← Back to C-Kernel-Engine Docs Doxygen Source Documentation
 
Loading...
Searching...
No Matches
ck_features.h
Go to the documentation of this file.
1/**
2 * @file ck_features.h
3 * @brief CPU feature detection and dispatch macros
4 *
5 * Defines standardized macros for SIMD instruction set detection
6 * and kernel dispatch. Use these instead of CPU model checks.
7 *
8 * Feature Priority (best available):
9 * AMX (512-bit tile ops, Intel Sapphire Rapids+)
10 * AVX-512 (512-bit vector, Intel Skylake-X+)
11 * AVX2 (256-bit with FMA, Intel Haswell+)
12 * AVX (256-bit, Intel Sandy Bridge+)
13 * NEON/SVE2 (ARM)
14 * DSA (PowerPC)
15 * Reference (fallback)
16 */
17
18#ifndef CK_FEATURES_H
19#define CK_FEATURES_H
20
21#include <stdint.h>
22
23/*============================================================================
24 * Compiler Feature Detection
25 * These are set by the compiler based on -march flags
26 *============================================================================*/
27
28/* Intel/x86 */
29#if defined(__AMX__)
30 #define CK_HAS_AMX 1
31#endif
32
33#if defined(__AVX512F__) && defined(__AVX512BW__) && defined(__AVX512DQ__)
34 #define CK_HAS_AVX512 1
35#endif
36
37#if defined(__AVX2__) && defined(__FMA__)
38 #define CK_HAS_AVX2_FMA 1
39#endif
40
41#if defined(__AVX__)
42 #define CK_HAS_AVX 1
43#endif
44
45#if defined(__VNNI__)
46 #define CK_HAS_VNNI 1
47#endif
48
49#if defined(__AVX512VNNI__)
50 #define CK_HAS_AVX512VNNI 1
51#endif
52
53/* ARM */
54#if defined(__aarch64__)
55 #if defined(__ARM_FEATURE_SVE2)
56 #define CK_HAS_SVE2 1
57 #endif
58 #if defined(__ARM_FEATURE_NEON)
59 #define CK_HAS_NEON 1
60 #endif
61#endif
62
63/* PowerPC */
64#if defined(__ALTIVEC__)
65 #define CK_HAS_ALTIVEC 1
66#endif
67
68#if defined(__VSX__)
69 #define CK_HAS_VSX 1
70#endif
71
72/* RISC-V */
73#if defined(__riscv_vector)
74 #define CK_HAS_RVV 1
75#endif
76
77/*============================================================================
78 * Runtime Feature Detection (CPUID/MSR)
79 * For compiled binaries that need runtime dispatch
80 *============================================================================*/
81
82#if defined(__x86_64__) || defined(__i386__)
83 #include <cpuid.h>
84
85 static inline uint32_t ck_cpuid_max_leaf(void) {
86 uint32_t eax, ecx = 0, ebx = 0, edx = 0;
87 __cpuid(0, eax, ebx, ecx, edx);
88 return eax;
89 }
90
91 static inline void ck_cpuid_leaf1(uint32_t *eax, uint32_t *ebx,
92 uint32_t *ecx, uint32_t *edx) {
93 __cpuid(1, *eax, *ebx, *ecx, *edx);
94 }
95
96 static inline void ck_cpuid_leaf7(uint32_t ecx_val, uint32_t *eax,
97 uint32_t *ebx, uint32_t *ecx,
98 uint32_t *edx) {
99 int info[4];
100 __cpuidex(info, 7, ecx_val);
101 *eax = info[0];
102 *ebx = info[1];
103 *ecx = info[2];
104 *edx = info[3];
105 }
106
107 /* Check if OS supports XSAVE/XRSTORE (needed for AVX-512/AMX state) */
108 static inline int ck_os_has_xtile(void) {
109 uint32_t eax, ebx, ecx, edx;
110 ck_cpuid_leaf1(&eax, &ebx, &ecx, &edx);
111 /* ECX bit 26 = XSAVE, ECX bit 27 = OSXSAVE */
112 return (ecx & (1 << 27)) && (ecx & (1 << 26));
113 }
114#endif
115
116/*============================================================================
117 * Feature Flag Macros
118 * Use these in dispatch functions instead of CPU model checks
119 *============================================================================*/
120
121/* Best available vector width */
122#if defined(CK_HAS_AMX)
123 #define CK_VECTOR_WIDTH 512
124 #define CK_HAS_BEST_VECTOR 1
125#elif defined(CK_HAS_AVX512)
126 #define CK_VECTOR_WIDTH 512
127 #define CK_HAS_BEST_VECTOR 1
128#elif defined(CK_HAS_AVX2_FMA)
129 #define CK_VECTOR_WIDTH 256
130 #define CK_HAS_BEST_VECTOR 1
131#elif defined(CK_HAS_AVX)
132 #define CK_VECTOR_WIDTH 256
133 #define CK_HAS_BEST_VECTOR 1
134#elif defined(CK_HAS_NEON)
135 #define CK_VECTOR_WIDTH 128
136 #define CK_HAS_BEST_VECTOR 1
137#else
138 #define CK_VECTOR_WIDTH 32 /* Scalar fallback */
139 #define CK_HAS_BEST_VECTOR 0
140#endif
141
142/* AI acceleration features */
143#if defined(CK_HAS_AMX)
144 #define CK_HAS_AI_ACCEL 1
145#elif defined(CK_HAS_AVX512VNNI)
146 #define CK_HAS_AI_ACCEL 1
147#elif defined(CK_HAS_VNNI)
148 #define CK_HAS_AI_ACCEL 1
149#else
150 #define CK_HAS_AI_ACCEL 0
151#endif
152
153/*============================================================================
154 * Kernel Dispatch Macros
155 * Use these for clean dispatch in kernel functions
156 *============================================================================*/
157
158/**
159 * @brief Dispatch to best available GEMM kernel
160 *
161 * Usage:
162 * CK_GEMM_DISPATCH(y, W, x, M, K);
163 * expands to appropriate kernel call
164 */
165#if defined(CK_HAS_AMX)
166 #define CK_GEMM_DISPATCH(...) gemm_amx(__VA_ARGS__)
167#elif defined(CK_HAS_AVX512)
168 #define CK_GEMM_DISPATCH(...) gemm_avx512(__VA_ARGS__)
169#elif defined(CK_HAS_AVX2_FMA)
170 #define CK_GEMM_DISPATCH(...) gemm_avx2(__VA_ARGS__)
171#elif defined(CK_HAS_AVX)
172 #define CK_GEMM_DISPATCH(...) gemm_avx(__VA_ARGS__)
173#else
174 #define CK_GEMM_DISPATCH(...) gemm_ref(__VA_ARGS__)
175#endif
176
177/**
178 * @brief Dispatch to best available GEMV kernel
179 */
180#if defined(CK_HAS_AMX)
181 #define CK_GEMV_DISPATCH(...) gemv_amx(__VA_ARGS__)
182#elif defined(CK_HAS_AVX512)
183 #define CK_GEMV_DISPATCH(...) gemv_avx512(__VA_ARGS__)
184#elif defined(CK_HAS_AVX2_FMA)
185 #define CK_GEMV_DISPATCH(...) gemv_avx2(__VA_ARGS__)
186#elif defined(CK_HAS_AVX)
187 #define CK_GEMV_DISPATCH(...) gemv_avx(__VA_ARGS__)
188#else
189 #define CK_GEMV_DISPATCH(...) gemv_ref(__VA_ARGS__)
190#endif
191
192/**
193 * @brief Dispatch to best available quantized GEMV kernel
194 * For INT8/INT4 quantization with VNNI/AMX acceleration
195 */
196#if defined(CK_HAS_AMX)
197 #define CK_QGEMM_DISPATCH(...) qgemm_amx(__VA_ARGS__)
198#elif defined(CK_HAS_AVX512VNNI)
199 #define CK_QGEMM_DISPATCH(...) qgemm_avx512vnni(__VA_ARGS__)
200#elif defined(CK_HAS_VNNI)
201 #define CK_QGEMM_DISPATCH(...) qgemm_vnni(__VA_ARGS__)
202#elif defined(CK_HAS_AVX2_FMA)
203 #define CK_QGEMM_DISPATCH(...) qgemm_avx2(__VA_ARGS__)
204#else
205 #define CK_QGEMM_DISPATCH(...) qgemm_ref(__VA_ARGS__)
206#endif
207
208/*============================================================================
209 * Capability Reporting
210 *============================================================================*/
211
212/**
213 * @brief CPU capability information structure
214 */
215typedef struct {
216 const char *name;
217 int width; /* Vector width in bits */
218 int has_fma; /* Fused multiply-add */
219 int has_ai_accel; /* AI-specific instructions (VNNI/AMX) */
220 const char *best_kernel; /* Recommended kernel name */
221} ck_capability_t;
222
223/**
224 * @brief Get current platform capabilities
225 */
226static inline ck_capability_t ck_get_capabilities(void) {
227 ck_capability_t cap = {
228 .name = "unknown",
229 .width = 32,
230 .has_fma = 0,
231 .has_ai_accel = 0,
232 .best_kernel = "gemm_ref"
233 };
234
235#if defined(CK_HAS_AMX)
236 cap.name = "AMX (Intel Sapphire Rapids+)";
237 cap.width = 512;
238 cap.has_fma = 1;
239 cap.has_ai_accel = 1;
240 cap.best_kernel = "gemm_amx";
241#elif defined(CK_HAS_AVX512)
242 cap.name = "AVX-512 (Intel Skylake-X+)";
243 cap.width = 512;
244 cap.has_fma = 1;
245 cap.has_ai_accel = 1;
246 cap.best_kernel = "gemm_avx512";
247#elif defined(CK_HAS_AVX2_FMA)
248 cap.name = "AVX2+FMA (Intel Haswell+)";
249 cap.width = 256;
250 cap.has_fma = 1;
251 cap.has_ai_accel = 0;
252 cap.best_kernel = "gemm_avx2";
253#elif defined(CK_HAS_AVX)
254 cap.name = "AVX (Intel Sandy Bridge+)";
255 cap.width = 256;
256 cap.has_fma = 0;
257 cap.has_ai_accel = 0;
258 cap.best_kernel = "gemm_avx";
259#elif defined(CK_HAS_NEON)
260 cap.name = "NEON (ARM)";
261 cap.width = 128;
262 cap.has_fma = 1;
263 cap.has_ai_accel = 0;
264 cap.best_kernel = "gemm_neon";
265#elif defined(CK_HAS_ALTIVEC)
266 cap.name = "AltiVec (PowerPC)";
267 cap.width = 128;
268 cap.has_fma = 1;
269 cap.has_ai_accel = 0;
270 cap.best_kernel = "gemm_altivec";
271#endif
272
273 return cap;
274}
275
276#endif /* CK_FEATURES_H */
ck_get_capabilities
static ck_capability_t ck_get_capabilities(void)
Get current platform capabilities.
Definition ck_features.h:226
ck_capability_t
CPU capability information structure.
Definition ck_features.h:215
ck_capability_t::best_kernel
const char * best_kernel
Definition ck_features.h:220
ck_capability_t::has_fma
int has_fma
Definition ck_features.h:218
ck_capability_t::width
int width
Definition ck_features.h:217
ck_capability_t::has_ai_accel
int has_ai_accel
Definition ck_features.h:219
ck_capability_t::name
const char * name
Definition ck_features.h:216