Skip to content

GitLab

Commit d92fb7e8 authored by rocking's avatar rocking
Browse files

Merge commit 'a3c910ac' into gemm_softmax

parents bfc80764 a3c910ac
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 1194 additions and 815 deletions
include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp
View file @ d92fb7e8
...@@ -277,9 +277,12 @@ struct ThreadwiseTensorSliceTransfer_v3r1 ...@@ -277,9 +277,12 @@ struct ThreadwiseTensorSliceTransfer_v3r1
// sub-dword transpose between src_thread_scratch_ and dst_thread_scratch_ // sub-dword transpose between src_thread_scratch_ and dst_thread_scratch_
// TODO make this logic more generic for more sub-dword datatype // TODO make this logic more generic for more sub-dword datatype
if constexpr(SrcVectorDim != DstVectorDim && if constexpr(SrcVectorDim != DstVectorDim &&
is_same<half_t, remove_cvref_t<SrcData>>::value && ((is_same<half_t, remove_cvref_t<SrcData>>::value &&
is_same<half_t, remove_cvref_t<DstData>>::value && is_same<half_t, remove_cvref_t<DstData>>::value &&
SrcScalarPerVector % 2 == 0 && DstScalarPerVector % 2 == 0) SrcScalarPerVector % 2 == 0 && DstScalarPerVector % 2 == 0) ||
(is_same<int8_t, remove_cvref_t<SrcData>>::value &&
is_same<int8_t, remove_cvref_t<DstData>>::value &&
SrcScalarPerVector % 4 == 0 && DstScalarPerVector % 4 == 0)))
{ {
// each transpose does // each transpose does
// DstScalarPerVector # of src vectors in src_thread_scratch_ // DstScalarPerVector # of src vectors in src_thread_scratch_
......
include/ck/utility/transpose_vectors.hpp
View file @ d92fb7e8
...@@ -49,7 +49,7 @@ __device__ void transpose_fp16_2x2(const half2_t& x0, const half2_t& x1, half2_t ...@@ -49,7 +49,7 @@ __device__ void transpose_fp16_2x2(const half2_t& x0, const half2_t& x1, half2_t
template <index_t NX, index_t NY> template <index_t NX, index_t NY>
struct transpose_vectors<half_t, NX, NY> struct transpose_vectors<half_t, NX, NY>
{ {
// we got [NY * NX] ammount of S data to be transposed // we got [NY * NX] amount of S data to be transposed
static constexpr index_t s_per_x = NY; static constexpr index_t s_per_x = NY;
static constexpr index_t s_per_y = NX; static constexpr index_t s_per_y = NX;
...@@ -83,5 +83,86 @@ struct transpose_vectors<half_t, NX, NY> ...@@ -83,5 +83,86 @@ struct transpose_vectors<half_t, NX, NY>
} }
}; };
// transpose int8 4x4
__device__ void transpose_int8_4x4(const int8x4_t& x0,
const int8x4_t& x1,
const int8x4_t& x2,
const int8x4_t& x3,
int8x4_t& y0,
int8x4_t& y1,
int8x4_t& y2,
int8x4_t& y3)
{
int32_t t0, t1;
int32_t z0, z1, z2, z3;
constexpr int32_t m0 = 0x05010400;
constexpr int32_t m1 = 0x05040100;
constexpr int32_t m2 = 0x07060302;
constexpr int32_t m3 = 0x07030602;
// ex: v_perm_b32(0x 11 22 33 44, 0x 55 66 77 88, 0x 05 01 04 00) -> 0x33774488
// -- -- -- -- -- -- -- -- - - - -
// index 7 6 5 4 3 2 1 0 33 77 44 88
// index is reversed because of little endianness (least significant bits first)
// clang-format off
asm volatile("v_perm_b32 %0, %1, %2, %3" : "=v"(t0) : "v"(bit_cast<int32_t>(x1)), "v"(bit_cast<int32_t>(x0)), "s"(m0));
asm volatile("v_perm_b32 %0, %1, %2, %3" : "=v"(t1) : "v"(bit_cast<int32_t>(x3)), "v"(bit_cast<int32_t>(x2)), "s"(m0));
asm volatile("v_perm_b32 %0, %1, %2, %3" : "=v"(z0) : "v"(bit_cast<int32_t>(t1)), "v"(bit_cast<int32_t>(t0)), "s"(m1));
asm volatile("v_perm_b32 %0, %1, %2, %3" : "=v"(z1) : "v"(bit_cast<int32_t>(t1)), "v"(bit_cast<int32_t>(t0)), "s"(m2));
asm volatile("v_perm_b32 %0, %1, %2, %3" : "=v"(t0) : "v"(bit_cast<int32_t>(x1)), "v"(bit_cast<int32_t>(x0)), "s"(m3));
asm volatile("v_perm_b32 %0, %1, %2, %3" : "=v"(t1) : "v"(bit_cast<int32_t>(x3)), "v"(bit_cast<int32_t>(x2)), "s"(m3));
asm volatile("v_perm_b32 %0, %1, %2, %3" : "=v"(z2) : "v"(bit_cast<int32_t>(t1)), "v"(bit_cast<int32_t>(t0)), "s"(m1));
asm volatile("v_perm_b32 %0, %1, %2, %3" : "=v"(z3) : "v"(bit_cast<int32_t>(t1)), "v"(bit_cast<int32_t>(t0)), "s"(m2));
// clang-format on
y0 = bit_cast<int8x4_t>(z0);
y1 = bit_cast<int8x4_t>(z1);
y2 = bit_cast<int8x4_t>(z2);
y3 = bit_cast<int8x4_t>(z3);
}
template <index_t NX, index_t NY>
struct transpose_vectors<int8_t, NX, NY>
{
// we got [NY * NX] amount of S data to be transposed
static constexpr index_t s_per_x = NY;
static constexpr index_t s_per_y = NX;
using S = int8_t;
using VX = vector_type<int8_t, s_per_x>;
using VY = vector_type<int8_t, s_per_y>;
__device__ void operator()(const StaticallyIndexedArray<const VX&, NX>& vx_tuple,
StaticallyIndexedArray<VY&, NY>& vy_tuple)
{
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
static constexpr auto I4 = Number<4>{};
static_assert((NX % 4 == 0 && NY % 4 == 0), "wrong!");
// loop over 4x4 tile and transpose data from vx_tuple into vy_tuple
static_for<0, NY, 4>{}([&](auto iy) {
static_for<0, NX, 4>{}([&](auto ix) {
// reference to 4 int8 data from vx_tuple
const auto& x_s4_0 = vx_tuple[ix].template AsType<int8x4_t>()[iy / I4];
const auto& x_s4_1 = vx_tuple[ix + I1].template AsType<int8x4_t>()[iy / I4];
const auto& x_s4_2 = vx_tuple[ix + I2].template AsType<int8x4_t>()[iy / I4];
const auto& x_s4_3 = vx_tuple[ix + I3].template AsType<int8x4_t>()[iy / I4];
// reference to 4 int8 data from vy_tuple
auto& y_s4_0 = vy_tuple(iy).template AsType<int8x4_t>()(ix / I4);
auto& y_s4_1 = vy_tuple(iy + I1).template AsType<int8x4_t>()(ix / I4);
auto& y_s4_2 = vy_tuple(iy + I2).template AsType<int8x4_t>()(ix / I4);
auto& y_s4_3 = vy_tuple(iy + I3).template AsType<int8x4_t>()(ix / I4);
// transpose
transpose_int8_4x4(x_s4_0, x_s4_1, x_s4_2, x_s4_3, y_s4_0, y_s4_1, y_s4_2, y_s4_3);
});
});
}
};
} // namespace ck } // namespace ck
#endif #endif
library/CMakeLists.txt
View file @ d92fb7e8
add_subdirectory(src/host_tensor) add_subdirectory(src/host_tensor)
add_subdirectory(src/tensor_operation_instance/gpu) add_subdirectory(src/tensor_operation_instance/gpu)
add_subdirectory(src/utility)
library/include/ck/library/utility/fill.hpp 0 → 100644
View file @ d92fb7e8
#pragma once
#include <algorithm>
#include <random>
#include "data_type.hpp"
namespace ck {
namespace utils {
// template <typename T, class Enable = void>
// struct FillUniform;
// TODO: what's wrong with this specialization???
// err: segmentation fault in mt19937 - infinite loop like.
// template <typename T>
// struct FillUniform<T, typename std::enable_if<std::is_integral<T>::value &&
// !std::is_same<T, bhalf_t>::value>::type>
// {
// int a_{0};
// int b_{5};
// // T a_ = T{0};
// // T b_ = T{5};
// template <typename ForwardIter>
// void operator()(ForwardIter first, ForwardIter last) const
// {
// std::mt19937 gen{11939};
// std::uniform_int_distribution<int> dis(a_, b_);
// std::generate(first, last, [&dis, &gen]() { return ck::type_convert<T>(dis(gen)); });
// }
// };
// struct FillUniform<T, typename std::enable_if<std::is_floating_point<T>::value ||
// std::is_same<T, bhalf_t>::value>::type>
template <typename T>
struct FillUniform
{
float a_{0};
float b_{5};
template <typename ForwardIter>
void operator()(ForwardIter first, ForwardIter last) const
{
std::mt19937 gen{11939};
std::uniform_real_distribution<> dis(a_, b_);
std::generate(first, last, [&dis, &gen]() { return ck::type_convert<T>(dis(gen)); });
}
};
template <typename T>
struct FillMonotonicSeq
{
T init_value_{0};
T step_{1};
template <typename ForwardIter>
void operator()(ForwardIter first, ForwardIter last) const
{
std::generate(first, last, [=, n = init_value_]() mutable {
auto tmp = n;
n += step_;
return tmp;
});
}
};
template <typename T>
struct FillConstant
{
T value_{0};
template <typename ForwardIter>
void operator()(ForwardIter first, ForwardIter last) const
{
std::fill(first, last, value_);
}
};
} // namespace utils
} // namespace ck
library/include/ck/library/utility/op_instance_engine.hpp 0 → 100644
View file @ d92fb7e8
#pragma once
#include <cstdlib>
#include <limits>
#include <memory>
#include <stdexcept>
#include <tuple>
#include <utility>
#include <vector>
#include "check_err.hpp"
#include "device_base.hpp"
#include "functional2.hpp"
namespace ck {
namespace utils {
struct ProfileBestConfig
{
std::string best_op_name;
float best_avg_time = std::numeric_limits<float>::max();
float best_tflops = std::numeric_limits<float>::max();
float best_gb_per_sec = std::numeric_limits<float>::max();
};
/**
* @brief This class describes an operation instance(s).
*
* Op instance defines a particular specializations of operator
* template. Thanks to this specific input/output data types, data
* layouts and modifying elementwise operations it is able to create
* it's input/output tensors, provide pointers to instances which
* can execute it and all operation specific parameters.
*/
template <typename OutDataType, typename... InArgTypes>
class OpInstance
{
public:
template <typename T>
using TensorPtr = std::unique_ptr<Tensor<T>>;
using InTensorsTuple = std::tuple<TensorPtr<InArgTypes>...>;
using DeviceMemPtr = std::unique_ptr<DeviceMem>;
using DeviceBuffers = std::vector<DeviceMemPtr>;
OpInstance() = default;
OpInstance(const OpInstance&) = default;
OpInstance& operator=(const OpInstance&) = default;
virtual ~OpInstance(){};
virtual InTensorsTuple GetInputTensors() const = 0;
virtual TensorPtr<OutDataType> GetOutputTensor() const = 0;
virtual std::unique_ptr<tensor_operation::device::BaseInvoker>
MakeInvokerPointer(tensor_operation::device::BaseOperator*) const = 0;
virtual std::unique_ptr<tensor_operation::device::BaseArgument>
MakeArgumentPointer(tensor_operation::device::BaseOperator*,
const DeviceBuffers&,
const DeviceMemPtr&) const = 0;
virtual std::size_t GetFlops() const = 0;
virtual std::size_t GetBtype() const = 0;
};
/**
* @brief A generic operation instance run engine.
*/
template <typename OutDataType, typename... InArgTypes>
class OpInstanceRunEngine
{
public:
using OpInstanceT = OpInstance<InArgTypes..., OutDataType>;
template <typename T>
using TensorPtr = std::unique_ptr<Tensor<T>>;
using DeviceMemPtr = std::unique_ptr<DeviceMem>;
using InTensorsTuple = std::tuple<TensorPtr<InArgTypes>...>;
using DeviceBuffers = std::vector<DeviceMemPtr>;
using InArgsTypesTuple = std::tuple<InArgTypes...>;
OpInstanceRunEngine() = delete;
template <typename ReferenceOp = std::function<void()>>
OpInstanceRunEngine(const OpInstanceT& op_instance,
const ReferenceOp& reference_op = ReferenceOp{})
: op_instance_{op_instance}
{
in_tensors_ = op_instance_.GetInputTensors();
out_tensor_ = op_instance_.GetOutputTensor();
if constexpr(std::is_invocable_v<ReferenceOp,
const Tensor<InArgTypes>&...,
Tensor<OutDataType>&>)
{
ref_output_ = op_instance_.GetOutputTensor();
CallRefOpUnpackArgs(reference_op, std::make_index_sequence<kNInArgs_>{});
}
AllocateDeviceInputTensors(std::make_index_sequence<kNInArgs_>{});
out_device_buffer_ =
std::make_unique<DeviceMem>(sizeof(OutDataType) * out_tensor_->mDesc.GetElementSpace());
out_device_buffer_->SetZero();
}
virtual ~OpInstanceRunEngine(){};
template <typename OpInstancePtr>
bool Test(const std::vector<OpInstancePtr>& op_ptrs)
{
bool res{true};
for(auto& op_ptr : op_ptrs)
{
auto invoker = op_instance_.MakeInvokerPointer(op_ptr.get());
auto argument = op_instance_.MakeArgumentPointer(
op_ptr.get(), in_device_buffers_, out_device_buffer_);
if(op_ptr->IsSupportedArgument(argument.get()))
{
invoker->Run(argument.get());
out_device_buffer_->FromDevice(out_tensor_->mData.data());
if(!ref_output_)
{
throw std::runtime_error(
"OpInstanceRunEngine::Test: Reference value not availabe."
" You have to provide reference function.");
}
// TODO: enable flexible use of custom check_error functions
res = res && check_err(out_tensor_->mData, ref_output_->mData);
out_device_buffer_->SetZero();
}
}
return res;
}
template <typename OpInstancePtr>
ProfileBestConfig Profile(const std::vector<OpInstancePtr>& op_ptrs,
int nrepeat = 100,
bool do_verification = false,
bool do_log = false)
{
bool res{true};
ProfileBestConfig best_config;
for(auto& op_ptr : op_ptrs)
{
auto invoker = op_instance_.MakeInvokerPointer(op_ptr.get());
auto argument = op_instance_.MakeArgumentPointer(
op_ptr.get(), in_device_buffers_, out_device_buffer_);
if(op_ptr->IsSupportedArgument(argument.get()))
{
std::string op_name = op_ptr->GetTypeString();
float avg_time = invoker->Run(argument.get(), nrepeat);
std::size_t flops = op_instance_.GetFlops();
std::size_t num_btype = op_instance_.GetBtype();
float tflops = static_cast<float>(flops) / 1.E9 / avg_time;
float gb_per_sec = num_btype / 1.E6 / avg_time;
std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s, " << op_name << std::endl;
if(tflops < best_config.best_tflops)
{
best_config.best_op_name = op_name;
best_config.best_tflops = tflops;
best_config.best_gb_per_sec = gb_per_sec;
best_config.best_avg_time = avg_time;
}
if(do_verification)
{
out_device_buffer_->FromDevice(out_tensor_->mData.data());
if(!ref_output_)
{
throw std::runtime_error(
"OpInstanceRunEngine::Profile: Reference value not availabe."
" You have to provide reference function.");
}
// TODO: enable flexible use of custom check_error functions
res = res && CheckErr(out_tensor_->mData, ref_output_->mData);
if(do_log) {}
}
out_device_buffer_->SetZero();
}
}
return best_config;
}
void SetAtol(double a) { atol_ = a; }
void SetRtol(double r) { rtol_ = r; }
private:
template <typename F, std::size_t... Is>
void CallRefOpUnpackArgs(const F& f, std::index_sequence<Is...>) const
{
f(*std::get<Is>(in_tensors_)..., *ref_output_);
}
template <std::size_t... Is>
void AllocateDeviceInputTensors(std::index_sequence<Is...>)
{
(AllocateDeviceInputTensorsImpl<Is>(), ...);
}
template <std::size_t Index>
void AllocateDeviceInputTensorsImpl()
{
const auto& ts = std::get<Index>(in_tensors_);
in_device_buffers_
.emplace_back(
std::make_unique<DeviceMem>(sizeof(std::tuple_element_t<Index, InArgsTypesTuple>) *
ts->mDesc.GetElementSpace()))
->ToDevice(ts->mData.data());
}
static constexpr std::size_t kNInArgs_ = std::tuple_size_v<InTensorsTuple>;
const OpInstanceT& op_instance_;
double rtol_{1e-5};
double atol_{1e-8};
InTensorsTuple in_tensors_;
TensorPtr<OutDataType> out_tensor_;
TensorPtr<OutDataType> ref_output_;
DeviceBuffers in_device_buffers_;
DeviceMemPtr out_device_buffer_;
template <typename T>
bool CheckErr(const std::vector<T>& dev_out, const std::vector<T>& ref_out) const
{
return ck::utils::check_err(dev_out, ref_out, "Error: incorrect results!", atol_, rtol_);
}
};
} // namespace utils
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instance.cpp
View file @ d92fb7e8
...@@ -28,19 +28,19 @@ using device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances = std::tuple< ...@@ -28,19 +28,19 @@ using device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances = std::tuple<
//#####################| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| //#####################| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//#####################| | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| //#####################| | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | //#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 256, 128, 32, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 256, 32, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 256, 16, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 128, 128, 128, 32, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>, DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 128, 128, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 32, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 128, 128, 64, 32, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>, DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 128, 128, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 128, 64, 128, 32, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>, DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 128, 64, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 64, 64, 64, 32, 4, 4, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>, DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 64, 64, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 64, 32, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 64, 16, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 128, 32, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>, DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 128, 128, 32, 32, 4, 4, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>, DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 128, 128, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 128, 32, 128, 32, 4, 4, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>, DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 128, 32, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 64, 64, 32, 32, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>, DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 64, 64, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>,
DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 64, 32, 64, 32, 4, 4, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4> DeviceGemm_Xdl_CShuffle< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 64, 32, 64, 16, 4, 4, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>
// clang-format on // clang-format on
>; >;
......
library/src/utility/CMakeLists.txt 0 → 100644
View file @ d92fb7e8
include_directories(BEFORE
${PROJECT_SOURCE_DIR}/include/ck
${PROJECT_SOURCE_DIR}/include/ck/tensor_operation/gpu/device
${PROJECT_SOURCE_DIR}/include/ck/tensor_operation/gpu/element
${PROJECT_SOURCE_DIR}/include/ck/utility
${PROJECT_SOURCE_DIR}/library/include/ck/library/host_tensor
${PROJECT_SOURCE_DIR}/library/include/ck/library/reference_tensor_operation/cpu
${PROJECT_SOURCE_DIR}/library/include/ck/library/utility
)
set(CONV_FWD_UTIL_SOURCE
conv_fwd_util.cpp
)
add_library(conv_fwd_util SHARED ${CONV_FWD_UTIL_SOURCE})
target_link_libraries(conv_fwd_util PRIVATE host_tensor)
target_compile_features(conv_fwd_util PUBLIC)
set_target_properties(conv_fwd_util PROPERTIES POSITION_INDEPENDENT_CODE ON)
target_include_directories(conv_fwd_util SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
clang_tidy_check(conv_fwd_util)
library/src/utility/conv_fwd_util.cpp 0 → 100644
View file @ d92fb7e8
#include "conv_fwd_util.hpp"
namespace ck {
namespace utils {
namespace conv {
/**
* @brief Calculate number of FLOPs for Convolution
*
* @param[in] N Batch size.
* @param[in] C Number of input channels.
* @param[in] K Number of output channels.
* @param[in] filter_spatial_lengths Filter spatial dimensions lengths.
* @param[in] output_spatial_lengths Convolution output spatial dimensions
* lengths.
*
* @return The number of flops.
*/
std::size_t get_flops(ck::index_t N,
ck::index_t C,
ck::index_t K,
const std::vector<ck::index_t>& filter_spatial_lengths,
const std::vector<ck::index_t>& output_spatial_lengths)
{
// 2 * N * K * <output spatial lengths product> * C * <filter spatial lengths product>
return static_cast<std::size_t>(2) * N * K *
std::accumulate(std::begin(output_spatial_lengths),
std::end(output_spatial_lengths),
static_cast<std::size_t>(1),
std::multiplies<std::size_t>()) *
C *
std::accumulate(std::begin(filter_spatial_lengths),
std::end(filter_spatial_lengths),
static_cast<std::size_t>(1),
std::multiplies<std::size_t>());
}
ConvParams::ConvParams()
: num_dim_spatial(2),
N(128),
K(256),
C(192),
filter_spatial_lengths(2, 3),
input_spatial_lengths(2, 71),
conv_filter_strides(2, 2),
conv_filter_dilations(2, 1),
input_left_pads(2, 1),
input_right_pads(2, 1)
{
}
ConvParams::ConvParams(ck::index_t n_dim,
ck::index_t n_batch,
ck::index_t n_out_channels,
ck::index_t n_in_channels,
const std::vector<ck::index_t>& filters_len,
const std::vector<ck::index_t>& input_len,
const std::vector<ck::index_t>& strides,
const std::vector<ck::index_t>& dilations,
const std::vector<ck::index_t>& left_pads,
const std::vector<ck::index_t>& right_pads)
: num_dim_spatial(n_dim),
N(n_batch),
K(n_out_channels),
C(n_in_channels),
filter_spatial_lengths(filters_len),
input_spatial_lengths(input_len),
conv_filter_strides(strides),
conv_filter_dilations(dilations),
input_left_pads(left_pads),
input_right_pads(right_pads)
{
if(filter_spatial_lengths.size() != num_dim_spatial ||
input_spatial_lengths.size() != num_dim_spatial ||
conv_filter_strides.size() != num_dim_spatial ||
conv_filter_dilations.size() != num_dim_spatial ||
input_left_pads.size() != num_dim_spatial || input_right_pads.size() != num_dim_spatial)
{
throw(
std::runtime_error("ConvParams::GetOutputSpatialLengths: "
"parameter size is different from number of declared dimensions!"));
}
}
std::vector<ck::index_t> ConvParams::GetOutputSpatialLengths() const
{
if(filter_spatial_lengths.size() != num_dim_spatial ||
input_spatial_lengths.size() != num_dim_spatial ||
conv_filter_strides.size() != num_dim_spatial ||
conv_filter_dilations.size() != num_dim_spatial ||
input_left_pads.size() != num_dim_spatial || input_right_pads.size() != num_dim_spatial)
{
throw(
std::runtime_error("ConvParams::GetOutputSpatialLengths: "
"parameter size is different from number of declared dimensions!"));
}
std::vector<ck::index_t> out_spatial_len(num_dim_spatial, 0);
for(ck::index_t i = 0; i < num_dim_spatial; ++i)
{
// XEff = (X - 1) * conv_dilation_w + 1;
// Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
const ck::index_t idx_eff = (filter_spatial_lengths[i] - 1) * conv_filter_dilations[i] + 1;
out_spatial_len[i] =
(input_spatial_lengths[i] + input_left_pads[i] + input_right_pads[i] - idx_eff) /
conv_filter_strides[i] +
1;
}
return out_spatial_len;
}
ConvParams parse_conv_params(int num_dim_spatial, int arg_idx, char* const argv[])
{
ck::utils::conv::ConvParams params;
params.num_dim_spatial = num_dim_spatial;
params.N = std::stoi(argv[arg_idx++]);
params.K = std::stoi(argv[arg_idx++]);
params.C = std::stoi(argv[arg_idx++]);
params.filter_spatial_lengths.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.filter_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
}
params.input_spatial_lengths.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
}
params.conv_filter_strides.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.conv_filter_strides[i] = std::stoi(argv[arg_idx++]);
}
params.conv_filter_dilations.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.conv_filter_dilations[i] = std::stoi(argv[arg_idx++]);
}
params.input_left_pads.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_left_pads[i] = std::stoi(argv[arg_idx++]);
}
params.input_right_pads.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_right_pads[i] = std::stoi(argv[arg_idx++]);
}
return params;
}
HostTensorDescriptor get_output_host_tensor_descriptor(const std::vector<std::size_t>& dims,
int num_dim_spatial)
{
namespace tl = ck::tensor_layout::convolution;
switch(num_dim_spatial)
{
case 3: {
return ck::utils::conv::get_host_tensor_descriptor(dims, tl::NDHWK{});
}
case 2: {
return ck::utils::conv::get_host_tensor_descriptor(dims, tl::NHWK{});
}
case 1: {
return ck::utils::conv::get_host_tensor_descriptor(dims, tl::NWK{});
}
default: {
throw std::runtime_error("Unsupported number of spatial dimensions provided!");
}
}
}
HostTensorDescriptor get_filters_host_tensor_descriptor(const std::vector<std::size_t>& dims,
int num_dim_spatial)
{
namespace tl = ck::tensor_layout::convolution;
switch(num_dim_spatial)
{
case 3: {
return ck::utils::conv::get_host_tensor_descriptor(dims, tl::KZYXC{});
}
case 2: {
return ck::utils::conv::get_host_tensor_descriptor(dims, tl::KYXC{});
}
case 1: {
return ck::utils::conv::get_host_tensor_descriptor(dims, tl::KXC{});
}
default: {
throw std::runtime_error("Unsupported number of spatial dimensions provided!");
}
}
}
HostTensorDescriptor get_input_host_tensor_descriptor(const std::vector<std::size_t>& dims,
int num_dim_spatial)
{
namespace tl = ck::tensor_layout::convolution;
switch(num_dim_spatial)
{
case 3: {
return ck::utils::conv::get_host_tensor_descriptor(dims, tl::NDHWC{});
}
case 2: {
return ck::utils::conv::get_host_tensor_descriptor(dims, tl::NHWC{});
}
case 1: {
return ck::utils::conv::get_host_tensor_descriptor(dims, tl::NWC{});
}
default: {
throw std::runtime_error("Unsupported number of spatial dimensions provided!");
}
}
}
} // namespace conv
} // namespace utils
} // namespace ck
std::ostream& operator<<(std::ostream& os, const ck::utils::conv::ConvParams& p)
{
os << "ConvParams {"
<< "\nnum_dim_spatial: " << p.num_dim_spatial << "\nN: " << p.N << "\nK: " << p.K
<< "\nC: " << p.C << "\nfilter_spatial_lengths: " << p.filter_spatial_lengths
<< "\ninput_spatial_lengths: " << p.input_spatial_lengths
<< "\nconv_filter_strides: " << p.conv_filter_strides
<< "\nconv_filter_dilations: " << p.conv_filter_dilations
<< "\ninput_left_pads: " << p.input_left_pads
<< "\ninput_right_pads: " << p.input_right_pads;
return os;
}
profiler/CMakeLists.txt
View file @ d92fb7e8
...@@ -29,10 +29,10 @@ set(PROFILER_SOURCE ...@@ -29,10 +29,10 @@ set(PROFILER_SOURCE
src/profile_gemm_bias_relu_add.cpp src/profile_gemm_bias_relu_add.cpp
src/profile_gemm_reduce.cpp src/profile_gemm_reduce.cpp
src/profile_batched_gemm.cpp src/profile_batched_gemm.cpp
src/profile_conv_fwd.cpp
src/profile_conv_fwd_bias_relu.cpp src/profile_conv_fwd_bias_relu.cpp
src/profile_conv_fwd_bias_relu_add.cpp src/profile_conv_fwd_bias_relu_add.cpp
src/profile_conv_fwd_bias_relu_atomic_add.cpp src/profile_conv_fwd_bias_relu_atomic_add.cpp
src/profile_convnd_fwd.cpp
src/profile_convnd_bwd_data.cpp src/profile_convnd_bwd_data.cpp
src/profile_reduce.cpp src/profile_reduce.cpp
src/profile_grouped_gemm.cpp src/profile_grouped_gemm.cpp
...@@ -43,19 +43,21 @@ set(PROFILER_SOURCE ...@@ -43,19 +43,21 @@ set(PROFILER_SOURCE
add_executable(ckProfiler ${PROFILER_SOURCE}) add_executable(ckProfiler ${PROFILER_SOURCE})
target_link_libraries(ckProfiler PRIVATE host_tensor) target_link_libraries(ckProfiler PRIVATE host_tensor)
target_link_libraries(ckProfiler PRIVATE conv_fwd_util)
target_link_libraries(ckProfiler PRIVATE device_gemm_reduce_instance) target_link_libraries(ckProfiler PRIVATE device_gemm_reduce_instance)
target_link_libraries(ckProfiler PRIVATE device_gemm_instance) target_link_libraries(ckProfiler PRIVATE device_gemm_instance)
target_link_libraries(ckProfiler PRIVATE device_gemm_bias2d_instance) target_link_libraries(ckProfiler PRIVATE device_gemm_bias2d_instance)
target_link_libraries(ckProfiler PRIVATE device_gemm_bias_relu_instance) target_link_libraries(ckProfiler PRIVATE device_gemm_bias_relu_instance)
target_link_libraries(ckProfiler PRIVATE device_gemm_bias_relu_add_instance) target_link_libraries(ckProfiler PRIVATE device_gemm_bias_relu_add_instance)
target_link_libraries(ckProfiler PRIVATE device_batched_gemm_instance) target_link_libraries(ckProfiler PRIVATE device_batched_gemm_instance)
target_link_libraries(ckProfiler PRIVATE device_conv1d_fwd_instance)
target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_instance) target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_instance)
target_link_libraries(ckProfiler PRIVATE device_conv3d_fwd_instance)
target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_bias_relu_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_add_instance)
target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_bias_relu_atomic_add_instance) target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_bias_relu_atomic_add_instance)
target_link_libraries(ckProfiler PRIVATE device_convnd_bwd_data_instance) target_link_libraries(ckProfiler PRIVATE device_convnd_bwd_data_instance)
target_link_libraries(ckProfiler PRIVATE device_reduce_instance) target_link_libraries(ckProfiler PRIVATE device_reduce_instance)
target_link_libraries(ckProfiler PRIVATE device_reduce_instance)
target_link_libraries(ckProfiler PRIVATE device_grouped_gemm_instance) target_link_libraries(ckProfiler PRIVATE device_grouped_gemm_instance)
target_link_libraries(ckProfiler PRIVATE device_conv2d_bwd_weight_instance) target_link_libraries(ckProfiler PRIVATE device_conv2d_bwd_weight_instance)
target_link_libraries(ckProfiler PRIVATE device_batched_gemm_reduce_instance) target_link_libraries(ckProfiler PRIVATE device_batched_gemm_reduce_instance)
profiler/include/profile_batched_gemm_reduce_impl.hpp
View file @ d92fb7e8
...@@ -8,7 +8,7 @@ ...@@ -8,7 +8,7 @@
#include "tensor_layout.hpp" #include "tensor_layout.hpp"
#include "device_tensor.hpp" #include "device_tensor.hpp"
#include "element_wise_operation.hpp" #include "element_wise_operation.hpp"
#include "element_wise_reduce_operation.hpp" #include "reduction_operator.hpp"
#include "device_gemm_reduce.hpp" #include "device_gemm_reduce.hpp"
#include "reference_batched_gemm.hpp" #include "reference_batched_gemm.hpp"
...@@ -21,8 +21,7 @@ using DeviceGemmReduceNoOpPtr = ck::tensor_operation::device::DeviceGemmReducePt ...@@ -21,8 +21,7 @@ using DeviceGemmReduceNoOpPtr = ck::tensor_operation::device::DeviceGemmReducePt
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::ReduceSum, ck::tensor_operation::element_wise::UnarySquare<float, float, false>>;
ck::tensor_operation::element_wise::ReduceSquareSum>;
void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instances( void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instances(
std::vector<DeviceGemmReduceNoOpPtr>&); std::vector<DeviceGemmReduceNoOpPtr>&);
...@@ -120,17 +119,19 @@ bool profile_batched_gemm_reduce_impl(int do_verification, ...@@ -120,17 +119,19 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
b_g_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread); b_g_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
} }
using AElementOp = ck::tensor_operation::element_wise::PassThrough; using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough; using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough; using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using D0ReduceOp = ck::tensor_operation::element_wise::ReduceSum; using D0ReduceOp = ck::reduce::Add<float>;
using D1ReduceOp = ck::tensor_operation::element_wise::ReduceSquareSum; using D1ReduceOp = ck::reduce::Add<float>;
using D1ElementOp = ck::tensor_operation::element_wise::UnarySquare<float, float, false>;
const auto a_element_op = AElementOp{}; const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{}; const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{}; const auto c_element_op = CElementOp{};
const auto d0_reduce_op = D0ReduceOp{}; const auto d0_reduce_op = D0ReduceOp{};
const auto d1_reduce_op = D1ReduceOp{}; const auto d1_reduce_op = D1ReduceOp{};
const auto d1_element_op = D1ElementOp{};
if(do_verification) if(do_verification)
{ {
...@@ -154,17 +155,21 @@ bool profile_batched_gemm_reduce_impl(int do_verification, ...@@ -154,17 +155,21 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
{ {
for(int m = 0; m < M; ++m) for(int m = 0; m < M; ++m)
{ {
float d0_acc = d0_reduce_op.GetReduceZeroValue(); float d0_acc = d0_reduce_op.GetReductionZeroVal();
float d1_acc = d1_reduce_op.GetReduceZeroValue(); float d1_acc = d1_reduce_op.GetReductionZeroVal();
for(int n = 0; n < N; ++n) for(int n = 0; n < N; ++n)
{ {
d0_reduce_op.Reduce(d0_acc, c_g_m_n_host_result(batch, m, n)); float d0_val = ck::type_convert<float>(c_g_m_n_host_result(batch, m, n));
d1_reduce_op.Reduce(d1_acc, c_g_m_n_host_result(batch, m, n)); float d1_val;
d1_element_op(d1_val, d0_val);
d0_reduce_op(d0_acc, d0_val);
d1_reduce_op(d1_acc, d1_val);
} }
d0_g_m_host_result(batch, m) = d0_acc; d0_g_m_host_result(batch, m) = ck::type_convert<DDataType>(d0_acc);
d1_g_m_host_result(batch, m) = d1_acc; d1_g_m_host_result(batch, m) = ck::type_convert<DDataType>(d1_acc);
} }
} }
} }
...@@ -247,8 +252,7 @@ bool profile_batched_gemm_reduce_impl(int do_verification, ...@@ -247,8 +252,7 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
a_element_op, a_element_op,
b_element_op, b_element_op,
c_element_op, c_element_op,
d0_reduce_op, d1_element_op,
d1_reduce_op,
BatchCount); BatchCount);
auto invoker_ptr = gemm_ptr->MakeInvokerPointer(); auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
......
profiler/include/profile_conv_fwd_impl.hpp deleted 100644 → 0
View file @ bfc80764
#pragma once
#include "check_err.hpp"
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "tensor_layout.hpp"
#include "device_tensor.hpp"
#include "device_conv_fwd.hpp"
#include "element_wise_operation.hpp"
#include "reference_conv_fwd.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_conv2d_fwd_instance {
using DeviceConvFwdNoOpPtr = DeviceConvFwdPtr<ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>;
void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances(std::vector<DeviceConvFwdNoOpPtr>&);
void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instances(std::vector<DeviceConvFwdNoOpPtr>&);
void add_device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances(
std::vector<DeviceConvFwdNoOpPtr>&);
void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances(std::vector<DeviceConvFwdNoOpPtr>&);
void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances(std::vector<DeviceConvFwdNoOpPtr>&);
} // namespace device_conv2d_fwd_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
namespace ck {
namespace profiler {
template <int NDimSpatial,
typename InDataType,
typename WeiDataType,
typename OutDataType,
typename InLayout,
typename WeiLayout,
typename OutLayout>
void profile_conv_fwd_impl(int do_verification,
int init_method,
bool do_log,
int nrepeat,
ck::index_t N,
ck::index_t K,
ck::index_t C,
std::vector<ck::index_t> input_spatial_lengths,
std::vector<ck::index_t> filter_spatial_lengths,
std::vector<ck::index_t> output_spatial_lengths,
std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads)
{
const ck::index_t Y = filter_spatial_lengths[0];
const ck::index_t X = filter_spatial_lengths[1];
const ck::index_t Hi = input_spatial_lengths[0];
const ck::index_t Wi = input_spatial_lengths[1];
const ck::index_t Ho = output_spatial_lengths[0];
const ck::index_t Wo = output_spatial_lengths[1];
auto f_host_tensor_descriptor =
[](std::size_t N_, std::size_t C_, std::size_t H, std::size_t W, auto layout) {
if constexpr(is_same<decltype(layout), ck::tensor_layout::convolution::NCHW>::value ||
is_same<decltype(layout), ck::tensor_layout::convolution::KCYX>::value ||
is_same<decltype(layout), ck::tensor_layout::convolution::NKHW>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
std::vector<std::size_t>({C_ * H * W, H * W, W, 1}));
}
else if constexpr(is_same<decltype(layout), tensor_layout::convolution::NHWC>::value ||
is_same<decltype(layout), tensor_layout::convolution::KYXC>::value ||
is_same<decltype(layout), tensor_layout::convolution::NHWK>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
std::vector<std::size_t>({C_ * H * W, 1, W * C_, C_}));
}
};
Tensor<InDataType> in_n_c_hi_wi(f_host_tensor_descriptor(N, C, Hi, Wi, InLayout{}));
Tensor<WeiDataType> wei_k_c_y_x(f_host_tensor_descriptor(K, C, Y, X, WeiLayout{}));
Tensor<OutDataType> out_n_k_ho_wo_host_result(
f_host_tensor_descriptor(N, K, Ho, Wo, OutLayout{}));
Tensor<OutDataType> out_n_k_ho_wo_device_result(
f_host_tensor_descriptor(N, K, Ho, Wo, OutLayout{}));
std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.mDesc << std::endl;
std::cout << "wei_k_c_y_x: " << wei_k_c_y_x.mDesc << std::endl;
std::cout << "out_n_k_ho_wo: " << out_n_k_ho_wo_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
break;
default:
in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
}
using InElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{};
const auto out_element_op = OutElementOp{};
if(do_verification)
{
using ReferenceConvFwdInstance = ck::tensor_operation::host::ReferenceConvFwd<InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>;
auto ref_conv = ReferenceConvFwdInstance{};
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi,
wei_k_c_y_x,
out_n_k_ho_wo_host_result,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
in_element_op,
wei_element_op,
out_element_op);
ref_invoker.Run(ref_argument);
}
DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpace());
DeviceMem wei_device_buf(sizeof(WeiDataType) * wei_k_c_y_x.mDesc.GetElementSpace());
DeviceMem out_device_buf(sizeof(OutDataType) *
out_n_k_ho_wo_device_result.mDesc.GetElementSpace());
in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
wei_device_buf.ToDevice(wei_k_c_y_x.mData.data());
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DeviceConvFwdNoOpPtr =
ck::tensor_operation::device::DeviceConvFwdPtr<PassThrough, PassThrough, PassThrough>;
// add device Conv instances
std::vector<DeviceConvFwdNoOpPtr> conv_ptrs;
if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, float> &&
ck::is_same_v<ck::remove_cv_t<WeiDataType>, float> &&
ck::is_same_v<ck::remove_cv_t<OutDataType>, float>)
{
ck::tensor_operation::device::device_conv2d_fwd_instance::
add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances(conv_ptrs);
}
else if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, ck::half_t> &&
ck::is_same_v<ck::remove_cv_t<WeiDataType>, ck::half_t> &&
ck::is_same_v<ck::remove_cv_t<OutDataType>, ck::half_t>)
{
ck::tensor_operation::device::device_conv2d_fwd_instance::
add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instances(conv_ptrs);
ck::tensor_operation::device::device_conv2d_fwd_instance::
add_device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances(conv_ptrs);
}
else if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, bhalf_t> &&
ck::is_same_v<ck::remove_cv_t<WeiDataType>, bhalf_t> &&
ck::is_same_v<ck::remove_cv_t<OutDataType>, bhalf_t>)
{
ck::tensor_operation::device::device_conv2d_fwd_instance::
add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances(conv_ptrs);
}
else if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, int8_t> &&
ck::is_same_v<ck::remove_cv_t<WeiDataType>, int8_t> &&
ck::is_same_v<ck::remove_cv_t<OutDataType>, int8_t>)
{
ck::tensor_operation::device::device_conv2d_fwd_instance::
add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances(conv_ptrs);
}
if(conv_ptrs.size() <= 0)
{
throw std::runtime_error("wrong! no device Conv instance found");
}
std::string best_conv_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device Conv instances
for(auto& conv_ptr : conv_ptrs)
{
auto argument_ptr = conv_ptr->MakeArgumentPointer(
static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
N,
K,
C,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
in_element_op,
wei_element_op,
out_element_op);
auto invoker_ptr = conv_ptr->MakeInvokerPointer();
if(conv_ptr->IsSupportedArgument(argument_ptr.get()))
{
std::string conv_name = conv_ptr->GetTypeString();
float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X;
std::size_t num_btype = sizeof(InDataType) * (N * C * Hi * Wi) +
sizeof(WeiDataType) * (K * C * Y * X) +
sizeof(OutDataType) * (N * K * Ho * Wo);
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, " << conv_name << std::endl;
if(tflops > best_tflops)
{
best_conv_name = conv_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
out_device_buf.FromDevice(out_n_k_ho_wo_device_result.mData.data());
ck::utils::check_err(out_n_k_ho_wo_device_result.mData,
out_n_k_ho_wo_host_result.mData);
if(do_log)
{
LogRangeAsType<float>(std::cout << "in : ", in_n_c_hi_wi.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "wei: ", wei_k_c_y_x.mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "out_host : ", out_n_k_ho_wo_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "out_device: ", out_n_k_ho_wo_device_result.mData, ",")
<< std::endl;
}
}
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_conv_name << std::endl;
}
} // namespace profiler
} // namespace ck
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