Skip to content

GitLab

Unverified Commit cd167e49 authored by Chao Liu's avatar Chao Liu Committed by GitHub
Browse files

Compile for gfx908 and gfx90a (#130) 

* adding compilation for multiple targets

* fix build

* clean

* update Jekinsfile

* update readme

* update Jenkins

* use ck::half_t instead of ushort for bf16

* rename enum classes

* clean

* rename

* clean
parent ecf337ba
Hide whitespace changes
Inline Side-by-side
Showing with 232 additions and 249 deletions
include/ck/utility/data_type_enum.hpp
View file @ cd167e49
......@@ -3,7 +3,7 @@
namespace ck {
enum struct DataTypeEnum_t
enum struct DataTypeEnum
{
Half = 0,
Float = 1,
......
include/ck/utility/data_type_enum_helper.hpp
View file @ cd167e49
......@@ -6,35 +6,35 @@
namespace ck {
template <DataTypeEnum_t DataTypeEnum>
template <DataTypeEnum DataTypeEnum>
struct get_datatype_from_enum;
template <>
struct get_datatype_from_enum<DataTypeEnum_t::Int8>
struct get_datatype_from_enum<DataTypeEnum::Int8>
{
using type = int8_t;
};
template <>
struct get_datatype_from_enum<DataTypeEnum_t::Int32>
struct get_datatype_from_enum<DataTypeEnum::Int32>
{
using type = int32_t;
};
template <>
struct get_datatype_from_enum<DataTypeEnum_t::Half>
struct get_datatype_from_enum<DataTypeEnum::Half>
{
using type = half_t;
};
template <>
struct get_datatype_from_enum<DataTypeEnum_t::Float>
struct get_datatype_from_enum<DataTypeEnum::Float>
{
using type = float;
};
template <>
struct get_datatype_from_enum<DataTypeEnum_t::Double>
struct get_datatype_from_enum<DataTypeEnum::Double>
{
using type = double;
};
......@@ -45,31 +45,31 @@ struct get_datatype_enum_from_type;
template <>
struct get_datatype_enum_from_type<int8_t>
{
static constexpr DataTypeEnum_t value = DataTypeEnum_t::Int8;
static constexpr DataTypeEnum value = DataTypeEnum::Int8;
};
template <>
struct get_datatype_enum_from_type<int32_t>
{
static constexpr DataTypeEnum_t value = DataTypeEnum_t::Int32;
static constexpr DataTypeEnum value = DataTypeEnum::Int32;
};
template <>
struct get_datatype_enum_from_type<half_t>
{
static constexpr DataTypeEnum_t value = DataTypeEnum_t::Half;
static constexpr DataTypeEnum value = DataTypeEnum::Half;
};
template <>
struct get_datatype_enum_from_type<float>
{
static constexpr DataTypeEnum_t value = DataTypeEnum_t::Float;
static constexpr DataTypeEnum value = DataTypeEnum::Float;
};
template <>
struct get_datatype_enum_from_type<double>
{
static constexpr DataTypeEnum_t value = DataTypeEnum_t::Double;
static constexpr DataTypeEnum value = DataTypeEnum::Double;
};
} // namespace ck
......
include/ck/utility/dynamic_buffer.hpp
View file @ cd167e49
#ifndef CK_BUFFER_HPP
#define CK_BUFFER_HPP
#pragma once
#include "amd_buffer_addressing.hpp"
#include "c_style_pointer_cast.hpp"
#include "config.hpp"
......@@ -8,7 +6,7 @@
namespace ck {
template <AddressSpaceEnum_t BufferAddressSpace,
template <AddressSpaceEnum BufferAddressSpace,
typename T,
typename ElementSpaceSize,
bool InvalidElementUseNumericalZeroValue>
......@@ -34,7 +32,7 @@ struct DynamicBuffer
{
}
__host__ __device__ static constexpr AddressSpaceEnum_t GetAddressSpace()
__host__ __device__ static constexpr AddressSpaceEnum GetAddressSpace()
{
return BufferAddressSpace;
}
......@@ -55,7 +53,7 @@ struct DynamicBuffer
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");
"wrong! X should contain multiple T");
#if CK_USE_AMD_BUFFER_LOAD
bool constexpr use_amd_buffer_addressing = true;
......@@ -63,7 +61,7 @@ struct DynamicBuffer
bool constexpr use_amd_buffer_addressing = false;
#endif
if constexpr(GetAddressSpace() == AddressSpaceEnum_t::Global && use_amd_buffer_addressing)
if constexpr(GetAddressSpace() == AddressSpaceEnum::Global && use_amd_buffer_addressing)
{
constexpr index_t t_per_x = scalar_per_x_vector / scalar_per_t_vector;
......@@ -81,50 +79,48 @@ struct DynamicBuffer
}
else
{
if constexpr(InvalidElementUseNumericalZeroValue)
if(is_valid_element)
{
#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};
return tmp;
#else
return is_valid_element ? *c_style_pointer_cast<const X*>(&p_data_[i]) : X{0};
return *c_style_pointer_cast<const X*>(&p_data_[i]);
#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
if constexpr(InvalidElementUseNumericalZeroValue)
{
return X{0};
}
else
{
return X{invalid_element_value_};
}
}
}
}
template <InMemoryDataOperationEnum_t Op,
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_t::Set)
if constexpr(Op == InMemoryDataOperationEnum::Set)
{
this->template Set<X>(i, is_valid_element, x);
}
else if constexpr(Op == InMemoryDataOperationEnum_t::AtomicAdd)
else if constexpr(Op == InMemoryDataOperationEnum::AtomicAdd)
{
this->template AtomicAdd<X>(i, is_valid_element, x);
}
else if constexpr(Op == InMemoryDataOperationEnum_t::Add)
else if constexpr(Op == InMemoryDataOperationEnum::Add)
{
auto tmp = this->template Get<X>(i, is_valid_element);
this->template Set<X>(i, is_valid_element, x + tmp);
......@@ -145,143 +141,120 @@ struct DynamicBuffer
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");
"wrong! X should contain 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_);
bool constexpr use_amd_buffer_addressing = true;
#else
if(is_valid_element)
{
#if CK_EXPERIMENTAL_USE_MEMCPY_FOR_VECTOR_ACCESS
X tmp = x;
bool constexpr use_amd_buffer_addressing = false;
#endif
__builtin_memcpy(&(p_data_[i]), &tmp, sizeof(X));
#if CK_WORKAROUND_SWDEV_XXXXXX_INT8_DS_WRITE_ISSUE
bool constexpr workaround_int8_ds_write_issue = true;
#else
*c_style_pointer_cast<X*>(&p_data_[i]) = x;
#endif
}
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>(
x, p_data_, i, is_valid_element, element_space_size_);
}
else if constexpr(GetAddressSpace() == AddressSpaceEnum_t::Lds)
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)
{
#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
// 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)
{
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);
}
// 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
else if constexpr(is_same<remove_cvref_t<T>, int8_t>::value &&
is_same<remove_cvref_t<X>, int8x2_t>::value)
{
#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
// 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);
}
#endif
}
}
else
......@@ -305,27 +278,49 @@ struct DynamicBuffer
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 need to be multiple T");
static_assert(GetAddressSpace() == AddressSpaceEnum_t::Global, "only support global mem");
"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);
#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);
#else
bool constexpr use_amd_buffer_addressing = false;
#endif
#if CK_USE_AMD_BUFFER_ATOMIC_ADD
constexpr index_t t_per_x = scalar_per_x_vector / scalar_per_t_vector;
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)
amd_buffer_atomic_add<remove_cvref_t<T>, t_per_x>(
x, p_data_, i, is_valid_element, element_space_size_);
}
else
{
atomicAdd(&p_data_[i], x);
if(is_valid_element)
{
// FIXME: atomicAdd is defined by HIP, need to avoid implicit type casting when
// calling it
atomicAdd(c_style_pointer_cast<X*>(&p_data_[i]), x);
}
}
#endif
}
__host__ __device__ static constexpr bool IsStaticBuffer() { return false; }
......@@ -333,14 +328,14 @@ struct DynamicBuffer
__host__ __device__ static constexpr bool IsDynamicBuffer() { return true; }
};
template <AddressSpaceEnum_t BufferAddressSpace, typename T, typename ElementSpaceSize>
template <AddressSpaceEnum 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,
AddressSpaceEnum BufferAddressSpace,
typename T,
typename ElementSpaceSize,
typename X,
......@@ -353,4 +348,3 @@ make_dynamic_buffer(T* p, ElementSpaceSize element_space_size, X invalid_element
}
} // namespace ck
#endif
include/ck/utility/utility.hpp → include/ck/utility/get_id.hpp
View file @ cd167e49
#ifndef CK_UTILITY_HPP
#define CK_UTILITY_HPP
#pragma once
#include "config.hpp"
namespace ck {
......@@ -16,5 +14,3 @@ __device__ index_t get_block_1d_id() { return blockIdx.x; }
__device__ index_t get_grid_size() { return gridDim.x; }
} // namespace ck
#endif
include/ck/utility/multi_index.hpp
View file @ cd167e49
......@@ -3,7 +3,7 @@
#include "common_header.hpp"
#if CK_USE_DYNAMICALLY_INDEXED_MULTI_INDEX
#if CK_EXPERIMENTAL_USE_DYNAMICALLY_INDEXED_MULTI_INDEX
#include "array_multi_index.hpp"
#else
#include "statically_indexed_array_multi_index.hpp"
......
include/ck/utility/reduction_enums.hpp
View file @ cd167e49
......@@ -28,7 +28,7 @@
namespace ck {
enum class ReduceTensorOp_t
enum struct ReduceTensorOp
{
ADD = 0,
MUL = 1,
......@@ -41,19 +41,19 @@ enum class ReduceTensorOp_t
// MUL_NO_ZEROS = 8,
};
enum class NanPropagation_t
enum struct NanPropagation
{
NOT_PROPAGATE_NAN = 0,
PROPAGATE_NAN = 1,
};
enum class ReduceTensorIndices_t
enum struct ReduceTensorIndices
{
NO_INDICES = 0,
FLATTENED_INDICES = 1,
};
enum class IndicesType_t
enum struct IndicesType
{
INDICES_32BIT = 0,
INDICES_64BIT = 1,
......
include/ck/utility/static_buffer.hpp
View file @ cd167e49
......@@ -6,7 +6,7 @@
namespace ck {
// static buffer for scalar
template <AddressSpaceEnum_t AddressSpace,
template <AddressSpaceEnum AddressSpace,
typename T,
index_t N,
bool InvalidElementUseNumericalZeroValue> // TODO remove this bool, no longer needed
......@@ -17,10 +17,7 @@ struct StaticBuffer : public StaticallyIndexedArray<T, N>
__host__ __device__ constexpr StaticBuffer() : base{} {}
__host__ __device__ static constexpr AddressSpaceEnum_t GetAddressSpace()
{
return AddressSpace;
}
__host__ __device__ static constexpr AddressSpaceEnum GetAddressSpace() { return AddressSpace; }
__host__ __device__ static constexpr bool IsStaticBuffer() { return true; }
......@@ -42,7 +39,7 @@ struct StaticBuffer : public StaticallyIndexedArray<T, N>
};
// static buffer for vector
template <AddressSpaceEnum_t AddressSpace,
template <AddressSpaceEnum AddressSpace,
typename S,
index_t NumOfVector,
index_t ScalarPerVector,
......@@ -59,10 +56,7 @@ struct StaticBufferTupleOfVector
__host__ __device__ constexpr StaticBufferTupleOfVector() : base{} {}
__host__ __device__ static constexpr AddressSpaceEnum_t GetAddressSpace()
{
return AddressSpace;
}
__host__ __device__ static constexpr AddressSpaceEnum GetAddressSpace() { return AddressSpace; }
__host__ __device__ static constexpr bool IsStaticBuffer() { return true; }
......@@ -158,7 +152,7 @@ struct StaticBufferTupleOfVector
}
};
template <AddressSpaceEnum_t AddressSpace, typename T, index_t N>
template <AddressSpaceEnum AddressSpace, typename T, index_t N>
__host__ __device__ constexpr auto make_static_buffer(Number<N>)
{
return StaticBuffer<AddressSpace, T, N, true>{};
......
include/ck/utility/synchronization.hpp
View file @ cd167e49
......@@ -7,7 +7,7 @@ namespace ck {
__device__ void block_sync_lds()
{
#if CK_BLOCK_SYNC_LDS_WITHOUT_SYNC_VMEM
#if CK_EXPERIMENTAL_BLOCK_SYNC_LDS_WITHOUT_SYNC_VMEM
asm volatile("\
s_waitcnt lgkmcnt(0) \n \
s_barrier \
......
library/include/ck/library/host_tensor/conv_common.hpp
View file @ cd167e49
......@@ -75,14 +75,14 @@ calculate_convolution_flops(const InDesc&, const WeiDesc& wei_desc, const OutDes
}
template <typename T>
inline auto activ(T v, const ck::ActivTypeEnum_t activ_type)
inline auto activ(T v, const ck::ActivTypeEnum activ_type)
{
const T alpha = 0.3;
switch(activ_type)
{
case ck::ActivTypeEnum_t::None: return v;
case ck::ActivTypeEnum_t::LeakyRelu: return (v >= 0 ? v : alpha * v);
case ck::ActivTypeEnum_t::Sigmoid: return (1 / (1 + exp(-v)));
case ck::ActivTypeEnum::None: return v;
case ck::ActivTypeEnum::LeakyRelu: return (v >= 0 ? v : alpha * v);
case ck::ActivTypeEnum::Sigmoid: return (1 / (1 + exp(-v)));
default: throw std::runtime_error("unsupported activ type"); break;
}
}
......
library/include/ck/library/host_tensor/device_tensor.hpp
View file @ cd167e49
#pragma once
#include "host_tensor.hpp"
#include "common_header.hpp"
template <typename TensorDesc>
void ostream_tensor_descriptor(TensorDesc, std::ostream& os = std::cout)
......
library/include/ck/library/host_tensor/host_reduce_util.hpp
View file @ cd167e49
......@@ -39,8 +39,8 @@ namespace ck {
namespace host_reduce {
using ck::NanPropagation_t;
using ck::ReduceTensorOp_t;
using ck::NanPropagation;
using ck::ReduceTensorOp;
template <typename T>
static inline bool float_equal_one(T);
......@@ -66,44 +66,44 @@ static inline bool float_equal_zero(half_float::half x)
return x == static_cast<half_float::half>(0.0f);
};
template <typename AccDataType, ReduceTensorOp_t ReduceOpId>
template <typename AccDataType, ReduceTensorOp ReduceOpId>
__host__ static inline std::function<void(AccDataType&)> PreUnaryOpFn(int)
{
using std::abs;
if constexpr(ReduceOpId == ReduceTensorOp_t::NORM1)
if constexpr(ReduceOpId == ReduceTensorOp::NORM1)
{
return ([&](AccDataType& a_) { a_ = abs(a_); });
}
else if constexpr(ReduceOpId == ReduceTensorOp_t::NORM2)
else if constexpr(ReduceOpId == ReduceTensorOp::NORM2)
{
return ([&](AccDataType& a_) { a_ = a_ * a_; });
}
else if constexpr(ReduceOpId == ReduceTensorOp_t::AMAX)
else if constexpr(ReduceOpId == ReduceTensorOp::AMAX)
{
return ([&](AccDataType& a_) { a_ = abs(a_); });
}
else
{
// ReduceTensorOp_t::AVG:
// ReduceTensorOp_t::ADD:
// ReduceTensorOp_t::MUL:
// ReduceTensorOp_t::MIN:
// ReduceTensorOp_t::MAX:
// ReduceTensorOp::AVG:
// ReduceTensorOp::ADD:
// ReduceTensorOp::MUL:
// ReduceTensorOp::MIN:
// ReduceTensorOp::MAX:
return ([&](AccDataType&) {});
};
};
template <typename AccDataType, ReduceTensorOp_t ReduceOpId>
template <typename AccDataType, ReduceTensorOp ReduceOpId>
__host__ static inline std::function<void(AccDataType&)> PosUnaryOpFn(int32_t divider)
{
using std::sqrt;
if constexpr(ReduceOpId == ReduceTensorOp_t::NORM2)
if constexpr(ReduceOpId == ReduceTensorOp::NORM2)
{
return ([&](AccDataType& a_) { a_ = sqrt(a_); });
}
else if constexpr(ReduceOpId == ReduceTensorOp_t::AVG)
else if constexpr(ReduceOpId == ReduceTensorOp::AVG)
{
return ([&, divider](AccDataType& a_) {
a_ = a_ / static_cast<AccDataType>(static_cast<float>(divider));
......@@ -111,36 +111,36 @@ __host__ static inline std::function<void(AccDataType&)> PosUnaryOpFn(int32_t di
}
else
{
// ReduceTensorOp_t::ADD:
// ReduceTensorOp_t::NORM1:
// ReduceTensorOp_t::MUL:
// ReduceTensorOp_t::MIN:
// ReduceTensorOp_t::MAX:
// ReduceTensorOp_t::AMAX:
// ReduceTensorOp::ADD:
// ReduceTensorOp::NORM1:
// ReduceTensorOp::MUL:
// ReduceTensorOp::MIN:
// ReduceTensorOp::MAX:
// ReduceTensorOp::AMAX:
return ([&](AccDataType&) {});
}
};
template <typename AccDataType, ReduceTensorOp_t ReduceOpId>
template <typename AccDataType, ReduceTensorOp ReduceOpId>
__host__ static inline std::function<void(AccDataType&, AccDataType)> ReduceOpFn()
{
if constexpr(ReduceOpId == ReduceTensorOp_t::ADD || ReduceOpId == ReduceTensorOp_t::AVG ||
ReduceOpId == ReduceTensorOp_t::NORM1 || ReduceOpId == ReduceTensorOp_t::NORM2)
if constexpr(ReduceOpId == ReduceTensorOp::ADD || ReduceOpId == ReduceTensorOp::AVG ||
ReduceOpId == ReduceTensorOp::NORM1 || ReduceOpId == ReduceTensorOp::NORM2)
{
return ([&](AccDataType& a_, AccDataType b_) { a_ = a_ + b_; });
}
else if constexpr(ReduceOpId == ReduceTensorOp_t::MUL)
else if constexpr(ReduceOpId == ReduceTensorOp::MUL)
{
return ([&](AccDataType& a_, AccDataType b_) { a_ = a_ * b_; });
}
else if constexpr(ReduceOpId == ReduceTensorOp_t::MIN)
else if constexpr(ReduceOpId == ReduceTensorOp::MIN)
{
return ([&](AccDataType& a_, AccDataType b_) {
if(a_ > b_)
a_ = b_;
});
}
else if constexpr(ReduceOpId == ReduceTensorOp_t::MAX || ReduceOpId == ReduceTensorOp_t::AMAX)
else if constexpr(ReduceOpId == ReduceTensorOp::MAX || ReduceOpId == ReduceTensorOp::AMAX)
{
return ([&](AccDataType& a_, AccDataType b_) {
if(a_ < b_)
......@@ -149,10 +149,10 @@ __host__ static inline std::function<void(AccDataType&, AccDataType)> ReduceOpFn
}
};
template <typename AccDataType, ReduceTensorOp_t ReduceOpId>
template <typename AccDataType, ReduceTensorOp ReduceOpId>
__host__ static inline std::function<void(AccDataType&, AccDataType, bool& changed)> ReduceOpFn2()
{
if constexpr(ReduceOpId == ReduceTensorOp_t::MIN)
if constexpr(ReduceOpId == ReduceTensorOp::MIN)
{
return ([&](AccDataType& a_, AccDataType b_, bool& changed) {
if(a_ > b_)
......@@ -164,7 +164,7 @@ __host__ static inline std::function<void(AccDataType&, AccDataType, bool& chang
changed = false;
});
}
else if constexpr(ReduceOpId == ReduceTensorOp_t::MAX || ReduceOpId == ReduceTensorOp_t::AMAX)
else if constexpr(ReduceOpId == ReduceTensorOp::MAX || ReduceOpId == ReduceTensorOp::AMAX)
{
return ([&](AccDataType& a_, AccDataType b_, bool& changed) {
if(a_ < b_)
......@@ -178,40 +178,40 @@ __host__ static inline std::function<void(AccDataType&, AccDataType, bool& chang
}
else
{
// ReduceTensorOp_t::ADD:
// ReduceTensorOp_t::MUL:
// ReduceTensorOp_t::AVG:
// ReduceTensorOp_t::NORM1:
// ReduceTensorOp_t::NORM2:
// ReduceTensorOp::ADD:
// ReduceTensorOp::MUL:
// ReduceTensorOp::AVG:
// ReduceTensorOp::NORM1:
// ReduceTensorOp::NORM2:
return (std::function<void(AccDataType&, AccDataType, bool&)>{});
};
};
template <typename AccDataType, ReduceTensorOp_t ReduceOpId>
template <typename AccDataType, ReduceTensorOp ReduceOpId>
__host__ static inline AccDataType ReduceOpZeroVal()
{
if constexpr(ReduceOpId == ReduceTensorOp_t::MUL)
if constexpr(ReduceOpId == ReduceTensorOp::MUL)
{
return (static_cast<AccDataType>(1.0f));
}
else if constexpr(ReduceOpId == ReduceTensorOp_t::MIN)
else if constexpr(ReduceOpId == ReduceTensorOp::MIN)
{
return (std::numeric_limits<AccDataType>::max());
}
else if constexpr(ReduceOpId == ReduceTensorOp_t::MAX)
else if constexpr(ReduceOpId == ReduceTensorOp::MAX)
{
return (std::numeric_limits<AccDataType>::lowest());
}
else if constexpr(ReduceOpId == ReduceTensorOp_t::AMAX)
else if constexpr(ReduceOpId == ReduceTensorOp::AMAX)
{
return (static_cast<AccDataType>(0.0f));
}
else
{
// ReduceTensorOp_t::ADD
// ReduceTensorOp_t::AVG
// ReduceTensorOp_t::NORM1
// ReduceTensorOp_t::NORM2
// ReduceTensorOp::ADD
// ReduceTensorOp::AVG
// ReduceTensorOp::NORM1
// ReduceTensorOp::NORM2
return (static_cast<AccDataType>(0.0f));
};
};
......
library/include/ck/library/host_tensor/host_reduction.hpp
View file @ cd167e49
......@@ -104,7 +104,7 @@ static size_t get_offset_from_index(const std::vector<size_t>& strides,
template <typename InDataType,
typename AccDataType,
typename OutDataType,
ck::ReduceTensorOp_t ReduceOpId,
ck::ReduceTensorOp ReduceOpId,
int Rank,
int NumReduceDim,
bool PropagateNan,
......
library/include/ck/library/obselete_driver_offline/device_convolution_add_forward_implicit_gemm_v5r1_dlops_nc0hwc1_kc0yxc1_nk0hwk1.hpp
View file @ cd167e49
......@@ -6,7 +6,7 @@
template <typename TInWei,
typename TAcc,
typename TOut,
ck::ActivTypeEnum_t activ_type,
ck::ActivTypeEnum activ_type,
typename InLengths,
typename WeiLengths,
typename AddLengths,
......
library/include/ck/library/obselete_driver_offline/device_convolution_backward_data_implicit_gemm_v4r1_xdlops_nhwc_kyxc_nhwk.hpp
View file @ cd167e49
......@@ -231,7 +231,7 @@ void device_convolution_backward_data_implicit_gemm_v4r1_xdlops_nhwc_kyxc_nhwk(
TInWei,
TAcc,
TOut,
InMemoryDataOperationEnum_t::Set,
InMemoryDataOperationEnum::Set,
decltype(wei_gemmk0_gemmm_gemmk1_grid_desc),
decltype(out_gemmk0_gemmn_gemmk1_grid_desc),
decltype(in_gemmm_gemmn_grid_desc),
......
library/include/ck/library/obselete_driver_offline/device_convolution_backward_data_implicit_gemm_v4r1r2_xdlops_nhwc_kyxc_nhwk.hpp
View file @ cd167e49
......@@ -338,7 +338,7 @@ void device_convolution_backward_data_implicit_gemm_v4r1r2_xdlops_nhwc_kyxc_nhwk
TInWei,
TAcc,
TOut,
InMemoryDataOperationEnum_t::Set,
InMemoryDataOperationEnum::Set,
decltype(out_gemmk0_gemmm_gemmk1_grid_desc),
decltype(wei_gemmk0_gemmn_gemmk1_grid_desc),
decltype(in_gemmm_gemmn_grid_desc),
......
library/include/ck/library/obselete_driver_offline/device_convolution_backward_data_implicit_gemm_v4r1r2_xdlops_nhwc_kyxc_nhwk_1x1.hpp
View file @ cd167e49
......@@ -307,7 +307,7 @@ void device_convolution_backward_data_implicit_gemm_v4r1r2_xdlops_nhwc_kyxc_nhwk
TInWei,
TAcc,
TOut,
InMemoryDataOperationEnum_t::Set,
InMemoryDataOperationEnum::Set,
decltype(out_gemmk0_gemmm_gemmk1_grid_desc),
decltype(wei_gemmk0_gemmn_gemmk1_grid_desc),
decltype(in_gemmm_gemmn_grid_desc),
......
library/include/ck/library/obselete_driver_offline/device_convolution_backward_weight_implicit_gemm_v4r4r2_xdlops_atomic_nchw_kcyx_nkhw.hpp
View file @ cd167e49
......@@ -171,7 +171,7 @@ void device_convolution_backward_weight_implicit_gemm_v4r4r2_xdlops_atomic_nchw_
TIn,
TAcc,
TWei,
InMemoryDataOperationEnum_t::AtomicAdd,
InMemoryDataOperationEnum::AtomicAdd,
decltype(out_gemmk0_gemmm_gemmk1_grid_desc),
decltype(in_gemmk0_gemmn_gemmk1_grid_desc),
decltype(wei_gemmm_gemmn_grid_desc),
......
library/include/ck/library/obselete_driver_offline/device_convolution_backward_weight_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nkhw.hpp
View file @ cd167e49
......@@ -168,7 +168,7 @@ void device_convolution_backward_weight_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nk
TIn,
TAcc,
TWei,
InMemoryDataOperationEnum_t::Set,
InMemoryDataOperationEnum::Set,
decltype(out_gemmk0_gemmm_gemmk1_grid_desc),
decltype(in_gemmk0_gemmn_gemmk1_grid_desc),
decltype(wei_gemmm_gemmn_grid_desc),
......
library/include/ck/library/obselete_driver_offline/device_convolution_backward_weight_implicit_gemm_v4r4r4_xdlops_atomic_nhwc_kyxc_nhwk.hpp
View file @ cd167e49
......@@ -200,7 +200,7 @@ void device_convolution_backward_weight_implicit_gemm_v4r4r4_xdlops_atomic_nhwc_
TIn,
TAcc,
TWei,
InMemoryDataOperationEnum_t::AtomicAdd,
InMemoryDataOperationEnum::AtomicAdd,
decltype(in_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc),
decltype(out_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc),
decltype(wei_gemmm_gemmn_grid_desc),
......
library/include/ck/library/obselete_driver_offline/device_convolution_backward_weight_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk.hpp
View file @ cd167e49
......@@ -199,7 +199,7 @@ void device_convolution_backward_weight_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nh
TIn,
TAcc,
TWei,
InMemoryDataOperationEnum_t::Set,
InMemoryDataOperationEnum::Set,
decltype(in_gemmk0_gemmm_gemmk1_grid_desc),
decltype(out_gemmk0_gemmn_gemmk1_grid_desc),
decltype(wei_gemmm_gemmn_grid_desc),
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment