mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-04-19 14:29:05 +00:00
5d37d7bff4
* delete obselete files * move files * build * update cmake * update cmake * fix build * reorg examples * update cmake for example and test
357 lines
15 KiB
C++
357 lines
15 KiB
C++
#ifndef CK_BUFFER_HPP
|
|
#define CK_BUFFER_HPP
|
|
|
|
#include "amd_buffer_addressing.hpp"
|
|
#include "c_style_pointer_cast.hpp"
|
|
#include "config.hpp"
|
|
#include "enable_if.hpp"
|
|
|
|
namespace ck {
|
|
|
|
template <AddressSpaceEnum_t BufferAddressSpace,
|
|
typename T,
|
|
typename ElementSpaceSize,
|
|
bool InvalidElementUseNumericalZeroValue>
|
|
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_t 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,
|
|
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 need to be 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_t::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>(
|
|
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>(
|
|
p_data_, i, is_valid_element, element_space_size_, invalid_element_value_);
|
|
}
|
|
}
|
|
else
|
|
{
|
|
if constexpr(InvalidElementUseNumericalZeroValue)
|
|
{
|
|
#if CK_EXPERIMENTAL_USE_MEMCPY_FOR_VECTOR_ACCESS
|
|
X tmp;
|
|
|
|
__builtin_memcpy(&tmp, &(p_data_[i]), sizeof(X));
|
|
|
|
return is_valid_element ? tmp : X{0};
|
|
#else
|
|
return is_valid_element ? *c_style_pointer_cast<const X*>(&p_data_[i]) : X{0};
|
|
#endif
|
|
}
|
|
else
|
|
{
|
|
#if CK_EXPERIMENTAL_USE_MEMCPY_FOR_VECTOR_ACCESS
|
|
X tmp;
|
|
|
|
__builtin_memcpy(&tmp, &(p_data_[i]), sizeof(X));
|
|
|
|
return is_valid_element ? tmp : X{invalid_element_value_};
|
|
#else
|
|
return is_valid_element ? *c_style_pointer_cast<const X*>(&p_data_[i])
|
|
: X{invalid_element_value_};
|
|
#endif
|
|
}
|
|
}
|
|
}
|
|
|
|
template <InMemoryDataOperationEnum_t 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_t::Set)
|
|
{
|
|
this->template Set<X>(i, is_valid_element, x);
|
|
}
|
|
else if constexpr(Op == InMemoryDataOperationEnum_t::AtomicAdd)
|
|
{
|
|
this->template AtomicAdd<X>(i, is_valid_element, x);
|
|
}
|
|
else if constexpr(Op == InMemoryDataOperationEnum_t::Add)
|
|
{
|
|
auto tmp = this->template Get<X>(i, is_valid_element);
|
|
this->template Set<X>(i, is_valid_element, x + tmp);
|
|
// tmp += x;
|
|
// this->template Set<X>(i, is_valid_element, tmp);
|
|
}
|
|
}
|
|
|
|
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 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 need to be multiple T");
|
|
|
|
if constexpr(GetAddressSpace() == AddressSpaceEnum_t::Global)
|
|
{
|
|
#if CK_USE_AMD_BUFFER_STORE
|
|
constexpr index_t t_per_x = scalar_per_x_vector / scalar_per_t_vector;
|
|
|
|
amd_buffer_store<remove_cvref_t<T>, t_per_x>(
|
|
x, p_data_, i, is_valid_element, element_space_size_);
|
|
#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
|
|
}
|
|
#endif
|
|
}
|
|
else if constexpr(GetAddressSpace() == AddressSpaceEnum_t::Lds)
|
|
{
|
|
if(is_valid_element)
|
|
{
|
|
#if !CK_WORKAROUND_SWDEV_XXXXXX_INT8_DS_WRITE_ISSUE
|
|
#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
|
|
#else
|
|
// 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
|
|
if constexpr(is_same<typename scalar_type<remove_cvref_t<T>>::type, int8_t>::value)
|
|
{
|
|
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 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
|
|
}
|
|
#endif
|
|
}
|
|
}
|
|
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)
|
|
{
|
|
// 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 need to be multiple T");
|
|
|
|
static_assert(GetAddressSpace() == AddressSpaceEnum_t::Global, "only support global mem");
|
|
|
|
#if CK_USE_AMD_BUFFER_ATOMIC_ADD
|
|
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)
|
|
{
|
|
atomicAdd(&p_data_[i], x);
|
|
}
|
|
#endif
|
|
}
|
|
|
|
__host__ __device__ static constexpr bool IsStaticBuffer() { return false; }
|
|
|
|
__host__ __device__ static constexpr bool IsDynamicBuffer() { return true; }
|
|
};
|
|
|
|
template <AddressSpaceEnum_t BufferAddressSpace, typename T, typename ElementSpaceSize>
|
|
__host__ __device__ constexpr auto make_dynamic_buffer(T* p, ElementSpaceSize element_space_size)
|
|
{
|
|
return DynamicBuffer<BufferAddressSpace, T, ElementSpaceSize, true>{p, element_space_size};
|
|
}
|
|
|
|
template <
|
|
AddressSpaceEnum_t BufferAddressSpace,
|
|
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>{
|
|
p, element_space_size, invalid_element_value};
|
|
}
|
|
|
|
} // namespace ck
|
|
#endif
|