Skip to content

[fix] Now non-swizzled stride can be less than lda*A_col #951

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

Closed
mengzcai wants to merge 1 commit into from
Closed

[fix] Now non-swizzled stride can be less than lda*A_col #951

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
1 change: 0 additions & 1 deletion projects/hipblaslt/clients/gtest/matmul_gtest.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -855,7 +855,6 @@ Tests:
K: [128, 129]
lda: [129, 2440]
batch_count: [1, 5]
stride_a: 0
Copy link
Copy Markdown
Contributor Author

@mengzcai mengzcai Jul 30, 2025

Choose a reason for hiding this comment

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

TODO:

  1. support stride_a is 0 when swizzle
  2. support arbitrary stride_a for both hipblaslt-bench and hipblaslt-test when swizzled

transA: T
transB: N
alpha: 1
Expand Down
18 changes: 12 additions & 6 deletions projects/hipblaslt/clients/include/testing_matmul.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1376,28 +1376,34 @@ void testing_matmul_with_bias(const Arguments& arg,
do_batched[i] = (arg.batch_count > 1);
num_batches[i] = (do_batched[i] ? arg.batch_count : 1);

stride_a[i] = (do_batched[i] && arg.stride_a[i] >= lda[i] * A_col[i]) ? arg.stride_a[i]
: lda[i] * A_col[i];
stride_a[i] = do_batched[i] ? arg.stride_a[i] : lda[i] * A_col[i];
stride_b[i] = do_batched[i] ? arg.stride_b[i] : ldb[i] * B_col[i];
stride_c[i] = do_batched[i] ? arg.stride_c[i] : ldc[i] * N[i];
stride_d[i] = do_batched[i] ? arg.stride_c[i] : ldd[i] * N[i];
stride_e[i] = do_batched[i] ? arg.stride_e[i] : lde[i] * N[i];

size_A[i] = stride_a[i] * num_batches[i];
size_A[i]
= stride_a[i] == 0 ? lda[i] * A_col[i] * num_batches[i] : stride_a[i] * num_batches[i];
size_dA[i] = size_A[i];
stride_da[i] = stride_a[i];
if(arg.swizzle_a && isSwizzleSupported(TiA))
{
//TODO:
// 1. support stride_a is 0 when swizzle
// 2. support --any_stride for hipblaslt-bench when swizzle
// 3. support arbitrary stride_a for both hipblaslt-bench and hipblaslt-test when swizzled
stride_a[i] = lda[i] * A_col[i];
stride_da[i] = 0;
size_t MiM = 16, MiK = 0, __ = 0, PackK = 0;
calculateKforSwizzling(TiA, arg, MiK, __, PackK);
size_t K_block = MiK * PackK;
size_t stride_swizzle
= ((M[i] + MiM - 1) / MiM) * MiM * ((K[i] + K_block - 1) / K_block) * K_block;
if(do_batched[i] && arg.stride_a[i] >= (int64_t)stride_swizzle)
if(do_batched[i] && stride_a[i] > 0 && stride_a[i] != lda[i] * A_col[i])
{
stride_da[i] = arg.stride_a[i];
stride_swizzle = (size_t)arg.stride_a[i];
hipblaslt_cerr << "Error stride_a: swizzle_a does not yet support arbitrary stride_a!" << std::endl;
stride_da[i] = stride_a[i];
stride_swizzle = (size_t)stride_a[i];
}
Comment thread
solaslin marked this conversation as resolved.
size_dA[i] = num_batches[i] * stride_swizzle;
}
Expand Down
Loading