Skip to content

fix(ck): fix gfx950 bpreshuffle determinism and illegal-type codegen #8639

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
andriy-ca merged 9 commits into from
Jul 6, 2026
Merged

fix(ck): fix gfx950 bpreshuffle determinism and illegal-type codegen #8639

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
d56a67a
Stabilize CK blockscale bpreshuffle GEMM
Jun 19, 2026
6c769bb
Address bpreshuffle review comments
Jun 19, 2026
b1cb121
Apply pre-commit formatting
Jun 19, 2026
ea529cd
Simplify bpreshuffle test comments
Jun 19, 2026
3d353a9
Remove unnecessary bpreshuffle formatting guard
Jun 20, 2026
1bc7f50
Apply CK profiler clang-format
Jun 20, 2026
92c4415
Keep bpreshuffle profiler defaults readable
Jun 20, 2026
11b6ccb
Merge branch 'develop' into users/hdt98/gemm_a8w8_blockscale_bpreshuf…
andriy-ca Jul 3, 2026
be0058e
fix(ck): address bpreshuffle review comments
hdt98 Jul 3, 2026
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
114 changes: 54 additions & 60 deletions .../tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_blockscale_b_preshuffle_v1.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -285,6 +285,15 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1<BlockGemmPipelineS
{
ignore = b_block_desc;
ignore = b_block_buf;

static_assert(sizeof(AccDataType) == 4 && is_floating_point<AccDataType>::value,
"bpreshuffle accumulator anchor requires 32-bit float AccDataType");

// The empty asm is a read/write VGPR use of the updated accumulator.
// It keeps each post-scale accumulator definition visible to the
// optimizer, which enables determinism for bitwise-stable repeated launches.
auto anchor_accumulator_value = [&](auto& value) { asm volatile("" : "+v"(value)); };
Comment thread
hdt98 marked this conversation as resolved.

// __builtin_amdgcn_sched_barrier(0);
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
a_thread_desc_.GetElementSpaceSize());
Expand Down Expand Up @@ -451,14 +460,9 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1<BlockGemmPipelineS
c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})
.template AsType<AccDataType>()(Number<t>{}) = 0;
});
vector_type<AccDataType, 2> c_scale_thread_vec;
constexpr index_t cscale_offset = CScaleThreadDesc{}.CalculateOffset(
make_tuple(kscale0, m0, n0 * num_scale_n_block / NRepeat));

c_scale_thread_vec.template AsType<AccDataType>()(Number<0>{}) =
c_scale_thread_buf[Number<cscale_offset>{}];
c_scale_thread_vec.template AsType<AccDataType>()(Number<1>{}) =
c_scale_thread_buf[Number<cscale_offset>{}];
const auto c_scale_thread = c_scale_thread_buf[Number<cscale_offset>{}];

static_for<0, KRepeat / num_scale_k_block, 1>{}([&](auto k0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
Expand Down Expand Up @@ -501,16 +505,18 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1<BlockGemmPipelineS
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));

static_for<0, xdlops_gemm.GetRegSizePerXdlops() / 2, 1>{}([&](auto t) {
using pk_fma_type = typename vector_type<AccDataType, 2>::type;

c_thread_buf.GetVectorTypeReference(Number<c_offset>{})
.template AsType<pk_fma_type>()(t) = __builtin_elementwise_fma(
c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})
.template AsType<pk_fma_type>()[t],
c_scale_thread_vec.template AsType<pk_fma_type>()[Number<0>{}],
// Keep the post-scale FMA scalar; the old packed 2-lane update hits
Comment thread
hdt98 marked this conversation as resolved.
// ROCm 7.2 gfx950 illegal-type legalization/codegen issues.
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
auto& c_acc_vec =
c_thread_buf.GetVectorTypeReference(Number<c_offset>{})
.template AsType<pk_fma_type>()[t]);
.template AsType<AccDataType>();
const auto c_partial_acc =
c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})
.template AsType<AccDataType>()(t);
c_acc_vec(t) = __builtin_elementwise_fma(
c_partial_acc, c_scale_thread, c_acc_vec(t));
anchor_accumulator_value(c_acc_vec(t));
});
});

Expand Down Expand Up @@ -606,14 +612,9 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1<BlockGemmPipelineS
c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})
.template AsType<AccDataType>()(Number<t>{}) = 0;
});
vector_type<AccDataType, 2> c_scale_thread_vec;
constexpr index_t cscale_offset = CScaleThreadDesc{}.CalculateOffset(
make_tuple(kscale0, m0, n0 * num_scale_n_block / NRepeat));

c_scale_thread_vec.template AsType<AccDataType>()(Number<0>{}) =
c_scale_thread_buf[Number<cscale_offset>{}];
c_scale_thread_vec.template AsType<AccDataType>()(Number<1>{}) =
c_scale_thread_buf[Number<cscale_offset>{}];
const auto c_scale_thread = c_scale_thread_buf[Number<cscale_offset>{}];

static_for<0, KRepeat / num_scale_k_block, 1>{}([&](auto k0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
Expand Down Expand Up @@ -653,16 +654,17 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1<BlockGemmPipelineS
});
constexpr index_t c_offset = c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));

static_for<0, xdlops_gemm.GetRegSizePerXdlops() / 2, 1>{}([&](auto t) {
using pk_fma_type = typename vector_type<AccDataType, 2>::type;

c_thread_buf.GetVectorTypeReference(Number<c_offset>{})
.template AsType<pk_fma_type>()(t) = __builtin_elementwise_fma(
// Keep the post-scale FMA scalar; the old packed 2-lane update hits ROCm 7.2
// gfx950 illegal-type legalization/codegen issues.
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
auto& c_acc_vec = c_thread_buf.GetVectorTypeReference(Number<c_offset>{})
.template AsType<AccDataType>();
const auto c_partial_acc =
c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})
.template AsType<pk_fma_type>()[t],
c_scale_thread_vec.template AsType<pk_fma_type>()[Number<0>{}],
c_thread_buf.GetVectorTypeReference(Number<c_offset>{})
.template AsType<pk_fma_type>()[t]);
.template AsType<AccDataType>()(t);
c_acc_vec(t) =
__builtin_elementwise_fma(c_partial_acc, c_scale_thread, c_acc_vec(t));
anchor_accumulator_value(c_acc_vec(t));
});
});

Expand Down Expand Up @@ -701,14 +703,9 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1<BlockGemmPipelineS
c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})
.template AsType<AccDataType>()(Number<t>{}) = 0;
});
vector_type<AccDataType, 2> c_scale_thread_vec;
constexpr index_t cscale_offset = CScaleThreadDesc{}.CalculateOffset(
make_tuple(kscale0, m0, n0 * num_scale_n_block / NRepeat));

c_scale_thread_vec.template AsType<AccDataType>()(Number<0>{}) =
c_scale_thread_buf[Number<cscale_offset>{}];
c_scale_thread_vec.template AsType<AccDataType>()(Number<1>{}) =
c_scale_thread_buf[Number<cscale_offset>{}];
const auto c_scale_thread = c_scale_thread_buf[Number<cscale_offset>{}];

static_for<0, KRepeat / num_scale_k_block, 1>{}([&](auto k0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
Expand Down Expand Up @@ -748,16 +745,17 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1<BlockGemmPipelineS
});
constexpr index_t c_offset = c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));

static_for<0, xdlops_gemm.GetRegSizePerXdlops() / 2, 1>{}([&](auto t) {
using pk_fma_type = typename vector_type<AccDataType, 2>::type;

c_thread_buf.GetVectorTypeReference(Number<c_offset>{})
.template AsType<pk_fma_type>()(t) = __builtin_elementwise_fma(
// Keep the post-scale FMA scalar; the old packed 2-lane update hits ROCm 7.2
// gfx950 illegal-type legalization/codegen issues.
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
auto& c_acc_vec = c_thread_buf.GetVectorTypeReference(Number<c_offset>{})
.template AsType<AccDataType>();
const auto c_partial_acc =
c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})
.template AsType<pk_fma_type>()[t],
c_scale_thread_vec.template AsType<pk_fma_type>()[Number<0>{}],
c_thread_buf.GetVectorTypeReference(Number<c_offset>{})
.template AsType<pk_fma_type>()[t]);
.template AsType<AccDataType>()(t);
c_acc_vec(t) =
__builtin_elementwise_fma(c_partial_acc, c_scale_thread, c_acc_vec(t));
anchor_accumulator_value(c_acc_vec(t));
});
});
}
Expand All @@ -771,14 +769,9 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1<BlockGemmPipelineS
c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})
.template AsType<AccDataType>()(Number<t>{}) = 0;
});
vector_type<AccDataType, 2> c_scale_thread_vec;
constexpr index_t cscale_offset = CScaleThreadDesc{}.CalculateOffset(
make_tuple(kscale0, m0, n0 * num_scale_n_block / NRepeat));

c_scale_thread_vec.template AsType<AccDataType>()(Number<0>{}) =
c_scale_thread_buf[Number<cscale_offset>{}];
c_scale_thread_vec.template AsType<AccDataType>()(Number<1>{}) =
c_scale_thread_buf[Number<cscale_offset>{}];
const auto c_scale_thread = c_scale_thread_buf[Number<cscale_offset>{}];

static_for<0, KRepeat / num_scale_k_block, 1>{}([&](auto k0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
Expand Down Expand Up @@ -819,16 +812,17 @@ struct BlockwiseGemmXdlops_pipeline_blockscale_bpreshuffle_v1<BlockGemmPipelineS

constexpr index_t c_offset = c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));

static_for<0, xdlops_gemm.GetRegSizePerXdlops() / 2, 1>{}([&](auto t) {
using pk_fma_type = typename vector_type<AccDataType, 2>::type;

c_thread_buf.GetVectorTypeReference(Number<c_offset>{})
.template AsType<pk_fma_type>()(t) = __builtin_elementwise_fma(
// Keep the post-scale FMA scalar; the old packed 2-lane update hits ROCm 7.2
// gfx950 illegal-type legalization/codegen issues.
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
auto& c_acc_vec = c_thread_buf.GetVectorTypeReference(Number<c_offset>{})
.template AsType<AccDataType>();
const auto c_partial_acc =
c_thread_buf_per_scale.GetVectorTypeReference(Number<0>{})
.template AsType<pk_fma_type>()[t],
c_scale_thread_vec.template AsType<pk_fma_type>()[Number<0>{}],
c_thread_buf.GetVectorTypeReference(Number<c_offset>{})
.template AsType<pk_fma_type>()[t]);
.template AsType<AccDataType>()(t);
c_acc_vec(t) =
__builtin_elementwise_fma(c_partial_acc, c_scale_thread, c_acc_vec(t));
anchor_accumulator_value(c_acc_vec(t));
});
});
}
Expand Down
Loading
Loading