File size: 20,290 Bytes
b024d42 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 | // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/ck.hpp"
#include "ck/utility/data_type.hpp"
#include "enable_if.hpp"
#include "c_style_pointer_cast.hpp"
#include "amd_buffer_addressing.hpp"
#include "generic_memory_space_atomic.hpp"
namespace ck {
// T may be scalar or vector
// X may be scalar or vector
// T and X have same scalar type
// X contains multiple T
template <AddressSpaceEnum BufferAddressSpace,
typename T,
typename ElementSpaceSize,
bool InvalidElementUseNumericalZeroValue,
AmdBufferCoherenceEnum coherence = AmdBufferCoherenceEnum::DefaultCoherence>
struct DynamicBuffer
{
using type = T;
T* p_data_;
ElementSpaceSize element_space_size_;
T invalid_element_value_ = T{0};
__host__ __device__ constexpr DynamicBuffer(T* p_data, ElementSpaceSize element_space_size)
: p_data_{p_data}, element_space_size_{element_space_size}
{
}
__host__ __device__ constexpr DynamicBuffer(T* p_data,
ElementSpaceSize element_space_size,
T invalid_element_value)
: p_data_{p_data},
element_space_size_{element_space_size},
invalid_element_value_{invalid_element_value}
{
}
__host__ __device__ static constexpr AddressSpaceEnum GetAddressSpace()
{
return BufferAddressSpace;
}
__host__ __device__ constexpr const T& operator[](index_t i) const { return p_data_[i]; }
__host__ __device__ constexpr T& operator()(index_t i) { return p_data_[i]; }
template <typename X,
typename enable_if<is_same<typename scalar_type<remove_cvref_t<X>>::type,
typename scalar_type<remove_cvref_t<T>>::type>::value ||
!is_native_type<X>(),
bool>::type = false>
__host__ __device__ constexpr auto Get(index_t i, bool is_valid_element) const
{
// X contains multiple T
constexpr index_t scalar_per_t_vector = scalar_type<remove_cvref_t<T>>::vector_size;
constexpr index_t scalar_per_x_vector = scalar_type<remove_cvref_t<X>>::vector_size;
static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
"wrong! X should contain multiple T");
#if CK_USE_AMD_BUFFER_LOAD
bool constexpr use_amd_buffer_addressing = true;
#else
bool constexpr use_amd_buffer_addressing = false;
#endif
if constexpr(GetAddressSpace() == AddressSpaceEnum::Global && use_amd_buffer_addressing)
{
constexpr index_t t_per_x = scalar_per_x_vector / scalar_per_t_vector;
if constexpr(InvalidElementUseNumericalZeroValue)
{
return amd_buffer_load_invalid_element_return_zero<remove_cvref_t<T>,
t_per_x,
coherence>(
p_data_, i, is_valid_element, element_space_size_);
}
else
{
return amd_buffer_load_invalid_element_return_customized_value<remove_cvref_t<T>,
t_per_x,
coherence>(
p_data_, i, is_valid_element, element_space_size_, invalid_element_value_);
}
}
else
{
if(is_valid_element)
{
#if CK_EXPERIMENTAL_USE_MEMCPY_FOR_VECTOR_ACCESS
X tmp;
__builtin_memcpy(&tmp, &(p_data_[i]), sizeof(X));
return tmp;
#else
return *c_style_pointer_cast<const X*>(&p_data_[i]);
#endif
}
else
{
if constexpr(InvalidElementUseNumericalZeroValue)
{
return X{0};
}
else
{
return X{invalid_element_value_};
}
}
}
}
template <InMemoryDataOperationEnum Op,
typename X,
typename enable_if<is_same<typename scalar_type<remove_cvref_t<X>>::type,
typename scalar_type<remove_cvref_t<T>>::type>::value,
bool>::type = false>
__host__ __device__ void Update(index_t i, bool is_valid_element, const X& x)
{
if constexpr(Op == InMemoryDataOperationEnum::Set)
{
this->template Set<X>(i, is_valid_element, x);
}
else if constexpr(Op == InMemoryDataOperationEnum::AtomicAdd)
{
this->template AtomicAdd<X>(i, is_valid_element, x);
}
else if constexpr(Op == InMemoryDataOperationEnum::AtomicMax)
{
this->template AtomicMax<X>(i, is_valid_element, x);
}
else if constexpr(Op == InMemoryDataOperationEnum::Add)
{
auto tmp = this->template Get<X>(i, is_valid_element);
using scalar_t = typename scalar_type<remove_cvref_t<T>>::type;
// handle bfloat addition
if constexpr(is_same_v<scalar_t, bhalf_t>)
{
if constexpr(is_scalar_type<X>::value)
{
// Scalar type
auto result =
type_convert<X>(type_convert<float>(x) + type_convert<float>(tmp));
this->template Set<X>(i, is_valid_element, result);
}
else
{
// Vector type
constexpr auto vector_size = scalar_type<remove_cvref_t<X>>::vector_size;
const vector_type<scalar_t, vector_size> a_vector{tmp};
const vector_type<scalar_t, vector_size> b_vector{x};
static_for<0, vector_size, 1>{}([&](auto idx) {
auto result = type_convert<scalar_t>(
type_convert<float>(a_vector.template AsType<scalar_t>()[idx]) +
type_convert<float>(b_vector.template AsType<scalar_t>()[idx]));
this->template Set<scalar_t>(i + idx, is_valid_element, result);
});
}
}
else
{
this->template Set<X>(i, is_valid_element, x + tmp);
}
}
}
template <typename DstBuffer, index_t NumElemsPerThread>
__host__ __device__ void DirectCopyToLds(DstBuffer& dst_buf,
index_t src_offset,
index_t dst_offset,
bool is_valid_element) const
{
// Copy data from global to LDS memory using direct loads.
static_assert(GetAddressSpace() == AddressSpaceEnum::Global,
"Source data must come from a global memory buffer.");
static_assert(DstBuffer::GetAddressSpace() == AddressSpaceEnum::Lds,
"Destination data must be stored in an LDS memory buffer.");
amd_direct_load_global_to_lds<T, NumElemsPerThread>(p_data_,
src_offset,
dst_buf.p_data_,
dst_offset,
is_valid_element,
element_space_size_);
}
template <typename X,
typename enable_if<is_same<typename scalar_type<remove_cvref_t<X>>::type,
typename scalar_type<remove_cvref_t<T>>::type>::value ||
!is_native_type<X>(),
bool>::type = false>
__host__ __device__ void Set(index_t i, bool is_valid_element, const X& x)
{
// X contains multiple T
constexpr index_t scalar_per_t_vector = scalar_type<remove_cvref_t<T>>::vector_size;
constexpr index_t scalar_per_x_vector = scalar_type<remove_cvref_t<X>>::vector_size;
static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
"wrong! X should contain multiple T");
#if CK_USE_AMD_BUFFER_STORE
bool constexpr use_amd_buffer_addressing = true;
#else
bool constexpr use_amd_buffer_addressing = false;
#endif
#if CK_WORKAROUND_SWDEV_XXXXXX_INT8_DS_WRITE_ISSUE
bool constexpr workaround_int8_ds_write_issue = true;
#else
bool constexpr workaround_int8_ds_write_issue = false;
#endif
if constexpr(GetAddressSpace() == AddressSpaceEnum::Global && use_amd_buffer_addressing)
{
constexpr index_t t_per_x = scalar_per_x_vector / scalar_per_t_vector;
amd_buffer_store<remove_cvref_t<T>, t_per_x, coherence>(
x, p_data_, i, is_valid_element, element_space_size_);
}
else if constexpr(GetAddressSpace() == AddressSpaceEnum::Lds &&
is_same<typename scalar_type<remove_cvref_t<T>>::type, int8_t>::value &&
workaround_int8_ds_write_issue)
{
if(is_valid_element)
{
// HACK: compiler would lower IR "store<i8, 16> address_space(3)" into inefficient
// ISA, so I try to let compiler emit IR "store<i32, 4>" which would be lower to
// ds_write_b128
// TODO: remove this after compiler fix
static_assert((is_same<remove_cvref_t<T>, int8_t>::value &&
is_same<remove_cvref_t<X>, int8_t>::value) ||
(is_same<remove_cvref_t<T>, int8_t>::value &&
is_same<remove_cvref_t<X>, int8x2_t>::value) ||
(is_same<remove_cvref_t<T>, int8_t>::value &&
is_same<remove_cvref_t<X>, int8x4_t>::value) ||
(is_same<remove_cvref_t<T>, int8_t>::value &&
is_same<remove_cvref_t<X>, int8x8_t>::value) ||
(is_same<remove_cvref_t<T>, int8_t>::value &&
is_same<remove_cvref_t<X>, int8x16_t>::value) ||
(is_same<remove_cvref_t<T>, int8x4_t>::value &&
is_same<remove_cvref_t<X>, int8x4_t>::value) ||
(is_same<remove_cvref_t<T>, int8x8_t>::value &&
is_same<remove_cvref_t<X>, int8x8_t>::value) ||
(is_same<remove_cvref_t<T>, int8x16_t>::value &&
is_same<remove_cvref_t<X>, int8x16_t>::value),
"wrong! not implemented for this combination, please add "
"implementation");
if constexpr(is_same<remove_cvref_t<T>, int8_t>::value &&
is_same<remove_cvref_t<X>, int8_t>::value)
{
// HACK: cast pointer of x is bad
// TODO: remove this after compiler fix
*c_style_pointer_cast<int8_t*>(&p_data_[i]) =
*c_style_pointer_cast<const int8_t*>(&x);
}
else if constexpr(is_same<remove_cvref_t<T>, int8_t>::value &&
is_same<remove_cvref_t<X>, int8x2_t>::value)
{
// HACK: cast pointer of x is bad
// TODO: remove this after compiler fix
*c_style_pointer_cast<int16_t*>(&p_data_[i]) =
*c_style_pointer_cast<const int16_t*>(&x);
}
else if constexpr(is_same<remove_cvref_t<T>, int8_t>::value &&
is_same<remove_cvref_t<X>, int8x4_t>::value)
{
// HACK: cast pointer of x is bad
// TODO: remove this after compiler fix
*c_style_pointer_cast<int32_t*>(&p_data_[i]) =
*c_style_pointer_cast<const int32_t*>(&x);
}
else if constexpr(is_same<remove_cvref_t<T>, int8_t>::value &&
is_same<remove_cvref_t<X>, int8x8_t>::value)
{
// HACK: cast pointer of x is bad
// TODO: remove this after compiler fix
*c_style_pointer_cast<int32x2_t*>(&p_data_[i]) =
*c_style_pointer_cast<const int32x2_t*>(&x);
}
else if constexpr(is_same<remove_cvref_t<T>, int8_t>::value &&
is_same<remove_cvref_t<X>, int8x16_t>::value)
{
// HACK: cast pointer of x is bad
// TODO: remove this after compiler fix
*c_style_pointer_cast<int32x4_t*>(&p_data_[i]) =
*c_style_pointer_cast<const int32x4_t*>(&x);
}
else if constexpr(is_same<remove_cvref_t<T>, int8x4_t>::value &&
is_same<remove_cvref_t<X>, int8x4_t>::value)
{
// HACK: cast pointer of x is bad
// TODO: remove this after compiler fix
*c_style_pointer_cast<int32_t*>(&p_data_[i]) =
*c_style_pointer_cast<const int32_t*>(&x);
}
else if constexpr(is_same<remove_cvref_t<T>, int8x8_t>::value &&
is_same<remove_cvref_t<X>, int8x8_t>::value)
{
// HACK: cast pointer of x is bad
// TODO: remove this after compiler fix
*c_style_pointer_cast<int32x2_t*>(&p_data_[i]) =
*c_style_pointer_cast<const int32x2_t*>(&x);
}
else if constexpr(is_same<remove_cvref_t<T>, int8x16_t>::value &&
is_same<remove_cvref_t<X>, int8x16_t>::value)
{
// HACK: cast pointer of x is bad
// TODO: remove this after compiler fix
*c_style_pointer_cast<int32x4_t*>(&p_data_[i]) =
*c_style_pointer_cast<const int32x4_t*>(&x);
}
}
}
else
{
if(is_valid_element)
{
#if CK_EXPERIMENTAL_USE_MEMCPY_FOR_VECTOR_ACCESS
X tmp = x;
__builtin_memcpy(&(p_data_[i]), &tmp, sizeof(X));
#else
*c_style_pointer_cast<X*>(&p_data_[i]) = x;
#endif
}
}
}
template <typename X,
typename enable_if<is_same<typename scalar_type<remove_cvref_t<X>>::type,
typename scalar_type<remove_cvref_t<T>>::type>::value,
bool>::type = false>
__host__ __device__ void AtomicAdd(index_t i, bool is_valid_element, const X& x)
{
using scalar_t = typename scalar_type<remove_cvref_t<T>>::type;
// X contains multiple T
constexpr index_t scalar_per_t_vector = scalar_type<remove_cvref_t<T>>::vector_size;
constexpr index_t scalar_per_x_vector = scalar_type<remove_cvref_t<X>>::vector_size;
static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
"wrong! X should contain multiple T");
static_assert(GetAddressSpace() == AddressSpaceEnum::Global, "only support global mem");
#if CK_USE_AMD_BUFFER_ATOMIC_ADD_INTEGER && CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT
bool constexpr use_amd_buffer_addressing =
is_same_v<remove_cvref_t<scalar_t>, int32_t> ||
is_same_v<remove_cvref_t<scalar_t>, float> ||
(is_same_v<remove_cvref_t<scalar_t>, half_t> && scalar_per_x_vector % 2 == 0) ||
(is_same_v<remove_cvref_t<scalar_t>, bhalf_t> && scalar_per_x_vector % 2 == 0);
#elif CK_USE_AMD_BUFFER_ATOMIC_ADD_INTEGER && (!CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT)
bool constexpr use_amd_buffer_addressing = is_same_v<remove_cvref_t<scalar_t>, int32_t>;
#elif(!CK_USE_AMD_BUFFER_ATOMIC_ADD_INTEGER) && CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT
bool constexpr use_amd_buffer_addressing =
is_same_v<remove_cvref_t<scalar_t>, float> ||
(is_same_v<remove_cvref_t<scalar_t>, half_t> && scalar_per_x_vector % 2 == 0) ||
(is_same_v<remove_cvref_t<scalar_t>, bhalf_t> && scalar_per_x_vector % 2 == 0);
#else
bool constexpr use_amd_buffer_addressing = false;
#endif
if constexpr(use_amd_buffer_addressing)
{
constexpr index_t t_per_x = scalar_per_x_vector / scalar_per_t_vector;
amd_buffer_atomic_add<remove_cvref_t<T>, t_per_x>(
x, p_data_, i, is_valid_element, element_space_size_);
}
else
{
if(is_valid_element)
{
atomic_add<X>(c_style_pointer_cast<X*>(&p_data_[i]), x);
}
}
}
template <typename X,
typename enable_if<is_same<typename scalar_type<remove_cvref_t<X>>::type,
typename scalar_type<remove_cvref_t<T>>::type>::value,
bool>::type = false>
__host__ __device__ void AtomicMax(index_t i, bool is_valid_element, const X& x)
{
// X contains multiple T
constexpr index_t scalar_per_t_vector = scalar_type<remove_cvref_t<T>>::vector_size;
constexpr index_t scalar_per_x_vector = scalar_type<remove_cvref_t<X>>::vector_size;
static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
"wrong! X should contain multiple T");
static_assert(GetAddressSpace() == AddressSpaceEnum::Global, "only support global mem");
#if CK_USE_AMD_BUFFER_ATOMIC_MAX_FLOAT64
using scalar_t = typename scalar_type<remove_cvref_t<T>>::type;
bool constexpr use_amd_buffer_addressing = is_same_v<remove_cvref_t<scalar_t>, double>;
#else
bool constexpr use_amd_buffer_addressing = false;
#endif
if constexpr(use_amd_buffer_addressing)
{
constexpr index_t t_per_x = scalar_per_x_vector / scalar_per_t_vector;
amd_buffer_atomic_max<remove_cvref_t<T>, t_per_x>(
x, p_data_, i, is_valid_element, element_space_size_);
}
else if(is_valid_element)
{
atomic_max<X>(c_style_pointer_cast<X*>(&p_data_[i]), x);
}
}
__host__ __device__ static constexpr bool IsStaticBuffer() { return false; }
__host__ __device__ static constexpr bool IsDynamicBuffer() { return true; }
};
template <AddressSpaceEnum BufferAddressSpace,
AmdBufferCoherenceEnum coherence = AmdBufferCoherenceEnum::DefaultCoherence,
typename T,
typename ElementSpaceSize>
__host__ __device__ constexpr auto make_dynamic_buffer(T* p, ElementSpaceSize element_space_size)
{
return DynamicBuffer<BufferAddressSpace, T, ElementSpaceSize, true, coherence>{
p, element_space_size};
}
template <
AddressSpaceEnum BufferAddressSpace,
AmdBufferCoherenceEnum coherence = AmdBufferCoherenceEnum::DefaultCoherence,
typename T,
typename ElementSpaceSize,
typename X,
typename enable_if<is_same<remove_cvref_t<T>, remove_cvref_t<X>>::value, bool>::type = false>
__host__ __device__ constexpr auto
make_dynamic_buffer(T* p, ElementSpaceSize element_space_size, X invalid_element_value)
{
return DynamicBuffer<BufferAddressSpace, T, ElementSpaceSize, false, coherence>{
p, element_space_size, invalid_element_value};
}
} // namespace ck
|