Skip to content

GitLab

Unverified Commit b53e9d08 authored by zjing14's avatar zjing14 Committed by GitHub
Browse files

Batched GEMM for fp16 (#79) 

* prepare host for batched_gemm

* init commit of batched kernels

* fixed

* refine transform with freeze

* m/n padding

* fixed a bug; clean

* add small tiles

* clean

* clean code

* clean code

* add nt, tn, tt layout

* add missing file

* use StaticBufferTupleOfVector instead

* add reference_batched_gemm

* fixed a macro
parent 6f928a08
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 2098 additions and 13 deletions
composable_kernel/include/tensor_operation/blockwise_gemm_xdlops.hpp
View file @ b53e9d08
......@@ -37,10 +37,11 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
static constexpr index_t MWaves = MPerBlock / (MRepeat * MPerXDL);
static constexpr index_t NWaves = NPerBlock / (NRepeat * NPerXDL);
StaticBufferOfVectorTypeV2<AddressSpaceEnum_t::Vgpr,
vector_type<FloatAcc, xdlops_gemm.GetRegSizePerXdlops()>,
MRepeat * NRepeat,
true>
StaticBufferTupleOfVector<AddressSpaceEnum_t::Vgpr,
FloatAcc,
MRepeat * NRepeat,
xdlops_gemm.GetRegSizePerXdlops(),
true>
c_thread_buf_;
__host__ __device__ constexpr auto& GetCThreadBuffer() { return c_thread_buf_; }
......@@ -140,6 +141,19 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M0, M1, M2, N));
}
__host__ __device__ static constexpr auto GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()
{
constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
constexpr auto N = c_m0_m1_m2_n_tblk_lens[I3];
return make_naive_tensor_descriptor_packed(
make_tuple(I1, Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M0, M1, M2, N));
}
__host__ __device__ static constexpr auto GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()
{
constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2 =
......@@ -153,6 +167,21 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_block_desc_m0_n0_m1_n1_m2_n2);
}
__host__ __device__ static constexpr auto GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()
{
constexpr auto c_block_desc_g_m0_n0_m1_n1_m2_n2 =
make_naive_tensor_descriptor_packed(make_tuple(I1,
Number<MRepeat>{},
Number<NRepeat>{},
Number<MWaves>{},
Number<NWaves>{},
Number<MPerXDL>{},
Number<NPerXDL>{}));
return xdlops_gemm.MakeCDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(
c_block_desc_g_m0_n0_m1_n1_m2_n2);
}
template <typename CGridDesc_M_N>
__host__ __device__ static constexpr auto
MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_M_N& c_grid_desc_m_n)
......@@ -170,6 +199,26 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m0_n0_m1_n1_m2_n2);
}
template <typename CGridDesc_G_M_N>
__host__ __device__ static constexpr auto
MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_G_M_N& c_grid_desc_g_m_n)
{
const auto G = c_grid_desc_g_m_n.GetLength(I0);
const auto M = c_grid_desc_g_m_n.GetLength(I1);
const auto N = c_grid_desc_g_m_n.GetLength(I2);
const auto c_grid_desc_g_m0_n0_m1_n1_m2_n2 = transform_tensor_descriptor(
c_grid_desc_g_m_n,
make_tuple(make_pass_through_transform(G),
make_unmerge_transform(make_tuple(M / (MWaves * MPerXDL), MWaves, MPerXDL)),
make_unmerge_transform(make_tuple(N / (NWaves * NPerXDL), NWaves, NPerXDL))),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1, 3, 5>{}, Sequence<2, 4, 6>{}));
return xdlops_gemm.MakeCDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(
c_grid_desc_g_m0_n0_m1_n1_m2_n2);
}
__host__ __device__ static constexpr auto MakeABlockDescriptor_K0_M0_M1_M2_K1()
{
return transform_tensor_descriptor(
......@@ -239,11 +288,13 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
using mfma_input_type =
typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset = c_thread_desc_.CalculateOffset(make_tuple(m0, n0));
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.template Run(a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVector(Number<c_offset>{}));
xdlops_gemm.template Run(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
......@@ -258,9 +309,9 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
static constexpr auto b_thread_desc_ =
make_naive_tensor_descriptor_packed(make_tuple(Number<K0>{}, I1, I1, I1, Number<K1>{}));
// C[M, N]
static constexpr auto c_thread_desc_ =
make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{}, Number<NRepeat>{}));
// C[M, N, NumRegXdlops]
static constexpr auto c_thread_desc_ = make_naive_tensor_descriptor_packed(
make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, xdlops_gemm.GetRegSizePerXdlops()));
using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
FloatAB,
......
composable_kernel/include/tensor_operation/xdlops_gemm.hpp
View file @ b53e9d08
......@@ -614,6 +614,43 @@ struct XdlopsGemm
Sequence<7>{}));
}
template <typename CDesc_G_M0_N0_M1_N1_M2_N2>
__host__ __device__ static constexpr auto MakeCDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(
const CDesc_G_M0_N0_M1_N1_M2_N2& c_desc_g_m0_n0_m1_n1_m2_n2)
{
const auto G = c_desc_g_m0_n0_m1_n1_m2_n2.GetLength(I0);
const auto M0 = c_desc_g_m0_n0_m1_n1_m2_n2.GetLength(I1);
const auto N0 = c_desc_g_m0_n0_m1_n1_m2_n2.GetLength(I2);
const auto M1 = c_desc_g_m0_n0_m1_n1_m2_n2.GetLength(I3);
const auto N1 = c_desc_g_m0_n0_m1_n1_m2_n2.GetLength(I4);
return transform_tensor_descriptor(
c_desc_g_m0_n0_m1_n1_m2_n2,
make_tuple(make_pass_through_transform(G),
make_pass_through_transform(M0),
make_pass_through_transform(N0),
make_pass_through_transform(M1),
make_pass_through_transform(N1),
make_unmerge_transform(make_tuple(mfma_instr.num_groups_per_blk,
mfma_instr.num_input_blks,
mfma_instr.group_size)),
make_pass_through_transform(mfma_instr.num_threads_per_blk)),
make_tuple(Sequence<0>{},
Sequence<1>{},
Sequence<2>{},
Sequence<3>{},
Sequence<4>{},
Sequence<5>{},
Sequence<6>{}),
make_tuple(Sequence<0>{},
Sequence<1>{},
Sequence<2>{},
Sequence<3>{},
Sequence<4>{},
Sequence<5, 6, 7>{},
Sequence<8>{}));
}
__device__ static constexpr index_t GetRegSizePerXdlops()
{
return MPerXdlops * NPerXdlops / mfma_instr.wave_size;
......
composable_kernel/include/utility/static_buffer.hpp
View file @ b53e9d08
......@@ -149,6 +149,13 @@ struct StaticBufferTupleOfVector
return base::operator()(i_v);
}
__host__ __device__ void Clear()
{
const index_t numScalars = NumOfVector * ScalarPerVector;
static_for<0, Number<numScalars>{}, 1>{}([&](auto i) { SetAsType(i, S{0}); });
}
};
template <AddressSpaceEnum_t AddressSpace, typename T, index_t N>
......
composable_kernel/include/utility/static_buffer_of_vector_type_v2.hpp
View file @ b53e9d08
......@@ -104,6 +104,11 @@ struct StaticBufferOfVectorTypeV2 : public StaticallyIndexedArray<T, N>
[&](auto i) { GetElement(i, true) = invalid_element_value_; });
}
__host__ __device__ void Fill(VecBaseType v)
{
static_for<0, GetNumElements(), 1>{}([&](auto i) { GetElement(i, true) = v; });
}
__host__ __device__ static constexpr bool IsStaticBuffer() { return true; }
__host__ __device__ static constexpr bool IsDynamicBuffer() { return false; }
......
device_operation/CMakeLists.txt
View file @ b53e9d08
......@@ -48,6 +48,13 @@ set(DEVICE_GEMM_BIAS_RELU_ADD_INSTANCE_SOURCE
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_c_shuffle_bias_relu_add_f16_f16_f16_km_nk_mn_instance.cpp;
)
set(DEVICE_BATCHED_GEMM_INSTANCE_SOURCE
${PROJECT_SOURCE_DIR}/device_operation/src/device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instance.cpp;
)
# device_conv2d_fwd_instance
set(DEVICE_CONV2D_FWD_INSTANCE_SOURCE
${PROJECT_SOURCE_DIR}/device_operation/src/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instance.cpp;
......@@ -73,6 +80,7 @@ set(DEVICE_CONV2D_FWD_BIAS_RELU_ATOMIC_ADD_INSTANCE_SOURCE
add_library(device_gemm_instance SHARED ${DEVICE_GEMM_INSTANCE_SOURCE})
add_library(device_gemm_bias_relu_instance SHARED ${DEVICE_GEMM_BIAS_RELU_INSTANCE_SOURCE})
add_library(device_gemm_bias_relu_add_instance SHARED ${DEVICE_GEMM_BIAS_RELU_ADD_INSTANCE_SOURCE})
add_library(device_batched_gemm_instance SHARED ${DEVICE_BATCHED_GEMM_INSTANCE_SOURCE})
add_library(device_conv2d_fwd_instance SHARED ${DEVICE_CONV2D_FWD_INSTANCE_SOURCE})
add_library(device_conv2d_fwd_bias_relu_instance SHARED ${DEVICE_CONV2D_FWD_BIAS_RELU_INSTANCE_SOURCE})
add_library(device_conv2d_fwd_bias_relu_add_instance SHARED ${DEVICE_CONV2D_FWD_BIAS_RELU_ADD_INSTANCE_SOURCE})
......@@ -81,6 +89,7 @@ add_library(device_conv2d_fwd_bias_relu_atomic_add_instance SHARED ${DEVICE_CONV
target_include_directories(device_gemm_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
target_include_directories(device_gemm_bias_relu_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
target_include_directories(device_gemm_bias_relu_add_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
target_include_directories(device_batched_gemm_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
target_include_directories(device_conv2d_fwd_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
target_include_directories(device_conv2d_fwd_bias_relu_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
target_include_directories(device_conv2d_fwd_bias_relu_add_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
......@@ -89,6 +98,7 @@ target_include_directories(device_conv2d_fwd_bias_relu_atomic_add_instance SYSTE
target_compile_features(device_gemm_instance PUBLIC)
target_compile_features(device_gemm_bias_relu_instance PUBLIC)
target_compile_features(device_gemm_bias_relu_add_instance PUBLIC)
target_compile_features(device_batched_gemm_instance PUBLIC)
target_compile_features(device_conv2d_fwd_instance PUBLIC)
target_compile_features(device_conv2d_fwd_bias_relu_instance PUBLIC)
target_compile_features(device_conv2d_fwd_bias_relu_add_instance PUBLIC)
......@@ -97,6 +107,7 @@ target_compile_features(device_conv2d_fwd_bias_relu_atomic_add_instance PUBLIC)
set_target_properties(device_gemm_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
set_target_properties(device_gemm_bias_relu_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
set_target_properties(device_gemm_bias_relu_add_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
set_target_properties(device_batched_gemm_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
set_target_properties(device_conv2d_fwd_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
set_target_properties(device_conv2d_fwd_bias_relu_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
set_target_properties(device_conv2d_fwd_bias_relu_add_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
......@@ -105,6 +116,7 @@ set_target_properties(device_conv2d_fwd_bias_relu_atomic_add_instance PROPERTIES
install(TARGETS device_gemm_instance LIBRARY DESTINATION lib)
install(TARGETS device_gemm_bias_relu_instance LIBRARY DESTINATION lib)
install(TARGETS device_gemm_bias_relu_add_instance LIBRARY DESTINATION lib)
install(TARGETS device_batched_gemm_instance LIBRARY DESTINATION lib)
install(TARGETS device_conv2d_fwd_instance LIBRARY DESTINATION lib)
install(TARGETS device_conv2d_fwd_bias_relu_instance LIBRARY DESTINATION lib)
install(TARGETS device_conv2d_fwd_bias_relu_add_instance LIBRARY DESTINATION lib)
......
device_operation/include/device_batched_gemm_xdl.hpp 0 → 100644
View file @ b53e9d08
This diff is collapsed.
device_operation/src/device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instance.cpp 0 → 100644
View file @ b53e9d08
#include <stdlib.h>
#include "config.hpp"
#include "device_batched_gemm_xdl.hpp"
#include "element_wise_operation.hpp"
#include "device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_batched_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
using device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instances =
std::tuple<
// clang-format off
//##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
//##########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar|
//##########| | | | | | | | Operation| Operation| Operation| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
//##########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceBatchedGemmXdl< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 8, 1>
// clang-format on
>;
void add_device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instances(
std::vector<DeviceGemmPtr<PassThrough, PassThrough, PassThrough>>& instances)
{
add_device_operation_instances(instances,
device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instances{});
}
} // namespace device_batched_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
device_operation/src/device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instance.cpp 0 → 100644
View file @ b53e9d08
#include <stdlib.h>
#include "config.hpp"
#include "device_batched_gemm_xdl.hpp"
#include "element_wise_operation.hpp"
#include "device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_batched_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
using device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instances =
std::tuple<
// clang-format off
//##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
//##########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar|
//##########| | | | | | | | Operation| Operation| Operation| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
//##########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceBatchedGemmXdl< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 8, 1>
// clang-format on
>;
void add_device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instances(
std::vector<DeviceGemmPtr<PassThrough, PassThrough, PassThrough>>& instances)
{
add_device_operation_instances(instances,
device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instances{});
}
} // namespace device_batched_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
device_operation/src/device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instance.cpp 0 → 100644
View file @ b53e9d08
#include <stdlib.h>
#include "config.hpp"
#include "device_batched_gemm_xdl.hpp"
#include "element_wise_operation.hpp"
#include "device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_batched_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
using device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instances =
std::tuple<
// clang-format off
//####################| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
//####################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar|
//####################| | | | | | | | Operation| Operation| Operation| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
//####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 128, 32, 32, 4, 8, 16, 16, 2, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 64, 32, 32, 4, 8, 32, 32, 1, 1, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 64, 16, 16, 4, 8, 16, 16, 1, 1, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, 8, 1>
// clang-format on
>;
void add_device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instances(
std::vector<DeviceGemmPtr<PassThrough, PassThrough, PassThrough>>& instances)
{
add_device_operation_instances(instances,
device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instances{});
}
} // namespace device_batched_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
device_operation/src/device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instance.cpp 0 → 100644
View file @ b53e9d08
#include <stdlib.h>
#include "config.hpp"
#include "device_batched_gemm_xdl.hpp"
#include "element_wise_operation.hpp"
#include "device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_batched_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
using device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instances = std::tuple<
// clang-format off
//#################| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
//#################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar|
//#################| | | | | | | | Operation| Operation| Operation| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
//#################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 8, 1>,
DeviceBatchedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 8, 1>
// clang-format on
>;
void add_device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instances(
std::vector<DeviceGemmPtr<PassThrough, PassThrough, PassThrough>>& instances)
{
add_device_operation_instances(instances,
device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instances{});
}
} // namespace device_batched_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
profiler/CMakeLists.txt
View file @ b53e9d08
......@@ -23,6 +23,7 @@ set(PROFILER_SOURCE
src/profile_conv_fwd_bias_relu.cpp
src/profile_conv_fwd_bias_relu_add.cpp
src/profile_conv_fwd_bias_relu_atomic_add.cpp
src/profile_batched_gemm.cpp
)
add_executable(ckProfiler ${PROFILER_SOURCE})
......@@ -35,3 +36,4 @@ target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_instance)
target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_bias_relu_instance)
target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_bias_relu_add_instance)
target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_bias_relu_atomic_add_instance)
target_link_libraries(ckProfiler PRIVATE device_batched_gemm_instance)
profiler/include/profile_batched_gemm_impl.hpp 0 → 100644
View file @ b53e9d08
#pragma once
#include "reference_batched_gemm.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_batched_gemm_instance {
using DeviceGemmNoOpPtr =
ck::tensor_operation::device::DeviceGemmPtr<ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>;
void add_device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instances(std::vector<DeviceGemmNoOpPtr>&);
} // namespace device_batched_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
namespace ck {
namespace profiler {
template <typename ADataType,
typename BDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename CLayout>
void profile_batched_gemm_impl(int do_verification,
int init_method,
bool do_log,
int nrepeat,
int M,
int N,
int K,
int StrideA,
int StrideB,
int StrideC,
int BatchCount = 1)
{
auto f_host_tensor_descriptor = [](std::size_t batch_count,
std::size_t row,
std::size_t col,
std::size_t stride,
auto layout) {
if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
std::vector<std::size_t>({row * stride, stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
std::vector<std::size_t>({col * stride, 1, stride}));
}
};
Tensor<ADataType> a_g_m_k(f_host_tensor_descriptor(BatchCount, M, K, StrideA, ALayout{}));
Tensor<BDataType> b_g_k_n(f_host_tensor_descriptor(BatchCount, K, N, StrideB, BLayout{}));
Tensor<CDataType> c_g_m_n_host_result(
f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
Tensor<CDataType> c_g_m_n_device_result(
f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
std::cout << "b_g_k_n: " << b_g_k_n.mDesc << std::endl;
std::cout << "c_g_m_n: " << c_g_m_n_host_result.mDesc << std::endl;
std::size_t num_thread = std::thread::hardware_concurrency();
switch(init_method)
{
case 0: break;
case 1:
a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread);
b_g_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
break;
default:
a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}, num_thread);
b_g_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
}
// set zero to c_device_buf
c_g_m_n_device_result.GenerateTensorValue(GeneratorTensor_0<CDataType>{}, num_thread);
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
if(do_verification)
{
using ReferenceBatchedGemmInstance =
ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
BDataType,
CDataType,
AElementOp,
BElementOp,
CElementOp>;
auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
auto ref_invoker = ref_batched_gemm.MakeInvoker();
auto ref_argument = ref_batched_gemm.MakeArgument(
a_g_m_k, b_g_k_n, c_g_m_n_host_result, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
}
DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpace());
DeviceMem b_device_buf(sizeof(BDataType) * b_g_k_n.mDesc.GetElementSpace());
DeviceMem c_device_buf(sizeof(CDataType) * c_g_m_n_device_result.mDesc.GetElementSpace());
a_device_buf.ToDevice(a_g_m_k.mData.data());
b_device_buf.ToDevice(b_g_k_n.mData.data());
c_device_buf.ToDevice(c_g_m_n_device_result.mData.data());
// add device GEMM instances
std::vector<ck::tensor_operation::device::device_batched_gemm_instance::DeviceGemmNoOpPtr>
gemm_ptrs;
if constexpr(is_same<ADataType, half_t>::value && is_same<BDataType, half_t>::value &&
is_same<CDataType, half_t>::value)
{
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_batched_gemm_instance::
add_device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_batched_gemm_instance::
add_device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_batched_gemm_instance::
add_device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_batched_gemm_instance::
add_device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instances(gemm_ptrs);
}
}
if(gemm_ptrs.size() <= 0)
{
throw std::runtime_error("wrong! no device GEMM instance found");
}
std::string best_gemm_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device GEMM instances
for(auto& gemm_ptr : gemm_ptrs)
{
auto argument_ptr =
gemm_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
StrideC,
ck::tensor_operation::element_wise::PassThrough{},
ck::tensor_operation::element_wise::PassThrough{},
ck::tensor_operation::element_wise::PassThrough{},
BatchCount);
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
{
std::string gemm_name = gemm_ptr->GetTypeString();
float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
std::size_t flop = std::size_t(2) * BatchCount * M * N * K;
std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(BDataType) * K * M +
sizeof(CDataType) * M * N) *
BatchCount;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s, " << gemm_name << std::endl;
if(tflops > best_tflops)
{
best_gemm_name = gemm_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
c_device_buf.FromDevice(c_g_m_n_device_result.mData.data());
check_error(c_g_m_n_host_result, c_g_m_n_device_result);
if(do_log)
{
LogRangeAsType<float>(std::cout << "a : ", a_g_m_k.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "b: ", b_g_k_n.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "c_host: ", c_g_m_n_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "c_device: ", c_g_m_n_device_result.mData, ",")
<< std::endl;
}
}
}
else
{
std::cout << "this device GEMM instance does not support this GEMM problem"
<< std::endl;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_gemm_name << std::endl;
}
} // namespace profiler
} // namespace ck
profiler/src/profile_batched_gemm.cpp 0 → 100644
View file @ b53e9d08
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "config.hpp"
#include "print.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_gemm.hpp"
#include "device_tensor.hpp"
#include "device_base.hpp"
#include "device_batched_gemm_xdl.hpp"
#include "profile_batched_gemm_impl.hpp"
enum GemmMatrixLayout
{
MK_KN_MN, // 0
MK_NK_MN, // 1
KM_KN_MN, // 2
KM_NK_MN, // 3
MK_KN_NM, // 4
MK_NK_NM, // 5
KM_KN_NM, // 6
KM_NK_NM, // 7
};
enum GemmDataType
{
F32_F32_F32, // 0
F16_F16_F16, // 1
};
int profile_batched_gemm(int argc, char* argv[])
{
if(!(argc == 15))
{
printf("arg1: tensor operation (batched_gemm: Batched GEMM)\n");
printf("arg2: data type (0: fp32; 1: fp16)\n");
printf("arg3: matrix layout (0: A[g, m, k] * B[g, k, n] = C[g, m, n];\n");
printf(" 1: A[g, m, k] * B[g, n, k] = C[g, m, n];\n");
printf(" 2: A[g, k, m] * B[g, k, n] = C[g, m, n];\n");
printf(" 3: A[g, k, m] * B[g, n, k] = C[g, m, n])\n");
printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg8: print tensor value (0: no; 1: yes)\n");
printf("arg7: run kernel # of times (>1)\n");
printf("arg8 to 14: M, N, K, StrideA, StrideB, StrideC, BatchCount\n");
exit(1);
}
const int data_type = static_cast<GemmDataType>(std::stoi(argv[2]));
const int layout = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
const bool do_verification = std::stoi(argv[4]);
const int init_method = std::stoi(argv[5]);
const bool do_log = std::stoi(argv[6]);
const int nrepeat = std::stoi(argv[7]);
const int M = std::stoi(argv[8]);
const int N = std::stoi(argv[9]);
const int K = std::stoi(argv[10]);
const int StrideA = std::stoi(argv[11]);
const int StrideB = std::stoi(argv[12]);
const int StrideC = std::stoi(argv[13]);
const int BatchCount = std::stoi(argv[14]);
if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_batched_gemm_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
nrepeat,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_batched_gemm_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
nrepeat,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
{
ck::profiler::profile_batched_gemm_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
nrepeat,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
{
ck::profiler::profile_batched_gemm_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
nrepeat,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC);
}
else
{
throw std::runtime_error("wrong! this GEMM data_type & layout is not implemented");
}
return 1;
}
profiler/src/profiler.cpp
View file @ b53e9d08
......@@ -6,6 +6,7 @@
#include <half.hpp>
int profile_gemm(int, char*[]);
int profile_batched_gemm(int, char*[]);
int profile_gemm_bias_relu(int, char*[]);
int profile_gemm_bias_relu_add(int, char*[]);
int profile_conv_fwd(int, char*[]);
......@@ -19,14 +20,18 @@ int main(int argc, char* argv[])
{
return profile_gemm(argc, argv);
}
if(strcmp(argv[1], "gemm_bias_relu") == 0)
else if(strcmp(argv[1], "gemm_bias_relu") == 0)
{
return profile_gemm_bias_relu(argc, argv);
}
if(strcmp(argv[1], "gemm_bias_relu_add") == 0)
else if(strcmp(argv[1], "gemm_bias_relu_add") == 0)
{
return profile_gemm_bias_relu_add(argc, argv);
}
else if(strcmp(argv[1], "batched_gemm") == 0)
{
return profile_batched_gemm(argc, argv);
}
else if(strcmp(argv[1], "conv_fwd") == 0)
{
return profile_conv_fwd(argc, argv);
......
reference_operation/include/reference_batched_gemm.hpp 0 → 100644
View file @ b53e9d08
#ifndef REFERENCE_BATCHED_GEMM_HPP
#define REFERENCE_BATCHED_GEMM_HPP
#include <iostream>
#include <sstream>
#include "device_base.hpp"
#include "host_tensor.hpp"
namespace ck {
namespace tensor_operation {
namespace host {
template <typename ADataType,
typename BDataType,
typename CDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
struct ReferenceBatchedGemm : public device::BaseOperator
{
// Argument
struct Argument : public device::BaseArgument
{
Argument(const Tensor<ADataType>& a_g_m_k,
const Tensor<BDataType>& b_g_k_n,
Tensor<CDataType>& c_g_m_n,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op)
: a_g_m_k_{a_g_m_k},
b_g_k_n_{b_g_k_n},
c_g_m_n_{c_g_m_n},
a_element_op_{a_element_op},
b_element_op_{b_element_op},
c_element_op_{c_element_op}
{
}
const Tensor<ADataType>& a_g_m_k_;
const Tensor<BDataType>& b_g_k_n_;
Tensor<CDataType>& c_g_m_n_;
AElementwiseOperation a_element_op_;
BElementwiseOperation b_element_op_;
CElementwiseOperation c_element_op_;
};
// Invoker
struct Invoker : public device::BaseInvoker
{
using Argument = ReferenceBatchedGemm::Argument;
float Run(const Argument& arg)
{
auto f_gmk_gkn_gmn = [&](auto g, auto m, auto n) {
const int K = arg.a_g_m_k_.mDesc.GetLengths()[2];
float v_acc = 0;
for(int k = 0; k < K; ++k)
{
float v_a;
float v_b;
arg.a_element_op_(v_a, static_cast<const float>(arg.a_g_m_k_(g, m, k)));
arg.b_element_op_(v_b, static_cast<const float>(arg.b_g_k_n_(g, k, n)));
v_acc += v_a * v_b;
}
float v_c;
arg.c_element_op_(v_c, v_acc);
arg.c_g_m_n_(g, m, n) = v_c;
};
make_ParallelTensorFunctor(f_gmk_gkn_gmn,
arg.c_g_m_n_.mDesc.GetLengths()[0],
arg.c_g_m_n_.mDesc.GetLengths()[1],
arg.c_g_m_n_.mDesc.GetLengths()[2])(
std::thread::hardware_concurrency());
return 0;
}
float Run(const device::BaseArgument* p_arg, int) override
{
return Run(*dynamic_cast<const Argument*>(p_arg));
}
};
static constexpr bool IsValidCompilationParameter()
{
// TODO: properly implement this check
return true;
}
bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
static auto MakeArgument(const Tensor<ADataType>& a_g_m_k,
const Tensor<BDataType>& b_g_k_n,
Tensor<CDataType>& c_g_m_n,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op)
{
return Argument{a_g_m_k, b_g_k_n, c_g_m_n, a_element_op, b_element_op, c_element_op};
}
static auto MakeInvoker() { return Invoker{}; }
virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
{
return std::make_unique<Invoker>(Invoker{});
}
std::string GetTypeString() const override
{
auto str = std::stringstream();
// clang-format off
str << "ReferenceBatchedGemm"
<< std::endl;
// clang-format on
return str.str();
}
};
} // namespace host
} // namespace tensor_operation
} // namespace ck
#endif
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