xref: /aosp_15_r20/external/XNNPACK/src/bf16-gemm/gen/1x8c2-minmax-neonbf16-bfdot-lane-ld128.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/bf16-gemm/c2-neonbf16-bfdot-lane-ld128.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_1x8c2__neonbf16_bfdot_lane_ld128(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_1x8c2__neonbf16_bfdot_lane_ld128(
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 <= 1);
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 
42   const bfloat16_t* w = (const bfloat16_t*) w_ptr;
43   do {
44     float32x4_t vacc0x0123 = vcvt_f32_bf16(vld1_bf16(w)); w += 4;
45     float32x4_t vacc0x4567 = vcvt_f32_bf16(vld1_bf16(w)); w += 4;
46 
47     size_t k = kc;
48     for (; k >= 8 * sizeof(bfloat16_t); k -= 8 * sizeof(bfloat16_t)) {
49       const bfloat16x8_t va0 = vld1q_bf16(a0); a0 += 8;
50 
51       const bfloat16x8_t vb0123c01 = vld1q_bf16(w); w += 8;
52       const bfloat16x8_t vb4567c01 = vld1q_bf16(w); w += 8;
53 
54       vacc0x0123 = vbfdotq_laneq_f32(vacc0x0123, vb0123c01, va0, 0);
55       vacc0x4567 = vbfdotq_laneq_f32(vacc0x4567, vb4567c01, va0, 0);
56       const bfloat16x8_t vb0123c23 = vld1q_bf16(w); w += 8;
57       const bfloat16x8_t vb4567c23 = vld1q_bf16(w); w += 8;
58 
59       vacc0x0123 = vbfdotq_laneq_f32(vacc0x0123, vb0123c23, va0, 1);
60       vacc0x4567 = vbfdotq_laneq_f32(vacc0x4567, vb4567c23, va0, 1);
61       const bfloat16x8_t vb0123c45 = vld1q_bf16(w); w += 8;
62       const bfloat16x8_t vb4567c45 = vld1q_bf16(w); w += 8;
63 
64       vacc0x0123 = vbfdotq_laneq_f32(vacc0x0123, vb0123c45, va0, 2);
65       vacc0x4567 = vbfdotq_laneq_f32(vacc0x4567, vb4567c45, va0, 2);
66       const bfloat16x8_t vb0123c67 = vld1q_bf16(w); w += 8;
67       const bfloat16x8_t vb4567c67 = vld1q_bf16(w); w += 8;
68 
69       vacc0x0123 = vbfdotq_laneq_f32(vacc0x0123, vb0123c67, va0, 3);
70       vacc0x4567 = vbfdotq_laneq_f32(vacc0x4567, vb4567c67, va0, 3);
71     }
72     if XNN_UNLIKELY(k != 0) {
73       const bfloat16x8_t va0 = vld1q_bf16(a0); a0 = (const bfloat16_t*) ((uintptr_t) a0 + k);
74 
75       const bfloat16x8_t vb0123c01 = vld1q_bf16(w); w += 8;
76       const bfloat16x8_t vb4567c01 = vld1q_bf16(w); w += 8;
77 
78       const uint32x4_t va0c01 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_low_bf16(va0)), 0);
79 
80       const uint32x4_t vm0123c01 = vreinterpretq_u32_u16(vceqq_u16(vreinterpretq_u16_bf16(vb0123c01), vmovq_n_u16(0)));
81       const uint32x4_t vm4567c01 = vreinterpretq_u32_u16(vceqq_u16(vreinterpretq_u16_bf16(vb4567c01), vmovq_n_u16(0)));
82 
83       const uint32x4_t va0x0123c01 = vbicq_u32(va0c01, vm0123c01);
84       vacc0x0123 = vbfdotq_f32(vacc0x0123, vb0123c01, vreinterpretq_bf16_u32(va0x0123c01));
85       const uint32x4_t va0x4567c01 = vbicq_u32(va0c01, vm4567c01);
86       vacc0x4567 = vbfdotq_f32(vacc0x4567, vb4567c01, vreinterpretq_bf16_u32(va0x4567c01));
87 
88       if (k > 2 * sizeof(bfloat16_t)) {
89         const bfloat16x8_t vb0123c23 = vld1q_bf16(w); w += 8;
90         const bfloat16x8_t vb4567c23 = vld1q_bf16(w); w += 8;
91 
92         const uint32x4_t va0c23 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_low_bf16(va0)), 1);
93 
94         const uint32x4_t vm0123c23 = vreinterpretq_u32_u16(vceqq_u16(vreinterpretq_u16_bf16(vb0123c23), vmovq_n_u16(0)));
95         const uint32x4_t vm4567c23 = vreinterpretq_u32_u16(vceqq_u16(vreinterpretq_u16_bf16(vb4567c23), vmovq_n_u16(0)));
96 
97         const uint32x4_t va0x0123c23 = vbicq_u32(va0c23, vm0123c23);
98         vacc0x0123 = vbfdotq_f32(vacc0x0123, vb0123c23, vreinterpretq_bf16_u32(va0x0123c23));
99         const uint32x4_t va0x4567c23 = vbicq_u32(va0c23, vm4567c23);
100         vacc0x4567 = vbfdotq_f32(vacc0x4567, vb4567c23, vreinterpretq_bf16_u32(va0x4567c23));
101 
102         if (k > 4 * sizeof(bfloat16_t)) {
103           const bfloat16x8_t vb0123c45 = vld1q_bf16(w); w += 8;
104           const bfloat16x8_t vb4567c45 = vld1q_bf16(w); w += 8;
105 
106           const uint32x4_t va0c45 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_high_bf16(va0)), 0);
107 
108           const uint32x4_t vm0123c45 = vreinterpretq_u32_u16(vceqq_u16(vreinterpretq_u16_bf16(vb0123c45), vmovq_n_u16(0)));
109           const uint32x4_t vm4567c45 = vreinterpretq_u32_u16(vceqq_u16(vreinterpretq_u16_bf16(vb4567c45), vmovq_n_u16(0)));
110 
111           const uint32x4_t va0x0123c45 = vbicq_u32(va0c45, vm0123c45);
112           vacc0x0123 = vbfdotq_f32(vacc0x0123, vb0123c45, vreinterpretq_bf16_u32(va0x0123c45));
113           const uint32x4_t va0x4567c45 = vbicq_u32(va0c45, vm4567c45);
114           vacc0x4567 = vbfdotq_f32(vacc0x4567, vb4567c45, vreinterpretq_bf16_u32(va0x4567c45));
115 
116           if (k > 6 * sizeof(bfloat16_t)) {
117             const bfloat16x8_t vb0123c67 = vld1q_bf16(w); w += 8;
118             const bfloat16x8_t vb4567c67 = vld1q_bf16(w); w += 8;
119 
120             const uint32x4_t va0c67 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_high_bf16(va0)), 1);
121 
122             const uint32x4_t vm0123c67 = vreinterpretq_u32_u16(vceqq_u16(vreinterpretq_u16_bf16(vb0123c67), vmovq_n_u16(0)));
123             const uint32x4_t vm4567c67 = vreinterpretq_u32_u16(vceqq_u16(vreinterpretq_u16_bf16(vb4567c67), vmovq_n_u16(0)));
124 
125             const uint32x4_t va0x0123c67 = vbicq_u32(va0c67, vm0123c67);
126             vacc0x0123 = vbfdotq_f32(vacc0x0123, vb0123c67, vreinterpretq_bf16_u32(va0x0123c67));
127             const uint32x4_t va0x4567c67 = vbicq_u32(va0c67, vm4567c67);
128             vacc0x4567 = vbfdotq_f32(vacc0x4567, vb4567c67, vreinterpretq_bf16_u32(va0x4567c67));
129           }
130         }
131       }
132     }
133 
134     const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max);
135     vacc0x0123 = vminq_f32(vacc0x0123, vmax);
136     vacc0x4567 = vminq_f32(vacc0x4567, vmax);
137 
138     const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min);
139     vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
140     vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);
141 
142     bfloat16x4_t vout0x0123 = vcvt_bf16_f32(vacc0x0123);
143     bfloat16x4_t vout0x4567 = vcvt_bf16_f32(vacc0x4567);
144 
145     if XNN_LIKELY(nc >= 8) {
146       vst1_bf16(c0, vout0x0123);
147       vst1_bf16(c0 + 4, vout0x4567);
148       c0 = (bfloat16_t*) ((uintptr_t) c0 + cn_stride);
149 
150       a0 = (const bfloat16_t*) ((uintptr_t) a0 - kc);
151 
152       nc -= 8;
153     } else {
154       if (nc & 4) {
155         vst1_bf16(c0, vout0x0123); c0 += 4;
156 
157         vout0x0123 = vout0x4567;
158       }
159       if (nc & 2) {
160         vst1_lane_u32((void*) c0, vreinterpret_u32_bf16(vout0x0123), 0); c0 += 2;
161 
162         vout0x0123 = vreinterpret_bf16_u16(vext_u16(vreinterpret_u16_bf16(vout0x0123), vreinterpret_u16_bf16(vout0x0123), 2));
163       }
164       if (nc & 1) {
165         vst1_lane_bf16(c0, vout0x0123, 0);
166       }
167 
168       nc = 0;
169     }
170   } while (nc != 0);
171 }
172