f32(vacc0x1, va0x1o, vb1o);
167 vacc1x1 = vfmaq_f32(vacc1x1, va1x1o, vb1o);
168 vacc0x2 = vfmaq_f32(vacc0x2, va0x2o, vb2o);
169 vacc1x2 = vfmaq_f32(vacc1x2, va1x2o, vb2o);
170 vacc0x3 = vfmaq_f32(vacc0x3, va0x3o, vb3o);
171 vacc1x3 = vfmaq_f32(vacc1x3, va1x3o, vb3o);
172 }
173
174 #if XNN_ARCH_ARM64
175 const float32x4_t vacc0x01 = vpaddq_f32(vacc0x0, vacc0x1);
176 const float32x4_t vacc1x01 = vpaddq_f32(vacc1x0, vacc1x1);
177 const float32x4_t vacc0x23 = vpaddq_f32(vacc0x2, vacc0x3);
178 const float32x4_t vacc1x23 = vpaddq_f32(vacc1x2, vacc1x3);
179
180 float32x4_t vacc0x0123 = vpaddq_f32(vacc0x01, vacc0x23);
181 float32x4_t vacc1x0123 = vpaddq_f32(vacc1x01, vacc1x23);
182 #else
183 const float32x2_t vsum0x0 = vadd_f32(vget_low_f32(vacc0x0), vget_high_f32(vacc0x0));
184 const float32x2_t vsum1x0 = vadd_f32(vget_low_f32(vacc1x0), vget_high_f32(vacc1x0));
185 const float32x2_t vsum0x1 = vadd_f32(vget_low_f32(vacc0x1), vget_high_f32(vacc0x1));
186 const float32x2_t vsum1x1 = vadd_f32(vget_low_f32(vacc1x1), vget_high_f32(vacc1x1));
187 const float32x2_t vsum0x2 = vadd_f32(vget_low_f32(vacc0x2), vget_high_f32(vacc0x2));
188 const float32x2_t vsum1x2 = vadd_f32(vget_low_f32(vacc1x2), vget_high_f32(vacc1x2));
189 const float32x2_t vsum0x3 = vadd_f32(vget_low_f32(vacc0x3), vget_high_f32(vacc0x3));
190 const float32x2_t vsum1x3 = vadd_f32(vget_low_f32(vacc1x3), vget_high_f32(vacc1x3));
191
192 float32x4_t vacc0x0123 = vcombine_f32(vpadd_f32(vsum0x0, vsum0x1), vpadd_f32(vsum0x2, vsum0x3));
193 float32x4_t vacc1x0123 = vcombine_f32(vpadd_f32(vsum1x0, vsum1x1), vpadd_f32(vsum1x2, vsum1x3));
194 #endif
195
196 const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max);
197 vacc0x0123 = vminq_f32(vacc0x0123, vmax);
198 vacc1x0123 = vminq_f32(vacc1x0123, vmax);
199
200 const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min);
201 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
202 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
203
204 uint16x4_t vout0x0123 = vshrn_n_u32(vreinterpretq_u32_f32(vacc0x0123), 16);
205 uint16x4_t vout1x0123 = vshrn_n_u32(vreinterpretq_u32_f32(vacc1x0123), 16);
206
207 if XNN_LIKELY(nc >= 4) {
208 vst1_u16(c0, vout0x0123);
209 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
210 vst1_u16(c1, vout1x0123);
211 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
212
213 a0 = (const uint16_t*) ((uintptr_t) a0 - kc);
214 a1 = (const uint16_t*) ((uintptr_t) a1 - kc);
215
216 nc -= 4;
217 } else {
218 if (nc & 2) {
219 vst1_lane_u32((void*) c0, vreinterpret_u32_u16(vout0x0123), 0); c0 += 2;
220 vst1_lane_u32((void*) c1, vreinterpret_u32_u16(vout1x0123), 0); c1 += 2;
221
222 vout0x0123 = vext_u16(vout0x0123, vout0x0123, 2);
223 vout1x0123 = vext_u16(vout1x0123, vout1x0123, 2);
224 }
225 if (nc & 1) {
226 vst1_lane_u16(c0, vout0x0123, 0);
227 vst1_lane_u16(c1, vout1x0123, 0);
228 }
229
230 nc = 0;
231 }
232 } while (nc != 0);
233 }
234