Skip to content

GitLab

Commit 54e3ce2e authored by Paul's avatar Paul
Browse files

Format

parent 05605886
Show whitespace changes
Inline Side-by-side
Showing with 71 additions and 69 deletions
src/targets/gpu/kernels/include/migraphx/kernels/buffer_address.hpp
View file @ 54e3ce2e
......@@ -27,105 +27,104 @@ __device__ vec<int32_t, 4> make_wave_buffer_resource(T* p_wave, index_int bytes)
return result.content;
}
template<class T>
template <class T>
struct raw_buffer_load;
#define MIGRAPHX_BUFFER_ADDR_VISIT_TYPES(m) \
m(i8, int8_t) \
m(i16, int16_t) \
m(f16, half) \
m(f32, float)
#define MIGRAPHX_BUFFER_ADDR_VISIT_TYPES(m) m(i8, int8_t) m(i16, int16_t) m(f16, half) m(f32, float)
#define MIGRAPHX_RAW_BUFFER_LOAD(llvmtype, ...) \
__device__ __VA_ARGS__ \
llvm_amdgcn_raw_buffer_load_ ## llvmtype(vec<int32_t, 4> srsrc, \
__device__ __VA_ARGS__ llvm_amdgcn_raw_buffer_load_##llvmtype( \
vec<int32_t, 4> srsrc, \
int32_t voffset, \
int32_t soffset, \
int32_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load." #llvmtype); \
template<> \
struct raw_buffer_load<__VA_ARGS__> { \
template <> \
struct raw_buffer_load<__VA_ARGS__> \
{ \
static __device__ __VA_ARGS__ apply(vec<int32_t, 4> srsrc, \
int32_t voffset, \
int32_t soffset, \
int32_t glc_slc) { \
return llvm_amdgcn_raw_buffer_load_ ## llvmtype(srsrc, voffset, soffset, glc_slc); \
int32_t glc_slc) \
{ \
return llvm_amdgcn_raw_buffer_load_##llvmtype(srsrc, voffset, soffset, glc_slc); \
} \
};
#define MIGRAPHX_RAW_BUFFER_LOAD_VEC(llvmtype, cpptype) \
MIGRAPHX_RAW_BUFFER_LOAD(llvmtype, cpptype) \
MIGRAPHX_RAW_BUFFER_LOAD(v2 ## llvmtype, vec<cpptype, 2>) \
MIGRAPHX_RAW_BUFFER_LOAD(v4 ## llvmtype, vec<cpptype, 4>)
MIGRAPHX_RAW_BUFFER_LOAD(v2##llvmtype, vec<cpptype, 2>) \
MIGRAPHX_RAW_BUFFER_LOAD(v4##llvmtype, vec<cpptype, 4>)
MIGRAPHX_BUFFER_ADDR_VISIT_TYPES(MIGRAPHX_RAW_BUFFER_LOAD_VEC)
#define MIGRAPHX_RAW_BUFFER_STORE(llvmtype, ...) \
__device__ void \
llvm_amdgcn_raw_buffer_store_ ## llvmtype(__VA_ARGS__ vdata, vec<int32_t, 4> srsrc, \
__device__ void llvm_amdgcn_raw_buffer_store_##llvmtype( \
__VA_ARGS__ vdata, \
vec<int32_t, 4> srsrc, \
int32_t voffset, \
int32_t soffset, \
int32_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store." #llvmtype); \
__device__ void raw_buffer_store(__VA_ARGS__ vdata, vec<int32_t, 4> srsrc, \
__device__ void raw_buffer_store(__VA_ARGS__ vdata, \
vec<int32_t, 4> srsrc, \
int32_t voffset, \
int32_t soffset, \
int32_t glc_slc) { \
llvm_amdgcn_raw_buffer_store_ ## llvmtype(vdata, srsrc, voffset, soffset, glc_slc); \
int32_t glc_slc) \
{ \
llvm_amdgcn_raw_buffer_store_##llvmtype(vdata, srsrc, voffset, soffset, glc_slc); \
}
#define MIGRAPHX_RAW_BUFFER_STORE_VEC(llvmtype, cpptype) \
MIGRAPHX_RAW_BUFFER_STORE(llvmtype, cpptype) \
MIGRAPHX_RAW_BUFFER_STORE(v2 ## llvmtype, vec<cpptype, 2>) \
MIGRAPHX_RAW_BUFFER_STORE(v4 ## llvmtype, vec<cpptype, 4>)
MIGRAPHX_RAW_BUFFER_STORE(v2##llvmtype, vec<cpptype, 2>) \
MIGRAPHX_RAW_BUFFER_STORE(v4##llvmtype, vec<cpptype, 4>)
MIGRAPHX_BUFFER_ADDR_VISIT_TYPES(MIGRAPHX_RAW_BUFFER_STORE_VEC)
template<class T, index_int N>
template <class T, index_int N>
struct raw_buffer_load<vec<T, N>>
{
static __device__ vec<T, N> apply(vec<int32_t, 4> srsrc,
int32_t voffset,
int32_t soffset,
int32_t glc_slc)
static __device__ vec<T, N>
apply(vec<int32_t, 4> srsrc, int32_t voffset, int32_t soffset, int32_t glc_slc)
{
static_assert(N % 2 == 0, "Invalid vector size");
union type
{
vec<T, N> data;
vec<T, N/2> reg[2];
vec<T, N / 2> reg[2];
};
type result;
auto offset = sizeof(T) * (N/2);
result.reg[0] = raw_buffer_load<vec<T, N/2>>::apply(srsrc, voffset+offset, soffset, glc_slc);
result.reg[1] = raw_buffer_load<vec<T, N/2>>::apply(srsrc, voffset, soffset, glc_slc);
auto offset = sizeof(T) * (N / 2);
result.reg[0] =
raw_buffer_load<vec<T, N / 2>>::apply(srsrc, voffset + offset, soffset, glc_slc);
result.reg[1] = raw_buffer_load<vec<T, N / 2>>::apply(srsrc, voffset, soffset, glc_slc);
return result.data;
}
};
template<class T, index_int N>
__device__ void raw_buffer_store(vec<T, N> vdata, vec<int32_t, 4> srsrc,
int32_t voffset,
int32_t soffset,
int32_t glc_slc) {
template <class T, index_int N>
__device__ void raw_buffer_store(
vec<T, N> vdata, vec<int32_t, 4> srsrc, int32_t voffset, int32_t soffset, int32_t glc_slc)
{
union type
{
vec<T, N> data;
vec<T, N/2> reg[2];
vec<T, N / 2> reg[2];
};
type x;
x.data = vdata;
auto offset = sizeof(T) * (N/2);
auto offset = sizeof(T) * (N / 2);
raw_buffer_store(x.reg[0], srsrc, voffset, soffset, glc_slc);
raw_buffer_store(x.reg[1], srsrc, voffset+offset, soffset, glc_slc);
raw_buffer_store(x.reg[1], srsrc, voffset + offset, soffset, glc_slc);
}
template<class T>
template <class T>
__device__ T buffer_load(const T* p, index_int offset, index_int size, address_space::global)
{
auto resource = make_wave_buffer_resource(p, size * sizeof(T));
return raw_buffer_load<T>::apply(resource, offset * sizeof(T), 0, 0);
}
template<class T>
template <class T>
__device__ void buffer_store(T data, T* p, index_int offset, index_int size, address_space::global)
{
auto resource = make_wave_buffer_resource(p, size * sizeof(T));
......@@ -134,18 +133,17 @@ __device__ void buffer_store(T data, T* p, index_int offset, index_int size, add
#endif
template<class T, class AddressSpace>
template <class T, class AddressSpace>
__device__ T buffer_load(const T* p, index_int offset, index_int, AddressSpace)
{
return p[offset];
}
template<class T, class AddressSpace>
template <class T, class AddressSpace>
__device__ void buffer_store(T data, T* p, index_int offset, index_int, AddressSpace)
{
p[offset] = data;
}
} // namespace migraphx
#endif // MIGRAPHX_GUARD_KERNELS_BUFFER_ADDRESS_HPP
src/targets/gpu/kernels/include/migraphx/kernels/types.hpp
View file @ 54e3ce2e
......@@ -41,8 +41,12 @@ using half2 = migraphx::vec<half, 2>;
struct address_space
{
struct global {};
struct local {};
struct global
{
};
struct local
{
};
};
} // namespace migraphx
......
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