ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-18 12:00:07 +00:00
Code Issues Packages Projects Releases Wiki Activity

[CK_TILE] Enable MXFP6 for MX GEMM op (#5095)

Browse Source 
## Motivation

Add support for MXFP6 in the MX GEMM op in CK-Tile.

Depends on https://github.com/ROCm/rocm-libraries/pull/4594

## Technical Details

<!-- Explain the changes along with any relevant GitHub links. -->

## Test Plan

<!-- Explain any relevant testing done to verify this PR. -->

## Test Result

<!-- Briefly summarize test outcomes. -->

## Submission Checklist

- [ ] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
This commit is contained in:
Sami Remes
2026-03-20 03:07:47 +02:00
committed by GitHub
parent a61238b69f
commit a699df9fdc
13 changed files with 160 additions and 31 deletions
Download Patch File Download Diff File Expand all files Collapse all files

30
include/ck_tile/core/numeric/pk_fp6.hpp
View File

@@ -22,7 +22,10 @@ struct pk_fp6_t
static constexpr index_t vector_size = (packed_size * num_bits_elem) / num_bits_vec_elem;
element_type data_[vector_size]; // packed data
using type = pk_fp6_t<packed_size>;
CK_TILE_HOST_DEVICE constexpr explicit pk_fp6_t(int value = 0)
CK_TILE_HOST_DEVICE constexpr pk_fp6_t() : data_{element_type{}} {}
CK_TILE_HOST_DEVICE constexpr explicit pk_fp6_t(int value)
{
for(size_t i = 0; i < vector_size; ++i)
{
@@ -59,13 +62,14 @@ struct pk_fp6_t
const int bit_offset = bit_pos % num_bits_vec_elem;
const int overhang = bit_offset + num_bits_elem - num_bits_vec_elem;
int32_t bits = pk.data_[arr_idx] >> bit_offset;
uint32_t bits = static_cast<uint32_t>(pk.data_[arr_idx]) >> bit_offset;
if(overhang > 0 && (arr_idx + 1) < vector_size)
{
bits |= (pk.data_[arr_idx + 1] & ((1u << overhang) - 1)) << (num_bits_elem - overhang);
bits |= (static_cast<uint32_t>(pk.data_[arr_idx + 1]) & ((1u << overhang) - 1))
<< (num_bits_elem - overhang);
}
return bits & 0x3F;
return static_cast<int32_t>(bits & 0x3F);
}
CK_TILE_HOST_DEVICE int32_t unpack(const index_t i) const { return unpack(*this, i); }
@@ -97,6 +101,22 @@ struct pk_fp6_t
}
return sign == 1 ? -1 * result : result;
}
CK_TILE_HOST static int32_t float_to_fp6_e2m3(float val)
{
int32_t best = 0;
float best_err = 1e30f;
for(int32_t i = 0; i < 64; i++)
{
float err = std::fabs(val - fp6_e2m3_to_float(i));
if(err < best_err)
{
best = i;
best_err = err;
}
}
return best;
}
};
using pk_fp6x16_t = pk_fp6_t<16>;
@@ -105,5 +125,7 @@ template <>
struct numeric_traits<pk_fp6x16_t>
{
static constexpr int PackedSize = 16;
static constexpr int exp = 2;
static constexpr int mant = 3;
};
} // namespace ck_tile

16
include/ck_tile/core/tensor/buffer_view.hpp
View File

@@ -19,6 +19,14 @@
namespace ck_tile {
// buffer_load_dwordx3 to LDS uses a fixed 16-byte per-thread stride,
// padding each 12-byte element to 16 bytes in LDS.
template <typename T>
CK_TILE_HOST_DEVICE constexpr index_t lds_padded_sizeof()
{
return (sizeof(T) == 12) ? 16 : sizeof(T);
}
// T may be scalar or vector
// X may be scalar or vector
// T and X have same scalar type
@@ -840,7 +848,10 @@ struct buffer_view<address_space_enum::lds,
{
using buf_t = ext_vector_t<typename vector_traits<remove_cvref_t<T>>::scalar_type,
scalar_per_t_vector * scalar_per_x_vector>;
auto rtn = *c_style_pointer_cast<const buf_t*>(&p_data_[i + linear_offset]);
constexpr index_t padded_stride = lds_padded_sizeof<T>();
const char* base =
reinterpret_cast<const char*>(p_data_) + (i + linear_offset) * padded_stride;
auto rtn = *c_style_pointer_cast<const buf_t*>(base);
return bit_cast<X>(rtn);
}
#endif
@@ -872,7 +883,8 @@ struct buffer_view<address_space_enum::lds,
bool /*is_valid_element*/,
bool_constant<pre_nop> = {}) const
{
smem_load<sizeof(X)>{}(dst, v_offset * sizeof(T), i_offset * sizeof(T));
constexpr index_t padded_stride = lds_padded_sizeof<T>();
smem_load<sizeof(X)>{}(dst, v_offset * padded_stride, i_offset * padded_stride);
}
template <typename X,

18
include/ck_tile/core/tensor/tile_scatter_gather.hpp
View File

@@ -631,21 +631,24 @@ struct tile_scatter_gather
// issues * warps * lanes
static_assert(LdsTileWindow::get_num_of_dimension() == 3); // TODO: hard coded
// buffer load with dwordx3 requires 128-bit alignment
constexpr index_t lds_stride = lds_padded_sizeof<LdsDataType>();
const index_t size_per_buf =
lds_tile.get_bottom_tensor_view().get_tensor_descriptor().calculate_offset(
make_tuple(number<0>{}, number<0>{}, number<0>{})) *
sizeof(LdsDataType);
lds_stride;
const index_t size_per_wave =
lds_tile.get_bottom_tensor_view().get_tensor_descriptor().calculate_offset(
make_tuple(number<0>{}, number<1>{}, number<0>{})) *
sizeof(LdsDataType) -
lds_stride -
size_per_buf;
const index_t size_per_issue =
lds_tile.get_bottom_tensor_view().get_tensor_descriptor().calculate_offset(
make_tuple(number<1>{}, number<0>{}, number<0>{})) *
sizeof(LdsDataType) -
lds_stride -
size_per_buf;
const index_t m0_init_value = size_per_buf + size_per_wave * get_warp_id();
@@ -780,9 +783,12 @@ struct tile_scatter_gather
make_tensor_coordinate(tensor_descriptor, lds_bottom_tensor_thread_idx);
// Calculate SMEM address using base pointer
CK_TILE_LDS_ADDR LdsDataType* smem = lds_base_ptr +
lds_coord.get_offset() / Traits::PackedSize +
lds_ys_offset / Traits::PackedSize;
// Use byte arithmetic for dwordx3 padding (12-byte elements use 16-byte LDS stride)
CK_TILE_LDS_ADDR LdsDataType* smem =
reinterpret_cast<CK_TILE_LDS_ADDR LdsDataType*>(
reinterpret_cast<CK_TILE_LDS_ADDR char*>(lds_base_ptr) +
(lds_coord.get_offset() + lds_ys_offset) / Traits::PackedSize *
lds_padded_sizeof<LdsDataType>());
const auto dram_ys_offset = [&]() {
if constexpr(static_move_ys)

14
include/ck_tile/core/tensor/tile_window.hpp
View File

@@ -501,21 +501,23 @@ struct tile_window_with_static_distribution
// issues * warps * lanes
static_assert(LdsTileWindow::get_num_of_dimension() == 3); // TODO: hard coded
constexpr index_t lds_stride = lds_padded_sizeof<LdsDataType>();
const index_t size_per_buf =
lds_tile.get_bottom_tensor_view().get_tensor_descriptor().calculate_offset(
make_tuple(number<0>{}, number<0>{}, number<0>{})) *
sizeof(LdsDataType);
lds_stride;
const index_t size_per_wave =
lds_tile.get_bottom_tensor_view().get_tensor_descriptor().calculate_offset(
make_tuple(number<0>{}, number<1>{}, number<0>{})) *
sizeof(LdsDataType) -
lds_stride -
size_per_buf;
const index_t size_per_issue =
lds_tile.get_bottom_tensor_view().get_tensor_descriptor().calculate_offset(
make_tuple(number<1>{}, number<0>{}, number<0>{})) *
sizeof(LdsDataType) -
lds_stride -
size_per_buf;
// Use VALU so the compiler can optimize redundant/repeated computations
@@ -628,8 +630,12 @@ struct tile_window_with_static_distribution
make_tensor_coordinate(tensor_descriptor, lds_bottom_tensor_thread_idx);
// Calculate SMEM address using base pointer
// Use byte arithmetic for dwordx3 padding (12-byte elements use 16-byte LDS stride)
CK_TILE_LDS_ADDR LdsDataType* smem =
lds_base_ptr + (lds_coord.get_offset() + lds_ys_offset) / Traits::PackedSize;
reinterpret_cast<CK_TILE_LDS_ADDR LdsDataType*>(
reinterpret_cast<CK_TILE_LDS_ADDR char*>(lds_base_ptr) +
(lds_coord.get_offset() + lds_ys_offset) / Traits::PackedSize *
lds_padded_sizeof<LdsDataType>());
const auto dram_ys_offset = [&]() {
if constexpr(static_move_ys)