Skip to content

[TIR] Check dynamic shared memory in VerifyGPUCode #10923

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

Merged
masahi merged 1 commit into from
Apr 7, 2022
Merged

[TIR] Check dynamic shared memory in VerifyGPUCode #10923

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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/tir/analysis/verify_gpu_code.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,6 +30,7 @@
#include <tvm/tir/stmt.h>
#include <tvm/tir/stmt_functor.h>

#include "../../runtime/thread_storage_scope.h"
#include "../transforms/ir_utils.h"

namespace tvm {
Expand Down Expand Up @@ -61,11 +62,12 @@ class GPUCodeVerifier : public StmtExprVisitor {
void VisitStmt_(const AllocateNode* op) final {
StmtVisitor::VisitStmt_(op);
auto scope = GetPtrStorageScope(op->buffer_var);
runtime::StorageScope storage_scope = runtime::StorageScope::Create(scope);
// visit an allocation of a buffer in shared memory, record its size
if (scope == "local") {
if (storage_scope.rank == runtime::StorageRank::kLocal) {
size_t size = static_cast<size_t>(op->ConstantAllocationSize());
local_memory_per_block_ += size * op->dtype.bytes() * op->dtype.lanes();
} else if (scope == "shared") {
} else if (storage_scope.rank == runtime::StorageRank::kShared) {
size_t size = static_cast<size_t>(op->ConstantAllocationSize());
Copy link
Member

@masahi masahi Apr 7, 2022

Choose a reason for hiding this comment

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

Note that op->ConstantAllocationSize() is 0 for symbolic shape dyn-shared memory. But VerifyGPUCode cannot do anything with symbolic shape, so this should be fine.

shared_memory_per_block_ += size * op->dtype.bytes() * op->dtype.lanes();
}
Expand Down
9 changes: 5 additions & 4 deletions tests/python/unittest/test_tir_analysis_verify_gpu_code.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,7 +32,7 @@ def _fverify(f, *_):

@tvm.testing.requires_gpu
def test_shared_memory():
def check_shared_memory(dtype):
def check_shared_memory(storage_scope, dtype):
N = 1024
M = 128

Expand All @@ -43,7 +43,7 @@ def check_shared_memory(dtype):
B = te.compute((N,), lambda i: A[i], name="B")

s = te.create_schedule([B.op])
AA = s.cache_read(A, "shared", [B])
AA = s.cache_read(A, storage_scope, [B])
o, i = s[B].split(s[B].op.axis[0], M)
s[AA].compute_at(s[B], o)
s[B].bind(o, te.thread_axis("blockIdx.x"))
Expand Down Expand Up @@ -90,8 +90,9 @@ def check_shared_memory(dtype):
tvm.build(s, [A, B], target)
assert valid[0]

check_shared_memory("float32")
check_shared_memory("int8x4")
check_shared_memory("shared", "float32")
check_shared_memory("shared", "int8x4")
check_shared_memory("shared.dyn", "float32")


@tvm.testing.requires_gpu
Expand Down