Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[NCHW][SWDEV-312112] fix local buffer alignment bug on transpose+nhwc kernel's solver #1324

Merged
merged 8 commits into from
Dec 4, 2021
Merged

[NCHW][SWDEV-312112] fix local buffer alignment bug on transpose+nhwc kernel's solver #1324

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
4512949
[nchw]fix local buffer alignment bug on transpose+nhwc kernel's solver
shaojiewang Dec 3, 2021
3bc6fdf
add more alignment bytes
shaojiewang Dec 3, 2021
936cc72
90a seems strange
shaojiewang Dec 3, 2021
ec7934c
[skip ci]90a seems strange
shaojiewang Dec 3, 2021
9f82aa0
[skip ci]check offset
shaojiewang Dec 3, 2021
a410d51
[skip ci]check offset
shaojiewang Dec 3, 2021
cffe2cc
checkoffset
shaojiewang Dec 3, 2021
dfd8a41
add some comments
shaojiewang Dec 3, 2021
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 4 additions & 2 deletions src/conv/invokers/impl_gemm_dynamic.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -579,7 +579,8 @@ InvokerFactory MakeImplGemmDynamicForwardXdlopsNHWCInvokerFactory(
trans_output_idx = idx++;
}

const size_t cast_offset = is_nchw ? (trans_output_offset + trans_output_size) : 0;
// 4 bytes alignment to do atomic add
const size_t cast_offset = is_nchw ? (((trans_output_offset + trans_output_size + 3) >> 2) << 2) : 0;
atamazov marked this conversation as resolved.
Show resolved Hide resolved
const size_t cast_size = need_cast ? miopen::GetTypeSize(miopenFloat) * n * k * ho * wo : 0;

const int kID_trans_start = isGfx90aFp16altSupport ? 2 : 1;
Expand Down Expand Up @@ -886,7 +887,8 @@ InvokerFactory MakeImplGemmDynamicBackwardDataXdlopsNHWCInvokerFactory(
trans_output_idx = idx++;
}

const size_t cast_offset = is_nchw ? (trans_output_offset + trans_output_size) : 0;
// 4 bytes alignment to do atomic add
const size_t cast_offset = is_nchw ? (((trans_output_offset + trans_output_size + 3) >> 2) << 2) : 0;
const size_t cast_size = need_cast ? miopen::GetTypeSize(miopenFloat) * n * c * hi * wi : 0;

const int kID_trans_start = isGfx90aFp16altSupport ? 2 : 1;
Expand Down
3 changes: 3 additions & 0 deletions src/solver/conv_asm_implicit_gemm_gtc_bwd_nhwc.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -927,6 +927,9 @@ ConvAsmImplicitGemmGTCDynamicBwdXdlopsNHWC::GetWorkspaceSize(const ConvolutionCo
workspace_size += trans_weight.GetSize();
if(!trans_output.IsSkippable())
workspace_size += trans_output.GetSize();

// 4 bytes alignment to do atomic add
workspace_size = ((workspace_size + 3) >> 2) << 2;
}

if(!ctx.IsFp32())
Expand Down
3 changes: 3 additions & 0 deletions src/solver/conv_asm_implicit_gemm_gtc_fwd_nhwc.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -777,6 +777,9 @@ ConvAsmImplicitGemmGTCDynamicFwdXdlopsNHWC::GetWorkspaceSize(const ConvolutionCo
workspace_size += trans_weight.GetSize();
if(!trans_output.IsSkippable())
workspace_size += trans_output.GetSize();

// 4 bytes alignment to do atomic add
workspace_size = ((workspace_size + 3) >> 2) << 2;
}

if(!ctx.IsFp32())
Expand Down
6 changes: 5 additions & 1 deletion src/solver/conv_asm_implicit_gemm_gtc_wrw_nhwc.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -903,6 +903,9 @@ ConvAsmImplicitGemmGTCDynamicWrwXdlopsNHWC::GetWorkspaceSize(const ConvolutionCo
workspace_size += trans_weight.GetSize();
if(!trans_output.IsSkippable())
workspace_size += trans_output.GetSize();

// 4 bytes alignment to do atomic add
workspace_size = ((workspace_size + 3) >> 2) << 2;
}

if(!ctx.IsFp32())
Expand Down Expand Up @@ -1071,7 +1074,8 @@ ConvSolution ConvAsmImplicitGemmGTCDynamicWrwXdlopsNHWC::GetSolution(

MIOPEN_LOG_I2(SolverDbId(*this) << ": " << config.ToString() << msg.str());

const size_t cast_offset = is_nchw ? (trans_output_offset + trans_output_size) : 0;
// 4 bytes alignment to do atomic add
const size_t cast_offset = is_nchw ? (((trans_output_offset + trans_output_size + 3) >> 2) << 2) : 0;
const size_t cast_size = need_cast ?
miopen::GetTypeSize(miopenFloat) * k * (c / group) * y * x : 0;

Expand Down