1 // Auto-generated file. Do not edit!
2 // Template: src/bf16-gemm/c8-neonbf16.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_3x4c8__neonbf16_bfmlal(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_3x4c8__neonbf16_bfmlal(
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 <= 3);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(bfloat16_t) == 0);
35 assert(a != NULL);
36 assert(w_ptr != NULL);
37 assert(c != NULL);
38
39 const bfloat16_t* a0 = (const bfloat16_t*) a;
40 bfloat16_t* c0 = (bfloat16_t*) c;
41 const bfloat16_t* a1 = (const bfloat16_t*) ((uintptr_t) a0 + a_stride);
42 bfloat16_t* c1 = (bfloat16_t*) ((uintptr_t) c0 + cm_stride);
43 if XNN_UNPREDICTABLE(mr < 2) {
44 a1 = a0;
45 c1 = c0;
46 }
47 const bfloat16_t* a2 = (const bfloat16_t*) ((uintptr_t) a1 + a_stride);
48 bfloat16_t* c2 = (bfloat16_t*) ((uintptr_t) c1 + cm_stride);
49 if XNN_UNPREDICTABLE(mr <= 2) {
50 a2 = a1;
51 c2 = c1;
52 }
53
54 const bfloat16_t* w = (const bfloat16_t*) w_ptr;
55 do {
56 float32x4_t vacc0x0 = vcvt_f32_bf16(vld1_lane_bf16(w, vreinterpret_bf16_u16(vdup_n_u16(0)), 0)); w += 1;
57 float32x4_t vacc0x1 = vcvt_f32_bf16(vld1_lane_bf16(w, vreinterpret_bf16_u16(vdup_n_u16(0)), 0)); w += 1;
58 float32x4_t vacc0x2 = vcvt_f32_bf16(vld1_lane_bf16(w, vreinterpret_bf16_u16(vdup_n_u16(0)), 0)); w += 1;
59 float32x4_t vacc0x3 = vcvt_f32_bf16(vld1_lane_bf16(w, vreinterpret_bf16_u16(vdup_n_u16(0)), 0)); w += 1;
60 float32x4_t vacc1x0 = vacc0x0;
61 float32x4_t vacc1x1 = vacc0x1;
62 float32x4_t vacc1x2 = vacc0x2;
63 float32x4_t vacc1x3 = vacc0x3;
64 float32x4_t vacc2x0 = vacc0x0;
65 float32x4_t vacc2x1 = vacc0x1;
66 float32x4_t vacc2x2 = vacc0x2;
67 float32x4_t vacc2x3 = vacc0x3;
68
69 size_t k = kc;
70 for (; k >= 8 * sizeof(bfloat16_t); k -= 8 * sizeof(bfloat16_t)) {
71 const bfloat16x8_t va0 = vld1q_bf16(a0); a0 += 8;
72 const bfloat16x8_t va1 = vld1q_bf16(a1); a1 += 8;
73 const bfloat16x8_t va2 = vld1q_bf16(a2); a2 += 8;
74
75 const bfloat16x8_t vb0 = vld1q_bf16(w); w += 8;
76 const bfloat16x8_t vb1 = vld1q_bf16(w); w += 8;
77 const bfloat16x8_t vb2 = vld1q_bf16(w); w += 8;
78 const bfloat16x8_t vb3 = vld1q_bf16(w); w += 8;
79
80 vacc0x0 = vbfmlalbq_f32(vacc0x0, va0, vb0);
81 vacc1x0 = vbfmlalbq_f32(vacc1x0, va1, vb0);
82 vacc2x0 = vbfmlalbq_f32(vacc2x0, va2, vb0);
83 vacc0x1 = vbfmlalbq_f32(vacc0x1, va0, vb1);
84 vacc1x1 = vbfmlalbq_f32(vacc1x1, va1, vb1);
85 vacc2x1 = vbfmlalbq_f32(vacc2x1, va2, vb1);
86 vacc0x2 = vbfmlalbq_f32(vacc0x2, va0, vb2);
87 vacc1x2 = vbfmlalbq_f32(vacc1x2, va1, vb2);
88 vacc2x2 = vbfmlalbq_f32(vacc2x2, va2, vb2);
89 vacc0x3 = vbfmlalbq_f32(vacc0x3, va0, vb3);
90 vacc1x3 = vbfmlalbq_f32(vacc1x3, va1, vb3);
91 vacc2x3 = vbfmlalbq_f32(vacc2x3, va2, vb3);
92
93 vacc0x0 = vbfmlaltq_f32(vacc0x0, va0, vb0);
94 vacc1x0 = vbfmlaltq_f32(vacc1x0, va1, vb0);
95 vacc2x0 = vbfmlaltq_f32(vacc2x0, va2, vb0);
96 vacc0x1 = vbfmlaltq_f32(vacc0x1, va0, vb1);
97 vacc1x1 = vbfmlaltq_f32(vacc1x1, va1, vb1);
98 vacc2x1 = vbfmlaltq_f32(vacc2x1, va2, vb1);
99 vacc0x2 = vbfmlaltq_f32(vacc0x2, va0, vb2);
100 vacc1x2 = vbfmlaltq_f32(vacc1x2, va1, vb2);
101 vacc2x2 = vbfmlaltq_f32(vacc2x2, va2, vb2);
102 vacc0x3 = vbfmlaltq_f32(vacc0x3, va0, vb3);
103 vacc1x3 = vbfmlaltq_f32(vacc1x3, va1, vb3);
104 vacc2x3 = vbfmlaltq_f32(vacc2x3, va2, vb3);
105 }
106 if XNN_UNLIKELY(k != 0) {
107 const bfloat16x8_t va0 = vld1q_bf16(a0); a0 = (const bfloat16_t*) ((uintptr_t) a0 + k);
108 const bfloat16x8_t va1 = vld1q_bf16(a1); a1 = (const bfloat16_t*) ((uintptr_t) a1 + k);
109 const bfloat16x8_t va2 = vld1q_bf16(a2); a2 = (const bfloat16_t*) ((uintptr_t) a2 + k);
110
111 const bfloat16x8_t vb0 = vld1q_bf16(w); w += 8;
112 const bfloat16x8_t vb1 = vld1q_bf16(w); w += 8;
113 const bfloat16x8_t vb2 = vld1q_bf16(w); w += 8;
114 const bfloat16x8_t vb3 = vld1q_bf16(w); w += 8;
115
116 const uint16x8_t vm0 = vceqq_u16(vreinterpretq_u16_bf16(vb0), vmovq_n_u16(0));
117 const uint16x8_t vm1 = vceqq_u16(vreinterpretq_u16_bf16(vb1), vmovq_n_u16(0));
118 const uint16x8_t vm2 = vceqq_u16(vreinterpretq_u16_bf16(vb2), vmovq_n_u16(0));
119 const uint16x8_t vm3 = vceqq_u16(vreinterpretq_u16_bf16(vb3), vmovq_n_u16(0));
120
121 const bfloat16x8_t va0x0 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va0), vm0));
122 vacc0x0 = vbfmlalbq_f32(vacc0x0, va0x0, vb0);
123 vacc0x0 = vbfmlaltq_f32(vacc0x0, va0x0, vb0);
124 const bfloat16x8_t va1x0 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va1), vm0));
125 vacc1x0 = vbfmlalbq_f32(vacc1x0, va1x0, vb0);
126 vacc1x0 = vbfmlaltq_f32(vacc1x0, va1x0, vb0);
127 const bfloat16x8_t va2x0 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va2), vm0));
128 vacc2x0 = vbfmlalbq_f32(vacc2x0, va2x0, vb0);
129 vacc2x0 = vbfmlaltq_f32(vacc2x0, va2x0, vb0);
130 const bfloat16x8_t va0x1 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va0), vm1));
131 vacc0x1 = vbfmlalbq_f32(vacc0x1, va0x1, vb1);
132 vacc0x1 = vbfmlaltq_f32(vacc0x1, va0x1, vb1);
133 const bfloat16x8_t va1x1 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va1), vm1));
134 vacc1x1 = vbfmlalbq_f32(vacc1x1, va1x1, vb1);
135 vacc1x1 = vbfmlaltq_f32(vacc1x1, va1x1, vb1);
136 const bfloat16x8_t va2x1 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va2), vm1));
137 vacc2x1 = vbfmlalbq_f32(vacc2x1, va2x1, vb1);
138 vacc2x1 = vbfmlaltq_f32(vacc2x1, va2x1, vb1);
139 const bfloat16x8_t va0x2 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va0), vm2));
140 vacc0x2 = vbfmlalbq_f32(vacc0x2, va0x2, vb2);
141 vacc0x2 = vbfmlaltq_f32(vacc0x2, va0x2, vb2);
142 const bfloat16x8_t va1x2 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va1), vm2));
143 vacc1x2 = vbfmlalbq_f32(vacc1x2, va1x2, vb2);
144 vacc1x2 = vbfmlaltq_f32(vacc1x2, va1x2, vb2);
145 const bfloat16x8_t va2x2 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va2), vm2));
146 vacc2x2 = vbfmlalbq_f32(vacc2x2, va2x2, vb2);
147 vacc2x2 = vbfmlaltq_f32(vacc2x2, va2x2, vb2);
148 const bfloat16x8_t va0x3 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va0), vm3));
149 vacc0x3 = vbfmlalbq_f32(vacc0x3, va0x3, vb3);
150 vacc0x3 = vbfmlaltq_f32(vacc0x3, va0x3, vb3);
151 const bfloat16x8_t va1x3 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va1), vm3));
152 vacc1x3 = vbfmlalbq_f32(vacc1x3, va1x3, vb3);
153 vacc1x3 = vbfmlaltq_f32(vacc1x3, va1x3, vb3);
154 const bfloat16x8_t va2x3 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va2), vm3));
155 vacc2x3 = vbfmlalbq_f32(vacc2x3, va2x3, vb3);
156 vacc2x3 = vbfmlaltq_f32(vacc2x3, va2x3, vb3);
157 }
158
159 #if XNN_ARCH_ARM64
160 const float32x4_t vacc0x01 = vpaddq_f32(vacc0x0, vacc0x1);
161 const float32x4_t vacc1x01 = vpaddq_f32(vacc1x0, vacc1x1);
162 const float32x4_t vacc2x01 = vpaddq_f32(vacc2x0, vacc2x1);
163 const float32x4_t vacc0x23 = vpaddq_f32(vacc0x2, vacc0x3);
164 const float32x4_t vacc1x23 = vpaddq_f32(vacc1x2, vacc1x3);
165 const float32x4_t vacc2x23 = vpaddq_f32(vacc2x2, vacc2x3);
166
167 float32x4_t vacc0x0123 = vpaddq_f32(vacc0x01, vacc0x23);
168 float32x4_t vacc1x0123 = vpaddq_f32(vacc1x01, vacc1x23);
169 float32x4_t vacc2x0123 = vpaddq_f32(vacc2x01, vacc2x23);
170 #else
171 const float32x2_t vsum0x0 = vadd_f32(vget_low_f32(vacc0x0), vget_high_f32(vacc0x0));
172 const float32x2_t vsum1x0 = vadd_f32(vget_low_f32(vacc1x0), vget_high_f32(vacc1x0));
173 const float32x2_t vsum2x0 = vadd_f32(vget_low_f32(vacc2x0), vget_high_f32(vacc2x0));
174 const float32x2_t vsum0x1 = vadd_f32(vget_low_f32(vacc0x1), vget_high_f32(vacc0x1));
175 const float32x2_t vsum1x1 = vadd_f32(vget_low_f32(vacc1x1), vget_high_f32(vacc1x1));
176 const float32x2_t vsum2x1 = vadd_f32(vget_low_f32(vacc2x1), vget_high_f32(vacc2x1));
177 const float32x2_t vsum0x2 = vadd_f32(vget_low_f32(vacc0x2), vget_high_f32(vacc0x2));
178 const float32x2_t vsum1x2 = vadd_f32(vget_low_f32(vacc1x2), vget_high_f32(vacc1x2));
179 const float32x2_t vsum2x2 = vadd_f32(vget_low_f32(vacc2x2), vget_high_f32(vacc2x2));
180 const float32x2_t vsum0x3 = vadd_f32(vget_low_f32(vacc0x3), vget_high_f32(vacc0x3));
181 const float32x2_t vsum1x3 = vadd_f32(vget_low_f32(vacc1x3), vget_high_f32(vacc1x3));
182 const float32x2_t vsum2x3 = vadd_f32(vget_low_f32(vacc2x3), vget_high_f32(vacc2x3));
183
184 float32x4_t vacc0x0123 = vcombine_f32(vpadd_f32(vsum0x0, vsum0x1), vpadd_f32(vsum0x2, vsum0x3));
185 float32x4_t vacc1x0123 = vcombine_f32(vpadd_f32(vsum1x0, vsum1x1), vpadd_f32(vsum1x2, vsum1x3));
186 float32x4_t vacc2x0123 = vcombine_f32(vpadd_f32(vsum2x0, vsum2x1), vpadd_f32(vsum2x2, vsum2x3));
187 #endif
188
189 const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max);
190 vacc0x0123 = vminq_f32(vacc0x0123, vmax);
191 vacc1x0123 = vminq_f32(vacc1x0123, vmax);
192 vacc2x0123 = vminq_f32(vacc2x0123, vmax);
193
194 const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min);
195 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
196 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
197 vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
198
199 bfloat16x4_t vout0x0123 = vcvt_bf16_f32(vacc0x0123);
200 bfloat16x4_t vout1x0123 = vcvt_bf16_f32(vacc1x0123);
201 bfloat16x4_t vout2x0123 = vcvt_bf16_f32(vacc2x0123);
202
203 if XNN_LIKELY(nc >= 4) {
204 vst1_bf16(c0, vout0x0123);
205 c0 = (bfloat16_t*) ((uintptr_t) c0 + cn_stride);
206 vst1_bf16(c1, vout1x0123);
207 c1 = (bfloat16_t*) ((uintptr_t) c1 + cn_stride);
208 vst1_bf16(c2, vout2x0123);
209 c2 = (bfloat16_t*) ((uintptr_t) c2 + cn_stride);
210
211 a0 = (const bfloat16_t*) ((uintptr_t) a0 - kc);
212 a1 = (const bfloat16_t*) ((uintptr_t) a1 - kc);
213 a2 = (const bfloat16_t*) ((uintptr_t) a2 - kc);
214
215 nc -= 4;
216 } else {
217 if (nc & 2) {
218 vst1_lane_u32((void*) c0, vreinterpret_u32_bf16(vout0x0123), 0); c0 += 2;
219 vst1_lane_u32((void*) c1, vreinterpret_u32_bf16(vout1x0123), 0); c1 += 2;
220 vst1_lane_u32((void*) c2, vreinterpret_u32_bf16(vout2x0123), 0); c2 += 2;
221
222 vout0x0123 = vreinterpret_bf16_u16(vext_u16(vreinterpret_u16_bf16(vout0x0123), vreinterpret_u16_bf16(vout0x0123), 2));
223 vout1x0123 = vreinterpret_bf16_u16(vext_u16(vreinterpret_u16_bf16(vout1x0123), vreinterpret_u16_bf16(vout1x0123), 2));
224 vout2x0123 = vreinterpret_bf16_u16(vext_u16(vreinterpret_u16_bf16(vout2x0123), vreinterpret_u16_bf16(vout2x0123), 2));
225 }
226 if (nc & 1) {
227 vst1_lane_bf16(c0, vout0x0123, 0);
228 vst1_lane_bf16(c1, vout1x0123, 0);
229 vst1_lane_bf16(c2, vout2x0123, 0);
230 }
231
232 nc = 0;
233 }
234 } while (nc != 0);
235 }
236