Skip to content

GitLab

Commit b6d3aa5d authored by Andriy Roshchenko's avatar Andriy Roshchenko
Browse files

Merge branch 'gfx950' into andriy/lwpck-2430

parents a634647d b6f7cddd
Hide whitespace changes
Inline Side-by-side
Showing with 906 additions and 163 deletions
example/ck_tile/CMakeLists.txt
View file @ b6d3aa5d
...@@ -12,3 +12,4 @@ add_subdirectory(09_topk_softmax) ...@@ -12,3 +12,4 @@ add_subdirectory(09_topk_softmax)
add_subdirectory(10_rmsnorm2d) add_subdirectory(10_rmsnorm2d)
add_subdirectory(11_add_rmsnorm2d_rdquant) add_subdirectory(11_add_rmsnorm2d_rdquant)
add_subdirectory(12_smoothquant) add_subdirectory(12_smoothquant)
add_subdirectory(13_moe_sorting)
include/ck/ck.hpp
View file @ b6d3aa5d
...@@ -63,13 +63,15 @@ CK_DECLARE_ENV_VAR_BOOL(CK_LOGGING) ...@@ -63,13 +63,15 @@ CK_DECLARE_ENV_VAR_BOOL(CK_LOGGING)
#define __gfx101__ #define __gfx101__
#endif #endif
#if defined(__gfx1030__) || defined(__gfx1031__) || defined(__gfx1032__) || \ #if defined(__gfx1030__) || defined(__gfx1031__) || defined(__gfx1032__) || \
defined(__gfx1034__) || defined(__gfx1035__) || defined(__gfx1036__) defined(__gfx1034__) || defined(__gfx1035__) || defined(__gfx1036__) || \
defined(__gfx10_3_generic__)
#define __gfx103__ #define __gfx103__
#endif #endif
#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__) #if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || \
defined(__gfx1103__) || defined(__gfx11_generic__)
#define __gfx11__ #define __gfx11__
#endif #endif
#if defined(__gfx1200__) || defined(__gfx1201__) #if defined(__gfx1200__) || defined(__gfx1201__) || defined(__gfx12_generic__)
#define __gfx12__ #define __gfx12__
#endif #endif
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_wmma_cshuffle.hpp
View file @ b6d3aa5d
...@@ -381,10 +381,6 @@ struct DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle ...@@ -381,10 +381,6 @@ struct DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle
{ {
tildes = {i_ztilde, i_ytilde, i_xtilde}; tildes = {i_ztilde, i_ytilde, i_xtilde};
} }
else
{
throw std::runtime_error("wrong! only implemented for 2D and 3D now");
}
const auto a_grid_desc_ak0_m_ak1 = const auto a_grid_desc_ak0_m_ak1 =
transform_conv_to_gemm.template MakeADescriptor_AK0_M_AK1<ALayout>( transform_conv_to_gemm.template MakeADescriptor_AK0_M_AK1<ALayout>(
...@@ -750,6 +746,12 @@ struct DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle ...@@ -750,6 +746,12 @@ struct DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle
} }
} }
// check number of dimension, only implemented for 2D and 3D now
if(NDimSpatial != 2 && NDimSpatial != 3)
{
return false;
}
return true; return true;
} }
......
include/ck/utility/amd_wmma.hpp
View file @ b6d3aa5d
...@@ -9,7 +9,8 @@ ...@@ -9,7 +9,8 @@
// TODO: Add arch limitation // TODO: Add arch limitation
namespace ck { namespace ck {
#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__) #if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || \
defined(__gfx1103__) || defined(__gfx11_generic__)
#define __gfx11__ #define __gfx11__
#endif #endif
/********************************WAVE32 MODE***********************************************/ /********************************WAVE32 MODE***********************************************/
...@@ -260,7 +261,7 @@ struct intrin_wmma_i32_16x16x16_iu8_w64<16, 16, neg_a, neg_b, clamp> ...@@ -260,7 +261,7 @@ struct intrin_wmma_i32_16x16x16_iu8_w64<16, 16, neg_a, neg_b, clamp>
// gfx12 // gfx12
/********************************WAVE32 MODE***********************************************/ /********************************WAVE32 MODE***********************************************/
#if defined(__gfx1200__) || defined(__gfx1201__) #if defined(__gfx1200__) || defined(__gfx1201__) || defined(__gfx12_generic__)
#define __gfx12__ #define __gfx12__
#endif #endif
......
include/ck_tile/core/config.hpp
View file @ b6d3aa5d
...@@ -11,13 +11,15 @@ ...@@ -11,13 +11,15 @@
#define __gfx94__ #define __gfx94__
#endif #endif
#if defined(__gfx1030__) || defined(__gfx1031__) || defined(__gfx1032__) || \ #if defined(__gfx1030__) || defined(__gfx1031__) || defined(__gfx1032__) || \
defined(__gfx1034__) || defined(__gfx1035__) || defined(__gfx1036__) defined(__gfx1034__) || defined(__gfx1035__) || defined(__gfx1036__) || \
defined(__gfx10_3_generic__)
#define __gfx103__ #define __gfx103__
#endif #endif
#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__) #if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || \
defined(__gfx1103__) || defined(__gfx11_generic__)
#define __gfx11__ #define __gfx11__
#endif #endif
#if defined(__gfx1200__) || defined(__gfx1201__) #if defined(__gfx1200__) || defined(__gfx1201__) || defined(__gfx12_generic__)
#define __gfx12__ #define __gfx12__
#endif #endif
......
include/ck_tile/core/tensor/shuffle_tile.hpp
View file @ b6d3aa5d
...@@ -170,7 +170,7 @@ CK_TILE_DEVICE void shuffle_tile(OutTensor& out, const InTensor& in) ...@@ -170,7 +170,7 @@ CK_TILE_DEVICE void shuffle_tile(OutTensor& out, const InTensor& in)
} }
else else
{ {
// NOT implemented static_assert(false, "The shuffle should always happen!");
} }
} }
......
include/ck_tile/host.hpp
View file @ b6d3aa5d
...@@ -23,6 +23,7 @@ ...@@ -23,6 +23,7 @@
#include "ck_tile/host/reference/reference_gemm.hpp" #include "ck_tile/host/reference/reference_gemm.hpp"
#include "ck_tile/host/reference/reference_im2col.hpp" #include "ck_tile/host/reference/reference_im2col.hpp"
#include "ck_tile/host/reference/reference_layernorm2d_fwd.hpp" #include "ck_tile/host/reference/reference_layernorm2d_fwd.hpp"
#include "ck_tile/host/reference/reference_moe_sorting.hpp"
#include "ck_tile/host/reference/reference_permute.hpp" #include "ck_tile/host/reference/reference_permute.hpp"
#include "ck_tile/host/reference/reference_reduce.hpp" #include "ck_tile/host/reference/reference_reduce.hpp"
#include "ck_tile/host/reference/reference_rmsnorm2d_fwd.hpp" #include "ck_tile/host/reference/reference_rmsnorm2d_fwd.hpp"
......
include/ck_tile/host/reference/reference_moe_sorting.hpp 0 → 100644
View file @ b6d3aa5d
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/host/host_tensor.hpp"
namespace ck_tile {
template <typename WeightType, typename IndexType = index_t>
CK_TILE_HOST void reference_moe_sorting(const HostTensor<IndexType>& topk_ids,
const HostTensor<WeightType>& weights,
HostTensor<IndexType>& p_sorted_token_ids,
HostTensor<WeightType>& sorted_weight,
HostTensor<IndexType>& sorted_expert_ids,
index_t& unit_cnt,
const index_t experts,
const index_t unit_size)
{
const index_t num_token = topk_ids.mDesc.get_lengths()[0];
const index_t topk = topk_ids.mDesc.get_lengths()[1];
std::vector<std::vector<IndexType>> expert_tokens(experts,
std::vector<IndexType>(unit_size, num_token));
std::vector<std::vector<WeightType>> expert_token_weights(
experts, std::vector<WeightType>(unit_size, 0));
std::vector<IndexType> expert_slices(experts, 1);
std::vector<IndexType> expert_slice_idxs(experts, 0);
for(index_t t = 0; t < num_token; t++)
{
for(index_t k = 0; k < topk; k++)
{
IndexType e = topk_ids(t, k);
WeightType w = weights(t, k);
index_t idx = expert_slice_idxs[e];
if(idx > expert_slices[e] * unit_size - 1)
{
expert_slices[e]++;
index_t new_size = expert_slices[e] * unit_size;
expert_tokens[e].resize(new_size);
expert_token_weights[e].resize(new_size);
for(index_t i = (expert_slices[e] - 1) * unit_size; i < new_size; i++)
{
expert_tokens[e][i] = num_token;
expert_token_weights[e][i] = 0;
}
}
expert_tokens[e][idx] = t;
expert_token_weights[e][idx] = w;
expert_slice_idxs[e]++;
}
}
IndexType* out_tokens = p_sorted_token_ids.data();
WeightType* out_weights = sorted_weight.data();
IndexType* out_expert_id = sorted_expert_ids.data();
for(index_t e = 0; e < experts; e++)
{
memcpy(out_tokens, expert_tokens[e].data(), sizeof(index_t) * expert_slices[e] * unit_size);
out_tokens += expert_slices[e] * unit_size;
memcpy(out_weights,
expert_token_weights[e].data(),
sizeof(WeightType) * expert_slices[e] * unit_size);
out_weights += expert_slices[e] * unit_size;
for(index_t s = 0; s < expert_slices[e]; s++)
{
out_expert_id[s] = e;
unit_cnt++;
}
out_expert_id += expert_slices[e];
}
unit_cnt *= unit_size;
return;
}
} // namespace ck_tile
include/ck_tile/ops/common/generic_2d_block_shape.hpp
View file @ b6d3aa5d
...@@ -38,9 +38,7 @@ namespace ck_tile { ...@@ -38,9 +38,7 @@ namespace ck_tile {
template <typename BlockTile_, // block size, seq<M, N> template <typename BlockTile_, // block size, seq<M, N>
typename WarpPerBlock_, // num warps along seq<M, N> typename WarpPerBlock_, // num warps along seq<M, N>
typename WarpTile_, // warp size, seq<M, N> typename WarpTile_, // warp size, seq<M, N>
typename Vector_, // contiguous pixels(vector size) along seq<M, N> typename Vector_> // contiguous pixels(vector size) along seq<M, N>)>
index_t BlockSize_ =
warpSize* reduce_on_sequence(WarpPerBlock_{}, multiplies{}, number<1>{})>
struct Generic2dBlockShape struct Generic2dBlockShape
{ {
// block size // block size
...@@ -68,10 +66,12 @@ struct Generic2dBlockShape ...@@ -68,10 +66,12 @@ struct Generic2dBlockShape
static_assert(Warp_M % Vector_M == 0); static_assert(Warp_M % Vector_M == 0);
static_assert(Warp_N % Vector_N == 0); static_assert(Warp_N % Vector_N == 0);
// num of threads along seq<M, N>, within each warp // num of threads along seq<M, N>, within each warp
static constexpr index_t ThreadPerWarp_M = Warp_M / Vector_M; static constexpr index_t ThreadPerWarp_M = Warp_M / Vector_M;
static constexpr index_t ThreadPerWarp_N = Warp_N / Vector_N; static constexpr index_t ThreadPerWarp_N = Warp_N / Vector_N;
static constexpr index_t ThreadPerBlock_M = Block_M / Repeat_M / Vector_M;
static constexpr index_t ThreadPerBlock_N = Block_N / Repeat_N / Vector_N;
static constexpr index_t BlockSize = BlockSize_; static constexpr index_t BlockSize = ThreadPerBlock_M * ThreadPerBlock_N;
}; };
} // namespace ck_tile } // namespace ck_tile
include/ck_tile/ops/fmha/block/page_block_navigator.hpp
View file @ b6d3aa5d
...@@ -230,7 +230,15 @@ struct PageBlockNavigator ...@@ -230,7 +230,15 @@ struct PageBlockNavigator
CK_TILE_HOST_DEVICE CK_TILE_HOST_DEVICE
DataType* get_block_ptr(index_t block_index) const DataType* get_block_ptr(index_t block_index) const
{ {
return physical_blocks + physical_block_indices[block_index] * block_stride + fixed_offset; if(block_index < num_blocks)
{
return physical_blocks + physical_block_indices[block_index] * block_stride +
fixed_offset;
}
else
{
return nullptr;
}
} }
CK_TILE_HOST_DEVICE int32_t get_block_index(const WindowOrigin& global_window_origin) const CK_TILE_HOST_DEVICE int32_t get_block_index(const WindowOrigin& global_window_origin) const
......
include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_qx_ks_vs_custom_policy.hpp
View file @ b6d3aa5d
...@@ -863,6 +863,8 @@ struct BlockFmhaPipelineQXKSVSCustomPolicy : BlockFmhaPipelineQXCustomPolicy<QLo ...@@ -863,6 +863,8 @@ struct BlockFmhaPipelineQXKSVSCustomPolicy : BlockFmhaPipelineQXCustomPolicy<QLo
constexpr index_t K0 = kKPerBlock / K1; constexpr index_t K0 = kKPerBlock / K1;
constexpr index_t N2 = get_warp_size() / K0; constexpr index_t N2 = get_warp_size() / K0;
constexpr index_t N1 = kBlockSize / get_warp_size(); constexpr index_t N1 = kBlockSize / get_warp_size();
static_assert(N2 != 0, "N2 is zero, which will lead to a division by zero error.");
static_assert(N1 != 0, "N1 is zero, which will lead to a division by zero error.");
constexpr index_t N0 = kNPerBlock / (N2 * N1); constexpr index_t N0 = kNPerBlock / (N2 * N1);
static_assert(N0 != 0); static_assert(N0 != 0);
......
include/ck_tile/ops/fused_moe/kernel/moe_sorting_kernel.hpp 0 → 100644
View file @ b6d3aa5d
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/ops/common.hpp"
#include "ck_tile/ops/elementwise.hpp"
#include "ck_tile/host/hip_check_error.hpp"
#include <string>
#include <type_traits>
namespace ck_tile {
struct MoeSortingHostArgs
{
const void* p_topk_ids;
const void* p_weights;
void* p_sorted_token_ids;
void* p_sorted_weights;
void* p_sorted_expert_ids;
void* p_total_tokens_post_pad;
void* p_moe_buf;
index_t tokens;
index_t unit_size;
index_t num_experts;
index_t topk;
index_t moe_buf_bytes;
};
template <typename Problem_>
struct MoeSortingKernel
{
using Problem = remove_cvref_t<Problem_>;
using IndexType = typename Problem::IndexType;
using WeightType = typename Problem::WeightType;
typedef MoeSortingHostArgs MoeSortingKargs;
using Hargs = MoeSortingHostArgs;
struct Kargs
{
const void* p_topk_ids;
const void* p_weights;
void* p_sorted_token_ids;
void* p_sorted_weights;
void* p_sorted_expert_ids;
void* p_total_tokens_post_pad;
void* p_moe_buf;
index_t tokens;
index_t num_experts;
index_t moe_buf_bytes;
index_t tokens_per_thread;
mdiv unit_size_mdiv;
mdiv topk_mdiv;
};
CK_TILE_HOST static constexpr auto GridSize(const Hargs& h)
{
// TODO: assume num-experts not too much
return dim3(1 + ck_tile::integer_divide_ceil(h.moe_buf_bytes, BlockSize(h).x * 16));
}
CK_TILE_HOST static constexpr auto BlockSize(const Hargs& h)
{
return dim3(ck_tile::integer_least_multiple(h.num_experts, ck_tile::get_warp_size()));
}
// in byte
CK_TILE_HOST static constexpr auto GetSmemSize(const Hargs& h)
{
const auto blocks = BlockSize(h);
return ((blocks.x + 1) * h.num_experts + (h.num_experts + 1)) * sizeof(index_t);
}
CK_TILE_HOST static constexpr auto MakeKargs(const Hargs& h)
{
Kargs k;
k.p_topk_ids = h.p_topk_ids;
k.p_weights = h.p_weights;
k.p_sorted_token_ids = h.p_sorted_token_ids;
k.p_sorted_weights = h.p_sorted_weights;
k.p_sorted_expert_ids = h.p_sorted_expert_ids;
k.p_moe_buf = h.p_moe_buf;
k.p_total_tokens_post_pad = h.p_total_tokens_post_pad;
k.tokens = h.tokens;
k.num_experts = h.num_experts;
k.moe_buf_bytes = h.moe_buf_bytes;
const auto blocks = BlockSize(h);
k.tokens_per_thread = integer_divide_ceil(h.tokens * h.topk, blocks.x);
k.unit_size_mdiv = mdiv{static_cast<uint32_t>(h.unit_size)};
k.topk_mdiv = mdiv{static_cast<uint32_t>(h.topk)};
return k;
}
CK_TILE_DEVICE index_t calc_index(index_t total_col, index_t row, index_t col) const
{
return row * total_col + col;
}
CK_TILE_DEVICE void moe_buf_set_zero_kernel(uint8x16_t* buf, index_t buf_bytes) const
{
const index_t offset = (blockIdx.x - 1) * blockDim.x + threadIdx.x;
if(offset < buf_bytes / 16)
{
buf[offset] = uint8x16_t{0};
}
}
CK_TILE_DEVICE void moe_align_block_size_kernel(const IndexType* __restrict__ topk_id,
const WeightType* __restrict__ weights,
index_t* p_sorted_token_ids,
WeightType* p_sorted_weights,
index_t* p_sorted_expert_ids,
index_t* p_total_tokens_post_pad,
const index_t num_experts,
const index_t tokens_per_thread,
const index_t numel,
const mdiv unit_size_mdiv,
const mdiv topk_mdiv,
void* smem) const
{
const index_t tid = static_cast<index_t>(threadIdx.x);
const index_t start_idx = tid * tokens_per_thread;
index_t* shared_mem = reinterpret_cast<index_t*>(smem);
index_t* tokens_cnts = shared_mem; // 2d: (blockDim.x + 1, num_experts)
index_t* cumsum = shared_mem + (blockDim.x + 1) * num_experts; // 1: (num_experts + 1)
for(int i = 0; i < num_experts; ++i)
{
tokens_cnts[calc_index(num_experts, tid + 1, i)] = 0;
}
#pragma unroll Problem_::InternalLoadUnroll
for(int i = start_idx; i < numel && i < start_idx + tokens_per_thread; ++i)
{
++tokens_cnts[calc_index(num_experts, tid + 1, topk_id[i])];
}
__syncthreads();
if(tid < num_experts)
{
tokens_cnts[calc_index(num_experts, 0, tid)] = 0;
for(int i = 1; i <= static_cast<index_t>(blockDim.x); ++i)
{
tokens_cnts[calc_index(num_experts, i, tid)] +=
tokens_cnts[calc_index(num_experts, i - 1, tid)];
}
}
// __syncthreads();
if(tid == 0)
{
cumsum[0] = 0;
for(int i = 1; i <= num_experts; ++i)
{
auto current_units = [&]() {
index_t x_ = tokens_cnts[calc_index(num_experts, blockDim.x, i - 1)] +
unit_size_mdiv.divisor - 1;
index_t y_ = unit_size_mdiv.div(x_);
return max(y_, 1) * unit_size_mdiv.divisor;
}();
cumsum[i] = cumsum[i - 1] + current_units;
}
*p_total_tokens_post_pad = cumsum[num_experts];
}
__syncthreads();
if(tid < num_experts)
{
for(int i = cumsum[tid]; i < cumsum[tid + 1]; i += unit_size_mdiv.divisor)
{
p_sorted_expert_ids[unit_size_mdiv.div(i)] = tid;
}
}
#pragma unroll Problem_::InternalLoadUnroll
for(int i = start_idx; i < numel && i < start_idx + tokens_per_thread; ++i)
{
index_t expert_id = topk_id[i];
index_t rank_post_pad =
tokens_cnts[calc_index(num_experts, tid, expert_id)] + cumsum[expert_id];
p_sorted_token_ids[rank_post_pad] = topk_mdiv.div(i);
p_sorted_weights[rank_post_pad] = weights[i];
++tokens_cnts[calc_index(num_experts, tid, expert_id)];
}
const index_t prefill_token = topk_mdiv.div(numel);
if(tid < num_experts)
{
index_t expert_offset =
cumsum[tid] + tokens_cnts[calc_index(num_experts, blockDim.x, tid)];
while(expert_offset < cumsum[tid + 1])
{
p_sorted_token_ids[expert_offset] = prefill_token;
p_sorted_weights[expert_offset] = static_cast<WeightType>(0.0);
expert_offset++;
}
}
}
CK_TILE_DEVICE void operator()(Kargs kargs) const
{
if(blockIdx.x > 0)
{
if(kargs.p_moe_buf)
{
moe_buf_set_zero_kernel(reinterpret_cast<uint8x16_t*>(kargs.p_moe_buf),
kargs.moe_buf_bytes);
}
return;
}
const size_t numel = kargs.tokens * kargs.topk_mdiv.divisor;
extern __shared__ char smem[];
return moe_align_block_size_kernel(static_cast<const IndexType*>(kargs.p_topk_ids),
static_cast<const WeightType*>(kargs.p_weights),
static_cast<IndexType*>(kargs.p_sorted_token_ids),
static_cast<WeightType*>(kargs.p_sorted_weights),
static_cast<IndexType*>(kargs.p_sorted_expert_ids),
static_cast<IndexType*>(kargs.p_total_tokens_post_pad),
kargs.num_experts,
kargs.tokens_per_thread,
numel,
kargs.unit_size_mdiv,
kargs.topk_mdiv,
smem);
}
};
} // namespace ck_tile
include/ck_tile/ops/fused_moe/pipeline/moe_sorting_pipeline.hpp 0 → 100644
View file @ b6d3aa5d
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/ops/fused_moe/pipeline/moe_sorting_policy.hpp"
#include <string>
#include <type_traits>
#ifndef TOPK_SOFTMAX_USE_RAW_TILE_WINDOW
#define TOPK_SOFTMAX_USE_RAW_TILE_WINDOW 0
#endif
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,
// index_t* p_sorted_token_ids,
// WeightType* p_sorted_weights,
// index_t* p_sorted_expert_ids,
// index_t* p_total_tokens_post_pad,
// const index_t num_experts,
// const index_t unit_size,
// const size_t numel,
// const index_t topk)
// {
// }
// };
} // namespace ck_tile
include/ck_tile/ops/fused_moe/pipeline/moe_sorting_policy.hpp 0 → 100644
View file @ b6d3aa5d
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/ops/softmax.hpp"
#include "ck_tile/ops/topk.hpp"
namespace ck_tile {
struct MoeSortingPolicy
{
};
} // namespace ck_tile
include/ck_tile/ops/fused_moe/pipeline/moe_sorting_problem.hpp 0 → 100644
View file @ b6d3aa5d
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include <string>
#include <type_traits>
namespace ck_tile {
template <typename IndexType_, typename WeightType_, index_t InternalLoadUnroll_>
struct MoeSortingProblem
{
// TODO: this kernel only support warp per row
using WeightType = remove_cvref_t<WeightType_>;
using IndexType = remove_cvref_t<IndexType_>;
static constexpr index_t WarpSize = get_warp_size();
static constexpr index_t WarpsPerBlock = 1;
static constexpr index_t InternalLoadUnroll = InternalLoadUnroll_;
};
} // namespace ck_tile
include/ck_tile/ops/gemm/kernel/gemm_kernel.hpp
View file @ b6d3aa5d
...@@ -115,12 +115,22 @@ struct GemmKernel ...@@ -115,12 +115,22 @@ struct GemmKernel
} }
}(); }();
auto a_pad_view = pad_tensor_view( auto a_pad_view = [&]() {
a_tensor_view, if constexpr(std::is_same_v<ALayout, tensor_layout::gemm::RowMajor>)
make_tuple(number<TilePartitioner::kM>{}, number<TilePartitioner::kK>{}), {
// somehow clang-format is splitting below line into multiple. return pad_tensor_view(
// clang-format off a_tensor_view,
sequence<false, GemmPipeline::kPadA>{}); make_tuple(number<TilePartitioner::kM>{}, number<TilePartitioner::kK>{}),
sequence<false, GemmPipeline::kPadK>{});
}
else
{
return pad_tensor_view(
a_tensor_view,
make_tuple(number<TilePartitioner::kM>{}, number<TilePartitioner::kK>{}),
sequence<GemmPipeline::kPadM, false>{});
}
}();
// clang-format on // clang-format on
auto a_block_window = make_tile_window( auto a_block_window = make_tile_window(
...@@ -128,12 +138,22 @@ struct GemmKernel ...@@ -128,12 +138,22 @@ struct GemmKernel
make_tuple(number<TilePartitioner::kM>{}, number<TilePartitioner::kK>{}), make_tuple(number<TilePartitioner::kM>{}, number<TilePartitioner::kK>{}),
{i_m, 0}); {i_m, 0});
auto b_pad_view = pad_tensor_view( auto b_pad_view = [&]() {
b_tensor_view, if constexpr(std::is_same_v<BLayout, tensor_layout::gemm::ColumnMajor>)
make_tuple(number<TilePartitioner::kN>{}, number<TilePartitioner::kK>{}), {
// clang-format off return pad_tensor_view(
sequence<false, GemmPipeline::kPadB>{}); b_tensor_view,
// clang-format on make_tuple(number<TilePartitioner::kN>{}, number<TilePartitioner::kK>{}),
sequence<false, GemmPipeline::kPadK>{});
}
else
{
return pad_tensor_view(
b_tensor_view,
make_tuple(number<TilePartitioner::kN>{}, number<TilePartitioner::kK>{}),
sequence<GemmPipeline::kPadN, false>{});
}
}();
auto b_block_window = make_tile_window( auto b_block_window = make_tile_window(
b_pad_view, b_pad_view,
...@@ -171,18 +191,28 @@ struct GemmKernel ...@@ -171,18 +191,28 @@ struct GemmKernel
} }
}(); }();
auto c_pad_view = pad_tensor_view( auto c_pad_view = [&]() {
c_tensor_view, if constexpr(std::is_same_v<CLayout, tensor_layout::gemm::RowMajor>)
make_tuple(number<TilePartitioner::kM>{}, number<TilePartitioner::kN>{}), {
// clang-format off return pad_tensor_view(
sequence<false, GemmPipeline::kPadC>{}); c_tensor_view,
// clang-format on make_tuple(number<TilePartitioner::kM>{}, number<TilePartitioner::kN>{}),
auto c_block_window = make_tile_window( sequence<false, GemmPipeline::kPadN>{});
}
else
{
return pad_tensor_view(
c_tensor_view,
make_tuple(number<TilePartitioner::kM>{}, number<TilePartitioner::kN>{}),
sequence<GemmPipeline::kPadM, false>{});
}
}();
auto CBlockWindow_pad = make_tile_window(
c_pad_view, c_pad_view,
make_tuple(number<TilePartitioner::kM>{}, number<TilePartitioner::kN>{}), make_tuple(number<TilePartitioner::kM>{}, number<TilePartitioner::kN>{}),
{i_m, i_n}); {i_m, i_n});
EpiloguePipeline{}(c_block_window, c_block_tile); EpiloguePipeline{}(CBlockWindow_pad, c_block_tile);
} }
}; };
......
include/ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_mem.hpp
View file @ b6d3aa5d
...@@ -113,9 +113,9 @@ struct GemmPipelineAgBgCrMem : public BaseGemmPipelineAgBgCrMem<Problem> ...@@ -113,9 +113,9 @@ struct GemmPipelineAgBgCrMem : public BaseGemmPipelineAgBgCrMem<Problem>
static constexpr index_t VectorSizeB = Problem::VectorSizeB; static constexpr index_t VectorSizeB = Problem::VectorSizeB;
static constexpr index_t VectorSizeC = Problem::VectorSizeC; static constexpr index_t VectorSizeC = Problem::VectorSizeC;
static constexpr bool kPadA = Problem::kPadA; static constexpr bool kPadM = Problem::kPadM;
static constexpr bool kPadB = Problem::kPadB; static constexpr bool kPadN = Problem::kPadN;
static constexpr bool kPadC = Problem::kPadC; static constexpr bool kPadK = Problem::kPadK;
// Where is the right place for HasHotLoop and TailNum ??? // Where is the right place for HasHotLoop and TailNum ???
static constexpr bool HasHotLoop = Problem::HasHotLoop; static constexpr bool HasHotLoop = Problem::HasHotLoop;
......
include/ck_tile/ops/gemm/pipeline/gemm_pipeline_agmem_bgmem_creg_v1.hpp
View file @ b6d3aa5d
...@@ -33,9 +33,9 @@ struct GemmPipelineAGmemBGmemCRegV1 ...@@ -33,9 +33,9 @@ struct GemmPipelineAGmemBGmemCRegV1
static constexpr index_t VectorSizeB = Problem::VectorSizeB; static constexpr index_t VectorSizeB = Problem::VectorSizeB;
static constexpr index_t VectorSizeC = Problem::VectorSizeC; static constexpr index_t VectorSizeC = Problem::VectorSizeC;
static constexpr bool kPadA = Problem::kPadA; static constexpr bool kPadM = Problem::kPadM;
static constexpr bool kPadB = Problem::kPadB; static constexpr bool kPadN = Problem::kPadN;
static constexpr bool kPadC = Problem::kPadC; static constexpr bool kPadK = Problem::kPadK;
CK_TILE_HOST_DEVICE static constexpr index_t GetStaticLdsSize() CK_TILE_HOST_DEVICE static constexpr index_t GetStaticLdsSize()
{ {
...@@ -101,11 +101,8 @@ struct GemmPipelineAGmemBGmemCRegV1 ...@@ -101,11 +101,8 @@ struct GemmPipelineAGmemBGmemCRegV1
Policy::template MakeADramTileDistribution<Problem>()); Policy::template MakeADramTileDistribution<Problem>());
// A LDS tile window for store // A LDS tile window for store
auto a_copy_lds_window = auto a_copy_lds_window = make_tile_window(
make_tile_window(a_lds_block, a_lds_block, make_tuple(number<kMPerBlock>{}, number<kKPerBlock>{}), {0, 0});
make_tuple(number<kMPerBlock>{}, number<kKPerBlock>{}),
{0, 0},
a_copy_dram_window.get_tile_distribution());
// B DRAM tile window for load // B DRAM tile window for load
auto b_copy_dram_window = auto b_copy_dram_window =
...@@ -115,11 +112,8 @@ struct GemmPipelineAGmemBGmemCRegV1 ...@@ -115,11 +112,8 @@ struct GemmPipelineAGmemBGmemCRegV1
Policy::template MakeBDramTileDistribution<Problem>()); Policy::template MakeBDramTileDistribution<Problem>());
// B LDS tile window for store // B LDS tile window for store
auto b_copy_lds_window = auto b_copy_lds_window = make_tile_window(
make_tile_window(b_lds_block, b_lds_block, make_tuple(number<kNPerBlock>{}, number<kKPerBlock>{}), {0, 0});
make_tuple(number<kNPerBlock>{}, number<kKPerBlock>{}),
{0, 0},
b_copy_dram_window.get_tile_distribution());
// A LDS tile for block GEMM // A LDS tile for block GEMM
auto a_lds_gemm_window = make_tile_window( auto a_lds_gemm_window = make_tile_window(
...@@ -149,12 +143,32 @@ struct GemmPipelineAGmemBGmemCRegV1 ...@@ -149,12 +143,32 @@ struct GemmPipelineAGmemBGmemCRegV1
tile_elementwise_inout([](auto& c) { c = 0; }, c_block_tile); tile_elementwise_inout([](auto& c) { c = 0; }, c_block_tile);
// LDS write 0 // LDS write 0
const auto a_block_tile_tmp = tile_elementwise_in(a_element_func, a_block_tile); if constexpr(std::is_same_v<ALayout, tensor_layout::gemm::ColumnMajor>)
store_tile(a_copy_lds_window, a_block_tile_tmp); {
auto a_shuffle_tmp = make_static_distributed_tensor<ADataType>(
Policy::template MakeShuffledARegBlockDescriptor<Problem>());
shuffle_tile(a_shuffle_tmp, a_block_tile);
const auto a_block_tile_tmp = tile_elementwise_in(a_element_func, a_shuffle_tmp);
store_tile(a_copy_lds_window, a_block_tile_tmp);
}
else
{
store_tile(a_copy_lds_window, tile_elementwise_in(a_element_func, a_block_tile));
}
// LDS write 0 // LDS write 0
const auto b_block_tile_tmp = tile_elementwise_in(b_element_func, b_block_tile); if constexpr(std::is_same_v<BLayout, tensor_layout::gemm::RowMajor>)
store_tile(b_copy_lds_window, b_block_tile_tmp); {
auto b_shuffle_tmp = make_static_distributed_tensor<BDataType>(
Policy::template MakeShuffledBRegBlockDescriptor<Problem>());
shuffle_tile(b_shuffle_tmp, b_block_tile);
const auto b_block_tile_tmp = tile_elementwise_in(b_element_func, b_shuffle_tmp);
store_tile(b_copy_lds_window, b_block_tile_tmp);
}
else
{
store_tile(b_copy_lds_window, tile_elementwise_in(b_element_func, b_block_tile));
}
} }
index_t iCounter = num_loop - 1; index_t iCounter = num_loop - 1;
...@@ -180,8 +194,19 @@ struct GemmPipelineAGmemBGmemCRegV1 ...@@ -180,8 +194,19 @@ struct GemmPipelineAGmemBGmemCRegV1
store_tile(a_copy_lds_window, a_block_tile_tmp); store_tile(a_copy_lds_window, a_block_tile_tmp);
// LDS write i + 1 // LDS write i + 1
const auto b_block_tile_tmp = tile_elementwise_in(b_element_func, b_block_tile); if constexpr(std::is_same_v<BLayout, tensor_layout::gemm::RowMajor>)
store_tile(b_copy_lds_window, b_block_tile_tmp); {
auto b_shuffle_tmp_loop = make_static_distributed_tensor<BDataType>(
Policy::template MakeShuffledBRegBlockDescriptor<Problem>());
shuffle_tile(b_shuffle_tmp_loop, b_block_tile);
store_tile(b_copy_lds_window,
tile_elementwise_in(b_element_func, b_shuffle_tmp_loop));
}
else
{
const auto b_block_tile_tmp = tile_elementwise_in(b_element_func, b_block_tile);
store_tile(b_copy_lds_window, b_block_tile_tmp);
}
iCounter--; iCounter--;
} }
......
include/ck_tile/ops/gemm/pipeline/gemm_pipeline_agmem_bgmem_creg_v1_default_policy.hpp
View file @ b6d3aa5d
...@@ -11,6 +11,7 @@ namespace ck_tile { ...@@ -11,6 +11,7 @@ namespace ck_tile {
// Default policy class should not be templated, put template on member functions instead // Default policy class should not be templated, put template on member functions instead
struct GemmPipelineAGmemBGmemCRegV1DefaultPolicy struct GemmPipelineAGmemBGmemCRegV1DefaultPolicy
{ {
#if 0 #if 0
// 2d // 2d
template <typename Problem> template <typename Problem>
...@@ -116,6 +117,20 @@ struct GemmPipelineAGmemBGmemCRegV1DefaultPolicy ...@@ -116,6 +117,20 @@ struct GemmPipelineAGmemBGmemCRegV1DefaultPolicy
return smem_size; return smem_size;
} }
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto GetSmemPackA()
{
using ADataType = remove_cvref_t<typename Problem::ADataType>;
return Problem::VectorLoadSize / sizeof(ADataType);
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto GetSmemPackB()
{
using BDataType = remove_cvref_t<typename Problem::BDataType>;
return Problem::VectorLoadSize / sizeof(BDataType);
}
#elif 1 #elif 1
// fake XOR // fake XOR
template <typename Problem> template <typename Problem>
...@@ -192,80 +207,269 @@ struct GemmPipelineAGmemBGmemCRegV1DefaultPolicy ...@@ -192,80 +207,269 @@ struct GemmPipelineAGmemBGmemCRegV1DefaultPolicy
CK_TILE_HOST_DEVICE static constexpr auto MakeADramTileDistribution() CK_TILE_HOST_DEVICE static constexpr auto MakeADramTileDistribution()
{ {
using ADataType = remove_cvref_t<typename Problem::ADataType>; using ADataType = remove_cvref_t<typename Problem::ADataType>;
using ALayout = remove_cvref_t<typename Problem::ALayout>;
constexpr index_t kBlockSize = Problem::kBlockSize;
constexpr index_t BlockSize = Problem::kBlockSize;
constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK; constexpr index_t MPerBlock = Problem::BlockGemmShape::kM;
constexpr index_t KPerBlock = Problem::BlockGemmShape::kK;
constexpr index_t K1 = 16 / sizeof(ADataType);
constexpr index_t K0 = kKPerBlock / K1; if constexpr(std::is_same_v<ALayout, ck_tile::tensor_layout::gemm::ColumnMajor>)
constexpr index_t M2 = get_warp_size() / K0; {
#if 1 // coalesce reading for each blocks constexpr index_t M1 = Problem::VectorLoadSize / sizeof(ADataType);
constexpr index_t M1 = kBlockSize / get_warp_size(); constexpr index_t M0 = MPerBlock / M1;
static_assert(M2 != 0, "M2 is zero, which will lead to a division by zero error."); constexpr index_t total_pixels = MPerBlock * KPerBlock / BlockSize;
static_assert(M1 != 0, "M1 is zero, which will lead to a division by zero error."); static_assert(total_pixels % M1 == 0);
constexpr index_t M0 = kMPerBlock / (M2 * M1); constexpr index_t K3 = total_pixels / M1;
constexpr index_t KPack = GetSmemPackA<Problem>();
return make_static_tile_distribution( static_assert(KPack % K3 == 0);
tile_distribution_encoding<sequence<1>, constexpr index_t K2 = KPack / K3;
tuple<sequence<M0, M1, M2>, sequence<K0, K1>>, if constexpr(get_warp_size() % (K2 * M0))
tuple<sequence<1>, sequence<1, 2>>, {
tuple<sequence<1>, sequence<2, 0>>, constexpr index_t K1 = get_warp_size() / (K2 * M0);
sequence<1, 2>, constexpr index_t K0 = BlockSize / get_warp_size();
sequence<0, 1>>{}); static_assert(KPerBlock == K0 * K1 * K2 * K3);
#else // coalesce reading for each warps return make_static_tile_distribution(
constexpr index_t M0 = kBlockSize / get_warp_size(); tile_distribution_encoding<sequence<1>,
constexpr index_t M1 = kMPerBlock / (M2 * M0); tuple<sequence<M0, M1>, sequence<K0, K1, K2, K3>>,
tuple<sequence<2>, sequence<2, 1, 2>>,
return make_static_tile_distribution( tuple<sequence<0>, sequence<1, 0, 2>>,
tile_distribution_encoding<sequence<1>, sequence<2, 1>,
tuple<sequence<M0, M1, M2>, sequence<K0, K1>>, sequence<3, 1>>{});
tuple<sequence<1>, sequence<1, 2>>, }
tuple<sequence<0>, sequence<2, 0>>, else
sequence<1, 2>, {
sequence<1, 1>>{}); constexpr index_t K1 = (K2 * M0) / get_warp_size();
#endif constexpr index_t K2_m = K2 / K1;
constexpr index_t K0 = BlockSize / get_warp_size() / K1;
static_assert(KPerBlock == K0 * K1 * K2_m * K3);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<M0, M1>, sequence<K0, K1, K2_m, K3>>,
tuple<sequence<2, 2>, sequence<1, 2>>,
tuple<sequence<0, 1>, sequence<0, 2>>,
sequence<2, 1>,
sequence<3, 1>>{});
}
}
else
{
constexpr index_t K1 = 16 / sizeof(ADataType);
constexpr index_t K0 = KPerBlock / K1;
constexpr index_t M2 = get_warp_size() / K0;
// coalesce reading for each blocks
if constexpr(get_warp_size() % (M2 * K0) == 0)
{
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);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<M0, M1, M2>, sequence<K0, K1>>,
tuple<sequence<1>, sequence<1, 2>>,
tuple<sequence<1>, sequence<2, 0>>,
sequence<1, 2>,
sequence<0, 1>>{});
}
else
{
constexpr index_t M0 = BlockSize / get_warp_size();
constexpr index_t M1 = MPerBlock / (M2 * M0);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<M0, M1, M2>, sequence<K0, K1>>,
tuple<sequence<1>, sequence<1, 2>>,
tuple<sequence<0>, sequence<2, 0>>,
sequence<1, 2>,
sequence<1, 1>>{});
}
}
} }
template <typename Problem> template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeBDramTileDistribution() CK_TILE_HOST_DEVICE static constexpr auto MakeBDramTileDistribution()
{ {
using BDataType = remove_cvref_t<typename Problem::BDataType>; using BDataType = remove_cvref_t<typename Problem::BDataType>;
using BLayout = remove_cvref_t<typename Problem::BLayout>;
constexpr index_t BlockSize = Problem::kBlockSize;
constexpr index_t NPerBlock = Problem::BlockGemmShape::kN;
constexpr index_t KPerBlock = Problem::BlockGemmShape::kK;
if constexpr(std::is_same_v<BLayout, ck_tile::tensor_layout::gemm::RowMajor>)
{
constexpr index_t N1 = Problem::VectorLoadSize / sizeof(BDataType);
constexpr index_t N0 = NPerBlock / N1;
constexpr index_t total_pixels = NPerBlock * KPerBlock / BlockSize;
static_assert(total_pixels % N1 == 0);
constexpr index_t K3 = total_pixels / N1;
constexpr index_t KPack = GetSmemPackB<Problem>();
static_assert(KPack % K3 == 0);
constexpr index_t K2 = KPack / K3;
if constexpr(get_warp_size() % (K2 * N0) == 0)
{
constexpr index_t K1 = get_warp_size() / (K2 * N0);
constexpr index_t K0 = BlockSize / get_warp_size();
static_assert(KPerBlock == K0 * K1 * K2 * K3);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<N0, N1>, sequence<K0, K1, K2, K3>>,
tuple<sequence<2>, sequence<2, 1, 2>>,
tuple<sequence<0>, sequence<1, 0, 2>>,
sequence<2, 1>,
sequence<3, 1>>{});
}
else
{
constexpr index_t K1 = (K2 * N0) / get_warp_size();
constexpr index_t K2_m = K2 / K1;
constexpr index_t K0 = BlockSize / get_warp_size() / K1;
static_assert(KPerBlock == K0 * K1 * K2_m * K3);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<N0, N1>, sequence<K0, K1, K2_m, K3>>,
tuple<sequence<2, 2>, sequence<1, 2>>,
tuple<sequence<0, 1>, sequence<0, 2>>,
sequence<2, 1>,
sequence<3, 1>>{});
}
}
else
{
constexpr index_t K1 = Problem::VectorLoadSize / sizeof(BDataType);
constexpr index_t K0 = KPerBlock / K1;
constexpr index_t N2 = get_warp_size() / K0;
// coalesce reading for each blocks
if constexpr(get_warp_size() % (N2 * K0) == 0)
{
constexpr index_t N1 = BlockSize / get_warp_size();
static_assert(N2 != 0, "N2 is zero, which will lead to a division by zero error.");
static_assert(N1 != 0, "N1 is zero, which will lead to a division by zero error.");
constexpr index_t N0 = NPerBlock / (N2 * N1);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<N0, N1, N2>, sequence<K0, K1>>,
tuple<sequence<1>, sequence<1, 2>>,
tuple<sequence<1>, sequence<2, 0>>,
sequence<1, 2>,
sequence<0, 1>>{});
}
// coalesce reading for each warps
else
{
constexpr index_t N0 = BlockSize / get_warp_size();
constexpr index_t N1 = NPerBlock / (N2 * N0);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<N0, N1, N2>, sequence<K0, K1>>,
tuple<sequence<1>, sequence<1, 2>>,
tuple<sequence<0>, sequence<2, 0>>,
sequence<1, 2>,
sequence<1, 1>>{});
}
}
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeShuffledBRegBlockDescriptor()
{
using BLayout = remove_cvref_t<typename Problem::BLayout>;
using BDataType = remove_cvref_t<typename Problem::BDataType>;
static_assert(std::is_same_v<BLayout, ck_tile::tensor_layout::gemm::RowMajor>);
constexpr index_t kBlockSize = Problem::kBlockSize; constexpr index_t kBlockSize = Problem::kBlockSize;
constexpr index_t kNPerBlock = Problem::BlockGemmShape::kN; constexpr index_t kNPerBlock = Problem::BlockGemmShape::kN;
constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK; constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
constexpr index_t K1 = 16 / sizeof(BDataType); constexpr index_t N1 = Problem::VectorLoadSize / sizeof(BDataType);
constexpr index_t K0 = kKPerBlock / K1; constexpr index_t N0 = kNPerBlock / N1;
constexpr index_t N2 = get_warp_size() / K0; constexpr index_t total_pixels = kNPerBlock * kKPerBlock / kBlockSize;
#if 1 // coalesce reading for each blocks static_assert(total_pixels % N1 == 0);
constexpr index_t N1 = kBlockSize / get_warp_size(); constexpr index_t K3 = total_pixels / N1;
static_assert(N2 != 0, "M2 is zero, which will lead to a division by zero error."); constexpr index_t kKPack = GetSmemPackB<Problem>();
static_assert(N1 != 0, "M1 is zero, which will lead to a division by zero error."); static_assert(kKPack % K3 == 0);
constexpr index_t N0 = kNPerBlock / (N2 * N1); constexpr index_t K2 = kKPack / K3; // TODO: this dimention could be outside single wave
constexpr index_t warp_size = get_warp_size();
return make_static_tile_distribution( if constexpr(warp_size % (K2 * N0) == 0)
tile_distribution_encoding<sequence<1>, {
tuple<sequence<N0, N1, N2>, sequence<K0, K1>>, constexpr index_t K1 = warp_size / (K2 * N0);
tuple<sequence<1>, sequence<1, 2>>, constexpr index_t K0 = kBlockSize / warp_size;
tuple<sequence<1>, sequence<2, 0>>,
sequence<1, 2>, return make_static_tile_distribution(
sequence<0, 1>>{}); tile_distribution_encoding<sequence<1>,
#else // coalesce reading for each warps tuple<sequence<N0, N1>, sequence<K0, K1, K2, K3>>,
constexpr index_t N0 = kBlockSize / get_warp_size(); tuple<sequence<2>, sequence<2, 1, 2>>,
constexpr index_t N1 = kNPerBlock / (N2 * N0); tuple<sequence<0>, sequence<1, 0, 2>>,
sequence<1, 2>,
return make_static_tile_distribution( sequence<1, 3>>{});
tile_distribution_encoding<sequence<1>, }
tuple<sequence<N0, N1, N2>, sequence<K0, K1>>, else
tuple<sequence<1>, sequence<1, 2>>, {
tuple<sequence<0>, sequence<2, 0>>, constexpr index_t K1 = (K2 * N0) / get_warp_size();
sequence<1, 2>, constexpr index_t K2_m = K2 / K1;
sequence<1, 1>>{}); constexpr index_t K0 = kBlockSize / get_warp_size() / K1;
#endif static_assert(kKPerBlock == K0 * K1 * K2_m * K3);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<N0, N1>, sequence<K0, K1, K2_m, K3>>,
tuple<sequence<2, 2>, sequence<1, 2>>,
tuple<sequence<0, 1>, sequence<0, 2>>,
sequence<1, 2>,
sequence<1, 3>>{});
}
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeShuffledARegBlockDescriptor()
{
using ALayout = remove_cvref_t<typename Problem::ALayout>;
using ADataType = remove_cvref_t<typename Problem::ADataType>;
static_assert(std::is_same_v<ALayout, ck_tile::tensor_layout::gemm::RowMajor>);
constexpr index_t kBlockSize = Problem::kBlockSize;
constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
constexpr index_t M1 = Problem::VectorLoadSize / sizeof(ADataType);
constexpr index_t M0 = kMPerBlock / M1;
constexpr index_t total_pixels = kMPerBlock * kKPerBlock / kBlockSize;
static_assert(total_pixels % M1 == 0);
constexpr index_t K3 = total_pixels / M1;
constexpr index_t kKPack = GetSmemPackA<Problem>();
static_assert(kKPack % K3 == 0);
constexpr index_t K2 = kKPack / K3; // TODO: this dimention could be outside single wave
constexpr index_t warp_size = get_warp_size();
if constexpr(warp_size % (K2 * M0) == 0)
{
constexpr index_t K1 = warp_size / (K2 * M0);
constexpr index_t K0 = kBlockSize / warp_size;
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<M0, M1>, sequence<K0, K1, K2, K3>>,
tuple<sequence<2>, sequence<2, 1, 2>>,
tuple<sequence<0>, sequence<1, 0, 2>>,
sequence<1, 2>,
sequence<1, 3>>{});
}
else
{
constexpr index_t K1 = (K2 * M0) / get_warp_size();
constexpr index_t K2_m = K2 / K1;
constexpr index_t K0 = kBlockSize / get_warp_size() / K1;
static_assert(kKPerBlock == K0 * K1 * K2_m * K3);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<M0, M1>, sequence<K0, K1, K2_m, K3>>,
tuple<sequence<2, 2>, sequence<1, 2>>,
tuple<sequence<0, 1>, sequence<0, 2>>,
sequence<1, 2>,
sequence<1, 3>>{});
}
} }
template <typename Problem> template <typename Problem>
......
include/ck_tile/ops/gemm/pipeline/gemm_pipeline_problem.hpp
View file @ b6d3aa5d
...@@ -3,40 +3,133 @@ ...@@ -3,40 +3,133 @@
#pragma once #pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_scheduler.hpp" #include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_scheduler.hpp"
namespace ck_tile { namespace ck_tile {
static constexpr int _VectorSize = 16;
template <typename ADataType_, template <typename ADataType_,
typename BDataType_, typename BDataType_,
typename CDataType_, typename CDataType_,
typename BlockGemmShape_, typename BlockGemmShape_,
typename TileGemmTraits_> typename TileGemmTraits_>
struct GemmPipelineProblem struct GemmPipelineProblemBase
{ {
using ADataType = remove_cvref_t<ADataType_>; using GemmTraits = remove_cvref_t<TileGemmTraits_>;
using BDataType = remove_cvref_t<BDataType_>;
using CDataType = remove_cvref_t<CDataType_>; using ADataType = remove_cvref_t<ADataType_>;
using BDataType = remove_cvref_t<BDataType_>;
using CDataType = remove_cvref_t<CDataType_>;
using BlockGemmShape = remove_cvref_t<BlockGemmShape_>; using BlockGemmShape = remove_cvref_t<BlockGemmShape_>;
using GemmTraits = remove_cvref_t<TileGemmTraits_>;
using ALayout = remove_cvref_t<typename GemmTraits::ALayout>; using ALayout = remove_cvref_t<typename GemmTraits::ALayout>;
using BLayout = remove_cvref_t<typename GemmTraits::BLayout>; using BLayout = remove_cvref_t<typename GemmTraits::BLayout>;
using CLayout = remove_cvref_t<typename GemmTraits::CLayout>; using CLayout = remove_cvref_t<typename GemmTraits::CLayout>;
static constexpr index_t kBlockSize = BlockGemmShape::NumWarps * get_warp_size(); static constexpr index_t VectorLoadSize = GemmTraits::_VectorSize;
static constexpr bool kPadA = GemmTraits::kPadA; static constexpr index_t kBlockSize = BlockGemmShape::NumWarps * get_warp_size();
static constexpr bool kPadB = GemmTraits::kPadB;
static constexpr bool kPadC = GemmTraits::kPadC; static constexpr bool kPadM = GemmTraits::kPadM;
static constexpr bool kPadN = GemmTraits::kPadN;
static constexpr bool kPadK = GemmTraits::kPadK;
CK_TILE_HOST_DEVICE static constexpr auto GetAlignmentA()
{
if constexpr(std::is_same_v<ALayout, ck_tile::tensor_layout::gemm::ColumnMajor>)
{
constexpr index_t pixels_per_thread =
BlockGemmShape::kM * BlockGemmShape::kK / kBlockSize;
return pixels_per_thread < VectorLoadSize / sizeof(ADataType)
? pixels_per_thread
: VectorLoadSize / sizeof(ADataType);
}
else
{
return VectorLoadSize / sizeof(ADataType);
}
}
CK_TILE_HOST_DEVICE static constexpr auto GetAlignmentB()
{
if constexpr(std::is_same_v<BLayout, ck_tile::tensor_layout::gemm::RowMajor>)
{
constexpr index_t pixels_per_thread =
BlockGemmShape::kN * BlockGemmShape::kK / kBlockSize;
return pixels_per_thread < VectorLoadSize / sizeof(BDataType)
? pixels_per_thread
: VectorLoadSize / sizeof(BDataType);
}
else
{
return VectorLoadSize / sizeof(BDataType);
}
}
CK_TILE_HOST_DEVICE static constexpr auto GetAlignmentC()
{
if constexpr(std::is_same_v<CLayout, ck_tile::tensor_layout::gemm::ColumnMajor>)
{
constexpr index_t N1 = kBlockSize / get_warp_size();
constexpr index_t N2 = std::min(BlockGemmShape::kN / N1, get_warp_size());
constexpr index_t M0 = get_warp_size() / N2;
constexpr index_t M1 = BlockGemmShape::kM / M0;
static constexpr index_t VectorSizeA = kPadA ? 1 : _VectorSize / sizeof(ADataType); return std::min(M1, static_cast<index_t>(VectorLoadSize / sizeof(CDataType)));
static constexpr index_t VectorSizeB = kPadB ? 1 : _VectorSize / sizeof(BDataType); }
static constexpr index_t VectorSizeC = kPadC ? 1 : _VectorSize / sizeof(CDataType); else
{
constexpr index_t M1 = kBlockSize / get_warp_size();
constexpr index_t M2 = std::min(BlockGemmShape::kM / M1, get_warp_size());
constexpr index_t N0 = get_warp_size() / M2;
constexpr index_t N1 = BlockGemmShape::kN / N0;
return std::min(N1, static_cast<index_t>(VectorLoadSize / sizeof(CDataType)));
}
}
static constexpr index_t VectorSizeA = []() {
if constexpr(std::is_same_v<ALayout, tensor_layout::gemm::RowMajor>)
{
return kPadK ? 1 : GetAlignmentA();
}
else
{
return kPadM ? 1 : GetAlignmentA();
}
}();
static constexpr index_t VectorSizeB = []() {
if constexpr(std::is_same_v<BLayout, tensor_layout::gemm::ColumnMajor>)
{
return kPadN ? 1 : GetAlignmentB();
}
else
{
return kPadK ? 1 : GetAlignmentB();
}
}();
static constexpr index_t VectorSizeC = []() {
if constexpr(std::is_same_v<CLayout, tensor_layout::gemm::RowMajor>)
{
return kPadN ? 1 : GetAlignmentC();
}
else
{
return kPadM ? 1 : GetAlignmentC();
}
}();
}; };
// Alias for GemmPipelineProblem
template <typename ADataType_,
typename BDataType_,
typename CDataType_,
typename BlockGemmShape_,
typename TileGemmTraits_>
using GemmPipelineProblem =
GemmPipelineProblemBase<ADataType_, BDataType_, CDataType_, BlockGemmShape_, TileGemmTraits_>;
template <typename ADataType_, template <typename ADataType_,
typename BDataType_, typename BDataType_,
typename CDataType_, typename CDataType_,
...@@ -45,30 +138,15 @@ template <typename ADataType_, ...@@ -45,30 +138,15 @@ template <typename ADataType_,
GemmPipelineScheduler Scheduler_ = GemmPipelineScheduler::Intrawave, GemmPipelineScheduler Scheduler_ = GemmPipelineScheduler::Intrawave,
bool HasHotLoop_ = true, bool HasHotLoop_ = true,
TailNumber TailNum_ = TailNumber::Full> TailNumber TailNum_ = TailNumber::Full>
struct UniversalGemmPipelineProblem struct UniversalGemmPipelineProblem : public GemmPipelineProblemBase<ADataType_,
BDataType_,
CDataType_,
BlockGemmShape_,
TileGemmTraits_>
{ {
using ADataType = remove_cvref_t<ADataType_>; static constexpr auto Scheduler = Scheduler_;
using BDataType = remove_cvref_t<BDataType_>; static constexpr auto HasHotLoop = HasHotLoop_;
using CDataType = remove_cvref_t<CDataType_>; static constexpr auto TailNum = TailNum_;
using BlockGemmShape = remove_cvref_t<BlockGemmShape_>;
using GemmTraits = remove_cvref_t<TileGemmTraits_>;
using ALayout = remove_cvref_t<typename GemmTraits::ALayout>;
using BLayout = remove_cvref_t<typename GemmTraits::BLayout>;
using CLayout = remove_cvref_t<typename GemmTraits::CLayout>;
static constexpr auto Scheduler = Scheduler_;
static constexpr auto HasHotLoop = HasHotLoop_;
static constexpr auto TailNum = TailNum_;
static constexpr index_t kBlockSize = BlockGemmShape::NumWarps * get_warp_size();
static constexpr bool kPadA = GemmTraits::kPadA;
static constexpr bool kPadB = GemmTraits::kPadB;
static constexpr bool kPadC = GemmTraits::kPadC;
static constexpr index_t VectorSizeA = kPadA ? _VectorSize / sizeof(ADataType) : 1;
static constexpr index_t VectorSizeB = kPadB ? _VectorSize / sizeof(BDataType) : 1;
static constexpr index_t VectorSizeC = kPadC ? _VectorSize / sizeof(CDataType) : 1;
}; };
} // namespace ck_tile } // namespace ck_tile
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment