Skip to content

GitLab

Commit e371df51 authored by Chao Liu's avatar Chao Liu
Browse files

use buffer load OOB check for padding

parent 7a929377
Hide whitespace changes
Inline Side-by-side
Showing with 346 additions and 157 deletions
composable_kernel/include/tensor_operation/threadwise_generic_tensor_slice_copy.hpp
View file @ e371df51
......@@ -112,17 +112,18 @@ struct ThreadwiseGenericTensorSliceCopy_v4r2
// Check src data's valid mapping situation, only check the first data in this src
// vector. It's user's responsiblity to make sure all data in the src vector
// has the valid/invalid mapping situation
if(src_coord.IsOffsetValidAssumingUpperIndexIsValid())
{
transfer_data<SrcData,
SrcDataPerRead,
SrcAddressSpace,
AddressSpace::Vgpr,
InMemoryDataOperation::Set,
SrcDataStride,
1>(
p_src, src_coord.GetOffset(), p_src_long_vector, buffer_offset);
}
transfer_data<SrcData,
SrcDataPerRead,
SrcAddressSpace,
AddressSpace::Vgpr,
InMemoryDataOperation::Set,
SrcDataStride,
1>(p_src,
src_coord.GetOffset(),
src_coord.IsOffsetValidAssumingUpperIndexIsValid(),
p_src_long_vector,
buffer_offset,
true);
}
// SrcData to DstData conversion
......@@ -146,17 +147,18 @@ struct ThreadwiseGenericTensorSliceCopy_v4r2
// Check dst data's valid mapping situation, only check the first data in this dst
// vector. It's user's responsiblity to make sure all data in the dst vector
// has the valid/invalid mapping situation
if(dst_coord.IsOffsetValidAssumingUpperIndexIsValid())
{
transfer_data<DstData,
DstDataPerWrite,
AddressSpace::Vgpr,
DstAddressSpace,
DstInMemOp,
1,
DstDataStride>(
p_dst_long_vector, buffer_offset, p_dst, dst_coord.GetOffset());
}
transfer_data<DstData,
DstDataPerWrite,
AddressSpace::Vgpr,
DstAddressSpace,
DstInMemOp,
1,
DstDataStride>(p_dst_long_vector,
buffer_offset,
true,
p_dst,
dst_coord.GetOffset(),
dst_coord.IsOffsetValidAssumingUpperIndexIsValid());
}
});
}
......@@ -266,18 +268,17 @@ struct ThreadwiseGenericTensorSliceCopy_v4r2
// src
// vector. It's user's responsiblity to make sure all data in the src vector
// has the valid/invalid mapping situation
if(src_coord.IsOffsetValidAssumingUpperIndexIsValid())
{
transfer_data<SrcData,
SrcDataPerRead,
SrcAddressSpace,
AddressSpace::Vgpr,
InMemoryDataOperation::Set>(p_src,
src_nonlinear_coord.GetOffset() +
src_linear_offset,
p_src_long_vector,
buffer_offset);
}
transfer_data<SrcData,
SrcDataPerRead,
SrcAddressSpace,
AddressSpace::Vgpr,
InMemoryDataOperation::Set>(
p_src,
src_nonlinear_coord.GetOffset() + src_linear_offset,
src_coord.IsOffsetValidAssumingUpperIndexIsValid(),
p_src_long_vector,
buffer_offset,
true);
}
// SrcData to DstData conversion
......@@ -305,15 +306,16 @@ struct ThreadwiseGenericTensorSliceCopy_v4r2
// dst
// vector. It's user's responsiblity to make sure all data in the dst vector
// has the valid/invalid mapping situation
if(dst_coord.IsOffsetValidAssumingUpperIndexIsValid())
{
transfer_data<DstData,
DstDataPerWrite,
AddressSpace::Vgpr,
DstAddressSpace,
DstInMemOp>(
p_dst_long_vector, buffer_offset, p_dst, dst_coord.GetOffset());
}
transfer_data<DstData,
DstDataPerWrite,
AddressSpace::Vgpr,
DstAddressSpace,
DstInMemOp>(p_dst_long_vector,
buffer_offset,
true,
p_dst,
dst_coord.GetOffset(),
dst_coord.IsOffsetValidAssumingUpperIndexIsValid());
}
});
});
......@@ -405,15 +407,17 @@ struct ThreadwiseGenericTensorSliceCopy_v4r2
// src
// vector. It's user's responsiblity to make sure all data in the src vector
// has the valid/invalid mapping situation
if(src_coord.IsOffsetValidAssumingUpperIndexIsValid())
{
transfer_data<SrcData,
SrcDataPerRead,
SrcAddressSpace,
AddressSpace::Vgpr,
InMemoryDataOperation::Set>(
p_src, src_coord.GetOffset(), p_src_long_vector, buffer_offset);
}
transfer_data<SrcData,
SrcDataPerRead,
SrcAddressSpace,
AddressSpace::Vgpr,
InMemoryDataOperation::Set>(
p_src,
src_coord.GetOffset(),
src_coord.IsOffsetValidAssumingUpperIndexIsValid(),
p_src_long_vector,
buffer_offset,
true);
}
// SrcData to DstData conversion
......@@ -450,18 +454,16 @@ struct ThreadwiseGenericTensorSliceCopy_v4r2
// dst
// vector. It's user's responsiblity to make sure all data in the dst vector
// has the valid/invalid mapping situation
if(dst_coord.IsOffsetValidAssumingUpperIndexIsValid())
{
transfer_data<DstData,
DstDataPerWrite,
AddressSpace::Vgpr,
DstAddressSpace,
DstInMemOp>(p_dst_long_vector,
buffer_offset,
p_dst,
dst_nonlinear_coord.GetOffset() +
dst_linear_offset);
}
transfer_data<DstData,
DstDataPerWrite,
AddressSpace::Vgpr,
DstAddressSpace,
DstInMemOp>(p_dst_long_vector,
buffer_offset,
true,
p_dst,
dst_nonlinear_coord.GetOffset() + dst_linear_offset,
dst_coord.IsOffsetValidAssumingUpperIndexIsValid());
}
});
});
......
composable_kernel/include/utility/amd_buffer_addressing.hpp
View file @ e371df51
......@@ -150,8 +150,11 @@ __llvm_amdgcn_buffer_atomic_add_f32(float vdata,
// 2) p_src to be a block-invariant pointer.
// It is user's responsibility to make sure that is true.
template <typename T, index_t VectorSize>
__device__ typename vector_type<T, VectorSize>::MemoryType amd_buffer_load(
const T* p_src_block, index_t src_thread_data_offset, index_t src_const_data_offset);
__device__ typename vector_type<T, VectorSize>::MemoryType
amd_buffer_load(const T* p_src_block,
index_t src_thread_data_offset,
index_t src_const_data_offset,
bool src_valid);
// buffer_store requires:
// 1) p_src must be in vgpr space, d_dst must be global memory
......@@ -161,18 +164,21 @@ template <typename T, index_t VectorSize>
__device__ void amd_buffer_store(const T* p_src,
T* p_dst_block,
index_t dst_thread_data_offset,
index_t dst_const_data_offset);
index_t dst_const_data_offset,
bool dst_valid);
template <typename T, index_t VectorSize>
__device__ void amd_buffer_atomic_add(const T* p_src,
T* p_dst_block,
index_t dst_thread_data_offset,
index_t dst_const_data_offset);
index_t dst_const_data_offset,
bool dst_valid);
template <>
__device__ float amd_buffer_load<float, 1>(const float* p_src_block,
index_t src_thread_data_offset,
index_t src_const_data_offset)
index_t src_const_data_offset,
bool src_valid)
{
BufferAddressConfig<float> src_block_config;
......@@ -187,13 +193,18 @@ __device__ float amd_buffer_load<float, 1>(const float* p_src_block,
index_t src_const_addr_offset = src_const_data_offset * sizeof(float);
return __llvm_amdgcn_buffer_load_f32(
src_block_config.data, 0, src_thread_addr_offset + src_const_addr_offset, false, false);
src_block_config.data,
0,
src_valid ? (src_thread_addr_offset + src_const_addr_offset) : -1,
false,
false);
}
template <>
__device__ float2_t amd_buffer_load<float, 2>(const float* p_src_block,
index_t src_thread_data_offset,
index_t src_const_data_offset)
index_t src_const_data_offset,
bool src_valid)
{
BufferAddressConfig<float> src_block_config;
......@@ -208,13 +219,18 @@ __device__ float2_t amd_buffer_load<float, 2>(const float* p_src_block,
index_t src_const_addr_offset = src_const_data_offset * sizeof(float);
return __llvm_amdgcn_buffer_load_f32x2(
src_block_config.data, 0, src_thread_addr_offset + src_const_addr_offset, false, false);
src_block_config.data,
0,
src_valid ? (src_thread_addr_offset + src_const_addr_offset) : -1,
false,
false);
}
template <>
__device__ float4_t amd_buffer_load<float, 4>(const float* p_src_block,
index_t src_thread_data_offset,
index_t src_const_data_offset)
index_t src_const_data_offset,
bool src_valid)
{
BufferAddressConfig<float> src_block_config;
......@@ -229,13 +245,18 @@ __device__ float4_t amd_buffer_load<float, 4>(const float* p_src_block,
index_t src_const_addr_offset = src_const_data_offset * sizeof(float);
return __llvm_amdgcn_buffer_load_f32x4(
src_block_config.data, 0, src_thread_addr_offset + src_const_addr_offset, false, false);
src_block_config.data,
0,
src_valid ? (src_thread_addr_offset + src_const_addr_offset) : -1,
false,
false);
}
template <>
__device__ half_t amd_buffer_load<half_t, 1>(const half_t* p_src_block,
index_t src_thread_data_offset,
index_t src_const_data_offset)
index_t src_const_data_offset,
bool src_valid)
{
BufferAddressConfig<half_t> src_block_config;
......@@ -251,16 +272,21 @@ __device__ half_t amd_buffer_load<half_t, 1>(const half_t* p_src_block,
index_t src_const_addr_offset = src_const_data_offset * sizeof(half_t);
return __llvm_amdgcn_buffer_load_f16(
src_block_config.data, 0, src_thread_addr_offset + src_const_addr_offset, false, false);
src_block_config.data,
0,
src_valid ? (src_thread_addr_offset + src_const_addr_offset) : -1,
false,
false);
#else
return p_src_block[src_thread_data_offset + src_const_data_offset];
return src_valid ? p_src_block[src_thread_data_offset + src_const_data_offset] : 0;
#endif
}
template <>
__device__ half2_t amd_buffer_load<half_t, 2>(const half_t* p_src_block,
index_t src_thread_data_offset,
index_t src_const_data_offset)
index_t src_const_data_offset,
bool src_valid)
{
BufferAddressConfig<half_t> src_block_config;
......@@ -276,10 +302,18 @@ __device__ half2_t amd_buffer_load<half_t, 2>(const half_t* p_src_block,
#if !CK_WORKAROUND_SWDEV_231101
return __llvm_amdgcn_buffer_load_f16x2(
src_block_config.data, 0, src_thread_addr_offset + src_const_addr_offset, false, false);
src_block_config.data,
0,
src_valid ? (src_thread_addr_offset + src_const_addr_offset) : -1,
false,
false);
#else
float dst_out_tmp = __llvm_amdgcn_buffer_load_f32(
src_block_config.data, 0, src_thread_addr_offset + src_const_addr_offset, false, false);
src_block_config.data,
0,
src_valid ? (src_thread_addr_offset + src_const_addr_offset) : -1,
false,
false);
return *reinterpret_cast<half2_t*>(&dst_out_tmp);
#endif
......@@ -288,7 +322,8 @@ __device__ half2_t amd_buffer_load<half_t, 2>(const half_t* p_src_block,
template <>
__device__ half4_t amd_buffer_load<half_t, 4>(const half_t* p_src_block,
index_t src_thread_data_offset,
index_t src_const_data_offset)
index_t src_const_data_offset,
bool src_valid)
{
BufferAddressConfig<half_t> src_block_config;
......@@ -304,10 +339,18 @@ __device__ half4_t amd_buffer_load<half_t, 4>(const half_t* p_src_block,
#if !CK_WORKAROUND_SWDEV_231101
return __llvm_amdgcn_buffer_load_f16x4(
src_block_config.data, 0, src_thread_addr_offset + src_const_addr_offset, false, false);
src_block_config.data,
0,
src_valid ? (src_thread_addr_offset + src_const_addr_offset) : -1,
false,
false);
#else
float2_t dst_out_tmp = __llvm_amdgcn_buffer_load_f32x2(
src_block_config.data, 0, src_thread_addr_offset + src_const_addr_offset, false, false);
src_block_config.data,
0,
src_valid ? (src_thread_addr_offset + src_const_addr_offset) : -1,
false,
false);
return *reinterpret_cast<half4_t*>(&dst_out_tmp);
#endif
......@@ -316,7 +359,8 @@ __device__ half4_t amd_buffer_load<half_t, 4>(const half_t* p_src_block,
template <>
__device__ half8_t amd_buffer_load<half_t, 8>(const half_t* p_src_block,
index_t src_thread_data_offset,
index_t src_const_data_offset)
index_t src_const_data_offset,
bool src_valid)
{
BufferAddressConfig<half_t> src_block_config;
......@@ -330,20 +374,21 @@ __device__ half8_t amd_buffer_load<half_t, 8>(const half_t* p_src_block,
index_t src_thread_addr_offset = src_thread_data_offset * sizeof(half_t);
index_t src_const_addr_offset = src_const_data_offset * sizeof(half_t);
#if !CK_WORKAROUND_SWDEV_231101
static_assert(false, "wrong! not supported");
#else
float4_t dst_out_tmp = __llvm_amdgcn_buffer_load_f32x4(
src_block_config.data, 0, src_thread_addr_offset + src_const_addr_offset, false, false);
src_block_config.data,
0,
src_valid ? (src_thread_addr_offset + src_const_addr_offset) : -1,
false,
false);
return *reinterpret_cast<half8_t*>(&dst_out_tmp);
#endif
}
template <>
__device__ ushort amd_buffer_load<ushort, 1>(const ushort* p_src_block,
index_t src_thread_data_offset,
index_t src_const_data_offset)
index_t src_const_data_offset,
bool src_valid)
{
BufferAddressConfig<ushort> src_block_config;
......@@ -359,16 +404,21 @@ __device__ ushort amd_buffer_load<ushort, 1>(const ushort* p_src_block,
index_t src_const_addr_offset = src_const_data_offset * sizeof(ushort);
return __llvm_amdgcn_buffer_load_bf16(
src_block_config.data, 0, src_thread_addr_offset + src_const_addr_offset, false, false);
src_block_config.data,
0,
src_valid ? (src_thread_addr_offset + src_const_addr_offset) : -1,
false,
false);
#else
return p_src_block[src_thread_data_offset + src_const_data_offset];
return src_valid ? p_src_block[src_thread_data_offset + src_const_data_offset] : 0;
#endif
}
template <>
__device__ ushort2_t amd_buffer_load<ushort, 2>(const ushort* p_src_block,
index_t src_thread_data_offset,
index_t src_const_data_offset)
index_t src_const_data_offset,
bool src_valid)
{
BufferAddressConfig<ushort> src_block_config;
......@@ -384,10 +434,18 @@ __device__ ushort2_t amd_buffer_load<ushort, 2>(const ushort* p_src_block,
#if !CK_WORKAROUND_SWDEV_231101
return __llvm_amdgcn_buffer_load_bf16x2(
src_block_config.data, 0, src_thread_addr_offset + src_const_addr_offset, false, false);
src_block_config.data,
0,
src_valid ? (src_thread_addr_offset + src_const_addr_offset) : -1,
false,
false);
#else
float dst_out_tmp = __llvm_amdgcn_buffer_load_f32(
src_block_config.data, 0, src_thread_addr_offset + src_const_addr_offset, false, false);
src_block_config.data,
0,
src_valid ? (src_thread_addr_offset + src_const_addr_offset) : -1,
false,
false);
return *reinterpret_cast<ushort2_t*>(&dst_out_tmp);
#endif
......@@ -396,7 +454,8 @@ __device__ ushort2_t amd_buffer_load<ushort, 2>(const ushort* p_src_block,
template <>
__device__ ushort4_t amd_buffer_load<ushort, 4>(const ushort* p_src_block,
index_t src_thread_data_offset,
index_t src_const_data_offset)
index_t src_const_data_offset,
bool src_valid)
{
BufferAddressConfig<ushort> src_block_config;
......@@ -412,10 +471,18 @@ __device__ ushort4_t amd_buffer_load<ushort, 4>(const ushort* p_src_block,
#if !CK_WORKAROUND_SWDEV_231101
return __llvm_amdgcn_buffer_load_bf16x4(
src_block_config.data, 0, src_thread_addr_offset + src_const_addr_offset, false, false);
src_block_config.data,
0,
src_valid ? (src_thread_addr_offset + src_const_addr_offset) : -1,
false,
false);
#else
float2_t dst_out_tmp = __llvm_amdgcn_buffer_load_f32x2(
src_block_config.data, 0, src_thread_addr_offset + src_const_addr_offset, false, false);
src_block_config.data,
0,
src_valid ? (src_thread_addr_offset + src_const_addr_offset) : -1,
false,
false);
return *reinterpret_cast<ushort4_t*>(&dst_out_tmp);
#endif
......@@ -424,7 +491,8 @@ __device__ ushort4_t amd_buffer_load<ushort, 4>(const ushort* p_src_block,
template <>
__device__ ushort8_t amd_buffer_load<ushort, 8>(const ushort* p_src_block,
index_t src_thread_data_offset,
index_t src_const_data_offset)
index_t src_const_data_offset,
bool src_valid)
{
BufferAddressConfig<ushort> src_block_config;
......@@ -438,21 +506,22 @@ __device__ ushort8_t amd_buffer_load<ushort, 8>(const ushort* p_src_block,
index_t src_thread_addr_offset = src_thread_data_offset * sizeof(ushort);
index_t src_const_addr_offset = src_const_data_offset * sizeof(ushort);
#if !CK_WORKAROUND_SWDEV_231101
static_assert(false, "wrong! not implemented");
#else
float4_t dst_out_tmp = __llvm_amdgcn_buffer_load_f32x4(
src_block_config.data, 0, src_thread_addr_offset + src_const_addr_offset, false, false);
src_block_config.data,
0,
src_valid ? (src_thread_addr_offset + src_const_addr_offset) : -1,
false,
false);
return *reinterpret_cast<ushort8_t*>(&dst_out_tmp);
#endif
}
template <>
__device__ void amd_buffer_store<float, 1>(const float* p_src,
float* p_dst_block,
index_t dst_thread_data_offset,
index_t dst_const_data_offset)
index_t dst_const_data_offset,
bool dst_valid)
{
BufferAddressConfig<float> dst_block_config;
......@@ -469,7 +538,8 @@ __device__ void amd_buffer_store<float, 1>(const float* p_src,
__llvm_amdgcn_buffer_store_f32(*p_src,
dst_block_config.data,
0,
dst_thread_addr_offset + dst_const_addr_offset,
dst_valid ? (dst_thread_addr_offset + dst_const_addr_offset)
: -1,
false,
false);
}
......@@ -478,7 +548,8 @@ template <>
__device__ void amd_buffer_store<float, 2>(const float* p_src,
float* p_dst_block,
index_t dst_thread_data_offset,
index_t dst_const_data_offset)
index_t dst_const_data_offset,
bool dst_valid)
{
BufferAddressConfig<float> dst_block_config;
......@@ -495,7 +566,8 @@ __device__ void amd_buffer_store<float, 2>(const float* p_src,
__llvm_amdgcn_buffer_store_f32x2(*reinterpret_cast<const float2_t*>(p_src),
dst_block_config.data,
0,
dst_thread_addr_offset + dst_const_addr_offset,
dst_valid ? (dst_thread_addr_offset + dst_const_addr_offset)
: -1,
false,
false);
}
......@@ -504,7 +576,8 @@ template <>
__device__ void amd_buffer_store<float, 4>(const float* p_src,
float* p_dst_block,
index_t dst_thread_data_offset,
index_t dst_const_data_offset)
index_t dst_const_data_offset,
bool dst_valid)
{
BufferAddressConfig<float> dst_block_config;
......@@ -521,7 +594,8 @@ __device__ void amd_buffer_store<float, 4>(const float* p_src,
__llvm_amdgcn_buffer_store_f32x4(*reinterpret_cast<const float4_t*>(p_src),
dst_block_config.data,
0,
dst_thread_addr_offset + dst_const_addr_offset,
dst_valid ? (dst_thread_addr_offset + dst_const_addr_offset)
: -1,
false,
false);
}
......@@ -530,7 +604,8 @@ template <>
__device__ void amd_buffer_store<half_t, 1>(const half_t* p_src,
half_t* p_dst_block,
index_t dst_thread_data_offset,
index_t dst_const_data_offset)
index_t dst_const_data_offset,
bool dst_valid)
{
BufferAddressConfig<half_t> dst_block_config;
......@@ -548,11 +623,15 @@ __device__ void amd_buffer_store<half_t, 1>(const half_t* p_src,
__llvm_amdgcn_buffer_store_f16(*p_src,
dst_block_config.data,
0,
dst_thread_addr_offset + dst_const_addr_offset,
dst_valid ? (dst_thread_addr_offset + dst_const_addr_offset)
: -1,
false,
false);
#else
p_dst_block[dst_thread_data_offset + dst_const_data_offset] = *p_src;
if(dst_valid)
{
p_dst_block[dst_thread_data_offset + dst_const_data_offset] = *p_src;
}
#endif
}
......@@ -560,7 +639,8 @@ template <>
__device__ void amd_buffer_store<half_t, 2>(const half_t* p_src,
half_t* p_dst_block,
index_t dst_thread_data_offset,
index_t dst_const_data_offset)
index_t dst_const_data_offset,
bool dst_valid)
{
BufferAddressConfig<half_t> dst_block_config;
......@@ -578,7 +658,8 @@ __device__ void amd_buffer_store<half_t, 2>(const half_t* p_src,
__llvm_amdgcn_buffer_store_f16x2(*reinterpret_cast<const half2_t*>(p_src),
dst_block_config.data,
0,
dst_thread_addr_offset + dst_const_addr_offset,
dst_valid ? (dst_thread_addr_offset + dst_const_addr_offset)
: -1,
false,
false);
#else
......@@ -587,7 +668,8 @@ __device__ void amd_buffer_store<half_t, 2>(const half_t* p_src,
__llvm_amdgcn_buffer_store_f32(*p_src_tmp,
dst_block_config.data,
0,
dst_thread_addr_offset + dst_const_addr_offset,
dst_valid ? (dst_thread_addr_offset + dst_const_addr_offset)
: -1,
false,
false);
#endif
......@@ -597,7 +679,8 @@ template <>
__device__ void amd_buffer_store<half_t, 4>(const half_t* p_src,
half_t* p_dst_block,
index_t dst_thread_data_offset,
index_t dst_const_data_offset)
index_t dst_const_data_offset,
bool dst_valid)
{
index_t dst_thread_addr_offset = dst_thread_data_offset * sizeof(half_t);
index_t dst_const_addr_offset = dst_const_data_offset * sizeof(half_t);
......@@ -615,7 +698,8 @@ __device__ void amd_buffer_store<half_t, 4>(const half_t* p_src,
__llvm_amdgcn_buffer_store_f16x4(*reinterpret_cast<const half4_t*>(p_src),
dst_block_config.data,
0,
dst_thread_addr_offset + dst_const_addr_offset,
dst_valid ? (dst_thread_addr_offset + dst_const_addr_offset)
: -1,
false,
false);
#else
......@@ -624,7 +708,8 @@ __device__ void amd_buffer_store<half_t, 4>(const half_t* p_src,
__llvm_amdgcn_buffer_store_f32x2(*p_src_tmp,
dst_block_config.data,
0,
dst_thread_addr_offset + dst_const_addr_offset,
dst_valid ? (dst_thread_addr_offset + dst_const_addr_offset)
: -1,
false,
false);
#endif
......@@ -634,7 +719,8 @@ template <>
__device__ void amd_buffer_store<ushort, 1>(const ushort* p_src,
ushort* p_dst_block,
index_t dst_thread_data_offset,
index_t dst_const_data_offset)
index_t dst_const_data_offset,
bool dst_valid)
{
BufferAddressConfig<ushort> dst_block_config;
......@@ -652,11 +738,15 @@ __device__ void amd_buffer_store<ushort, 1>(const ushort* p_src,
__llvm_amdgcn_buffer_store_bf16(*p_src,
dst_block_config.data,
0,
dst_thread_addr_offset + dst_const_addr_offset,
dst_valid ? (dst_thread_addr_offset + dst_const_addr_offset)
: -1,
false,
false);
#else
p_dst_block[dst_thread_data_offset + dst_const_data_offset] = *p_src;
if(dst_valid)
{
p_dst_block[dst_thread_data_offset + dst_const_data_offset] = *p_src;
}
#endif
}
......@@ -664,7 +754,8 @@ template <>
__device__ void amd_buffer_store<ushort, 2>(const ushort* p_src,
ushort* p_dst_block,
index_t dst_thread_data_offset,
index_t dst_const_data_offset)
index_t dst_const_data_offset,
bool dst_valid)
{
BufferAddressConfig<ushort> dst_block_config;
......@@ -682,7 +773,8 @@ __device__ void amd_buffer_store<ushort, 2>(const ushort* p_src,
__llvm_amdgcn_buffer_store_bf16x2(*p_src,
dst_block_config.data,
0,
dst_thread_addr_offset + dst_const_addr_offset,
dst_valid ? (dst_thread_addr_offset + dst_const_addr_offset)
: -1,
false,
false);
#else
......@@ -691,7 +783,8 @@ __device__ void amd_buffer_store<ushort, 2>(const ushort* p_src,
__llvm_amdgcn_buffer_store_f32(*p_src_tmp,
dst_block_config.data,
0,
dst_thread_addr_offset + dst_const_addr_offset,
dst_valid ? (dst_thread_addr_offset + dst_const_addr_offset)
: -1,
false,
false);
#endif
......@@ -701,7 +794,8 @@ template <>
__device__ void amd_buffer_store<ushort, 4>(const ushort* p_src,
ushort* p_dst_block,
index_t dst_thread_data_offset,
index_t dst_const_data_offset)
index_t dst_const_data_offset,
bool dst_valid)
{
BufferAddressConfig<ushort> dst_block_config;
......@@ -719,7 +813,8 @@ __device__ void amd_buffer_store<ushort, 4>(const ushort* p_src,
__llvm_amdgcn_buffer_store_bf16x4(*p_src,
dst_block_config.data,
0,
dst_thread_addr_offset + dst_const_addr_offset,
dst_valid ? (dst_thread_addr_offset + dst_const_addr_offset)
: -1,
false,
false);
#else
......@@ -728,7 +823,8 @@ __device__ void amd_buffer_store<ushort, 4>(const ushort* p_src,
__llvm_amdgcn_buffer_store_f32x2(*p_src_tmp,
dst_block_config.data,
0,
dst_thread_addr_offset + dst_const_addr_offset,
dst_valid ? (dst_thread_addr_offset + dst_const_addr_offset)
: -1,
false,
false);
#endif
......@@ -738,7 +834,8 @@ template <>
__device__ void amd_buffer_atomic_add<float, 1>(const float* p_src,
float* p_dst_block,
index_t dst_thread_data_offset,
index_t dst_const_data_offset)
index_t dst_const_data_offset,
bool dst_valid)
{
BufferAddressConfig<float> dst_block_config;
......@@ -752,20 +849,41 @@ __device__ void amd_buffer_atomic_add<float, 1>(const float* p_src,
index_t dst_thread_addr_offset = dst_thread_data_offset * sizeof(float);
index_t dst_const_addr_offset = dst_const_data_offset * sizeof(float);
__llvm_amdgcn_buffer_atomic_add_f32(
*p_src, dst_block_config.data, 0, dst_thread_addr_offset + dst_const_addr_offset, false);
__llvm_amdgcn_buffer_atomic_add_f32(*p_src,
dst_block_config.data,
0,
dst_valid ? (dst_thread_addr_offset + dst_const_addr_offset)
: -1,
false);
}
template <>
__device__ void amd_buffer_atomic_add<float, 2>(const float* p_src,
float* p_dst_block,
index_t dst_thread_data_offset,
index_t dst_const_data_offset)
index_t dst_const_data_offset,
bool dst_valid)
{
BufferAddressConfig<float> dst_block_config;
// fill in byte 0 - 1
dst_block_config.address[0] = p_dst_block;
// fill in byte 2
dst_block_config.range[2] = -1;
// fill in byte 3
dst_block_config.range[3] = 0x00027000;
index_t dst_thread_addr_offset = dst_thread_data_offset * sizeof(float);
index_t dst_const_addr_offset = dst_const_data_offset * sizeof(float);
for(index_t i = 0; i < 2; ++i)
{
amd_buffer_atomic_add<float, 1>(
&p_src[i], p_dst_block, dst_thread_data_offset, dst_const_data_offset + i);
__llvm_amdgcn_buffer_atomic_add_f32(
p_src[i],
dst_block_config.data,
0,
dst_valid ? (dst_thread_addr_offset + dst_const_addr_offset + i * sizeof(float)) : -1,
false);
}
}
......@@ -773,12 +891,29 @@ template <>
__device__ void amd_buffer_atomic_add<float, 4>(const float* p_src,
float* p_dst_block,
index_t dst_thread_data_offset,
index_t dst_const_data_offset)
index_t dst_const_data_offset,
bool dst_valid)
{
BufferAddressConfig<float> dst_block_config;
// fill in byte 0 - 1
dst_block_config.address[0] = p_dst_block;
// fill in byte 2
dst_block_config.range[2] = -1;
// fill in byte 3
dst_block_config.range[3] = 0x00027000;
index_t dst_thread_addr_offset = dst_thread_data_offset * sizeof(float);
index_t dst_const_addr_offset = dst_const_data_offset * sizeof(float);
for(index_t i = 0; i < 4; ++i)
{
amd_buffer_atomic_add<float, 1>(
&p_src[i], p_dst_block, dst_thread_data_offset, dst_const_data_offset + i);
__llvm_amdgcn_buffer_atomic_add_f32(
p_src[i],
dst_block_config.data,
0,
dst_valid ? (dst_thread_addr_offset + dst_const_addr_offset + i * sizeof(float)) : -1,
false);
}
}
......
composable_kernel/include/utility/in_memory_operation.amd.hpp.in
View file @ e371df51
......@@ -47,10 +47,25 @@ struct SetData
// This version is only for compatibility, don't use this version if possible
template <AddressSpace SrcAddressSpace, AddressSpace DstAddressSpace>
__device__ void Run(const T* p_src, index_t src_offset, T* p_dst, index_t dst_offset) const
__device__ void Run(const T* p_src,
index_t src_offset,
bool src_valid,
T* p_dst,
index_t dst_offset,
bool dst_valid) const
{
*reinterpret_cast<vector_t*>(&p_dst[dst_offset]) =
*reinterpret_cast<const vector_t*>(&p_src[src_offset]);
if(dst_valid)
{
if(src_valid)
{
*reinterpret_cast<vector_t*>(&p_dst[dst_offset]) =
*reinterpret_cast<const vector_t*>(&p_src[src_offset]);
}
else
{
*reinterpret_cast<vector_t*>(&p_dst[dst_offset]) = 0;
}
}
}
#if CK_USE_AMD_BUFFER_ADDRESSING
......@@ -61,11 +76,16 @@ struct SetData
template <>
__device__ void Run<AddressSpace::Global, AddressSpace::Vgpr>(const T* p_src,
index_t src_offset,
bool src_valid,
T* p_dst,
index_t dst_offset) const
index_t dst_offset,
bool dst_valid) const
{
*reinterpret_cast<vector_t*>(&p_dst[dst_offset]) =
amd_buffer_load<T, DataPerAccess>(p_src, src_offset, 0);
if(dst_valid)
{
*reinterpret_cast<vector_t*>(&p_dst[dst_offset]) =
amd_buffer_load<T, DataPerAccess>(p_src, src_offset, 0, src_valid);
}
}
// buffer_store requires:
......@@ -75,10 +95,15 @@ struct SetData
template <>
__device__ void Run<AddressSpace::Vgpr, AddressSpace::Global>(const T* p_src,
index_t src_offset,
bool src_valid,
T* p_dst,
index_t dst_offset) const
index_t dst_offset,
bool dst_valid) const
{
amd_buffer_store<T, DataPerAccess>(&(p_src[src_offset]), p_dst, dst_offset, 0);
const auto zeros = vector_t(0);
amd_buffer_store<T, DataPerAccess>(
src_valid ? &(p_src[src_offset]) : &zeros, p_dst, dst_offset, 0, dst_valid);
}
#endif
};
......@@ -90,10 +115,18 @@ struct AtomicAddData
// This version is only for compatibility, don't use this version if possible
template <AddressSpace SrcAddressSpace, AddressSpace DstAddressSpace>
__device__ void Run(const T* p_src, index_t src_offset, T* p_dst, index_t dst_offset) const
__device__ void Run(const T* p_src,
index_t src_offset,
bool src_valid,
T* p_dst,
index_t dst_offset,
bool dst_valid) const
{
atomic_add_impl(reinterpret_cast<vector_t*>(&p_dst[dst_offset]),
*reinterpret_cast<const vector_t*>(&p_src[src_offset]));
if(src_valid && dst_valid)
{
atomic_add_impl(reinterpret_cast<vector_t*>(&p_dst[dst_offset]),
*reinterpret_cast<const vector_t*>(&p_src[src_offset]));
}
}
#if CK_USE_AMD_BUFFER_ADDRESSING && CK_USE_AMD_BUFFER_ATOMIC_ADD
......@@ -104,10 +137,14 @@ struct AtomicAddData
template <>
__device__ void Run<AddressSpace::Vgpr, AddressSpace::Global>(const T* p_src,
index_t src_offset,
T* p_dst,
index_t dst_offset) const
bool src_valid T* p_dst,
index_t dst_offset,
bool dst_valid) const
{
amd_buffer_atomic_add<T, DataPerAccess>(&(p_src[src_offset]), p_dst, dst_offset, 0);
const auto zeros = vector_t(0);
amd_buffer_atomic_add<T, DataPerAccess>(
src_valid ? &(p_src[src_offset]) : &zeros, p_dst, dst_offset, 0, dst_valid);
}
#endif
};
......@@ -119,7 +156,12 @@ template <typename T,
InMemoryDataOperation DstInMemOp,
index_t SrcDataStride = 1,
index_t DstDataStride = 1>
__device__ void transfer_data(const T* p_src, index_t src_offset, T* p_dst, index_t dst_offset)
__device__ void transfer_data(const T* p_src,
index_t src_offset,
bool src_valid,
T* p_dst,
index_t dst_offset,
bool dst_valid)
{
static_assert(DstInMemOp == InMemoryDataOperation::Set ||
DstInMemOp == InMemoryDataOperation::AtomicAdd,
......@@ -131,27 +173,37 @@ __device__ void transfer_data(const T* p_src, index_t src_offset, T* p_dst, inde
// TODO: use static_if::ElseIf
static_if<DstInMemOp == InMemoryDataOperation::Set>{}([&](auto) {
SetData<T, DataPerAccess>{}.template Run<SrcAddressSpace, DstAddressSpace>(
p_src, src_offset, p_dst, dst_offset);
p_src, src_offset, src_valid, p_dst, dst_offset, dst_valid);
});
static_if<DstInMemOp == InMemoryDataOperation::AtomicAdd>{}([&](auto) {
AtomicAddData<T, DataPerAccess>{}.template Run<SrcAddressSpace, DstAddressSpace>(
p_src, src_offset, p_dst, dst_offset);
p_src, src_offset, src_valid, p_dst, dst_offset, dst_valid);
});
}
else
{
for(index_t i = 0; i < DataPerAccess; i++)
for(index_t i = 0; i < DataPerAccess; ++i)
{
// TODO: use static_if::ElseIf
static_if<DstInMemOp == InMemoryDataOperation::Set>{}([&](auto) {
SetData<T, 1>{}.template Run<SrcAddressSpace, DstAddressSpace>(
p_src, src_offset + i * SrcDataStride, p_dst, dst_offset + i * DstDataStride);
p_src,
src_offset + i * SrcDataStride,
src_valid,
p_dst,
dst_offset + i * DstDataStride,
dst_valid);
});
static_if<DstInMemOp == InMemoryDataOperation::AtomicAdd>{}([&](auto) {
AtomicAddData<T, 1>{}.template Run<SrcAddressSpace, DstAddressSpace>(
p_src, src_offset + i * SrcDataStride, p_dst, dst_offset + i * DstDataStride);
p_src,
src_offset + i * SrcDataStride,
src_valid,
p_dst,
dst_offset + i * DstDataStride,
dst_valid);
});
}
}
......
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