mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-04-19 22:39:03 +00:00
Add a scale op, related instances and examples (#1242)
* Add a scale op * Update the element op * Add instances * Add an example * Add a client example * Add a flag check * Revert flag check addition * Fix flag check * Update d strides in example * Update d strides in client example * Apply suggestions from code review Update copyright header Co-authored-by: Bartłomiej Kocot <barkocot@amd.com> * Move the example * Move the client example * Update element op * Update example with the new element op * Add scalar layout * Update example * Update kernel for scalar Ds * Revert kernel changes * Update element op * Update example to use scales' pointers * Format * Update instances * Update client example * Move element op to unary elements * Update element op to work with values instead of pointers * Update instances to take element op as an argument * Update examples to use random scale values --------- Co-authored-by: Bartłomiej Kocot <barkocot@amd.com>
This commit is contained in:
Rostyslav Geyyer
@@ -35,6 +35,10 @@ target_link_libraries(client_grouped_convnd_fwd_scaleadd_ab_int8 PRIVATE composa
|
||||
add_executable(client_grouped_convnd_fwd_bilinear_residual_fp16
|
||||
grouped_convnd_fwd_bilinear/grouped_conv_fwd_bilinear_residual_fp16.cpp)
|
||||
target_link_libraries(client_grouped_convnd_fwd_bilinear_residual_fp16 PRIVATE composable_kernel::device_conv_operations)
|
||||
# Fwd convscale
|
||||
add_executable(client_conv3d_fwd_convscale_fp8
|
||||
grouped_convnd_fwd_convscale/conv3d_fwd_convscale_fp8.cpp)
|
||||
target_link_libraries(client_conv3d_fwd_convscale_fp8 PRIVATE composable_kernel::device_conv_operations)
|
||||
# Bwd data bilinear
|
||||
add_executable(client_grouped_convnd_bwd_data_bilinear_residual_fp16
|
||||
grouped_convnd_bwd_data_bilinear/grouped_conv_bwd_data_bilinear_residual_fp16.cpp)
|
||||
|
||||
@@ -0,0 +1,316 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include <cstdlib>
|
||||
#include <iomanip>
|
||||
#include <iostream>
|
||||
#include <iterator>
|
||||
#include <numeric>
|
||||
#include <string>
|
||||
#include <vector>
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_convscale.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_abd.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
using ConvScale = ck::tensor_operation::element_wise::ConvScale;
|
||||
|
||||
struct SimpleDeviceMem
|
||||
{
|
||||
SimpleDeviceMem() = delete;
|
||||
|
||||
SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
|
||||
{
|
||||
(void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
|
||||
}
|
||||
|
||||
void* GetDeviceBuffer() { return p_mem_; }
|
||||
|
||||
~SimpleDeviceMem() { (void)hipFree(p_mem_); }
|
||||
|
||||
void* p_mem_;
|
||||
};
|
||||
|
||||
template <ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
|
||||
std::size_t
|
||||
GetFlops(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& output_lengths,
|
||||
const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& weights_lengths,
|
||||
const std::size_t& ds_size)
|
||||
{
|
||||
// G * N * C * <output spatial lengths product> * (2 * K * <filter spatial lengths product> +
|
||||
// <number of scale factors>)
|
||||
ck::index_t G = weights_lengths[0];
|
||||
ck::index_t N = output_lengths[1];
|
||||
ck::index_t K = weights_lengths[1];
|
||||
ck::index_t C = weights_lengths[2];
|
||||
|
||||
return G * N * C *
|
||||
std::accumulate(std::next(std::begin(output_lengths), NumNonSpatialDim),
|
||||
std::end(output_lengths),
|
||||
static_cast<std::size_t>(1),
|
||||
std::multiplies<>()) *
|
||||
(static_cast<std::size_t>(2) * K *
|
||||
std::accumulate(std::next(std::begin(weights_lengths), NumNonSpatialDim),
|
||||
std::end(weights_lengths),
|
||||
static_cast<std::size_t>(1),
|
||||
std::multiplies<>()) +
|
||||
ds_size);
|
||||
}
|
||||
|
||||
template <typename InDataType, ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
|
||||
std::size_t
|
||||
GetInputByte(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& input_lengths)
|
||||
{
|
||||
// sizeof(InDataType) * (G * N * C * <input spatial lengths product>) +
|
||||
return sizeof(InDataType) * std::accumulate(std::begin(input_lengths),
|
||||
std::end(input_lengths),
|
||||
static_cast<std::size_t>(1),
|
||||
std::multiplies<>());
|
||||
}
|
||||
|
||||
template <typename WeiDataType, ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
|
||||
std::size_t
|
||||
GetWeightByte(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& weights_lengths)
|
||||
{
|
||||
// sizeof(WeiDataType) * (G * K * C * <filter spatial lengths product>) +
|
||||
return sizeof(WeiDataType) * std::accumulate(std::begin(weights_lengths),
|
||||
std::end(weights_lengths),
|
||||
static_cast<std::size_t>(1),
|
||||
std::multiplies<>());
|
||||
}
|
||||
|
||||
template <typename OutDataType, ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
|
||||
std::size_t
|
||||
GetOutputByte(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& output_lengths)
|
||||
{
|
||||
// sizeof(OutDataType) * (G * N * K * <output spatial lengths product>);
|
||||
return sizeof(OutDataType) * std::accumulate(std::begin(output_lengths),
|
||||
std::end(output_lengths),
|
||||
static_cast<std::size_t>(1),
|
||||
std::multiplies<std::size_t>());
|
||||
}
|
||||
|
||||
template <ck::index_t NumDimSpatial,
|
||||
typename InDataType,
|
||||
typename WeiDataType,
|
||||
typename OutDataType,
|
||||
typename InLayout,
|
||||
typename WeiLayout,
|
||||
typename OutLayout,
|
||||
ck::index_t NumNonSpatialDim = 3,
|
||||
typename AComputeType = InDataType,
|
||||
typename BComputeType = AComputeType>
|
||||
bool run_grouped_conv_fwd_convscale(
|
||||
std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> in_lengths,
|
||||
std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> wei_lengths,
|
||||
std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> out_lengths)
|
||||
{
|
||||
std::size_t in_mem_size = GetInputByte<InDataType, NumDimSpatial>(in_lengths);
|
||||
std::size_t wei_mem_size = GetWeightByte<WeiDataType, NumDimSpatial>(wei_lengths);
|
||||
std::size_t out_mem_size = GetOutputByte<OutDataType, NumDimSpatial>(out_lengths);
|
||||
|
||||
SimpleDeviceMem in(in_mem_size);
|
||||
SimpleDeviceMem wei(wei_mem_size);
|
||||
SimpleDeviceMem out(out_mem_size);
|
||||
|
||||
float scale_in;
|
||||
float scale_wei;
|
||||
float scale_out;
|
||||
|
||||
std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> in_strides;
|
||||
std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> wei_strides;
|
||||
std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> out_strides;
|
||||
in_strides.fill(0);
|
||||
wei_strides.fill(0);
|
||||
out_strides.fill(0);
|
||||
in_strides.back() = 1;
|
||||
wei_strides.back() = 1;
|
||||
out_strides.back() = 1;
|
||||
|
||||
std::partial_sum(rbegin(in_lengths),
|
||||
std::prev(rend(in_lengths)),
|
||||
std::next(rbegin(in_strides)),
|
||||
std::multiplies<>{});
|
||||
std::partial_sum(rbegin(wei_lengths),
|
||||
std::prev(rend(wei_lengths)),
|
||||
std::next(rbegin(wei_strides)),
|
||||
std::multiplies<>{});
|
||||
std::partial_sum(rbegin(out_lengths),
|
||||
std::prev(rend(out_lengths)),
|
||||
std::next(rbegin(out_strides)),
|
||||
std::multiplies<>{});
|
||||
|
||||
// transpose NDHWGC/KZYXGC/NDHWGK to GNDHWC/GKZYXC/GNDHWK to GNCDHW/GKCZYX/GNKDHW
|
||||
std::rotate(std::next(rbegin(in_lengths)), std::next(rbegin(in_lengths), 2), rend(in_lengths));
|
||||
std::rotate(rbegin(in_lengths),
|
||||
std::next(rbegin(in_lengths)),
|
||||
std::next(rbegin(in_lengths), NumDimSpatial + 1));
|
||||
|
||||
std::rotate(std::next(rbegin(in_strides)), std::next(rbegin(in_strides), 2), rend(in_strides));
|
||||
std::rotate(rbegin(in_strides),
|
||||
std::next(rbegin(in_strides)),
|
||||
std::next(rbegin(in_strides), NumDimSpatial + 1));
|
||||
|
||||
std::rotate(rbegin(wei_lengths),
|
||||
std::next(rbegin(wei_lengths)),
|
||||
std::next(rbegin(wei_lengths), NumDimSpatial + 1));
|
||||
|
||||
std::rotate(rbegin(wei_strides),
|
||||
std::next(rbegin(wei_strides)),
|
||||
std::next(rbegin(wei_strides), NumDimSpatial + 1));
|
||||
|
||||
std::rotate(
|
||||
std::next(rbegin(out_lengths)), std::next(rbegin(out_lengths), 2), rend(out_lengths));
|
||||
std::rotate(rbegin(out_lengths),
|
||||
std::next(rbegin(out_lengths)),
|
||||
std::next(rbegin(out_lengths), NumDimSpatial + 1));
|
||||
|
||||
std::rotate(
|
||||
std::next(rbegin(out_strides)), std::next(rbegin(out_strides), 2), rend(out_strides));
|
||||
std::rotate(rbegin(out_strides),
|
||||
std::next(rbegin(out_strides)),
|
||||
std::next(rbegin(out_strides), NumDimSpatial + 1));
|
||||
|
||||
std::array<ck::index_t, NumDimSpatial> conv_filter_strides;
|
||||
std::array<ck::index_t, NumDimSpatial> conv_filter_dilations;
|
||||
std::array<ck::index_t, NumDimSpatial> input_left_pads;
|
||||
std::array<ck::index_t, NumDimSpatial> input_right_pads;
|
||||
conv_filter_strides.fill(1);
|
||||
conv_filter_dilations.fill(1);
|
||||
input_left_pads.fill(1);
|
||||
input_right_pads.fill(1);
|
||||
|
||||
std::size_t ds_size = 3; // 3 element-wise scale multipliers
|
||||
std::size_t flop = GetFlops<NumDimSpatial>(out_lengths, wei_lengths, ds_size);
|
||||
std::size_t num_bytes =
|
||||
in_mem_size + wei_mem_size + sizeof(float) + sizeof(float) + sizeof(float) + out_mem_size;
|
||||
|
||||
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
|
||||
InLayout,
|
||||
WeiLayout,
|
||||
ck::Tuple<>,
|
||||
OutLayout,
|
||||
InDataType,
|
||||
WeiDataType,
|
||||
ck::Tuple<>,
|
||||
OutDataType,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
ConvScale,
|
||||
AComputeType,
|
||||
BComputeType>;
|
||||
// get device op instances
|
||||
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
|
||||
DeviceOp>::GetInstances();
|
||||
|
||||
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
|
||||
|
||||
std::string best_op_name;
|
||||
int best_op_id = -1;
|
||||
float best_avg_time = std::numeric_limits<float>::max();
|
||||
float best_gb_per_sec = 0;
|
||||
float best_tflops = 0;
|
||||
|
||||
// profile device operation instances
|
||||
std::cout << "Run all instances and do timing" << std::endl;
|
||||
|
||||
for(int i = 0; i < op_ptrs.size(); ++i)
|
||||
{
|
||||
auto& op_ptr = op_ptrs[i];
|
||||
auto argument_ptr = op_ptr->MakeArgumentPointer(
|
||||
in.GetDeviceBuffer(),
|
||||
wei.GetDeviceBuffer(),
|
||||
std::array<const void*, 0>{},
|
||||
out.GetDeviceBuffer(),
|
||||
in_lengths,
|
||||
in_strides,
|
||||
wei_lengths,
|
||||
wei_strides,
|
||||
std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 0>{},
|
||||
std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 0>{},
|
||||
out_lengths,
|
||||
out_strides,
|
||||
conv_filter_strides,
|
||||
conv_filter_dilations,
|
||||
input_left_pads,
|
||||
input_right_pads,
|
||||
PassThrough{},
|
||||
PassThrough{},
|
||||
ConvScale{scale_in, scale_wei, scale_out});
|
||||
|
||||
auto invoker_ptr = op_ptr->MakeInvokerPointer();
|
||||
std::string op_name = op_ptr->GetTypeString();
|
||||
|
||||
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
|
||||
{
|
||||
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
|
||||
|
||||
float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
|
||||
float gb_per_sec = num_bytes / 1.E6 / avg_time;
|
||||
|
||||
std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
|
||||
<< gb_per_sec << " GB/s, " << op_name << std::endl;
|
||||
|
||||
if(tflops > best_tflops)
|
||||
{
|
||||
best_op_id = i;
|
||||
best_op_name = op_name;
|
||||
best_avg_time = avg_time;
|
||||
best_gb_per_sec = gb_per_sec;
|
||||
best_tflops = tflops;
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
std::cerr << op_name << " does not support this problem" << std::endl;
|
||||
}
|
||||
}
|
||||
|
||||
if(best_op_id < 0)
|
||||
{
|
||||
std::cerr << "no suitable instance" << std::endl;
|
||||
return false;
|
||||
}
|
||||
|
||||
std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
|
||||
<< " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
|
||||
|
||||
// run the best intance
|
||||
{
|
||||
auto& op_ptr = op_ptrs[best_op_id];
|
||||
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
|
||||
<< std::endl;
|
||||
auto argument_ptr = op_ptr->MakeArgumentPointer(
|
||||
in.GetDeviceBuffer(),
|
||||
wei.GetDeviceBuffer(),
|
||||
std::array<const void*, 0>{},
|
||||
out.GetDeviceBuffer(),
|
||||
in_lengths,
|
||||
in_strides,
|
||||
wei_lengths,
|
||||
wei_strides,
|
||||
std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 0>{},
|
||||
std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 0>{},
|
||||
out_lengths,
|
||||
out_strides,
|
||||
conv_filter_strides,
|
||||
conv_filter_dilations,
|
||||
input_left_pads,
|
||||
input_right_pads,
|
||||
PassThrough{},
|
||||
PassThrough{},
|
||||
ConvScale{scale_in, scale_wei, scale_out});
|
||||
|
||||
auto invoker_ptr = op_ptr->MakeInvokerPointer();
|
||||
|
||||
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
|
||||
{
|
||||
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
|
||||
}
|
||||
|
||||
std::cout << "Done" << std::endl;
|
||||
}
|
||||
return true;
|
||||
}
|
||||
@@ -0,0 +1,50 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include "common.hpp"
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
|
||||
using InDataType = ck::f8_t;
|
||||
using WeiDataType = ck::f8_t;
|
||||
using CShuffleDataType = float;
|
||||
using OutDataType = ck::f8_t;
|
||||
using AComputeDataType = ck::f8_t;
|
||||
using BComputeDataType = ck::f8_t;
|
||||
|
||||
using InLayout = ck::tensor_layout::convolution::NDHWGC;
|
||||
using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
|
||||
using OutLayout = ck::tensor_layout::convolution::NDHWGK;
|
||||
|
||||
static constexpr ck::index_t NumDimSpatial = 3;
|
||||
static constexpr ck::index_t G = 1;
|
||||
static constexpr ck::index_t N = 64;
|
||||
static constexpr ck::index_t K = 128;
|
||||
static constexpr ck::index_t C = 64;
|
||||
static constexpr ck::index_t Z = 3;
|
||||
static constexpr ck::index_t Y = 3;
|
||||
static constexpr ck::index_t X = 3;
|
||||
static constexpr ck::index_t Di = 28;
|
||||
static constexpr ck::index_t Hi = 28;
|
||||
static constexpr ck::index_t Wi = 3;
|
||||
static constexpr ck::index_t Do = 28;
|
||||
static constexpr ck::index_t Ho = 28;
|
||||
static constexpr ck::index_t Wo = 3;
|
||||
|
||||
int main()
|
||||
{
|
||||
return run_grouped_conv_fwd_convscale<NumDimSpatial,
|
||||
InDataType,
|
||||
WeiDataType,
|
||||
OutDataType,
|
||||
InLayout,
|
||||
WeiLayout,
|
||||
OutLayout,
|
||||
3,
|
||||
AComputeDataType,
|
||||
BComputeDataType>(
|
||||
{N, Di, Hi, Wi, G, C}, {G, K, Z, Y, X, C}, {N, Do, Ho, Wo, G, K})
|
||||
? EXIT_SUCCESS
|
||||
: EXIT_FAILURE;
|
||||
}
|
||||
Reference in New Issue
Block a user