ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-04 21:51:28 +00:00
Code Issues Packages Projects Releases Wiki Activity

[rocm-libraries] ROCm/rocm-libraries#6302 (commit 8d419e8)

Browse Source 
CK: Remove 41 commented-out dead code blocks (~200 lines)
 (#6302)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Depends on #6300

## Summary

Remove 41 commented-out code blocks across 33 files in Composable
Kernel, totaling ~200 lines.

Identified using an automated dead code scanning skill (`ck-dead-code`)
with a calibrated two-stage pipeline:
1. **Pre-filter**: Keyword-based scan found 1,338 `//`-commented blocks.
Calibrated heuristics (trained on 50-sample expert classification)
reduced to 89 high-confidence candidates — 93% noise reduction.
2. **Expert triage**: LLM expert classified each block in context as
CODE_REMOVE, CODE_KEEP, or NOT_CODE.

| Classification | Count |
|---------------|-------|
| Removed (this PR) | 41 |
| Kept (debug helpers, alt configs, reference impls) | 32 |
| Not code (false positives) | 16 |

Removed blocks include: superseded implementations, old test data,
abandoned stubs, unreachable code, and buggy dead code.
This commit is contained in:
Aviral Goel
2026-04-10 15:18:02 +00:00
committed by assistant-librarian[bot]
parent 4d0bbe5d17
commit e0dfe58d66
82 changed files with 22 additions and 2883 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
include/ck_tile/core/arch/amd_buffer_addressing.hpp
View File

@@ -2166,27 +2166,11 @@ CK_TILE_DEVICE void amd_buffer_store_impl(const thread_buffer<T, N> src_thread_d
}
else if constexpr(N == 8)
{
#if 0
thread_buffer<fp16_t, 8> tmp{src_thread_data};
llvm_amdgcn_raw_buffer_store_fp16x4(tmp.template get_as<fp16x4_t>()[number<0>{}],
dst_wave_buffer_resource,
dst_thread_addr_offset,
dst_wave_addr_offset,
static_cast<index_t>(coherence));
llvm_amdgcn_raw_buffer_store_fp16x4(tmp.template get_as<fp16x4_t>()[number<1>{}],
dst_wave_buffer_resource,
dst_thread_addr_offset,
dst_wave_addr_offset + 4 * sizeof(fp16_t),
static_cast<index_t>(coherence));
#else
llvm_amdgcn_raw_buffer_store_fp32x4(bit_cast<fp32x4_t>(src_thread_data),
dst_wave_buffer_resource,
dst_thread_addr_offset,
dst_wave_addr_offset,
static_cast<index_t>(coherence));
#endif
}
}
else if constexpr(std::is_same<T, bf16_t>::value) // bf16

16
include/ck_tile/core/arch/amd_buffer_addressing_builtins.hpp
View File

@@ -1992,27 +1992,11 @@ CK_TILE_DEVICE void amd_buffer_store_impl(const thread_buffer<T, N> src_thread_d
}
else if constexpr(N == 8)
{
#if 0
thread_buffer<fp16_t, 8> tmp{src_thread_data};
llvm_amdgcn_raw_buffer_store_fp16x4(tmp.template get_as<fp16x4_t>()[number<0>{}],
dst_wave_buffer_resource,
dst_thread_addr_offset,
dst_wave_addr_offset,
static_cast<index_t>(coherence));
llvm_amdgcn_raw_buffer_store_fp16x4(tmp.template get_as<fp16x4_t>()[number<1>{}],
dst_wave_buffer_resource,
dst_thread_addr_offset,
dst_wave_addr_offset + 4 * sizeof(fp16_t),
static_cast<index_t>(coherence));
#else
llvm_amdgcn_raw_buffer_store_fp32x4(bit_cast<fp32x4_t>(src_thread_data),
dst_wave_buffer_resource,
dst_thread_addr_offset,
dst_wave_addr_offset,
static_cast<index_t>(coherence));
#endif
}
}
else if constexpr(std::is_same<T, bf16_t>::value) // bf16

20
include/ck_tile/core/container/array.hpp
View File

@@ -84,19 +84,6 @@ struct array
data[i] = static_cast<value_type>(c);
}
// template <typename Y>
// CK_TILE_HOST_DEVICE constexpr array(const array& o)
// {
// // static_assert(ArrayType::size() == size(), "wrong! size not the same");
// __content = o.__content;
// }
// CK_TILE_HOST_DEVICE constexpr array& operator=(const array& o)
// {
// // static_assert(ArrayType::size() == size(), "wrong! size not the same");
// __content = o.__content;
// return *this;
// }
CK_TILE_HOST_DEVICE static constexpr auto size() { return N; }
CK_TILE_HOST_DEVICE static constexpr bool is_static() { return is_static_v<value_type>; }
@@ -247,13 +234,6 @@ CK_TILE_HOST_DEVICE constexpr details::return_type<D, Ts...> make_array(Ts&&...
return {std::forward<Ts>(ts)...};
}
// // make empty array
// template <typename T>
// CK_TILE_HOST_DEVICE constexpr auto make_array()
// {
// return array<T, 0>{};
// }
// compatible with old ck's initializer, make an array and fill it withe the last element from
// initializer_list
template <typename T, index_t Size>

26
include/ck_tile/core/container/sequence.hpp
View File

@@ -480,32 +480,6 @@ struct sequence_split
using right_type = decltype(Seq::extract(range1{}));
};
#if 0
// reverse sequence
template <typename Seq>
struct sequence_reverse
{
static constexpr index_t NSize = Seq{}.size();
using seq_split = sequence_split<Seq, NSize / 2>;
using type = typename sequence_merge<
typename sequence_reverse<typename seq_split::right_type>::type,
typename sequence_reverse<typename seq_split::left_type>::type>::type;
};
template <index_t I>
struct sequence_reverse<sequence<I>>
{
using type = sequence<I>;
};
template <index_t I0, index_t I1>
struct sequence_reverse<sequence<I0, I1>>
{
using type = sequence<I1, I0>;
};
#endif
namespace detail {
template <typename Id, index_t... Ns>
struct seq_reverse;

14
include/ck_tile/core/container/statically_indexed_array.hpp
View File

@@ -24,18 +24,4 @@ using statically_indexed_array = array<T, N>;
#endif
// consider always use ck_tile::array for this purpose
#if 0
template <typename X, typename... Xs>
CK_TILE_HOST_DEVICE constexpr auto make_statically_indexed_array(const X& x, const Xs&... xs)
{
return statically_indexed_array<X, sizeof...(Xs) + 1>(x, static_cast<X>(xs)...);
}
// make empty statically_indexed_array
template <typename X>
CK_TILE_HOST_DEVICE constexpr auto make_statically_indexed_array()
{
return statically_indexed_array<X, 0>();
}
#endif
} // namespace ck_tile

31
include/ck_tile/core/container/thread_buffer.hpp
View File

@@ -23,18 +23,6 @@ CK_TILE_HOST_DEVICE constexpr auto make_thread_buffer(Ts&&... ts)
}
#else
#if 0
template <typename T, index_t N>
using thread_buffer = array<T, N>;
template <typename... Ts>
CK_TILE_HOST_DEVICE constexpr auto make_thread_buffer(Ts&&... ts)
{
return make_array(ts...);
}
#endif
// clang-format off
template<typename T_, index_t N_>
struct thread_buffer {
@@ -103,25 +91,6 @@ struct thread_buffer {
return vx.data;
}
#if 0
template <typename X_,
index_t Is,
typename std::enable_if<has_same_scalar_type<value_type, X_>::value, bool>::type = false>
CK_TILE_HOST_DEVICE constexpr void _set_as(number<Is> is, X_ x)
{
using X = remove_cvref_t<X_>;
constexpr index_t kSPerX = vector_traits<X>::vector_size;
union {
X_ data;
tuple_array<value_type, kSPerX> sub_data;
} vx {x};
static_for<0, kSPerX, 1>{}(
[&](auto j) { operator()((is * number<sizeof(X_)/sizeof(value_type)>{}) + j) = vx.sub_data[j]; });
}
#endif
#define TB_COMMON_AS() \

10
include/ck_tile/core/container/tuple.hpp
View File

@@ -292,9 +292,6 @@ struct tuple : impl::tuple_base<make_index_sequence<sizeof...(T)>, T...>
// below function should be used under tuple_array<> type, no extra check will perform here
template <typename Tx> CK_TILE_HOST_DEVICE constexpr decltype(auto) get_as() { return reinterpret_cast<tuple_array<Tx, size()>&>(*this); }
template <typename Tx> CK_TILE_HOST_DEVICE constexpr decltype(auto) get_as() const { return reinterpret_cast<const tuple_array<Tx, size()>&>(*this); }
// below index is for index *AFTER* type convert, not before
//template <typename Tx> CK_TILE_HOST_DEVICE constexpr decltype(auto) get_as(index_t i) { TP_COM_(); return reinterpret_cast<tuple_array<Tx, size()>&>(*this).at(i); }
//template <typename Tx> CK_TILE_HOST_DEVICE constexpr decltype(auto) get_as(index_t i) const { TP_COM_(); return reinterpret_cast<const tuple_array<Tx, size()>&>(*this).at(i); }
template <typename Tx, index_t I> CK_TILE_HOST_DEVICE constexpr decltype(auto) get_as(number<I>) { TP_COM_(); return reinterpret_cast<tuple_array<Tx, size()>&>(*this).at(number<I>{}); }
template <typename Tx, index_t I> CK_TILE_HOST_DEVICE constexpr decltype(auto) get_as(number<I>) const { TP_COM_(); return reinterpret_cast<const tuple_array<Tx, size()>&>(*this).at(number<I>{}); }
@@ -333,13 +330,6 @@ struct vector_traits<tuple<T...>, void>
static constexpr index_t vector_size = sizeof...(T);
};
// template <class... T>
// CK_TILE_HOST_DEVICE constexpr
// tuple<T...>
// make_tuple(T const&... t)
// {
// return {t...};
// }
template <typename... Xs>
CK_TILE_HOST_DEVICE constexpr bool operator==(const tuple<Xs...>& a, const tuple<Xs...>& b)
{

87
include/ck_tile/core/numeric/half.hpp
View File

@@ -264,93 +264,6 @@ bool operator>(const half_t& x, const half_t& y) { return __hgt(x.to_fp16(), y.t
CK_TILE_DEVICE
bool operator>=(const half_t& x, const half_t& y) { return __hge(x.to_fp16(), y.to_fp16()); }
#if 0
CK_TILE_DEVICE
half_t operator+(const half_t& x, const half_t& y)
{
return half_t(__hadd(x.to_fp16(), y.to_fp16()));
}
CK_TILE_DEVICE
half_t operator-(const half_t& x) { return half_t(__hneg(x.to_fp16())); }
CK_TILE_DEVICE
half_t operator-(const half_t& x, const half_t& y)
{
return half_t(__hsub(x.to_fp16(), y.to_fp16()));
}
CK_TILE_DEVICE
half_t operator*(const half_t& x, const half_t& y)
{
return half_t(__hmul(x.to_fp16(), y.to_fp16()));
}
CK_TILE_DEVICE
half_t operator/(const half_t& x, const half_t& y)
{
return half_t(__hdiv(x.to_fp16(), y.to_fp16()));
}
CK_TILE_DEVICE
half_t& operator+=(half_t& x, const half_t& y)
{
x = half_t(__hadd(x.to_fp16(), y.to_fp16()));
return x;
}
CK_TILE_DEVICE
half_t& operator-=(half_t& x, const half_t& y)
{
x = half_t(__hsub(x.to_fp16(), y.to_fp16()));
return x;
}
CK_TILE_DEVICE
half_t& operator*=(half_t& x, const half_t& y)
{
x = half_t(__hmul(x.to_fp16(), y.to_fp16()));
return x;
}
CK_TILE_DEVICE
half_t& operator/=(half_t& x, const half_t& y)
{
x = half_t(__hdiv(x.to_fp16(), y.to_fp16()));
return x;
}
CK_TILE_DEVICE
half_t& operator++(half_t& x)
{
x = half_t(__hadd(x.to_fp16(), half_t(1.0f).to_fp16()));
return x;
}
CK_TILE_DEVICE
half_t& operator--(half_t& x)
{
x = half_t(__hsub(x.to_fp16(), half_t(1.0f).to_fp16()));
return x;
}
CK_TILE_DEVICE
half_t operator++(half_t& x, int)
{
half_t y(x);
x = half_t(__hadd(x.to_fp16(), half_t(1.0f).to_fp16()));
return y;
}
CK_TILE_DEVICE
half_t operator--(half_t& x, int)
{
half_t y(x);
x = half_t(__hsub(x.to_fp16(), half_t(1.0f).to_fp16()));
return y;
}
#endif
#if CK_TILE_USE_CUSTOM_DATA_TYPE
CK_TILE_ARITHMETIC_USING_FLOAT(CK_TILE_HOST, half_t)
#endif

21
include/ck_tile/core/numeric/int8.hpp
View File

@@ -73,27 +73,6 @@ struct numeric<int8_t>
CK_TILE_HOST_DEVICE static constexpr int8_t zero() { return 0; }
};
#if 0
template <>
struct numeric_traits<int8_t>
{
static constexpr int exp = 5;
static constexpr int mant = 10;
static constexpr int bias = 15;
static constexpr uint16_t nan_mask = 0x7C00;
static constexpr uint16_t head_mask = 0xFC00;
static constexpr uint16_t mant_mask = 0x3FF;
static constexpr uint16_t exp_mask = 0x1F;
static constexpr uint32_t Inf = 0x7C00;
static constexpr uint32_t NegInf = 0xFC00;
static constexpr uint32_t NaN = 0x7C01;
static constexpr uint32_t Neg0 = 0x8000;
static constexpr int PackedSize = 1;
using bitwise_type = uint16_t;
};
#endif
CK_TILE_HOST_DEVICE
constexpr float int8_to_float(const int8_t& x) { return static_cast<float>(x); }

4
include/ck_tile/core/tensor/sweep_tile.hpp
View File

@@ -295,10 +295,6 @@ struct tile_sweeper
F f;
};
// partial deduction is not allowed
// template <typename T, typename F, typename U>
// tile_sweeper(const F&, U = {})->tile_sweeper<T, F, U>;
// deduction guide
template <typename T,
typename F,

39
include/ck_tile/core/tensor/tile_distribution.hpp
View File

@@ -454,45 +454,6 @@ struct tile_distribution_detail
} // namespace detail
#if 0
// this returns a constexpr tile_distribution
template <typename StaticTileDistributionEncoding_>
CK_TILE_HOST_DEVICE constexpr auto make_tile_distribution(StaticTileDistributionEncoding_)
{
using DstrEncode = remove_cvref_t<StaticTileDistributionEncoding_>;
constexpr auto adaptor_impl =
detail::make_adaptor_encoding_for_tile_distribution(StaticTileDistributionEncoding_{});
constexpr auto ps_ys_to_xs_adaptor_impl = adaptor_impl.template at<0>();
constexpr auto ys_to_d_adaptor_impl = adaptor_impl.template at<1>();
constexpr index_t d_length = adaptor_impl.template at<2>();
constexpr auto rh_major_minor_to_hidden_ids_impl = adaptor_impl.template at<3>();
constexpr auto ps_ys_to_xs_adaptor =
CONSTRUCT_TENSOR_ADAPTOR_FROM_ENCODING(ps_ys_to_xs_adaptor_impl);
constexpr auto ys_to_d_adaptor = CONSTRUCT_TENSOR_ADAPTOR_FROM_ENCODING(ys_to_d_adaptor_impl);
constexpr auto ys_to_d_descriptor =
make_tensor_descriptor_from_adaptor(ys_to_d_adaptor, d_length);
//
constexpr index_t ndim_rh_major = DstrEncode::detail::ndim_rh_major_;
constexpr auto ndims_rhs_minor = DstrEncode::detail::ndims_rhs_minor_;
constexpr auto rh_major_minor_to_hidden_ids =
TO_TUPLE_OF_SEQUENCE(rh_major_minor_to_hidden_ids_impl, ndim_rh_major, ndims_rhs_minor);
return tile_distribution<
remove_cvref_t<decltype(ps_ys_to_xs_adaptor)>,
remove_cvref_t<decltype(ys_to_d_descriptor)>,
remove_cvref_t<DstrEncode>,
detail::tile_distribution_detail<remove_cvref_t<decltype(rh_major_minor_to_hidden_ids)>>>{
ps_ys_to_xs_adaptor, ys_to_d_descriptor};
}
#endif
// this returns a static tile_distribution
template <typename StaticTileDistributionEncoding_>
CK_TILE_HOST_DEVICE constexpr auto make_static_tile_distribution(StaticTileDistributionEncoding_)

63
include/ck_tile/ops/elementwise/unary_element_wise_operation.hpp
View File

@@ -745,14 +745,6 @@ struct PassThroughPack2
template <typename Y, typename X>
CK_TILE_HOST_DEVICE void operator()(Y& y, const X& x) const;
#if 0
CK_TILE_HOST_DEVICE constexpr void operator()(ck_tile::fp16x2_t& y, const ck_tile::f8x2_t& x) const
{
auto t = type_convert<float2_t>(x);
y = type_convert<fp16x2_t>(t);
}
#endif
CK_TILE_HOST_DEVICE constexpr void operator()(fp16x2_t& y, const pk_int4_t& x) const
{
uint8_t x_u8 = bit_cast<uint8_t>(x);
@@ -871,61 +863,6 @@ struct UnaryConvert
}
};
#if 0
struct ConvertBF16RTN
{
// convert to bf16 using round to nearest (rtn)
template <typename Y, typename X>
CK_TILE_HOST_DEVICE void operator()(Y& y, const X& x) const
{
// check Y datatype
static_assert(std::is_same_v<Y, ck_tile::bf16_t>, "Data type is not supported by this operation!");
// check X datatype
static_assert(std::is_same_v<X, float> || std::is_same_v<X, ck_tile::fp16_t>,
"Data type is not supported by this operation!");
y = bf16_convert_rtn<Y>(x);
}
};
struct ConvertF8SR
{
// convert to fp8 using stochastic rounding (SR)
template <typename Y, typename X>
CK_TILE_HOST_DEVICE void operator()(Y& y, const X& x) const
{
// check Y datatype
static_assert(std::is_same_v<Y, ck_tile::fp8_t> || std::is_same_v<Y, ck_tile::bf8_t>,
"Data type is not supported by this operation!");
// check X datatype
static_assert(std::is_same_v<X, float> || std::is_same_v<X, ck_tile::fp16_t>,
"Data type is not supported by this operation!");
y = f8_convert_sr<Y>(x);
}
};
struct ConvertF8RNE
{
// convert to fp8 using rounding to nearest even
template <typename Y, typename X>
CK_TILE_HOST_DEVICE void operator()(Y& y, const X& x) const
{
// check Y datatype
static_assert(std::is_same_v<Y, ck_tile::fp8_t> || std::is_same_v<Y, ck_tile::bf8_t>,
"Data type is not supported by this operation!");
// check X datatype
static_assert(std::is_same_v<X, float> || std::is_same_v<X, ck_tile::fp16_t>,
"Data type is not supported by this operation!");
y = f8_convert_rne<Y>(x);
}
};
#endif
struct Scale
{
static constexpr const char* name = "Scale";

10
include/ck_tile/ops/flatmm/kernel/grouped_flatmm_kernel.hpp
View File

@@ -339,16 +339,6 @@ struct GroupedFlatmmKernel : FlatmmKernel<TilePartitioner_, FlatmmPipeline_, Epi
{
return hostArgs;
}
// CK_TILE_HOST static constexpr auto
// MakeKernelArgs(const ContiguousGroupedFlatmmHostArgs& hostArgs)
// {
// return hostArgs;
// }
// CK_TILE_HOST static constexpr auto
// MakeKernelArgs(const MaskedGroupedFlatmmHostArgs& hostArgs)
// {
// return hostArgs;
// }
template <class ScaleM = FlatmmScalePointer<-1>,
class ScaleN = FlatmmScalePointer<-1>,

7
include/ck_tile/ops/flatmm/kernel/moe_flatmm_kernel.hpp
View File

@@ -483,13 +483,6 @@ struct MoeFlatmmKernel
if constexpr(std::is_same_v<BLayout, tensor_layout::gemm::RowMajor>)
{
// if(kargs.N % TilePartitioner::NPerBlock != 0 && FlatmmPipeline::kPadN == false)
// {
// std::cerr << "Can't support N that is not a multiple of NPerBlock"
// " without padding!"
// << std::endl;
// return false;
// }
if(kargs.N % FlatmmPipeline::GetVectorSizeB() != 0)
{
std::cerr << "N is not a multiple of vector load size for B tensor!" << std::endl;

4
include/ck_tile/ops/flatmm/pipeline/flatmm_pipeline_agmem_bgmem_creg_v1_policy.hpp
View File

@@ -392,10 +392,6 @@ struct UniversalFlatmmPipelineAgBgCrPolicy
constexpr index_t M1 = BlockSize / get_warp_size();
static_assert(M2 != 0, "M2 is zero, which will lead to a division by zero error.");
static_assert(M1 != 0, "M1 is zero, which will lead to a division by zero error.");
// constexpr index_t M0 = MPerBlock / (M2 * M1);
// static_assert(M0 * M1 * M2 == MPerBlock,
// "Incorrect M0, M2, M1 configuration! "
// "M0, M1, M2 must cover whole MPerBlock!");
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,

9
include/ck_tile/ops/flatmm/pipeline/mixed_prec_flatmm_pipeline_agmem_bgmem_creg_v1.hpp
View File

@@ -1151,11 +1151,6 @@ struct F16xMXF4FlatmmPipelineAGmemBGmemCRegV1
a_warp_tensor(number<AwarpIter>{}) =
load_tile(a_warp_windows_pong(number<AmIter>{})(number<AkIter>{}));
}
// barrier
// if constexpr((kIter == KIterPerWarp - 1) && (mIter == MIter_2nd_last))
// {
// block_sync_lds();
// }
});
}
});
@@ -1636,10 +1631,6 @@ struct F8xMXF4FlatmmPipelineAGmemBGmemCRegV1
? Aload_rep
: 0;
}
// if((kIter % KPerScaleLoad == 0) && (mIter == 0))
// {
// load_perM = load_perM + 1;
// }
SchedulerPerM(dsread_perM, dswrite_perM, load_perM);
}
}

13
include/ck_tile/ops/flatmm/pipeline/moe_flatmm_pipeline_agmem_bgmem_creg.hpp
View File

@@ -103,13 +103,8 @@ struct MoeFlatmmPipelineAGmemBGmemCRegV1
static constexpr index_t Aload_num_perK = dswrite_num_perK;
static constexpr index_t Aload_rep = dswrite_rep;
static constexpr index_t Bload_num_perK = kNPerBlock * WG::kK / NWarp / BK1 / WaveSize;
// static constexpr index_t ScaleBload_K1 = ContinuousScaleNPerThread *
// ContinuousScaleKPerThread; static constexpr index_t ScaleBload_num =
// kNPerBlock * kKPerBlock / NWarp / 32 / ScaleBload_K1 /
// WaveSize; // BlockN * BlockK / NWarp / ScalePerK / ScaleB_K1 / wavesize
// static constexpr index_t KPerScaleLoad = KIterPerWarp / ScaleBload_num;
static constexpr index_t HalfMIter = (MIterPerWarp + 1) / 2;
static constexpr index_t Bload_rep = (Bload_num_perK + HalfMIter - 1) / HalfMIter;
static constexpr index_t HalfMIter = (MIterPerWarp + 1) / 2;
static constexpr index_t Bload_rep = (Bload_num_perK + HalfMIter - 1) / HalfMIter;
static constexpr index_t mfma_perM_perK = NIterPerWarp * mfma_per_wg;
static constexpr index_t dswrite_mIter = (DsWritePreIssue - 1) % MIterPerWarp;
@@ -352,10 +347,6 @@ struct MoeFlatmmPipelineAGmemBGmemCRegV1
? Aload_rep
: 0;
}
// if((kIter % KPerScaleLoad == 0) && (mIter == 0))
// {
// load_perM = load_perM + 1;
// }
SchedulerPerM(dsread_perM, dswrite_perM, load_perM);
}
}

4
include/ck_tile/ops/flatmm/pipeline/mx_flatmm_pipeline_agmem_bgmem_creg_v1.hpp
View File

@@ -390,10 +390,6 @@ struct MXFlatmmPipelineAGmemBGmemCRegV1 : FlatmmPipelineAGmemBGmemCRegV1<Problem
? Aload_rep
: 0;
}
// if((kIter % KPerScaleLoad == 0) && (mIter == 0))
// {
// load_perM = load_perM + 1;
// }
SchedulerPerM(dsread_perM, dswrite_perM, load_perM);
}
}

3
include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_qx_ks_vs_custom_policy.hpp
View File

@@ -692,9 +692,6 @@ struct BlockFmhaPipelineQXKSVSCustomPolicy : BlockFmhaPipelineQXCustomPolicy<QLo
constexpr index_t LaneGroups = WarpSize / LanesPerK; // within a wave
constexpr index_t NumIssues = kNPerBlock / (LaneGroups * NumWarps);
static_assert(NumIssues == kNPerBlock * kKPerBlock / (kBlockSize * KVector));
// constexpr index_t SingleKSize = NumIssues * NumWarps * (WarpSize * KVector + kPad);
// constexpr index_t SingleVSize =
// MakeVLdsBlockDescriptor<Problem>().get_element_space_size();
constexpr index_t BufferSize =
GetSingleSmemElementSpaceSize<Problem>(); // max(SingleKSize, SingleVSize);

138
include/ck_tile/ops/fused_moe/kernel/moe_sorting_kernel.hpp
View File

@@ -456,9 +456,6 @@ struct MoeSortingKernel
template <typename T, typename F, index_t wave_size_ = get_warp_size()>
__device__ static constexpr T wave_reduce(T local, F reduce_f, number<wave_size_> = {})
{
// constexpr int wave_size = 64;
// constexpr int reduce_stage = 6; // 1<<6=64
// clang-format off
constexpr int reduce_stage = [](){
if constexpr(wave_size_ == 2) return 1;
else if constexpr(wave_size_ == 4) return 2;
@@ -1206,17 +1203,21 @@ CK_TILE_HOST_DEVICE index_t moe_sorting_mp_sem_smem_size()
template <typename T, typename F, index_t wave_size_ = get_warp_size()>
CK_TILE_DEVICE constexpr T moe_sorting_wave_reduce(T local, F reduce_f, number<wave_size_> = {})
{
// constexpr int wave_size = 64;
// constexpr int reduce_stage = 6; // 1<<6=64
// clang-format off
constexpr int reduce_stage = [](){
if constexpr(wave_size_ == 2) return 1;
else if constexpr(wave_size_ == 4) return 2;
else if constexpr(wave_size_ == 8) return 3;
else if constexpr(wave_size_ == 16) return 4;
else if constexpr(wave_size_ == 32) return 5;
else if constexpr(wave_size_ == 64) return 6;
else return 0;
constexpr int reduce_stage = []() {
if constexpr(wave_size_ == 2)
return 1;
else if constexpr(wave_size_ == 4)
return 2;
else if constexpr(wave_size_ == 8)
return 3;
else if constexpr(wave_size_ == 16)
return 4;
else if constexpr(wave_size_ == 32)
return 5;
else if constexpr(wave_size_ == 64)
return 6;
else
return 0;
}();
// clang-format on
T v_local = local;
@@ -3047,53 +3048,6 @@ struct MoeSortingMultiPhaseKernel_P23
x_r = x_v;
#endif
{
#if 0
#pragma unroll
for(int j = 0; j < index_pack / 2; j++)
{
int i_token = i * kBlockSize * index_pack + threadIdx.x + j * kBlockSize;
index_t x = x_d[j];
int i_topk = x - 1; // topk of this token
int i_show = x != 0 ? 1 : 0; // has this token or not
int cumsum = i_show;
impl::moe_sorting_wave_cumsum<int, get_warp_size()>(cumsum);
__syncthreads();
if(lane_id == get_warp_size() - 1)
{
s[4 + wave_id] = cumsum;
}
__syncthreads();
// reduce cross wave
static_for<0, kBlockSize / get_warp_size() - 1, 1>{}([&](auto i_w) {
IndexType prev = s[4 + i_w];
prev = wave_id > i_w ? prev : 0; // mask out
cumsum += prev;
});
cumsum += prev_cumsum; // add previous round cumsum
if(threadIdx.x == kBlockSize - 1)
{
s[0] = cumsum;
}
__syncthreads();
int position = cumsum - i_show;
prev_cumsum = s[0]; // update the last cumsum
if(i_show)
{
#if CK_TILE_REFERENCE_MOE_SORTING_MOCK_ID
p_sorted_token_ids[e_start + position] =
MOE_SORTING_MOCK_ID(i_token, i_topk);
#else
p_sorted_token_ids[e_start + position] = i_token;
#endif
p_sorted_weights[e_start + position] =
p_weights[i_token * kargs.topk_mdiv.divisor + i_topk];
}
}
#endif
{
d_t i_topk;
d_t i_show;
@@ -3151,68 +3105,6 @@ struct MoeSortingMultiPhaseKernel_P23
}
position += i_show[j];
});
#if 0
int i_token = i * kBlockSize * index_pack + threadIdx.x * 2 + j * kBlockSize * 2;
index_t x = x_d[j];
index_t x0 = static_cast<index_t>(x & 0xffff);
index_t x1 = static_cast<index_t>(x >> 16);
int i_topk_0 = x0 - 1; // topk of this token
int i_show_0 = x0 != 0 ? 1 : 0; // has this token or not
int i_topk_1 = x1 - 1; // topk of this token
int i_show_1 = x1 != 0 ? 1 : 0; // has this token or not
int cumsum = i_show_0 + i_show_1;
impl::moe_sorting_wave_cumsum<int, get_warp_size()>(cumsum);
__syncthreads();
if(lane_id == get_warp_size() - 1)
{
s[4 + wave_id] = cumsum;
}
__syncthreads();
// reduce cross wave
static_for<0, kBlockSize / get_warp_size() - 1, 1>{}([&](auto i_w) {
IndexType prev = s[4 + i_w];
prev = wave_id > i_w ? prev : 0; // mask out
cumsum += prev;
});
cumsum += prev_cumsum; // add previous round cumsum
if(threadIdx.x == kBlockSize - 1)
{
s[0] = cumsum;
}
__syncthreads();
int position_0 = cumsum - i_show_0 - i_show_1;
prev_cumsum = s[0]; // update the last cumsum
if(i_show_0)
{
#if CK_TILE_REFERENCE_MOE_SORTING_MOCK_ID
p_sorted_token_ids[e_start + position_0] =
MOE_SORTING_MOCK_ID(i_token, i_topk_0);
#else
p_sorted_token_ids[e_start + position_0] = i_token;
#endif
p_sorted_weights[e_start + position_0] =
p_weights[i_token * kargs.topk_mdiv.divisor + i_topk_0];
}
int position_1 = cumsum - i_show_1;
if(i_show_1)
{
#if CK_TILE_REFERENCE_MOE_SORTING_MOCK_ID
p_sorted_token_ids[e_start + position_1] =
MOE_SORTING_MOCK_ID(i_token + 1, i_topk_1);
#else
p_sorted_token_ids[e_start + position_1] = i_token + 1;
#endif
p_sorted_weights[e_start + position_1] =
p_weights[(i_token + 1) * kargs.topk_mdiv.divisor + i_topk_1];
}
#endif
}
}
}

8
include/ck_tile/ops/fused_moe/pipeline/moe_sorting_pipeline.hpp
View File

@@ -14,14 +14,6 @@
namespace ck_tile {
// template <typename Problem_, typename Policy_ = MoeSortingPolicy>
// struct MoeSortingPipeline
// {
// // TODO: this kernel only support warp per row
// using Problem = remove_cvref_t<Problem_>;
// using Policy = remove_cvref_t<Policy_>;
// using WeightType = typename Problem::WeightType;
// template <typename TopkIdWindow, typename WeightWindow>
// CK_TILE_DEVICE auto operator()(const TopkIdWindow& topk_id_window,
// const WeightWindow& weight_window,

3
include/ck_tile/ops/gemm/block/block_gemm_areg_bsmem_creg_one_warp_v1.hpp
View File

@@ -36,9 +36,6 @@ struct BlockGemmARegBSmemCRegOneWarpV1
std::is_same_v<CDataType, remove_cv_t<typename CBlockTensor::DataType>>,
"wrong!");
// constexpr index_t MPerBlock = ABlockTensorTmp{}.get_lengths()[number<0>{}];
// constexpr index_t NPerBlock = BBlockWindowTmp{}.get_window_lengths()[number<0>{}];
// constexpr index_t KPerBlock = ABlockTensorTmp{}.get_lengths()[number<1>{}];
constexpr index_t MPerBlock = BlockGemmShape::kM;
constexpr index_t NPerBlock = BlockGemmShape::kN;
constexpr index_t KPerBlock = BlockGemmShape::kK;

23
include/ck_tile/ops/gemm/block/block_gemm_areg_bsmem_creg_v1_default_policy.hpp
View File

@@ -19,30 +19,7 @@ struct BlockGemmARegBSmemCRegV1DefaultPolicy
std::is_same_v<typename Problem::BDataType, half_t> &&
std::is_same_v<typename Problem::CDataType, float>)
{
#if 0
constexpr index_t kBlockSize = Problem::kBlockSize;
constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
constexpr index_t kNPerBlock = Problem::BlockGemmShape::kN;
constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
static_assert(kBlockSize % get_warp_size() == 0, "wrong!");
constexpr index_t NumWarp = kBlockSize / get_warp_size();
// FIXME
if constexpr(NumWarp == 4 && kMPerBlock % 128 == 0 &&
kNPerBlock % 128 == 0 % kKPerBlock % 16 == 0)
{
return make_tuple(WarpGemmMfmaF16F16F32M32N32K8{}, 4, 1);
}
else
{
return make_tuple(WarpGemmMfmaF16F16F32M32N32K8{}, 4, 1);
}
#else
return make_tuple(WarpGemmMfmaF16F16F32M32N32K8TransposedCDistribution{}, 4, 1);
#endif
}
else if constexpr(std::is_same_v<typename Problem::ADataType, bf16_t> &&
std::is_same_v<typename Problem::BDataType, bf16_t> &&

23
include/ck_tile/ops/gemm/block/block_gemm_areg_bsmem_creg_v2_default_policy.hpp
View File

@@ -16,30 +16,7 @@ struct BlockGemmARegBSmemCRegV2DefaultPolicy
CK_TILE_HOST_DEVICE static constexpr auto GetWarpGemmMWarpNWarp()
{
#if 0
constexpr index_t kBlockSize = Problem::kBlockSize;
constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
constexpr index_t kNPerBlock = Problem::BlockGemmShape::kN;
constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
static_assert(kBlockSize % get_warp_size() == 0, "wrong!");
constexpr index_t NumWarp = kBlockSize / get_warp_size();
// FIXME
if constexpr(NumWarp == 4 && kMPerBlock % 128 == 0 &&
kNPerBlock % 128 == 0 % kKPerBlock % 16 == 0)
{
return make_tuple(WarpGemmMfmaF16F16F32M32N32K8{}, 4, 1);
}
else
{
return make_tuple(WarpGemmMfmaF16F16F32M32N32K8{}, 4, 1);
}
#else
return make_tuple(WarpGemmMfmaF16F16F32M32N32K8TransposedCDistribution{}, 4, 1);
#endif
}
};

23
include/ck_tile/ops/gemm/block/block_gemm_asmem_breg_creg_v1_default_policy.hpp
View File

@@ -19,30 +19,7 @@ struct BlockGemmASmemBRegCRegV1DefaultPolicy
std::is_same_v<typename Problem::BDataType, half_t> &&
std::is_same_v<typename Problem::CDataType, float>)
{
#if 0
constexpr index_t kBlockSize = Problem::kBlockSize;
constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
constexpr index_t kNPerBlock = Problem::BlockGemmShape::kN;
constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
static_assert(kBlockSize % get_warp_size() == 0, "wrong!");
constexpr index_t NumWarp = kBlockSize / get_warp_size();
// FIXME
if constexpr(NumWarp == 4 && kMPerBlock % 128 == 0 &&
kNPerBlock % 128 == 0 % kKPerBlock % 16 == 0)
{
return make_tuple(WarpGemmMfmaF16F16F32M32N32K8{}, 4, 1);
}
else
{
return make_tuple(WarpGemmMfmaF16F16F32M32N32K8{}, 4, 1);
}
#else
return make_tuple(WarpGemmMfmaF16F16F32M32N32K8TransposedCDistribution{}, 4, 1);
#endif
}
else if constexpr(std::is_same_v<typename Problem::ADataType, bf16_t> &&
std::is_same_v<typename Problem::BDataType, bf16_t> &&

15
include/ck_tile/ops/norm_reduce/block/block_norm_reduce.hpp
View File

@@ -120,10 +120,6 @@ struct BlockNormReduceSync
constexpr index_t idim_p_lane = NDimP - 1;
// const auto ps_idx = make_array<index_t>(get_warp_id(), get_lane_id());
// const auto rs_idx =
// mean_tensor.get_tile_distribution().calculate_rs_index_from_ps_index(ps_idx);
constexpr index_t thread_buf_size = MeanDistributedTensor_::get_thread_buffer_size();
static_assert(thread_buf_size == VarDistributedTensor_::get_thread_buffer_size());
@@ -360,17 +356,6 @@ struct BlockNormReduceCrossWarpSync
template <typename BlockShape>
CK_TILE_DEVICE constexpr index_t block_tile_welford_calculate_max_count(int row_size)
{
#if 0
using S = BlockShape;
index_t LastloopN = row_size % S::Block_N == 0 ? S::Block_N : row_size % S::Block_N;
constexpr index_t NThread = S::WarpPerBlock_N * S::ThreadPerWarp_N;
index_t iNLane = get_thread_id() % NThread;
index_t iN0 = LastloopN / (S::Vector_N * S::ThreadPerWarp_N);
index_t iN1 = (LastloopN % (S::Vector_N * S::ThreadPerWarp_N)) / S::Vector_N;
index_t N2 = (LastloopN % (S::Vector_N * S::ThreadPerWarp_N)) % S::Vector_N;
index_t iN3 = iNLane < iN1 ? S::Vector_N : iNLane == iN1 ? N2 : 0;
return iN0 * S::Vector_N + iN3;
#endif
using S_ = BlockShape;
constexpr index_t ThreadsPerBlock_N = S_::WarpPerBlock_N * S_::ThreadPerWarp_N;

26
include/ck_tile/ops/reduce/block/block_reduce2d.hpp
View File

@@ -140,28 +140,6 @@ struct BlockReduce2d
ReducePacksPerXDim{});
}
#if 0
constexpr auto I0 = number<0>{};
constexpr auto I1 = number<1>{};
constexpr auto spans = XDistributedTensor_::get_distributed_spans();
// FIXME: hard coded to reduce 2nd axis
sweep_tile_span(spans[I0], [&](auto dstr_idx_i0) {
constexpr auto y_dstr_idx = make_tuple(dstr_idx_i0);
auto y = y_tensor[y_dstr_idx];
sweep_tile_span(spans[I1], [&](auto dstr_idx_i1) {
constexpr auto in_dstr_idx = make_tuple(dstr_idx_i0, dstr_idx_i1);
const auto x = ck_tile::type_convert<ComputeDataType>(x_tensor[in_dstr_idx]);
y = reduce_func(y, x);
});
y_tensor(y_dstr_idx) = y;
});
#endif
template <typename XDistributedTensor_>
CK_TILE_DEVICE static auto MakeYBlockTile()
{
@@ -240,10 +218,6 @@ struct BlockReduce2dSync
constexpr index_t idim_p_lane = NDimP - 1;
// const auto ps_idx = make_array<index_t>(get_warp_id(), get_lane_id());
// const auto rs_idx =
// y_tensor.get_tile_distribution().calculate_rs_index_from_ps_index(ps_idx);
constexpr index_t thread_buf_size = YDistributedTensor_::get_thread_buffer_size();
// loop over thread data