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_2x4c8__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_2x4c8__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 <= 2);
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
48 const bfloat16_t* w = (const bfloat16_t*) w_ptr;
49 do {
50 float32x4_t vacc0x0 = vcvt_f32_bf16(vld1_lane_bf16(w, vreinterpret_bf16_u16(vdup_n_u16(0)), 0)); w += 1;
51 float32x4_t vacc0x1 = vcvt_f32_bf16(vld1_lane_bf16(w, vreinterpret_bf16_u16(vdup_n_u16(0)), 0)); w += 1;
52 float32x4_t vacc0x2 = vcvt_f32_bf16(vld1_lane_bf16(w, vreinterpret_bf16_u16(vdup_n_u16(0)), 0)); w += 1;
53 float32x4_t vacc0x3 = vcvt_f32_bf16(vld1_lane_bf16(w, vreinterpret_bf16_u16(vdup_n_u16(0)), 0)); w += 1;
54 float32x4_t vacc1x0 = vacc0x0;
55 float32x4_t vacc1x1 = vacc0x1;
56 float32x4_t vacc1x2 = vacc0x2;
57 float32x4_t vacc1x3 = vacc0x3;
58
59 size_t k = kc;
60 for (; k >= 8 * sizeof(bfloat16_t); k -= 8 * sizeof(bfloat16_t)) {
61 const bfloat16x8_t va0 = vld1q_bf16(a0); a0 += 8;
62 const bfloat16x8_t va1 = vld1q_bf16(a1); a1 += 8;
63
64 const bfloat16x8_t vb0 = vld1q_bf16(w); w += 8;
65 const bfloat16x8_t vb1 = vld1q_bf16(w); w += 8;
66 const bfloat16x8_t vb2 = vld1q_bf16(w); w += 8;
67 const bfloat16x8_t vb3 = vld1q_bf16(w); w += 8;
68
69 vacc0x0 = vbfmlalbq_f32(vacc0x0, va0, vb0);
70 vacc1x0 = vbfmlalbq_f32(vacc1x0, va1, vb0);
71 vacc0x1 = vbfmlalbq_f32(vacc0x1, va0, vb1);
72 vacc1x1 = vbfmlalbq_f32(vacc1x1, va1, vb1);
73 vacc0x2 = vbfmlalbq_f32(vacc0x2, va0, vb2);
74 vacc1x2 = vbfmlalbq_f32(vacc1x2, va1, vb2);
75 vacc0x3 = vbfmlalbq_f32(vacc0x3, va0, vb3);
76 vacc1x3 = vbfmlalbq_f32(vacc1x3, va1, vb3);
77
78 vacc0x0 = vbfmlaltq_f32(vacc0x0, va0, vb0);
79 vacc1x0 = vbfmlaltq_f32(vacc1x0, va1, vb0);
80 vacc0x1 = vbfmlaltq_f32(vacc0x1, va0, vb1);
81 vacc1x1 = vbfmlaltq_f32(vacc1x1, va1, vb1);
82 vacc0x2 = vbfmlaltq_f32(vacc0x2, va0, vb2);
83 vacc1x2 = vbfmlaltq_f32(vacc1x2, va1, vb2);
84 vacc0x3 = vbfmlaltq_f32(vacc0x3, va0, vb3);
85 vacc1x3 = vbfmlaltq_f32(vacc1x3, va1, vb3);
86 }
87 if XNN_UNLIKELY(k != 0) {
88 const bfloat16x8_t va0 = vld1q_bf16(a0); a0 = (const bfloat16_t*) ((uintptr_t) a0 + k);
89 const bfloat16x8_t va1 = vld1q_bf16(a1); a1 = (const bfloat16_t*) ((uintptr_t) a1 + k);
90
91 const bfloat16x8_t vb0 = vld1q_bf16(w); w += 8;
92 const bfloat16x8_t vb1 = vld1q_bf16(w); w += 8;
93 const bfloat16x8_t vb2 = vld1q_bf16(w); w += 8;
94 const bfloat16x8_t vb3 = vld1q_bf16(w); w += 8;
95
96 const uint16x8_t vm0 = vceqq_u16(vreinterpretq_u16_bf16(vb0), vmovq_n_u16(0));
97 const uint16x8_t vm1 = vceqq_u16(vreinterpretq_u16_bf16(vb1), vmovq_n_u16(0));
98 const uint16x8_t vm2 = vceqq_u16(vreinterpretq_u16_bf16(vb2), vmovq_n_u16(0));
99 const uint16x8_t vm3 = vceqq_u16(vreinterpretq_u16_bf16(vb3), vmovq_n_u16(0));
100
101 const bfloat16x8_t va0x0 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va0), vm0));
102 vacc0x0 = vbfmlalbq_f32(vacc0x0, va0x0, vb0);
103 vacc0x0 = vbfmlaltq_f32(vacc0x0, va0x0, vb0);
104 const bfloat16x8_t va1x0 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va1), vm0));
105 vacc1x0 = vbfmlalbq_f32(vacc1x0, va1x0, vb0);
106 vacc1x0 = vbfmlaltq_f32(vacc1x0, va1x0, vb0);
107 const bfloat16x8_t va0x1 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va0), vm1));
108 vacc0x1 = vbfmlalbq_f32(vacc0x1, va0x1, vb1);
109 vacc0x1 = vbfmlaltq_f32(vacc0x1, va0x1, vb1);
110 const bfloat16x8_t va1x1 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va1), vm1));
111 vacc1x1 = vbfmlalbq_f32(vacc1x1, va1x1, vb1);
112 vacc1x1 = vbfmlaltq_f32(vacc1x1, va1x1, vb1);
113 const bfloat16x8_t va0x2 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va0), vm2));
114 vacc0x2 = vbfmlalbq_f32(vacc0x2, va0x2, vb2);
115 vacc0x2 = vbfmlaltq_f32(vacc0x2, va0x2, vb2);
116 const bfloat16x8_t va1x2 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va1), vm2));
117 vacc1x2 = vbfmlalbq_f32(vacc1x2, va1x2, vb2);
118 vacc1x2 = vbfmlaltq_f32(vacc1x2, va1x2, vb2);
119 const bfloat16x8_t va0x3 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va0), vm3));
120 vacc0x3 = vbfmlalbq_f32(vacc0x3, va0x3, vb3);
121 vacc0x3 = vbfmlaltq_f32(vacc0x3, va0x3, vb3);
122 const bfloat16x8_t va1x3 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va1), vm3));
123 vacc1x3 = vbfmlalbq_f32(vacc1x3, va1x3, vb3);
124 vacc1x3 = vbfmlaltq_f32(vacc1x3, va1x3, vb3);
125 }
126
127 #if XNN_ARCH_ARM64
128 const float32x4_t vacc0x01 = vpaddq_f32(vacc0x0, vacc0x1);
129 const float32x4_t vacc1x01 = vpaddq_f32(vacc1x0, vacc1x1);
130 const float32x4_t vacc0x23 = vpaddq_f32(vacc0x2, vacc0x3);
131 const float32x4_t vacc1x23 = vpaddq_f32(vacc1x2, vacc1x3);
132
133 float32x4_t vacc0x0123 = vpaddq_f32(vacc0x01, vacc0x23);
134 float32x4_t vacc1x0123 = vpaddq_f32(vacc1x01, vacc1x23);
135 #else
136 const float32x2_t vsum0x0 = vadd_f32(vget_low_f32(vacc0x0), vget_high_f32(vacc0x0));
137 const float32x2_t vsum1x0 = vadd_f32(vget_low_f32(vacc1x0), vget_high_f32(vacc1x0));
138 const float32x2_t vsum0x1 = vadd_f32(vget_low_f32(vacc0x1), vget_high_f32(vacc0x1));
139 const float32x2_t vsum1x1 = vadd_f32(vget_low_f32(vacc1x1), vget_high_f32(vacc1x1));
140 const float32x2_t vsum0x2 = vadd_f32(vget_low_f32(vacc0x2), vget_high_f32(vacc0x2));
141 const float32x2_t vsum1x2 = vadd_f32(vget_low_f32(vacc1x2), vget_high_f32(vacc1x2));
142 const float32x2_t vsum0x3 = vadd_f32(vget_low_f32(vacc0x3), vget_high_f32(vacc0x3));
143 const float32x2_t vsum1x3 = vadd_f32(vget_low_f32(vacc1x3), vget_high_f32(vacc1x3));
144
145 float32x4_t vacc0x0123 = vcombine_f32(vpadd_f32(vsum0x0, vsum0x1), vpadd_f32(vsum0x2, vsum0x3));
146 float32x4_t vacc1x0123 = vcombine_f32(vpadd_f32(vsum1x0, vsum1x1), vpadd_f32(vsum1x2, vsum1x3));
147 #endif
148
149 const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max);
150 vacc0x0123 = vminq_f32(vacc0x0123, vmax);
151 vacc1x0123 = vminq_f32(vacc1x0123, vmax);
152
153 const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min);
154 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
155 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
156
157 bfloat16x4_t vout0x0123 = vcvt_bf16_f32(vacc0x0123);
158 bfloat16x4_t vout1x0123 = vcvt_bf16_f32(vacc1x0123);
159
160 if XNN_LIKELY(nc >= 4) {
161 vst1_bf16(c0, vout0x0123);
162 c0 = (bfloat16_t*) ((uintptr_t) c0 + cn_stride);
163 vst1_bf16(c1, vout1x0123);
164 c1 = (bfloat16_t*) ((uintptr_t) c1 + cn_stride);
165
166 a0 = (const bfloat16_t*) ((uintptr_t) a0 - kc);
167 a1 = (const bfloat16_t*) ((uintptr_t) a1 - kc);
168
169 nc -= 4;
170 } else {
171 if (nc & 2) {
172 vst1_lane_u32((void*) c0, vreinterpret_u32_bf16(vout0x0123), 0); c0 += 2;
173 vst1_lane_u32((void*) c1, vreinterpret_u32_bf16(vout1x0123), 0); c1 += 2;
174
175 vout0x0123 = vreinterpret_bf16_u16(vext_u16(vreinterpret_u16_bf16(vout0x0123), vreinterpret_u16_bf16(vout0x0123), 2));
176 vout1x0123 = vreinterpret_bf16_u16(vext_u16(vreinterpret_u16_bf16(vout1x0123), vreinterpret_u16_bf16(vout1x0123), 2));
177 }
178 if (nc & 1) {
179 vst1_lane_bf16(c0, vout0x0123, 0);
180 vst1_lane_bf16(c1, vout1x0123, 0);
181 }
182
183 nc = 0;
184 }
185 } while (nc != 0);
186 }
187