KleidiAI Coverage Report

Directory: ./
Coverage: low: ≥ 0% medium: ≥ 75.0% high: ≥ 90.0%
Coverage Exec / Excl / Total
Lines: 80.8% 21 / 15 / 41
Functions: 60.0% 3 / 0 / 5
Branches: 100.0% 6 / 30 / 36
kai/ukernels/matmul/pack/kai_lhs_quant_pack_bf16p8x4_f32_neon.c
Line Branch Exec Source
1 //
2 // SPDX-FileCopyrightText: Copyright 2024-2025 Arm Limited and/or its affiliates <open-source-office@arm.com>
3 //
4 // SPDX-License-Identifier: Apache-2.0
5 //
6
7 #if !defined(__aarch64__) || !defined(__ARM_FEATURE_BF16_VECTOR_ARITHMETIC)
8 #error This file must be compiled for AArch64, FEAT_BF16.
9 #else // Architectural features check.
10
11 #define MAX_MR 8
12
13 #include "kai_lhs_quant_pack_bf16p8x4_f32_neon.h"
14
15 #include <arm_neon.h>
16 #include <stddef.h>
17 #include <stdint.h>
18
19 #include "kai/kai_common.h"
20
21 static const size_t kai_mr = 8;
22 static const size_t kai_kr = 4;
23 static const size_t kai_sr = 1;
24
25 ✗ size_t kai_get_m_step_lhs_quant_pack_bf16p8x4_f32_neon(size_t mr) {
26 − KAI_ASSUME(mr == kai_mr);
27 ✗ return kai_mr;
28 }
29
30 600 size_t kai_get_lhs_offset_lhs_quant_pack_bf16p8x4_f32_neon(size_t m_idx, size_t lhs_stride) {
31 600 return m_idx * lhs_stride;
32 }
33
34 ✗ size_t kai_get_lhs_packed_offset_lhs_quant_pack_bf16p8x4_f32_neon(
35 size_t m_idx, size_t k, size_t mr, size_t kr, size_t sr) {
36 ✗ KAI_UNUSED(sr);
37 − KAI_ASSUME(mr == kai_mr);
38 − KAI_ASSUME(kr == kai_kr);
39 − KAI_ASSUME(sr == kai_sr);
40 − KAI_ASSUME(m_idx % mr == 0);
41
42 ✗ return m_idx * kai_roundup(k, kr) * sizeof(uint16_t);
43 }
44
45 600 size_t kai_get_lhs_packed_size_lhs_quant_pack_bf16p8x4_f32_neon(size_t m, size_t k, size_t mr, size_t kr, size_t sr) {
46 600 KAI_UNUSED(sr);
47 − KAI_ASSUME(mr == kai_mr);
48 − KAI_ASSUME(kr == kai_kr);
49 − KAI_ASSUME(sr == kai_sr);
50
51 600 return kai_roundup(m, mr) * kai_roundup(k, kr) * sizeof(uint16_t);
52 }
53
54 600 void kai_run_lhs_quant_pack_bf16p8x4_f32_neon(
55 size_t m, size_t k, size_t mr, size_t kr, size_t sr, size_t m_idx_start, const void* lhs, size_t lhs_stride,
56 void* lhs_packed) {
57 − KAI_ASSUME(mr == kai_mr);
58 − KAI_ASSUME(sr == kai_sr);
59 − KAI_ASSUME(kr == kai_kr);
60 600 KAI_UNUSED(sr);
61 − KAI_ASSUME(lhs != NULL);
62 − KAI_ASSUME(lhs_packed != NULL);
63
64 − KAI_ASSUME(m_idx_start == 0);
65 − KAI_ASSUME(mr <= MAX_MR);
66
67 600 const size_t block_height = mr;
68 600 const size_t row_offset = 0;
69
70 600 const void* in[MAX_MR];
71
72
2/2
✓ Branch 0 taken 600 times.
✓ Branch 1 taken 6228 times.
 6828 for (size_t block_y = 0; block_y < m; block_y += block_height) {
73
2/2
✓ Branch 0 taken 420 times.
✓ Branch 1 taken 5808 times.
 6228 const size_t height = KAI_MIN(m - block_y, block_height);
74 6228 void* out = (char*)lhs_packed + block_y * kai_roundup(k, kr) * sizeof(uint16_t);
75 6228 size_t width = k;
76
77
2/2
✓ Branch 0 taken 47604 times.
✓ Branch 1 taken 6228 times.
 53832 for (size_t y = 0; y < height; y++) {
78 47604 in[y] = (const char*)lhs + (block_y + y) * lhs_stride;
79 47604 }
80
81 12456 __asm__ __volatile__(
82 "ldr x28, [%x[in], #0x0]\n"
83 "ldr x27, [%x[in], #0x8]\n"
84 "cmp %x[height], #0x8\n"
85 "ldr x26, [%x[in], #0x10]\n"
86 "ldr x25, [%x[in], #0x18]\n"
87 "ldr x24, [%x[in], #0x20]\n"
88 "ldr x23, [%x[in], #0x28]\n"
89 "ldr x22, [%x[in], #0x30]\n"
90 "ldr x21, [%x[in], #0x38]\n"
91 "add x28, x28, %x[row_offset], LSL #2\n"
92 "add x27, x27, %x[row_offset], LSL #2\n"
93 "add x26, x26, %x[row_offset], LSL #2\n"
94 "add x25, x25, %x[row_offset], LSL #2\n"
95 "add x24, x24, %x[row_offset], LSL #2\n"
96 "add x23, x23, %x[row_offset], LSL #2\n"
97 "add x22, x22, %x[row_offset], LSL #2\n"
98 "add x21, x21, %x[row_offset], LSL #2\n"
99 "beq 1f\n"
100 "cmp %x[height], #0x2\n"
101 "mov x21, x28\n"
102 "csel x27, x27, x28, GE\n"
103 "csel x26, x26, x28, GT\n"
104 "cmp %x[height], #0x4\n"
105 "csel x25, x25, x28, GE\n"
106 "csel x24, x24, x28, GT\n"
107 "cmp %x[height], #0x6\n"
108 "csel x23, x23, x28, GE\n"
109 "csel x22, x22, x28, GT\n"
110 "1:" // no_pointer_adj
111 "cmp %x[width], #0x4\n"
112 "prfm pldl1keep, [x28, #0x0]\n"
113 "prfm pldl1keep, [x27, #0x0]\n"
114 "prfm pldl1keep, [x26, #0x0]\n"
115 "prfm pldl1keep, [x25, #0x0]\n"
116 "prfm pldl1keep, [x24, #0x0]\n"
117 "prfm pldl1keep, [x23, #0x0]\n"
118 "prfm pldl1keep, [x22, #0x0]\n"
119 "prfm pldl1keep, [x21, #0x0]\n"
120 "prfm pldl1keep, [x28, #0x40]\n"
121 "prfm pldl1keep, [x27, #0x40]\n"
122 "prfm pldl1keep, [x26, #0x40]\n"
123 "prfm pldl1keep, [x25, #0x40]\n"
124 "prfm pldl1keep, [x24, #0x40]\n"
125 "prfm pldl1keep, [x23, #0x40]\n"
126 "prfm pldl1keep, [x22, #0x40]\n"
127 "prfm pldl1keep, [x21, #0x40]\n"
128 "blt 3f\n"
129 "2:" // Main loop head
130 "ldr q19, [x28], #0x10\n"
131 "ldr q18, [x26], #0x10\n"
132 "subs %x[width], %x[width], #0x4\n"
133 "ldr q17, [x24], #0x10\n"
134 "ldr q16, [x22], #0x10\n"
135 "cmp %x[width], #0x4\n"
136 "ldr q23, [x27], #0x10\n"
137 "ldr q22, [x25], #0x10\n"
138 "ldr q21, [x23], #0x10\n"
139 "ldr q20, [x21], #0x10\n"
140 ".inst 0x0ea16a73 // bfcvtn v19.4h, v19.4s\n"
141 ".inst 0x0ea16a52 // bfcvtn v18.4h, v18.4s\n"
142 ".inst 0x0ea16a31 // bfcvtn v17.4h, v17.4s\n"
143 ".inst 0x0ea16a10 // bfcvtn v16.4h, v16.4s\n"
144 "prfm pldl1keep, [x28, #0x70]\n"
145 "prfm pldl1keep, [x27, #0x70]\n"
146 "prfm pldl1keep, [x26, #0x70]\n"
147 "prfm pldl1keep, [x25, #0x70]\n"
148 "prfm pldl1keep, [x24, #0x70]\n"
149 "prfm pldl1keep, [x23, #0x70]\n"
150 ".inst 0x4ea16af3 // bfcvtn2 v19.8h, v23.4s\n"
151 ".inst 0x4ea16ad2 // bfcvtn2 v18.8h, v22.4s\n"
152 "prfm pldl1keep, [x22, #0x70]\n"
153 "prfm pldl1keep, [x21, #0x70]\n"
154 ".inst 0x4ea16ab1 // bfcvtn2 v17.8h, v21.4s\n"
155 ".inst 0x4ea16a90 // bfcvtn2 v16.8h, v20.4s\n"
156 "str q19, [%x[out_ptr], #0x0]\n"
157 "str q18, [%x[out_ptr], #0x10]\n"
158 "str q17, [%x[out_ptr], #0x20]\n"
159 "str q16, [%x[out_ptr], #0x30]\n"
160 "add %x[out_ptr], %x[out_ptr], #0x40\n"
161 "bge 2b\n"
162 "3:" // Main loop skip
163 "cbz %x[width], 6f\n"
164 "tbz %x[width], #1, 4f\n"
165 "ldr d19, [x28], #0x8\n"
166 "ldr d23, [x27], #0x8\n"
167 "mov x20, #0x1\n"
168 "ldr d18, [x26], #0x8\n"
169 "ldr d22, [x25], #0x8\n"
170 "ldr d17, [x24], #0x8\n"
171 "ldr d21, [x23], #0x8\n"
172 "ldr d16, [x22], #0x8\n"
173 "ldr d20, [x21], #0x8\n"
174 "tbz %x[width], #0, 5f\n"
175 "ld1 { v19.s }[2], [x28]\n"
176 "ld1 { v23.s }[2], [x27]\n"
177 "ld1 { v18.s }[2], [x26]\n"
178 "ld1 { v22.s }[2], [x25]\n"
179 "ld1 { v17.s }[2], [x24]\n"
180 "ld1 { v21.s }[2], [x23]\n"
181 "ld1 { v16.s }[2], [x22]\n"
182 "ld1 { v20.s }[2], [x21]\n"
183 "b 5f\n"
184 "4:" // odd_loads_1_0
185 "ldr s19, [x28, #0x0]\n"
186 "ldr s23, [x27, #0x0]\n"
187 "mov x20, #0x1\n"
188 "ldr s18, [x26, #0x0]\n"
189 "ldr s22, [x25, #0x0]\n"
190 "ldr s17, [x24, #0x0]\n"
191 "ldr s21, [x23, #0x0]\n"
192 "ldr s16, [x22, #0x0]\n"
193 "ldr s20, [x21, #0x0]\n"
194 "5:" // Odd load end
195 ".inst 0x0ea16a73 // bfcvtn v19.4h, v19.4s\n"
196 ".inst 0x0ea16a52 // bfcvtn v18.4h, v18.4s\n"
197 ".inst 0x0ea16a31 // bfcvtn v17.4h, v17.4s\n"
198 ".inst 0x0ea16a10 // bfcvtn v16.4h, v16.4s\n"
199 ".inst 0x4ea16af3 // bfcvtn2 v19.8h, v23.4s\n"
200 ".inst 0x4ea16ad2 // bfcvtn2 v18.8h, v22.4s\n"
201 ".inst 0x4ea16ab1 // bfcvtn2 v17.8h, v21.4s\n"
202 ".inst 0x4ea16a90 // bfcvtn2 v16.8h, v20.4s\n"
203 "str q19, [%x[out_ptr], #0x0]\n"
204 "str q18, [%x[out_ptr], #0x10]\n"
205 "str q17, [%x[out_ptr], #0x20]\n"
206 "str q16, [%x[out_ptr], #0x30]\n"
207 "add %x[out_ptr], %x[out_ptr], #0x40\n"
208 "6:" // Odds skip
209 : [out_ptr] "+&r"(out), [width] "+&r"(width)
210 6228 : [height] "r"(height), [in] "r"(in), [row_offset] "r"(row_offset)
211 : "cc", "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "x20", "x21", "x22", "x23", "x24",
212 "x25", "x26", "x27", "x28");
213 6228 }
214 600 }
215
216 #endif // Architectural features check.
217