File size: 19,458 Bytes
57c3a10 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 | /*
* Copyright (c) 2023 by FlashInfer team.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#ifndef FLASHINFER_UTILS_CUH_
#define FLASHINFER_UTILS_CUH_
#include <cuda_bf16.h>
#include <cuda_device_runtime_api.h>
#include <cuda_fp16.h>
#include <cuda_fp8.h>
#include <cuda_runtime.h>
#include <cstdint>
#include <iostream>
#include <type_traits>
#include <vector>
#include "exception.h"
#define STR_HELPER(x) #x
#define STR(x) STR_HELPER(x)
// macro to turn off fp16 qk reduction to reduce binary
#ifndef FLASHINFER_ALWAYS_DISUSE_FP16_QK_REDUCTION
#define FLASHINFER_ALWAYS_DISUSE_FP16_QK_REDUCTION 0
#endif
#ifndef NDEBUG
#define FLASHINFER_CUDA_CALL(func, ...) \
{ \
cudaError_t e = (func); \
if (e != cudaSuccess) { \
std::cerr << "CUDA Error: " << cudaGetErrorString(e) << " (" << e << ") " << __FILE__ \
<< ": line " << __LINE__ << " at function " << STR(func) << std::endl; \
return e; \
} \
}
#else
#define FLASHINFER_CUDA_CALL(func, ...) \
{ \
cudaError_t e = (func); \
if (e != cudaSuccess) { \
return e; \
} \
}
#endif
#define DISPATCH_USE_FP16_QK_REDUCTION(use_fp16_qk_reduction, USE_FP16_QK_REDUCTION, ...) \
if (use_fp16_qk_reduction) { \
FLASHINFER_ERROR("FP16_QK_REDUCTION disabled at compile time"); \
} else { \
constexpr bool USE_FP16_QK_REDUCTION = false; \
__VA_ARGS__ \
}
#define DISPATCH_NUM_MMA_Q(num_mma_q, NUM_MMA_Q, ...) \
if (num_mma_q == 1) { \
constexpr size_t NUM_MMA_Q = 1; \
__VA_ARGS__ \
} else if (num_mma_q == 2) { \
constexpr size_t NUM_MMA_Q = 2; \
__VA_ARGS__ \
} else { \
std::ostringstream err_msg; \
err_msg << "Unsupported num_mma_q: " << num_mma_q; \
FLASHINFER_ERROR(err_msg.str()); \
}
#define DISPATCH_NUM_MMA_KV(max_mma_kv, NUM_MMA_KV, ...) \
if (max_mma_kv >= 8) { \
constexpr size_t NUM_MMA_KV = 8; \
__VA_ARGS__ \
} else if (max_mma_kv >= 4) { \
constexpr size_t NUM_MMA_KV = 4; \
__VA_ARGS__ \
} else if (max_mma_kv >= 2) { \
constexpr size_t NUM_MMA_KV = 2; \
__VA_ARGS__ \
} else if (max_mma_kv >= 1) { \
constexpr size_t NUM_MMA_KV = 1; \
__VA_ARGS__ \
} else { \
std::ostringstream err_msg; \
err_msg << "Unsupported max_mma_kv: " << max_mma_kv; \
FLASHINFER_ERROR(err_msg.str()); \
}
#define DISPATCH_CTA_TILE_Q(cta_tile_q, CTA_TILE_Q, ...) \
switch (cta_tile_q) { \
case 128: { \
constexpr uint32_t CTA_TILE_Q = 128; \
__VA_ARGS__ \
break; \
} \
case 64: { \
constexpr uint32_t CTA_TILE_Q = 64; \
__VA_ARGS__ \
break; \
} \
case 16: { \
constexpr uint32_t CTA_TILE_Q = 16; \
__VA_ARGS__ \
break; \
} \
default: { \
std::ostringstream err_msg; \
err_msg << "Unsupported cta_tile_q: " << cta_tile_q; \
FLASHINFER_ERROR(err_msg.str()); \
} \
}
#define DISPATCH_GQA_GROUP_SIZE(group_size, GROUP_SIZE, ...) \
if (group_size == 1) { \
constexpr size_t GROUP_SIZE = 1; \
__VA_ARGS__ \
} else if (group_size == 2) { \
constexpr size_t GROUP_SIZE = 2; \
__VA_ARGS__ \
} else if (group_size == 3) { \
constexpr size_t GROUP_SIZE = 3; \
__VA_ARGS__ \
} else if (group_size == 4) { \
constexpr size_t GROUP_SIZE = 4; \
__VA_ARGS__ \
} else if (group_size == 8) { \
constexpr size_t GROUP_SIZE = 8; \
__VA_ARGS__ \
} else { \
std::ostringstream err_msg; \
err_msg << "Unsupported group_size: " << group_size; \
FLASHINFER_ERROR(err_msg.str()); \
}
#define DISPATCH_MASK_MODE(mask_mode, MASK_MODE, ...) \
switch (mask_mode) { \
case MaskMode::kNone: { \
constexpr MaskMode MASK_MODE = MaskMode::kNone; \
__VA_ARGS__ \
break; \
} \
case MaskMode::kCausal: { \
constexpr MaskMode MASK_MODE = MaskMode::kCausal; \
__VA_ARGS__ \
break; \
} \
case MaskMode::kCustom: { \
constexpr MaskMode MASK_MODE = MaskMode::kCustom; \
__VA_ARGS__ \
break; \
} \
case MaskMode::kMultiItemScoring: { \
constexpr MaskMode MASK_MODE = MaskMode::kMultiItemScoring; \
__VA_ARGS__ \
break; \
} \
default: { \
std::ostringstream err_msg; \
err_msg << "Unsupported mask_mode: " << int(mask_mode); \
FLASHINFER_ERROR(err_msg.str()); \
} \
}
// convert head_dim to compile-time constant
#define DISPATCH_HEAD_DIM(head_dim, HEAD_DIM, ...) \
switch (head_dim) { \
case 64: { \
constexpr size_t HEAD_DIM = 64; \
__VA_ARGS__ \
break; \
} \
case 128: { \
constexpr size_t HEAD_DIM = 128; \
__VA_ARGS__ \
break; \
} \
case 256: { \
constexpr size_t HEAD_DIM = 256; \
__VA_ARGS__ \
break; \
} \
case 512: { \
constexpr size_t HEAD_DIM = 512; \
__VA_ARGS__ \
break; \
} \
default: { \
std::ostringstream err_msg; \
err_msg << "Unsupported head_dim: " << head_dim; \
FLASHINFER_ERROR(err_msg.str()); \
} \
}
#define DISPATCH_POS_ENCODING_MODE(pos_encoding_mode, POS_ENCODING_MODE, ...) \
switch (pos_encoding_mode) { \
case PosEncodingMode::kNone: { \
constexpr PosEncodingMode POS_ENCODING_MODE = PosEncodingMode::kNone; \
__VA_ARGS__ \
break; \
} \
case PosEncodingMode::kRoPELlama: { \
constexpr PosEncodingMode POS_ENCODING_MODE = PosEncodingMode::kRoPELlama; \
__VA_ARGS__ \
break; \
} \
case PosEncodingMode::kALiBi: { \
constexpr PosEncodingMode POS_ENCODING_MODE = PosEncodingMode::kALiBi; \
__VA_ARGS__ \
break; \
} \
default: { \
std::ostringstream err_msg; \
err_msg << "Unsupported pos_encoding_mode: " << int(pos_encoding_mode); \
FLASHINFER_ERROR(err_msg.str()); \
} \
}
#define DISPATCH_ALIGNED_VEC_SIZE(aligned_vec_size, ALIGNED_VEC_SIZE, ...) \
switch (aligned_vec_size) { \
case 16: { \
constexpr size_t ALIGNED_VEC_SIZE = 16; \
__VA_ARGS__ \
break; \
} \
case 8: { \
constexpr size_t ALIGNED_VEC_SIZE = 8; \
__VA_ARGS__ \
break; \
} \
case 4: { \
constexpr size_t ALIGNED_VEC_SIZE = 4; \
__VA_ARGS__ \
break; \
} \
case 2: { \
constexpr size_t ALIGNED_VEC_SIZE = 2; \
__VA_ARGS__ \
break; \
} \
case 1: { \
constexpr size_t ALIGNED_VEC_SIZE = 1; \
__VA_ARGS__ \
break; \
} \
default: { \
std::ostringstream err_msg; \
err_msg << "Unsupported aligned_vec_size: " << aligned_vec_size; \
FLASHINFER_ERROR(err_msg.str()); \
} \
}
#define DISPATCH_COMPUTE_CAP_DECODE_NUM_STAGES_SMEM(compute_capacity, NUM_STAGES_SMEM, ...) \
if (compute_capacity.first >= 8) { \
constexpr uint32_t NUM_STAGES_SMEM = 2; \
__VA_ARGS__ \
} else { \
constexpr uint32_t NUM_STAGES_SMEM = 1; \
__VA_ARGS__ \
}
namespace flashinfer {
template <typename T1, typename T2>
__forceinline__ __device__ __host__ T1 ceil_div(const T1 x, const T2 y) {
return (x + y - 1) / y;
}
template <typename T1, typename T2>
__forceinline__ __device__ __host__ T1 round_up(const T1 x, const T2 y) {
return ceil_div(x, y) * y;
}
inline std::pair<int, int> GetCudaComputeCapability() {
int device_id = 0;
cudaGetDevice(&device_id);
int major = 0, minor = 0;
cudaDeviceGetAttribute(&major, cudaDevAttrComputeCapabilityMajor, device_id);
cudaDeviceGetAttribute(&minor, cudaDevAttrComputeCapabilityMinor, device_id);
return std::make_pair(major, minor);
}
template <typename T>
inline void DebugPrintCUDAArray(T* device_ptr, size_t size, std::string prefix = "") {
std::vector<T> host_array(size);
std::cout << prefix;
cudaMemcpy(host_array.data(), device_ptr, size * sizeof(T), cudaMemcpyDeviceToHost);
for (size_t i = 0; i < size; ++i) {
std::cout << host_array[i] << " ";
}
std::cout << std::endl;
}
inline uint32_t FA2DetermineCtaTileQ(int64_t avg_packed_qo_len, uint32_t head_dim) {
if (avg_packed_qo_len > 64 && head_dim < 256) {
return 128;
} else {
auto compute_capacity = GetCudaComputeCapability();
if (compute_capacity.first >= 8) {
// Ampere or newer
if (avg_packed_qo_len > 16) {
// avg_packed_qo_len <= 64
return 64;
} else {
// avg_packed_qo_len <= 16
return 16;
}
} else {
// NOTE(Zihao): not enough shared memory on Turing for 1x4 warp layout
return 64;
}
}
}
#define LOOP_SPLIT_MASK(iter, COND1, COND2, ...) \
{ \
_Pragma("unroll 1") for (; (COND1); (iter) -= 1) { \
constexpr bool WITH_MASK = true; \
__VA_ARGS__ \
} \
_Pragma("unroll 1") for (; (COND2); (iter) -= 1) { \
constexpr bool WITH_MASK = false; \
__VA_ARGS__ \
} \
}
/*!
* \brief Return x - y if x > y, otherwise return 0.
*/
__device__ __forceinline__ uint32_t sub_if_greater_or_zero(uint32_t x, uint32_t y) {
return (x > y) ? x - y : 0U;
}
__device__ __forceinline__ void swap(uint32_t& a, uint32_t& b) {
uint32_t tmp = a;
a = b;
b = tmp;
}
__device__ __forceinline__ uint32_t dim2_offset(const uint32_t& dim_a, const uint32_t& idx_b,
const uint32_t& idx_a) {
return idx_b * dim_a + idx_a;
}
__device__ __forceinline__ uint32_t dim3_offset(const uint32_t& dim_b, const uint32_t& dim_a,
const uint32_t& idx_c, const uint32_t& idx_b,
const uint32_t& idx_a) {
return (idx_c * dim_b + idx_b) * dim_a + idx_a;
}
__device__ __forceinline__ uint32_t dim4_offset(const uint32_t& dim_c, const uint32_t& dim_b,
const uint32_t& dim_a, const uint32_t& idx_d,
const uint32_t& idx_c, const uint32_t& idx_b,
const uint32_t& idx_a) {
return ((idx_d * dim_c + idx_c) * dim_b + idx_b) * dim_a + idx_a;
}
#define DEFINE_HAS_MEMBER(member) \
template <typename T, typename = void> \
struct has_##member : std::false_type {}; \
template <typename T> \
struct has_##member<T, std::void_t<decltype(std::declval<T>().member)>> : std::true_type {}; \
template <typename T> \
inline constexpr bool has_##member##_v = has_##member<T>::value;
} // namespace flashinfer
#endif // FLASHINFER_UTILS_CUH_
|