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

[HOTFIX] Disable "granularity loss" W/A for #2492 and add a new, "tiny tensor" based one. #2695

Merged
merged 3 commits into from
Jan 26, 2024
Merged

[HOTFIX] Disable "granularity loss" W/A for #2492 and add a new, "tiny tensor" based one. #2695

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
0e1745b
disable 2492 granularity_loss workaround and enable tiny_tensor worka…
junliume Jan 25, 2024
39d3c4a
workaround_issue_2492_02(01) Macros to uppercase. Add doc for WORKARO…
atamazov Jan 25, 2024
daf36f7
Update src/solver/conv_winoRxS.cpp
junliume Jan 26, 2024
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
31 changes: 26 additions & 5 deletions src/solver/conv_winoRxS.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -64,9 +64,17 @@
/// tests use very small convolutions and Winograd algorithm is
/// ineffective with such small configs due to huge granularity loss,
/// we can disable Winograd without any performance implications.
#define WORKAROUND_ISSUE_2493 1

MIOPEN_DECLARE_ENV_VAR_BOOL(MIOPEN_DEBUG_WORKAROUND_ISSUE_2493)
#define WORKAROUND_ISSUE_2492_GRANULARITY_LOSS 0
/// \anchor disable_winograd_with_small_tensor
/// This is alternative W/A for the same problem, which
/// resolves issues shown at
/// https://github.com/ROCm/MIOpen/issues/2492#issuecomment-1910928563
/// The idea is to disable the solver for very small tensors.
/// PyTorch tests use sizes like 2x2x6x6 (NCHW), but let's add
/// some spare space for N anc C and disable starting from 4x4x6x6.
#define WORKAROUND_ISSUE_2492_TINY_TENSOR 1

MIOPEN_DECLARE_ENV_VAR_BOOL(MIOPEN_DEBUG_WORKAROUND_ISSUE_2492)

MIOPEN_DECLARE_ENV_VAR_BOOL(MIOPEN_DEBUG_AMD_WINOGRAD_RXS_F2X3)
MIOPEN_DECLARE_ENV_VAR_STR(MIOPEN_DEBUG_AMD_WINOGRAD_RXS_F2X3_PERF_VALS)
Expand Down Expand Up @@ -683,8 +691,8 @@ static bool IsApplicableBase(const ExecutionContext& ctx, const ProblemDescripti
return false;
// clang-format on

#if WORKAROUND_ISSUE_2493
if(!miopen::IsDisabled(ENV(MIOPEN_DEBUG_WORKAROUND_ISSUE_2493)) &&
#if WORKAROUND_ISSUE_2492_GRANULARITY_LOSS
if(!miopen::IsDisabled(ENV(MIOPEN_DEBUG_WORKAROUND_ISSUE_2492)) &&
!miopen::debug::IsWarmupOngoing)
{
constexpr double max_perf_drop_due_to_granularity = 200; // Times.
Expand All @@ -697,6 +705,19 @@ static bool IsApplicableBase(const ExecutionContext& ctx, const ProblemDescripti
}
#endif

#if WORKAROUND_ISSUE_2492_TINY_TENSOR
if(!miopen::IsDisabled(ENV(MIOPEN_DEBUG_WORKAROUND_ISSUE_2492)) &&
!miopen::debug::IsWarmupOngoing)
{
// Group count is not taken into account intentionally.
if(problem.GetInHeight() <= 6 //
&& problem.GetInWidth() <= 6 //
&& problem.GetBatchSize() <= 4 //
&& problem.GetInChannels() <= 4)
return false;
}
#endif

const auto n_inputs_per_group = problem.GetInChannels() / problem.GetGroupCount(),
n_outputs_per_group = problem.GetOutChannels() / problem.GetGroupCount();

Expand Down
16 changes: 10 additions & 6 deletions test/gtest/db_sync.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -44,9 +44,13 @@
#include <exception>
#include <unordered_set>

#define WORKAROUND_ISSUE_2493 1
/// \todo HACK
/// This should be set to 1 if either WORKAROUND_ISSUE_2492_GRANULARITY_LOSS
/// or WORKAROUND_ISSUE_2492_TINY_TENSOR is defined as non-zero in
/// src/solver/conv_winoRxS.cpp
#define WORKAROUND_ISSUE_2492 1

#if WORKAROUND_ISSUE_2493 && defined(_WIN32)
#if WORKAROUND_ISSUE_2492 && defined(_WIN32)
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The original W/A (#2507, #2510) has a typo (2492->2493) which is fixed here at once.

#define WIN32_LEAN_AND_MEAN
#include <Windows.h>
#endif
Expand Down Expand Up @@ -81,13 +85,13 @@ struct std::hash<KDBKey>
}
};

#if WORKAROUND_ISSUE_2493 && !defined(_WIN32)
#if WORKAROUND_ISSUE_2492 && !defined(_WIN32)
static void SetEnvironmentVariable(std::string_view name, std::string_view value)
{
const auto ret = setenv(name.data(), value.data(), 1);
ASSERT_TRUE(ret == 0);
}
#endif // WORKAROUND_ISSUE_2493
#endif // WORKAROUND_ISSUE_2492

#if WORKAROUND_ISSUE_1987
/// \todo Copied from src/db_record.cpp
Expand Down Expand Up @@ -444,8 +448,8 @@ TEST(DBSync, KDBTargetID)
if(miopen::IsEnabled(ENV(MIOPEN_TEST_DBSYNC)))
{
fs::path fdb_file_path, pdb_file_path, kdb_file_path;
#if WORKAROUND_ISSUE_2493
SetEnvironmentVariable("MIOPEN_DEBUG_WORKAROUND_ISSUE_2493", "0");
#if WORKAROUND_ISSUE_2492
SetEnvironmentVariable("MIOPEN_DEBUG_WORKAROUND_ISSUE_2492", "0");
#endif
SetupPaths(fdb_file_path, pdb_file_path, kdb_file_path, get_handle());
std::ignore = fdb_file_path;
Expand Down