Skip to content

GitLab

Commit d20c20a6 authored by Mirza Halilcevic's avatar Mirza Halilcevic
Browse files

Merge remote-tracking branch 'upstream/develop' into gemm_elementwise_gemm

parents 250a89f3 10158b0f
Expand all Show whitespace changes
Inline Side-by-side
Showing with 1873 additions and 151 deletions
include/ck_tile/ops/gemm/pipeline/block_gemm_pipeline_agmem_bgmem_creg_v2.hpp → include/ck_tile/ops/gemm/pipeline/gemm_pipeline_agmem_bgmem_creg_v2.hpp
View file @ d20c20a6
...@@ -4,15 +4,15 @@ ...@@ -4,15 +4,15 @@
#pragma once #pragma once
#include "ck_tile/core.hpp" #include "ck_tile/core.hpp"
#include "ck_tile/ops/gemm/pipeline/block_gemm_pipeline_agmem_bgmem_creg_v2_default_policy.hpp" #include "ck_tile/ops/gemm/pipeline/gemm_pipeline_agmem_bgmem_creg_v2_default_policy.hpp"
namespace ck_tile { namespace ck_tile {
// A Tile Window: global memory // A Tile Window: global memory
// B Tile Window: global memory // B Tile Window: global memory
// C Distributed tensor: register // C Distributed tensor: register
template <typename Problem, typename Policy = BlockGemmPipelineAGmemBGmemCRegV2DefaultPolicy> template <typename Problem, typename Policy = GemmPipelineAGmemBGmemCRegV2DefaultPolicy>
struct BlockGemmPipelineAGmemBGmemCRegV2 struct GemmPipelineAGmemBGmemCRegV2
{ {
using ADataType = remove_cvref_t<typename Problem::ADataType>; using ADataType = remove_cvref_t<typename Problem::ADataType>;
using BDataType = remove_cvref_t<typename Problem::BDataType>; using BDataType = remove_cvref_t<typename Problem::BDataType>;
......
include/ck_tile/ops/gemm/pipeline/block_gemm_pipeline_agmem_bgmem_creg_v2_default_policy.hpp → include/ck_tile/ops/gemm/pipeline/gemm_pipeline_agmem_bgmem_creg_v2_default_policy.hpp
View file @ d20c20a6
...@@ -7,12 +7,11 @@ ...@@ -7,12 +7,11 @@
namespace ck_tile { namespace ck_tile {
// Default policy for BlockGemmPipelineAGmemBGmemCRegV2 // Default policy for GemmPipelineAGmemBGmemCRegV2
// 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
// NOTE: policy should be binded to its corresponding operation. It's just a coincidence that // NOTE: policy should be binded to its corresponding operation. It's just a coincidence that
// BlockGemmPipelineAGmemBGmemCRegV2DefaultPolicy is the same as // GemmPipelineAGmemBGmemCRegV2DefaultPolicy is the same as
// BlockGemmPipelineAGmemBGmemCRegV1DefaultPolicy // GemmPipelineAGmemBGmemCRegV1DefaultPolicy
using BlockGemmPipelineAGmemBGmemCRegV2DefaultPolicy = using GemmPipelineAGmemBGmemCRegV2DefaultPolicy = GemmPipelineAGmemBGmemCRegV1DefaultPolicy;
BlockGemmPipelineAGmemBGmemCRegV1DefaultPolicy;
} // namespace ck_tile } // namespace ck_tile
include/ck_tile/ops/gemm/pipeline/block_gemm_pipeline_problem.hpp → include/ck_tile/ops/gemm/pipeline/gemm_pipeline_problem.hpp
View file @ d20c20a6
...@@ -13,20 +13,23 @@ template <typename ADataType_, ...@@ -13,20 +13,23 @@ template <typename ADataType_,
typename BDataType_, typename BDataType_,
typename CDataType_, typename CDataType_,
typename BlockGemmShape_, typename BlockGemmShape_,
bool kPadA_ = false, typename TileGemmTraits_>
bool kPadB_ = false, struct GemmPipelineProblem
bool kPadC_ = false>
struct BlockGemmPipelineProblem
{ {
using ADataType = remove_cvref_t<ADataType_>; using ADataType = remove_cvref_t<ADataType_>;
using BDataType = remove_cvref_t<BDataType_>; using BDataType = remove_cvref_t<BDataType_>;
using CDataType = remove_cvref_t<CDataType_>; using CDataType = remove_cvref_t<CDataType_>;
using BlockGemmShape = remove_cvref_t<BlockGemmShape_>; using BlockGemmShape = remove_cvref_t<BlockGemmShape_>;
using GemmTraits = remove_cvref_t<TileGemmTraits_>;
static constexpr index_t kBlockSize = BlockGemmShape::NumWarps * get_warp_size(); static constexpr index_t kBlockSize = BlockGemmShape::NumWarps * get_warp_size();
static constexpr bool kPadA = kPadA_; static constexpr bool kPadA = GemmTraits::kPadA;
static constexpr bool kPadB = kPadB_; static constexpr bool kPadB = GemmTraits::kPadB;
static constexpr bool kPadC = kPadC_; static constexpr bool kPadC = GemmTraits::kPadC;
using LayoutA = remove_cvref_t<typename GemmTraits::LayoutA>;
using LayoutB = remove_cvref_t<typename GemmTraits::LayoutB>;
using LayoutC = remove_cvref_t<typename GemmTraits::LayoutC>;
static constexpr index_t AlignmentA = kPadA ? 1 : VectorLoadSize / sizeof(ADataType); static constexpr index_t AlignmentA = kPadA ? 1 : VectorLoadSize / sizeof(ADataType);
static constexpr index_t AlignmentB = kPadB ? 1 : VectorLoadSize / sizeof(BDataType); static constexpr index_t AlignmentB = kPadB ? 1 : VectorLoadSize / sizeof(BDataType);
......
include/ck_tile/ops/gemm/pipeline/tile_gemm_traits.hpp 0 → 100644
View file @ d20c20a6
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
namespace ck_tile {
template <bool kPadA_,
bool kPadB_,
bool kPadC_,
typename LayoutA_,
typename LayoutB_,
typename LayoutC_>
struct TileGemmTraits
{
static constexpr bool kPadA = kPadA_;
static constexpr bool kPadB = kPadB_;
static constexpr bool kPadC = kPadC_;
using LayoutA = LayoutA_;
using LayoutB = LayoutB_;
using LayoutC = LayoutC_;
};
} // namespace ck_tile
include/ck_tile/ops/layernorm2d/kernel/layernorm2d_fwd_kernel.hpp
View file @ d20c20a6
...@@ -31,8 +31,14 @@ struct Layernorm2dFwd ...@@ -31,8 +31,14 @@ struct Layernorm2dFwd
static constexpr ck_tile::index_t kMPerBlock = Problem::BlockShape::kMPerBlock; static constexpr ck_tile::index_t kMPerBlock = Problem::BlockShape::kMPerBlock;
static constexpr ck_tile::index_t kNPerBlock = Problem::BlockShape::kNPerBlock; static constexpr ck_tile::index_t kNPerBlock = Problem::BlockShape::kNPerBlock;
static constexpr bool kPadM = Problem::kPadM;
static constexpr bool kPadN = Problem::kPadN;
static constexpr ck_tile::index_t kNThreadPerWarp = Problem::BlockShape::kNThreadPerWarp; static constexpr ck_tile::index_t kNThreadPerWarp = Problem::BlockShape::kNThreadPerWarp;
static constexpr ck_tile::index_t kNPerThread = Problem::BlockShape::kNPerThread;
static constexpr auto I0 = number<0>{};
static constexpr auto I1 = number<1>{};
struct Kargs struct Kargs
{ {
...@@ -96,19 +102,25 @@ struct Layernorm2dFwd ...@@ -96,19 +102,25 @@ struct Layernorm2dFwd
sequence<2>>{}); sequence<2>>{});
} }
template <typename Dstr> CK_TILE_DEVICE static int GetWelfordMaxCount(int N)
CK_TILE_DEVICE static constexpr auto GetNPerThread(Dstr)
{ {
constexpr auto nDstrSpan = Dstr::get_distributed_spans().template at<1>(); constexpr ck_tile::index_t kNThreadPerBlock = kNPerBlock / kNPerThread;
using Lengths = decltype(nDstrSpan.impl_);
ck_tile::index_t ret = 1; int thread_id_n = get_thread_id() % kNThreadPerBlock;
int max_count =
__builtin_amdgcn_readfirstlane(N < kNPerBlock ? 0 : kNPerThread * (N / kNPerBlock));
int n_per_block_tail_loop =
__builtin_amdgcn_readfirstlane(N - max_count * kNThreadPerBlock);
ck_tile::static_for<0, Lengths::size(), 1>{}( if(n_per_block_tail_loop > 0)
[&](auto idx) { ret *= Lengths::template at(idx); }); {
int thread_max_n = (thread_id_n + 1) * kNPerThread;
int delta = thread_max_n - n_per_block_tail_loop;
delta = clamp(thread_max_n - n_per_block_tail_loop, 0, kNPerThread);
max_count += kNPerThread - delta;
}
return ret; return max_count;
} }
template <typename DistributedTensor> template <typename DistributedTensor>
...@@ -129,42 +141,29 @@ struct Layernorm2dFwd ...@@ -129,42 +141,29 @@ struct Layernorm2dFwd
return out_dstr_tensor; return out_dstr_tensor;
} }
template <bool Cond = (kHasGamma && kHasBeta)> template <typename XBlockWindow,
CK_TILE_DEVICE std::enable_if_t<Cond> TwoPassLayernorm2dFwd(const XDataType* p_x, typename GammaBlockWindow,
const GammaDataType* p_gamma, typename BetaBlockWindow,
const BetaDataType* p_beta, typename YBlockWindow,
YDataType* p_y, typename MeanBlockWindow,
MeanDataType* p_mean, typename InvStdBlockWindow,
InvStdDataType* p_invStd, bool Cond = (kHasGamma && kHasBeta)>
const ComputeDataType epsilon, CK_TILE_DEVICE std::enable_if_t<Cond>
ck_tile::index_t M, TwoPassLayernorm2dFwd(XBlockWindow& x_block_window,
GammaBlockWindow& gamma_block_window,
BetaBlockWindow& beta_block_window,
YBlockWindow& y_block_window,
MeanBlockWindow& mean_block_window,
InvStdBlockWindow& inv_std_block_window,
ComputeDataType epsilon,
ck_tile::index_t N) const ck_tile::index_t N) const
{ {
constexpr auto I0 = number<0>{}; // TODO - Optimize tail loop to reduce move_tile_window()
constexpr auto I1 = number<1>{}; index_t num_n_tile_iteration =
__builtin_amdgcn_readfirstlane(integer_divide_ceil(N, kNPerBlock));
const auto x_m_n = make_naive_tensor_view<address_space_enum::global>( int welford_max_count = GetWelfordMaxCount(N);
p_x, make_tuple(M, N), make_tuple(N, 1), number<32>{}, number<1>{}); ThreadWelford<ComputeDataType, XDataType> thread_welford{welford_max_count};
const auto gamma_n = make_naive_tensor_view<address_space_enum::global>(
p_gamma, make_tuple(N), make_tuple(1), number<32>{}, number<1>{});
const auto beta_n = make_naive_tensor_view<address_space_enum::global>(
p_beta, make_tuple(N), make_tuple(1), number<32>{}, number<1>{});
const auto iM = get_block_id() * kMPerBlock;
constexpr auto xDstr = MakeXBlockTileDistribution();
auto x_block_window = make_tile_window(
x_m_n, make_tuple(number<kMPerBlock>{}, number<kNPerBlock>{}), {iM, 0}, xDstr);
index_t num_n_tile_iteration = __builtin_amdgcn_readfirstlane(N / kNPerBlock);
// TODO: padding - handle max_count if N % kNPerBlock != 0
constexpr auto NPerThread = GetNPerThread(xDstr);
ThreadWelford<ComputeDataType, XDataType> thread_welford{
type_convert<int>(NPerThread * N / kNPerBlock)};
using XTensorType = decltype(load_tile(x_block_window)); using XTensorType = decltype(load_tile(x_block_window));
auto mean_compute_block_tensor = auto mean_compute_block_tensor =
...@@ -190,44 +189,14 @@ struct Layernorm2dFwd ...@@ -190,44 +189,14 @@ struct Layernorm2dFwd
auto inv_std_compute_block_tensor = InvSqrt(var_compute_block_tensor, epsilon); auto inv_std_compute_block_tensor = InvSqrt(var_compute_block_tensor, epsilon);
if constexpr(kSaveMean) if constexpr(kSaveMean)
{
const auto mean_m = make_naive_tensor_view_packed<address_space_enum::global>(
p_mean, make_tuple(M), number<32>{});
auto mean_block_window =
make_tile_window(mean_m, make_tuple(number<kMPerBlock>{}), {iM});
store_tile(mean_block_window, cast_tile<MeanDataType>(mean_compute_block_tensor)); store_tile(mean_block_window, cast_tile<MeanDataType>(mean_compute_block_tensor));
}
if constexpr(kSaveInvStd) if constexpr(kSaveInvStd)
{ store_tile(inv_std_block_window,
const auto inv_std_m = make_naive_tensor_view_packed<address_space_enum::global>( cast_tile<InvStdDataType>(inv_std_compute_block_tensor));
p_invStd, make_tuple(M), number<32>{});
auto inv_std_block_window =
make_tile_window(inv_std_m, make_tuple(number<kMPerBlock>{}), {iM});
store_tile(inv_std_block_window, cast_tile<MeanDataType>(inv_std_compute_block_tensor));
}
// TODO: Extract normalize pipeline
const auto y_m_n = make_naive_tensor_view<address_space_enum::global>(
p_y, make_tuple(M, N), make_tuple(N, 1), number<32>{}, number<1>{});
auto y_block_window = make_tile_window(
y_m_n, make_tuple(number<kMPerBlock>{}, number<kNPerBlock>{}), {iM, 0});
constexpr auto gammaDstr = MakeGammaBetaBlockTileDistribution();
constexpr auto betaDstr = gammaDstr;
auto gamma_block_window =
make_tile_window(gamma_n, make_tuple(number<kNPerBlock>{}), {0}, gammaDstr);
auto beta_block_window = make_tile_window(
beta_n, make_tuple(number<kMPerBlock>{}, number<kNPerBlock>{}), {0}, betaDstr);
// reverse read x to reuse cache // reverse read x to reuse cache
ck_tile::index_t stride_to_right_most_window = N - kNPerBlock; ck_tile::index_t stride_to_right_most_window =
N % kNPerBlock == 0 ? N - kNPerBlock : N - N % kNPerBlock;
move_tile_window(x_block_window, {0, -kNPerBlock}); move_tile_window(x_block_window, {0, -kNPerBlock});
move_tile_window(gamma_block_window, {stride_to_right_most_window}); move_tile_window(gamma_block_window, {stride_to_right_most_window});
...@@ -274,16 +243,208 @@ struct Layernorm2dFwd ...@@ -274,16 +243,208 @@ struct Layernorm2dFwd
} }
} }
template <typename XBlockWindow,
typename GammaBlockWindow,
typename BetaBlockWindow,
typename YBlockWindow,
typename MeanBlockWindow,
typename InvStdBlockWindow,
bool Cond = (kHasGamma && kHasBeta)>
CK_TILE_DEVICE std::enable_if_t<Cond>
OnePassLayernorm2dFwd(XBlockWindow& x_block_window,
GammaBlockWindow& gamma_block_window,
BetaBlockWindow& beta_block_window,
YBlockWindow& y_block_window,
MeanBlockWindow& mean_block_window,
InvStdBlockWindow& inv_std_block_window,
ComputeDataType epsilon,
ck_tile::index_t N) const
{
int welford_max_count = GetWelfordMaxCount(N);
ThreadWelford<ComputeDataType, XDataType> thread_welford{welford_max_count};
using XTensorType = decltype(load_tile(x_block_window));
auto mean_compute_block_tensor =
thread_welford.template MakeInitialMeanVarDistributedTensor<XTensorType>();
auto var_compute_block_tensor =
thread_welford.template MakeInitialMeanVarDistributedTensor<XTensorType>();
clear_tile(mean_compute_block_tensor);
clear_tile(var_compute_block_tensor);
const auto x_block_tensor = load_tile(x_block_window);
thread_welford(x_block_tensor, mean_compute_block_tensor, var_compute_block_tensor);
// TODO: support cross warp Welford
WarpMergeWelford<ComputeDataType, true>{}(
mean_compute_block_tensor, var_compute_block_tensor, thread_welford.cur_count_);
auto inv_std_compute_block_tensor = InvSqrt(var_compute_block_tensor, epsilon);
if constexpr(kSaveMean)
store_tile(mean_block_window, cast_tile<MeanDataType>(mean_compute_block_tensor));
if constexpr(kSaveInvStd)
store_tile(inv_std_block_window,
cast_tile<InvStdDataType>(inv_std_compute_block_tensor));
// normalize
const auto gamma_block_tensor = load_tile(gamma_block_window);
const auto beta_block_tensor = load_tile(beta_block_window);
constexpr auto x_spans = decltype(x_block_tensor)::get_distributed_spans();
auto y_block_tensor =
make_static_distributed_tensor<YDataType>(x_block_tensor.get_tile_distribution());
sweep_tile_span(x_spans[I1], [&](auto idx1) {
constexpr auto j_idx = make_tuple(idx1);
const auto gamma = type_convert<ComputeDataType>(gamma_block_tensor[j_idx]);
const auto beta = type_convert<ComputeDataType>(beta_block_tensor[j_idx]);
sweep_tile_span(x_spans[I0], [&](auto idx0) {
constexpr auto i_idx = make_tuple(idx0);
constexpr auto i_j_idx = make_tuple(idx0, idx1);
const auto mean = mean_compute_block_tensor[i_idx];
const auto inv_std = inv_std_compute_block_tensor[i_idx];
const auto x = type_convert<ComputeDataType>(x_block_tensor[i_j_idx]);
auto y = (x - mean) * inv_std * gamma + beta;
y_block_tensor(i_j_idx) = type_convert<YDataType>(y);
});
});
store_tile(y_block_window, y_block_tensor);
}
CK_TILE_DEVICE void operator()(Kargs kargs) const CK_TILE_DEVICE void operator()(Kargs kargs) const
{ {
TwoPassLayernorm2dFwd(static_cast<const XDataType*>(kargs.p_x), const auto x_m_n = [&]() {
const auto x_dram_naive = make_naive_tensor_view<address_space_enum::global>(
static_cast<const XDataType*>(kargs.p_x),
make_tuple(kargs.M, kargs.N),
make_tuple(kargs.N, 1),
number<kNPerThread>{},
number<1>{});
return pad_tensor_view(x_dram_naive,
make_tuple(number<kMPerBlock>{}, number<kNPerBlock>{}),
sequence<kPadM, kPadN>{});
}();
const auto gamma_n = [&]() {
const auto gamma_dram_naive = make_naive_tensor_view<address_space_enum::global>(
static_cast<const GammaDataType*>(kargs.p_gamma), static_cast<const GammaDataType*>(kargs.p_gamma),
make_tuple(kargs.N),
make_tuple(1),
number<kNPerThread>{},
number<1>{});
return pad_tensor_view(
gamma_dram_naive, make_tuple(number<kNPerBlock>{}), sequence<kPadN>{});
}();
const auto beta_n = [&]() {
const auto gamma_dram_naive = make_naive_tensor_view<address_space_enum::global>(
static_cast<const BetaDataType*>(kargs.p_beta), static_cast<const BetaDataType*>(kargs.p_beta),
make_tuple(kargs.N),
make_tuple(1),
number<kNPerThread>{},
number<1>{});
return pad_tensor_view(
gamma_dram_naive, make_tuple(number<kNPerBlock>{}), sequence<kPadN>{});
}();
const auto iM = get_block_id() * kMPerBlock;
constexpr auto xDstr = MakeXBlockTileDistribution();
auto x_block_window = make_tile_window(
x_m_n, make_tuple(number<kMPerBlock>{}, number<kNPerBlock>{}), {iM, 0}, xDstr);
const auto y_m_n = [&]() {
const auto y_dram_naive = make_naive_tensor_view<address_space_enum::global>(
static_cast<YDataType*>(kargs.p_y), static_cast<YDataType*>(kargs.p_y),
make_tuple(kargs.M, kargs.N),
make_tuple(kargs.N, 1),
number<kNPerThread>{},
number<1>{});
return pad_tensor_view(y_dram_naive,
make_tuple(number<kMPerBlock>{}, number<kNPerBlock>{}),
sequence<kPadM, kPadN>{});
}();
auto y_block_window = make_tile_window(
y_m_n, make_tuple(number<kMPerBlock>{}, number<kNPerBlock>{}), {iM, 0});
constexpr auto gammaDstr = MakeGammaBetaBlockTileDistribution();
constexpr auto betaDstr = gammaDstr;
auto gamma_block_window =
make_tile_window(gamma_n, make_tuple(number<kNPerBlock>{}), {0}, gammaDstr);
auto beta_block_window = make_tile_window(
beta_n, make_tuple(number<kMPerBlock>{}, number<kNPerBlock>{}), {0}, betaDstr);
auto mean_block_window = [&]() {
if constexpr(kSaveMean)
{
const auto mean_m = [&]() {
const auto mean_dram_naive =
make_naive_tensor_view_packed<address_space_enum::global>(
static_cast<MeanDataType*>(kargs.p_mean), static_cast<MeanDataType*>(kargs.p_mean),
make_tuple(kargs.M),
number<1>{});
return pad_tensor_view(
mean_dram_naive, make_tuple(number<kMPerBlock>{}), sequence<kPadM>{});
}();
return make_tile_window(mean_m, make_tuple(number<kMPerBlock>{}), {iM});
}
else
return make_null_tile_window(make_tuple(number<kMPerBlock>{}));
}();
auto inv_std_block_window = [&]() {
if constexpr(kSaveInvStd)
{
const auto inv_std_m = [&]() {
const auto inv_std_dram_naive =
make_naive_tensor_view_packed<address_space_enum::global>(
static_cast<InvStdDataType*>(kargs.p_invStd), static_cast<InvStdDataType*>(kargs.p_invStd),
make_tuple(kargs.M),
number<1>{});
return pad_tensor_view(
inv_std_dram_naive, make_tuple(number<kMPerBlock>{}), sequence<kPadM>{});
}();
return make_tile_window(inv_std_m, make_tuple(number<kMPerBlock>{}), {iM});
}
else
return make_null_tile_window(make_tuple(number<kMPerBlock>{}));
}();
if(kargs.N <= kNPerBlock)
OnePassLayernorm2dFwd(x_block_window,
gamma_block_window,
beta_block_window,
y_block_window,
mean_block_window,
inv_std_block_window,
static_cast<const ComputeDataType>(kargs.epsilon),
kargs.N);
else
TwoPassLayernorm2dFwd(x_block_window,
gamma_block_window,
beta_block_window,
y_block_window,
mean_block_window,
inv_std_block_window,
static_cast<const ComputeDataType>(kargs.epsilon), static_cast<const ComputeDataType>(kargs.epsilon),
kargs.M,
kargs.N); kargs.N);
} }
}; };
......
include/ck_tile/ops/layernorm2d/pipeline/block_layernorm2d_fwd_problem.hpp
View file @ d20c20a6
...@@ -14,7 +14,9 @@ template <typename XDataType_, ...@@ -14,7 +14,9 @@ template <typename XDataType_,
typename YDataType_, typename YDataType_,
typename MeanDataType_, typename MeanDataType_,
typename InvStdDataType_, typename InvStdDataType_,
typename BlockShape_> typename BlockShape_,
bool kPadM_,
bool kPadN_>
struct BlockLayernorm2dFwdProblem struct BlockLayernorm2dFwdProblem
{ {
using XDataType = remove_cvref_t<XDataType_>; using XDataType = remove_cvref_t<XDataType_>;
...@@ -25,6 +27,8 @@ struct BlockLayernorm2dFwdProblem ...@@ -25,6 +27,8 @@ struct BlockLayernorm2dFwdProblem
using MeanDataType = remove_cvref_t<MeanDataType_>; using MeanDataType = remove_cvref_t<MeanDataType_>;
using InvStdDataType = remove_cvref_t<InvStdDataType_>; using InvStdDataType = remove_cvref_t<InvStdDataType_>;
using BlockShape = remove_cvref_t<BlockShape_>; using BlockShape = remove_cvref_t<BlockShape_>;
static constexpr bool kPadM = kPadM_;
static constexpr bool kPadN = kPadN_;
}; };
} // namespace ck_tile } // namespace ck_tile
library/src/tensor_operation_instance/gpu/CMakeLists.txt
View file @ d20c20a6
...@@ -37,11 +37,7 @@ function(add_instance_library INSTANCE_NAME) ...@@ -37,11 +37,7 @@ function(add_instance_library INSTANCE_NAME)
endforeach() endforeach()
endif() endif()
if(INSTANCES_ONLY) set(INST_TARGETS ${SUPPORTED_GPU_TARGETS})
set(INST_TARGETS ${DEFAULT_GPU_TARGETS})
else()
set(INST_TARGETS ${GPU_TARGETS})
endif()
# Do not build DL instances if DL_KERNELS macro is not set # Do not build DL instances if DL_KERNELS macro is not set
foreach(source IN LISTS ARGN) foreach(source IN LISTS ARGN)
...@@ -64,9 +60,9 @@ function(add_instance_library INSTANCE_NAME) ...@@ -64,9 +60,9 @@ function(add_instance_library INSTANCE_NAME)
list(REMOVE_ITEM ARGN "${source}") list(REMOVE_ITEM ARGN "${source}")
endif() endif()
endforeach() endforeach()
# Do not build mha instances if gfx94 targets are not on the target list # Do not build mha instances if gfx94 or gfx90a targets are not on the target list
foreach(source IN LISTS ARGN) foreach(source IN LISTS ARGN)
if(NOT INST_TARGETS MATCHES "gfx94" AND source MATCHES "mha") if(NOT INST_TARGETS MATCHES "gfx94" AND NOT INST_TARGETS MATCHES "gfx90a" AND source MATCHES "mha")
message("removing mha instance ${source} ") message("removing mha instance ${source} ")
list(REMOVE_ITEM ARGN "${source}") list(REMOVE_ITEM ARGN "${source}")
endif() endif()
...@@ -75,17 +71,13 @@ function(add_instance_library INSTANCE_NAME) ...@@ -75,17 +71,13 @@ function(add_instance_library INSTANCE_NAME)
if(ARGN) if(ARGN)
set(INST_OBJ) set(INST_OBJ)
foreach(source IN LISTS ARGN) foreach(source IN LISTS ARGN)
if(INSTANCES_ONLY) set(INST_TARGETS ${SUPPORTED_GPU_TARGETS})
set(INST_TARGETS ${DEFAULT_GPU_TARGETS})
else()
set(INST_TARGETS ${GPU_TARGETS})
endif()
if(source MATCHES "_xdl") if(source MATCHES "_xdl")
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201) list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201)
elseif(ARGN MATCHES "_wmma") elseif(ARGN MATCHES "_wmma")
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030) list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030)
elseif(ARGN MATCHES "mha") elseif(ARGN MATCHES "mha")
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx908 gfx90a gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201) list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx908 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201)
endif() endif()
set(offload_targets) set(offload_targets)
foreach(target IN LISTS INST_TARGETS) foreach(target IN LISTS INST_TARGETS)
...@@ -191,12 +183,7 @@ FOREACH(subdir_path ${dir_list}) ...@@ -191,12 +183,7 @@ FOREACH(subdir_path ${dir_list})
set(add_inst 1) set(add_inst 1)
endif() endif()
if(INSTANCES_ONLY) set(INST_TARGETS ${SUPPORTED_GPU_TARGETS})
set(INST_TARGETS ${DEFAULT_GPU_TARGETS})
else()
set(INST_TARGETS ${GPU_TARGETS})
endif()
if(("${cmake_instance}" MATCHES "quantization") AND (DEFINED DTYPES) AND (NOT DTYPES MATCHES "int8")) if(("${cmake_instance}" MATCHES "quantization") AND (DEFINED DTYPES) AND (NOT DTYPES MATCHES "int8"))
message("quantization instances will not be built!") message("quantization instances will not be built!")
...@@ -320,8 +307,7 @@ if(CK_DEVICE_CONV_INSTANCES) ...@@ -320,8 +307,7 @@ if(CK_DEVICE_CONV_INSTANCES)
endif() endif()
if(CK_DEVICE_MHA_INSTANCES) if(CK_DEVICE_MHA_INSTANCES)
set(gpu_list ${INST_TARGETS}) set(gpu_list ${INST_TARGETS})
list(FILTER gpu_list INCLUDE REGEX "^gfx94") if(gpu_list MATCHES "gfx94" OR gpu_list MATCHES "gfx90a")
if(gpu_list)
add_library(device_mha_operations STATIC ${CK_DEVICE_MHA_INSTANCES}) add_library(device_mha_operations STATIC ${CK_DEVICE_MHA_INSTANCES})
add_library(composablekernels::device_mha_operations ALIAS device_mha_operations) add_library(composablekernels::device_mha_operations ALIAS device_mha_operations)
target_compile_features(device_mha_operations PUBLIC) target_compile_features(device_mha_operations PUBLIC)
......
profiler/src/CMakeLists.txt
View file @ d20c20a6
...@@ -24,7 +24,7 @@ set(PROFILER_SOURCES ...@@ -24,7 +24,7 @@ set(PROFILER_SOURCES
profile_permute_scale.cpp profile_permute_scale.cpp
) )
if(GPU_TARGETS MATCHES "gfx9") if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES) if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
list(APPEND PROFILER_SOURCES profile_contraction_bilinear.cpp) list(APPEND PROFILER_SOURCES profile_contraction_bilinear.cpp)
list(APPEND PROFILER_SOURCES profile_contraction_scale.cpp) list(APPEND PROFILER_SOURCES profile_contraction_scale.cpp)
...@@ -49,7 +49,7 @@ if(GPU_TARGETS MATCHES "gfx9") ...@@ -49,7 +49,7 @@ if(GPU_TARGETS MATCHES "gfx9")
list(APPEND PROFILER_SOURCES profile_grouped_gemm_multiply_tile_loop.cpp) list(APPEND PROFILER_SOURCES profile_grouped_gemm_multiply_tile_loop.cpp)
endif() endif()
list(APPEND PROFILER_SOURCES profile_gemm_multiply_add.cpp) list(APPEND PROFILER_SOURCES profile_gemm_multiply_add.cpp)
if(GPU_TARGETS MATCHES "gfx94") if(SUPPORTED_GPU_TARGETS MATCHES "gfx94")
list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp) list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_ab_scale.cpp) list(APPEND PROFILER_SOURCES profile_gemm_ab_scale.cpp)
endif() endif()
...@@ -69,7 +69,7 @@ if(GPU_TARGETS MATCHES "gfx9") ...@@ -69,7 +69,7 @@ if(GPU_TARGETS MATCHES "gfx9")
endif() endif()
if(GPU_TARGETS MATCHES "gfx11" OR GPU_TARGETS MATCHES "gfx12" OR GPU_TARGETS MATCHES "gfx9") if(SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12" OR SUPPORTED_GPU_TARGETS MATCHES "gfx9")
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
list(APPEND PROFILER_SOURCES profile_gemm_bilinear.cpp) list(APPEND PROFILER_SOURCES profile_gemm_bilinear.cpp)
endif() endif()
...@@ -111,7 +111,7 @@ target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_column_to_image_inst ...@@ -111,7 +111,7 @@ target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_column_to_image_inst
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_transpose_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_transpose_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_permute_scale_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_permute_scale_instance)
if(GPU_TARGETS MATCHES "gfx9") if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES) if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance)
...@@ -135,7 +135,7 @@ if(GPU_TARGETS MATCHES "gfx9") ...@@ -135,7 +135,7 @@ if(GPU_TARGETS MATCHES "gfx9")
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_add_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_add_instance)
if(GPU_TARGETS MATCHES "gfx94") if(SUPPORTED_GPU_TARGETS MATCHES "gfx94")
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_ab_scale_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_ab_scale_instance)
endif() endif()
...@@ -159,7 +159,7 @@ if(GPU_TARGETS MATCHES "gfx9") ...@@ -159,7 +159,7 @@ if(GPU_TARGETS MATCHES "gfx9")
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convinvscale_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convinvscale_instance)
endif() endif()
if(GPU_TARGETS MATCHES "gfx9" OR GPU_TARGETS MATCHES "gfx11" OR GPU_TARGETS MATCHES "gfx12") if(SUPPORTED_GPU_TARGETS MATCHES "gfx9" OR SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12")
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bilinear_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bilinear_instance)
endif() endif()
......
script/cmake-ck-dev.sh
View file @ d20c20a6
...@@ -7,7 +7,8 @@ MY_PROJECT_SOURCE=$1 ...@@ -7,7 +7,8 @@ MY_PROJECT_SOURCE=$1
if [ $# -ge 2 ] ; then if [ $# -ge 2 ] ; then
GPU_TARGETS=$2 GPU_TARGETS=$2
REST_ARGS=${@:3} shift 2
REST_ARGS=$@
else else
GPU_TARGETS="gfx908;gfx90a;gfx940" GPU_TARGETS="gfx908;gfx90a;gfx940"
REST_ARGS= REST_ARGS=
......
script/cmake-ck-release.sh
View file @ d20c20a6
...@@ -7,7 +7,8 @@ MY_PROJECT_SOURCE=$1 ...@@ -7,7 +7,8 @@ MY_PROJECT_SOURCE=$1
if [ $# -ge 2 ] ; then if [ $# -ge 2 ] ; then
GPU_TARGETS=$2 GPU_TARGETS=$2
REST_ARGS=${@:3} shift 2
REST_ARGS=$@
else else
GPU_TARGETS="gfx908;gfx90a;gfx940" GPU_TARGETS="gfx908;gfx90a;gfx940"
REST_ARGS= REST_ARGS=
......
test/CMakeLists.txt
View file @ d20c20a6
This diff is collapsed.
test/data_type/CMakeLists.txt
View file @ d20c20a6
...@@ -18,4 +18,9 @@ if(result EQUAL 0) ...@@ -18,4 +18,9 @@ if(result EQUAL 0)
target_link_libraries(test_bf8 PRIVATE utility) target_link_libraries(test_bf8 PRIVATE utility)
endif() endif()
add_gtest_executable(test_custom_type test_custom_type.cpp)
if(result EQUAL 0)
target_link_libraries(test_custom_type PRIVATE utility)
endif()
add_gtest_executable(test_type_convert_const type_convert_const.cpp) add_gtest_executable(test_type_convert_const type_convert_const.cpp)
test/data_type/test_custom_type.cpp 0 → 100644
View file @ d20c20a6
This diff is collapsed.
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