| 1 | // Adapted from |
| 2 | // https://github.com/sgl-project/sglang/tree/main/sgl-kernel/csrc/cpu |
| 3 | // |
| 4 | // Synced from |
| 5 | // https://github.com/sgl-project/sglang/tree/7c248dde7fe1f3b5100966f8143f97a9932c22a4/sgl-kernel/csrc/cpu |
| 6 | // Sync date: 2026-07-29 |
| 7 | |
| 8 | // clang-format off |
| 9 | |
| 10 | #pragma once |
| 11 | |
| 12 | #include <ATen/ATen.h> |
| 13 | #include <ATen/Dispatch.h> |
| 14 | #include <ATen/Parallel.h> |
| 15 | #include <cstring> |
| 16 | |
| 17 | #if defined(_OPENMP) |
| 18 | #include <omp.h> |
| 19 | #endif |
| 20 | |
| 21 | // Adapted from sglang: dispatch utilities for FP8 W8A8 kernels |
| 22 | #define AT_DISPATCH_OUT_TYPES(TYPE, NAME, ...) \ |
| 23 | AT_DISPATCH_SWITCH( \ |
| 24 | TYPE, \ |
| 25 | NAME, \ |
| 26 | AT_PRIVATE_CASE_TYPE_USING_HINT(at::ScalarType::Float, out_t, __VA_ARGS__) \ |
| 27 | AT_PRIVATE_CASE_TYPE_USING_HINT(at::ScalarType::BFloat16, out_t, __VA_ARGS__) \ |
| 28 | AT_PRIVATE_CASE_TYPE_USING_HINT(at::ScalarType::Half, out_t, __VA_ARGS__)) |
| 29 | |
| 30 | #define AT_DISPATCH_CASE_ENUM_NO_RETURN(VALUE, TYPE, HINT, ...) \ |
| 31 | case VALUE: { \ |
| 32 | constexpr TYPE HINT = VALUE; \ |
| 33 | __VA_ARGS__; \ |
| 34 | break; \ |
| 35 | } |
| 36 | |
| 37 | #define AT_DISPATCH_BOOL_NO_RETURN(VALUE, NAME, HINT, ...) \ |
| 38 | [&]() { \ |
| 39 | switch (VALUE) { \ |
| 40 | AT_DISPATCH_CASE_ENUM_NO_RETURN(true, bool, HINT, __VA_ARGS__) \ |
| 41 | AT_DISPATCH_CASE_ENUM_NO_RETURN(false, bool, HINT, __VA_ARGS__) \ |
| 42 | } \ |
| 43 | }() |
| 44 | |
| 45 | #define AT_DISPATCH_QUANT_MODE_NO_RETURN(MODE, NAME, HINT, ...) \ |
| 46 | [&]() { \ |
| 47 | switch (MODE) { \ |
| 48 | AT_DISPATCH_CASE_ENUM_NO_RETURN(PER_TENSOR, int, HINT, __VA_ARGS__) \ |
| 49 | AT_DISPATCH_CASE_ENUM_NO_RETURN(PER_ROW, int, HINT, __VA_ARGS__) \ |
| 50 | AT_DISPATCH_CASE_ENUM_NO_RETURN(PER_GROUP, int, HINT, __VA_ARGS__) \ |
| 51 | } \ |
| 52 | }() |
| 53 | |
| 54 | namespace { |
| 55 | |
| 56 | // dispatch bool |
| 57 | #define AT_DISPATCH_BOOL(BOOL_V, BOOL_NAME, ...) \ |
| 58 | [&] { \ |
| 59 | if (BOOL_V) { \ |
| 60 | constexpr bool BOOL_NAME = true; \ |
| 61 | return __VA_ARGS__(); \ |
| 62 | } else { \ |
| 63 | constexpr bool BOOL_NAME = false; \ |
| 64 | return __VA_ARGS__(); \ |
| 65 | } \ |
| 66 | }() |
| 67 | |
| 68 | #define AT_DISPATCH_BOOL2(BOOL_V1, BOOL_NAME1, BOOL_V2, BOOL_NAME2, ...) \ |
| 69 | [&] { \ |
| 70 | if (BOOL_V1) { \ |
| 71 | constexpr bool BOOL_NAME1 = true; \ |
| 72 | if (BOOL_V2) { \ |
| 73 | constexpr bool BOOL_NAME2 = true; \ |
| 74 | return __VA_ARGS__(); \ |
| 75 | } else { \ |
| 76 | constexpr bool BOOL_NAME2 = false; \ |
| 77 | return __VA_ARGS__(); \ |
| 78 | } \ |
| 79 | } else { \ |
| 80 | constexpr bool BOOL_NAME1 = false; \ |
| 81 | if (BOOL_V2) { \ |
| 82 | constexpr bool BOOL_NAME2 = true; \ |
| 83 | return __VA_ARGS__(); \ |
| 84 | } else { \ |
| 85 | constexpr bool BOOL_NAME2 = false; \ |
| 86 | return __VA_ARGS__(); \ |
| 87 | } \ |
| 88 | } \ |
| 89 | }() |
| 90 | |
| 91 | // Half + BFloat16, plus one extra scalar type |
| 92 | #define AT_DISPATCH_CASE_REDUCED_FLOATING_TYPES_AND(SCALARTYPE, ...) \ |
| 93 | AT_DISPATCH_CASE_REDUCED_FLOATING_TYPES(__VA_ARGS__) \ |
| 94 | AT_DISPATCH_CASE(SCALARTYPE, __VA_ARGS__) |
| 95 | |
| 96 | #define AT_DISPATCH_REDUCED_FLOATING_TYPES_AND(SCALARTYPE, TYPE, NAME, ...) \ |
| 97 | AT_DISPATCH_SWITCH(TYPE, NAME, AT_DISPATCH_CASE_REDUCED_FLOATING_TYPES_AND(SCALARTYPE, __VA_ARGS__)) |
| 98 | |
| 99 | // dispatch: bfloat16, float16, int8_t, fp8_e4m3, uint8_t(mxfp4/int4) |
| 100 | #define CPU_DISPATCH_PACKED_TYPES(TYPE, ...) \ |
| 101 | [&] { \ |
| 102 | switch (TYPE) { \ |
| 103 | case at::ScalarType::BFloat16: { \ |
| 104 | using packed_t = at::BFloat16; \ |
| 105 | return __VA_ARGS__(); \ |
| 106 | } \ |
| 107 | case at::ScalarType::Half: { \ |
| 108 | using packed_t = at::Half; \ |
| 109 | return __VA_ARGS__(); \ |
| 110 | } \ |
| 111 | case at::ScalarType::Char: { \ |
| 112 | using packed_t = int8_t; \ |
| 113 | return __VA_ARGS__(); \ |
| 114 | } \ |
| 115 | case at::ScalarType::Float8_e4m3fn: { \ |
| 116 | using packed_t = at::Float8_e4m3fn; \ |
| 117 | return __VA_ARGS__(); \ |
| 118 | } \ |
| 119 | case at::ScalarType::Byte: { \ |
| 120 | using packed_t = uint8_t; \ |
| 121 | return __VA_ARGS__(); \ |
| 122 | } \ |
| 123 | default: \ |
| 124 | TORCH_CHECK(false, "Unsupported floating data type.\n"); \ |
| 125 | } \ |
| 126 | }() |
| 127 | |
| 128 | // Helper MICRO for CPU_DISPATCH_FLOATING_TYPES_EXT: |
| 129 | // TYPE1: the primary dtype (input, output, weight); |
| 130 | // TYPE2: defined as PARAM_T input |
| 131 | #define CPU_DISPATCH_TYPE1_WITH_PARAM(TYPE1, PARAM_T, ...) \ |
| 132 | switch (TYPE1) { \ |
| 133 | case at::ScalarType::BFloat16: { \ |
| 134 | using scalar_t = at::BFloat16; \ |
| 135 | using param_t = PARAM_T; \ |
| 136 | return __VA_ARGS__(); \ |
| 137 | } \ |
| 138 | case at::ScalarType::Half: { \ |
| 139 | using scalar_t = at::Half; \ |
| 140 | using param_t = PARAM_T; \ |
| 141 | return __VA_ARGS__(); \ |
| 142 | } \ |
| 143 | case at::ScalarType::Float: { \ |
| 144 | using scalar_t = float; \ |
| 145 | using param_t = PARAM_T; \ |
| 146 | return __VA_ARGS__(); \ |
| 147 | } \ |
| 148 | default: \ |
| 149 | TORCH_CHECK(false, "Unsupported floating data type."); \ |
| 150 | } |
| 151 | |
| 152 | // Helper MICRO for CPU_DISPATCH_REDUCED_FLOATING_TYPES_EXT: |
| 153 | // TYPE1: the primary dtype (input, output, weight); |
| 154 | // TYPE2: defined as PARAM_T input |
| 155 | #define CPU_DISPATCH_TYPE1_WITH_PARAM_REDUCED(TYPE1, PARAM_T, ...) \ |
| 156 | switch (TYPE1) { \ |
| 157 | case at::ScalarType::BFloat16: { \ |
| 158 | using scalar_t = at::BFloat16; \ |
| 159 | using param_t = PARAM_T; \ |
| 160 | return __VA_ARGS__(); \ |
| 161 | } \ |
| 162 | case at::ScalarType::Half: { \ |
| 163 | using scalar_t = at::Half; \ |
| 164 | using param_t = PARAM_T; \ |
| 165 | return __VA_ARGS__(); \ |
| 166 | } \ |
| 167 | default: \ |
| 168 | TORCH_CHECK(false, "Unsupported floating data type."); \ |
| 169 | } |
| 170 | |
| 171 | // Helper MICRO for CPU_DISPATCH_REDUCED_FLOATING_TYPES_EXT: |
| 172 | // TYPE1: the dtype both for scalar_t and param_t |
| 173 | #define CPU_DISPATCH_TYPE1_WITH_SAME_PARAM_REDUCED(TYPE1, ...) \ |
| 174 | switch (TYPE1) { \ |
| 175 | case at::ScalarType::BFloat16: { \ |
| 176 | using scalar_t = at::BFloat16; \ |
| 177 | using param_t = at::BFloat16; \ |
| 178 | return __VA_ARGS__(); \ |
| 179 | } \ |
| 180 | case at::ScalarType::Half: { \ |
| 181 | using scalar_t = at::Half; \ |
| 182 | using param_t = at::Half; \ |
| 183 | return __VA_ARGS__(); \ |
| 184 | } \ |
| 185 | default: \ |
| 186 | TORCH_CHECK(false, "Unsupported reduced floating data type."); \ |
| 187 | } |
| 188 | |
| 189 | // dispatch with mixed dtypes (TYPE1, TYPE2): |
| 190 | // TYPE1: the primary dtype (input, output, weight); |
| 191 | // TYPE2: the secondary dtype (bias, etc.). |
| 192 | #define CPU_DISPATCH_FLOATING_TYPES_EXT(TYPE1, TYPE2, ...) \ |
| 193 | [&] { \ |
| 194 | if (TYPE2 == at::kFloat) { \ |
| 195 | CPU_DISPATCH_TYPE1_WITH_PARAM(TYPE1, float, __VA_ARGS__) \ |
| 196 | } else if (TYPE2 == at::ScalarType::BFloat16) { \ |
| 197 | CPU_DISPATCH_TYPE1_WITH_PARAM(TYPE1, at::BFloat16, __VA_ARGS__) \ |
| 198 | } else if (TYPE2 == at::ScalarType::Half) { \ |
| 199 | CPU_DISPATCH_TYPE1_WITH_PARAM(TYPE1, at::Half, __VA_ARGS__) \ |
| 200 | } else { \ |
| 201 | TORCH_CHECK(false, "Unsupported floating data type."); \ |
| 202 | } \ |
| 203 | }() |
| 204 | |
| 205 | // dispatch with mixed dtypes (reduced one, no float for TYPE1) (TYPE1, TYPE2): |
| 206 | // TYPE1: the primary dtype (input, output, weight); |
| 207 | // TYPE2: the secondary dtype (bias, etc.). |
| 208 | #define CPU_DISPATCH_REDUCED_FLOATING_TYPES_EXT(TYPE1, TYPE2, ...) \ |
| 209 | [&] { \ |
| 210 | if (TYPE2 == at::kFloat) { \ |
| 211 | CPU_DISPATCH_TYPE1_WITH_PARAM_REDUCED(TYPE1, float, __VA_ARGS__) \ |
| 212 | } else { \ |
| 213 | TORCH_CHECK(TYPE1 == TYPE2); \ |
| 214 | CPU_DISPATCH_TYPE1_WITH_SAME_PARAM_REDUCED(TYPE1, __VA_ARGS__) \ |
| 215 | } \ |
| 216 | }() |
| 217 | |
| 218 | #define UNUSED(x) (void)(x) |
| 219 | |
| 220 | #define CHECK_CPU(x) TORCH_CHECK(x.device().type() == at::kCPU, #x " must be a CPU tensor") |
| 221 | |
| 222 | #define CHECK_CONTIGUOUS(x) TORCH_CHECK(x.is_contiguous(), #x " must be contiguous") |
| 223 | #define CHECK_LAST_DIM_CONTIGUOUS(x) \ |
| 224 | TORCH_CHECK(x.strides()[x.strides().size() - 1] == 1, #x "must be contiguous at last dimension") |
| 225 | |
| 226 | #define CHECK_INPUT(x) \ |
| 227 | CHECK_CPU(x); \ |
| 228 | CHECK_CONTIGUOUS(x) |
| 229 | #define CHECK_LAST_DIM_CONTIGUOUS_INPUT(x) \ |
| 230 | CHECK_CPU(x); \ |
| 231 | CHECK_LAST_DIM_CONTIGUOUS(x) |
| 232 | |
| 233 | #define CHECK_DIM(d, x) TORCH_CHECK(x.dim() == d, #x " must be a " #d "D tensor") |
| 234 | |
| 235 | #define CHECK_EQ(a, b) TORCH_CHECK((a) == (b), "CHECK_EQ(" #a ", " #b ") failed. ", a, " vs ", b) |
| 236 | #define CHECK_GT(a, b) TORCH_CHECK((a) > (b), "CHECK_GT(" #a ", " #b ") failed. ", a, " vs ", b) |
| 237 | #define CHECK_GE(a, b) TORCH_CHECK((a) >= (b), "CHECK_GE(" #a ", " #b ") failed. ", a, " vs ", b) |
| 238 | |
| 239 | template <bool is_only_lastdim_contiguous> |
| 240 | static inline void CHECK_INPUT_SHAPE_DTYPE(const at::Tensor& tensor, const at::IntArrayRef sizes, at::ScalarType st) { |
| 241 | TORCH_CHECK(tensor.sizes() == sizes, "Input tensor shape mismatch: expected ", sizes, ", got ", tensor.sizes()); |
| 242 | TORCH_CHECK(tensor.scalar_type() == st, "Input tensor dtype mismatch"); |
| 243 | if constexpr (is_only_lastdim_contiguous) { |
| 244 | CHECK_LAST_DIM_CONTIGUOUS_INPUT(tensor); |
| 245 | } else { |
| 246 | CHECK_INPUT(tensor); |
| 247 | } |
| 248 | } |
| 249 | |
| 250 | // [NB] Parallel Routines |
| 251 | // |
| 252 | // * at::parallel_for - applies for most of generic use cases, this will be compiled |
| 253 | // against openmp in default torch release. |
| 254 | // |
| 255 | // * parallel_for - same function as above, can choose payload partition scheme in |
| 256 | // balance211. |
| 257 | // |
| 258 | // * parallel_2d - parallel for 2 dimensions, used in GEMM, etc. |
| 259 | // this one will do payload balance across 2 dimensions. |
| 260 | // |
| 261 | |
| 262 | // grain size for each thread |
| 263 | constexpr int GRAIN_SIZE = 1024; |
| 264 | |
| 265 | template <typename T, typename std::enable_if<std::is_integral<T>::value, int>::type = 0> |
| 266 | inline T div_up(T x, T y) { |
| 267 | return (x + y - 1) / y; |
| 268 | } |
| 269 | |
| 270 | // you can only use at::get_thread_num() with at::parallel_for() |
| 271 | // as it is lazy initialized, otherwise it will always return 0. |
| 272 | inline int get_thread_num() { |
| 273 | #if defined(_OPENMP) |
| 274 | return omp_get_thread_num(); |
| 275 | #else |
| 276 | return 0; |
| 277 | #endif |
| 278 | } |
| 279 | |
| 280 | // balance payload across each thread |
| 281 | template <typename T> |
| 282 | inline void balance211(T n, T nth, T ith, T& n_start, T& n_end) { |
| 283 | // pytorch aten partition pattern |
| 284 | T n_my = div_up(n, nth); |
| 285 | n_start = ith * n_my; |
| 286 | n_end = std::min(n_start + n_my, n); |
| 287 | } |
| 288 | |
| 289 | template <typename func_t> |
| 290 | inline void parallel_for(int n, const func_t& f) { |
| 291 | #if defined(_OPENMP) |
| 292 | #pragma omp parallel |
| 293 | { |
| 294 | int nth = omp_get_num_threads(); |
| 295 | int ith = omp_get_thread_num(); |
| 296 | int tbegin, tend; |
| 297 | balance211(n, nth, ith, tbegin, tend); |
| 298 | f(tbegin, tend); |
| 299 | } |
| 300 | #else |
| 301 | f(0, n); |
| 302 | #endif |
| 303 | } |
| 304 | |
| 305 | // for 1d parallel, use `actual_nth` |
| 306 | // for 2d parallel, use even nths, e.g. 43->42 |
| 307 | int inline adjust_num_threads(int m) { |
| 308 | int actual_nth = at::get_num_threads(); |
| 309 | if (m == 1) { |
| 310 | return actual_nth; |
| 311 | } |
| 312 | return std::max(1, (actual_nth >> 1) * 2); |
| 313 | } |
| 314 | |
| 315 | template <typename func_t> |
| 316 | inline void parallel_2d(int m, int n, const func_t& f) { |
| 317 | // make sure we have even num_threads |
| 318 | int nth = adjust_num_threads(m); |
| 319 | |
| 320 | // [NOTE] thread blocking: |
| 321 | // |
| 322 | // 1) prefer square block per thread |
| 323 | // 2) use even number of CPU cores |
| 324 | // 3) use all `num_threads` cores |
| 325 | // |
| 326 | // we have: |
| 327 | // TM * TN = T |
| 328 | // BM / TM = BN / TN |
| 329 | // then: |
| 330 | // TM = ((BM / BN) * T) ^ 0.5 |
| 331 | // |
| 332 | float r = float(m) / n; |
| 333 | int nth_m = std::ceil(std::sqrt(r * nth)); |
| 334 | int nth_n = 1; |
| 335 | for (; nth_m > 0; --nth_m) { |
| 336 | nth_n = nth / nth_m; |
| 337 | if (nth_m * nth_n == nth) { |
| 338 | break; |
| 339 | } |
| 340 | } |
| 341 | |
| 342 | #if defined(_OPENMP) |
| 343 | #pragma omp parallel num_threads(nth) |
| 344 | { |
| 345 | int ith = omp_get_thread_num(); |
| 346 | int ith_m = ith / nth_n; |
| 347 | int ith_n = ith % nth_n; |
| 348 | |
| 349 | int thread_block_m = div_up(m, nth_m); |
| 350 | int thread_block_n = div_up(n, nth_n); |
| 351 | |
| 352 | int begin_m = ith_m * thread_block_m; |
| 353 | int end_m = std::min(m, begin_m + thread_block_m); |
| 354 | int begin_n = ith_n * thread_block_n; |
| 355 | int end_n = std::min(n, begin_n + thread_block_n); |
| 356 | |
| 357 | f(begin_m, end_m, begin_n, end_n); |
| 358 | } |
| 359 | #else |
| 360 | f(0, m, 0, n); |
| 361 | #endif |
| 362 | } |
| 363 | |
| 364 | // Like parallel_2d but with explicit nth_m x nth_n decomposition. |
| 365 | // Caller is responsible for choosing nth_m and nth_n. |
| 366 | // |
| 367 | // Cherry-picked from a newer sglang commit than this file's last full sync |
| 368 | // (needed by mhc.cpp's phase-1 tiling); not part of a full common.h resync. |
| 369 | template <typename func_t> |
| 370 | inline void parallel_2d_tiled(int m, int n, int nth_m, int nth_n, const func_t& f) { |
| 371 | const int nth = nth_m * nth_n; |
| 372 | #if defined(_OPENMP) |
| 373 | #pragma omp parallel num_threads(nth) |
| 374 | { |
| 375 | int ith = omp_get_thread_num(); |
| 376 | int ith_m = ith / nth_n; |
| 377 | int ith_n = ith % nth_n; |
| 378 | |
| 379 | int thread_block_m = div_up(m, nth_m); |
| 380 | int thread_block_n = div_up(n, nth_n); |
| 381 | |
| 382 | int begin_m = std::min(ith_m * thread_block_m, m); |
| 383 | int end_m = std::min(begin_m + thread_block_m, m); |
| 384 | int begin_n = std::min(ith_n * thread_block_n, n); |
| 385 | int end_n = std::min(begin_n + thread_block_n, n); |
| 386 | |
| 387 | f(begin_m, end_m, begin_n, end_n); |
| 388 | } |
| 389 | #else |
| 390 | f(0, m, 0, n); |
| 391 | #endif |
| 392 | } |
| 393 | |
| 394 | // limit max cache blocks |
| 395 | // when we need to do pre-unpack for weights, e.g. fp8 |
| 396 | #define MAX_CACHE_BLOCK_SIZE 4 |
| 397 | |
| 398 | template <typename T> |
| 399 | inline int get_cache_blocks(int chunk_size) { |
| 400 | // L2 2MB and ratio of 50% |
| 401 | const int L2_size = 2048 * 1024 >> 1; |
| 402 | return std::max(1, int(L2_size / (chunk_size * sizeof(T)))); |
| 403 | } |
| 404 | |
| 405 | template <> |
| 406 | inline int get_cache_blocks<at::Float8_e4m3fn>(int chunk_size) { |
| 407 | // fp8 uses bf16 as accumulate type |
| 408 | int cache_block_size = get_cache_blocks<at::BFloat16>(chunk_size); |
| 409 | return std::min(MAX_CACHE_BLOCK_SIZE, cache_block_size); |
| 410 | } |
| 411 | |
| 412 | template <> |
| 413 | inline int get_cache_blocks<uint8_t>(int chunk_size) { |
| 414 | // mxfp4 uses bf16 as accumulate type |
| 415 | int cache_block_size = get_cache_blocks<at::BFloat16>(chunk_size); |
| 416 | return std::min(MAX_CACHE_BLOCK_SIZE, cache_block_size); |
| 417 | } |
| 418 | |
| 419 | // 2d sequential loop in range : [mb0, mb1), [nb0, nb1) |
| 420 | template <typename T, typename func_t> |
| 421 | inline void loop_2d(int64_t mb0, int64_t mb1, int64_t nb0, int64_t nb1, int64_t chunk_size, const func_t& f) { |
| 422 | // get number of blocks for L2 in most inner loop |
| 423 | int64_t cache_blocks_nb = get_cache_blocks<T>(chunk_size); |
| 424 | |
| 425 | // loop order: [NB / cache_blocks_nb, MB, cache_blocks_nb] |
| 426 | // TODO: implement reverse order of [MB / cache_blocks_mb, NB, cache_blocks_mb] |
| 427 | for (int64_t nbb = nb0; nbb < nb1; nbb += cache_blocks_nb) { |
| 428 | for (int64_t mb = mb0; mb < mb1; ++mb) { |
| 429 | for (int64_t nb = nbb; nb < std::min(nbb + cache_blocks_nb, nb1); ++nb) { |
| 430 | f(mb, nb, nb - nbb); |
| 431 | } |
| 432 | } |
| 433 | } |
| 434 | } |
| 435 | |
| 436 | // data indexing for dimension collapse |
| 437 | template <typename T> |
| 438 | inline T data_index_init(T offset) { |
| 439 | return offset; |
| 440 | } |
| 441 | |
| 442 | template <typename T, typename... Args> |
| 443 | inline T data_index_init(T offset, T& x, const T& X, Args&&... args) { |
| 444 | offset = data_index_init(offset, std::forward<Args>(args)...); |
| 445 | x = offset % X; |
| 446 | return offset / X; |
| 447 | } |
| 448 | |
| 449 | inline bool data_index_step() { |
| 450 | return true; |
| 451 | } |
| 452 | |
| 453 | template <typename T, typename... Args> |
| 454 | inline bool data_index_step(T& x, const T& X, Args&&... args) { |
| 455 | if (data_index_step(std::forward<Args>(args)...)) { |
| 456 | x = ((x + 1) == X) ? 0 : (x + 1); |
| 457 | return x == 0; |
| 458 | } |
| 459 | return false; |
| 460 | } |
| 461 | |
| 462 | // forced unroll for perf critical path |
| 463 | |
| 464 | #if __has_attribute(always_inline) |
| 465 | #define ALWAYS_INLINE __attribute__((__always_inline__)) inline |
| 466 | #else |
| 467 | #define ALWAYS_INLINE inline |
| 468 | #endif |
| 469 | |
| 470 | template <int n> |
| 471 | struct Unroll { |
| 472 | template <typename Func, typename... Args> |
| 473 | ALWAYS_INLINE void operator()(const Func& f, Args... args) const { |
| 474 | Unroll<n - 1>{}(f, args...); |
| 475 | f(std::integral_constant<int, n - 1>{}, args...); |
| 476 | } |
| 477 | }; |
| 478 | |
| 479 | template <> |
| 480 | struct Unroll<1> { |
| 481 | template <typename Func, typename... Args> |
| 482 | ALWAYS_INLINE void operator()(const Func& f, Args... args) const { |
| 483 | f(std::integral_constant<int, 0>{}, args...); |
| 484 | } |
| 485 | }; |
| 486 | |
| 487 | // conditional data ptr for optional tensor |
| 488 | template <typename T> |
| 489 | inline T* conditional_data_ptr(const std::optional<at::Tensor>& opt) { |
| 490 | return opt.has_value() ? opt.value().data_ptr<T>() : nullptr; |
| 491 | } |
| 492 | |
| 493 | // Adapted from sglang: M blocking utilities for FP8 W8A8 kernels |
| 494 | inline int64_t get_m_block(int64_t M) { |
| 495 | if (M <= 48) { |
| 496 | return M; |
| 497 | } else if (M < 64) { |
| 498 | return 32; |
| 499 | } else if (M < 96) { |
| 500 | return 64; |
| 501 | } else { |
| 502 | return 128; |
| 503 | } |
| 504 | } |
| 505 | |
| 506 | inline std::tuple<bool, int64_t, int64_t, int64_t> get_m_blocking(int64_t M) { |
| 507 | bool parallel_on_M = M > 128; |
| 508 | int64_t block_m = get_m_block(M); |
| 509 | int64_t Mc = (M + block_m - 1) / block_m; |
| 510 | int64_t Mc_parallel = parallel_on_M ? Mc : 1; |
| 511 | return std::make_tuple(parallel_on_M, block_m, Mc, Mc_parallel); |
| 512 | } |
| 513 | |
| 514 | #if defined(CPU_CAPABILITY_AVX512) |
| 515 | template <typename T> |
| 516 | void zero_buffer(T* data, int64_t size) { |
| 517 | const int32_t vec_size = at::vec::Vectorized<T>::size(); |
| 518 | auto zero_vec = at::vec::Vectorized<T>(0); |
| 519 | int64_t d = 0; |
| 520 | for (; d < size - (size % vec_size); d += vec_size) { |
| 521 | zero_vec.store(data + d); |
| 522 | } |
| 523 | if (d < size) { |
| 524 | zero_vec.store(data + d, size - d); |
| 525 | } |
| 526 | } |
| 527 | #else |
| 528 | template <typename T> |
| 529 | void zero_buffer(T* data, int64_t size) { |
| 530 | memset(data, 0, sizeof(T) * size); |
| 531 | } |
| 532 | #endif |
| 533 | |
| 534 | #if defined(__x86_64__) |
| 535 | // CPU_CAPABILITY_AVX10_2 is defined by cmake only when the compiler accepts |
| 536 | // -mavx10.2; without it the builtin below is not merely absent but a hard |
| 537 | // error, so the AVX512 paths have to stand in. |
| 538 | bool avx10_2_available() { |
| 539 | #ifdef CPU_CAPABILITY_AVX10_2 |
| 540 | // __builtin_cpu_supports returns the masked feature bit rather than 0/1. |
| 541 | static const bool available = __builtin_cpu_supports("avx10.2"); |
| 542 | return available; |
| 543 | #else |
| 544 | return false; |
| 545 | #endif |
| 546 | } |
| 547 | #endif |
| 548 | |
| 549 | } // anonymous namespace |