youngkingdom/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
b234a8c024580194712d211124b09c663de678c1
cutlass/python/cutlass_cppgen/emit/pytorch.py
Jack Kosaian b234a8c024 Rename python/cutlass to python/cutlass_cppgen (#2652)
2025-09-18 14:26:57 -04:00

937 lines
37 KiB
Python
Raw Blame History

#################################################################################################
#
# Copyright (c) 2023 - 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: BSD-3-Clause
#
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions are met:
#
# 1. Redistributions of source code must retain the above copyright notice, this
# list of conditions and the following disclaimer.
#
# 2. Redistributions in binary form must reproduce the above copyright notice,
# this list of conditions and the following disclaimer in the documentation
# and/or other materials provided with the distribution.
#
# 3. Neither the name of the copyright holder nor the names of its
# contributors may be used to endorse or promote products derived from
# this software without specific prior written permission.
#
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#
#################################################################################################
"""
Utilities for generating source for building a PyTorch CUDA extension that using a CUTLASS kernel.
If specified, the extension can be JIT compiled via PyTorch's ``cpp_extension.load`` method.
Example usage with JIT compilation:
.. highlight:: python
.. code-block:: python
plan = cutlass_cppgen.op.Gemm(element=torch.float32, layout=cutlass_library.LayoutType.RowMajor)
op = plan.construct()
mod = cutlass_cppgen.emit.pytorch(op, 'cutlass_gemm', 80, jit=True)
# Generate inputs for the GEMM
A, B, C = [torch.ones((512, 512)).to('cuda') for _ in range(3)]
# Run the module
D = mod.run(A, B, C)
Example usage without JIT compilation:
.. highlight:: python
.. code-block:: python
plan = cutlass_cppgen.op.Gemm(element=torch.float32, layout=cutlass_cppgen.LayoutType.RowMajor)
op = plan.construct()
cutlass_cppgen.emit.pytorch(op, 'cutlass_gemm', 80, jit=False, sourcedir='output')
After this call, the directory ``output`` contains ``setup.py``,
``cutlass_gemm.cpp``, and ``cutlass_gemm_kernel.cu``. The module can be built from
within ``output`` by running: ``TORCH_CUDA_ARCH_LIST="8.0" python setup.py develop --user``.
The module can later be used in Python via:
.. highlight:: python
.. code-block:: python
import torch
import cutlass_gemm
# Generate inputs for the GEMM
A, B, C = [torch.ones((512, 512)).to('cuda') for _ in range(3)]
# Run the module
D = cutlass_gemm.run(A, B, C)
"""
import logging
import os
from cutlass_library import ConvKind, ConvKindNames, DataType, SubstituteTemplate
from cutlass_cppgen import CUTLASS_PATH, logger, swizzle
from cutlass_cppgen.backend.gemm_operation import GemmOperationGrouped, GemmOperationUniversal
from cutlass_cppgen.backend.conv2d_operation import Conv2dOperation
from cutlass_cppgen.backend.library import ApiVersion
from cutlass_cppgen.emit import common
from cutlass_cppgen.utils.datatypes import is_torch_available
if is_torch_available():
import torch
_PYTORCH_CUDA_TEMPLATE = common._CSTYLE_AUTOGEN_COMMENT + """
#include <cuda_runtime.h>
#include <torch/extension.h>
#include <ATen/ATen.h>
#include <ATen/cuda/CUDAContext.h>
#include "cutlass/cutlass.h"
#include "cutlass/util/device_memory.h"
// helper function allocating the memory
void* device_memory_allocation(size_t size, int device_id=0) {
if (size > 0) {
torch::Device device(torch::kCUDA, device_id);
cudaStream_t stream = at::cuda::getCurrentCUDAStream();
torch::TensorOptions options = torch::TensorOptions().dtype(torch::kI8).device(device);
at::Tensor device_tensor = torch::empty({(long)size,}, options);
return reinterpret_cast<void*>(device_tensor.data_ptr());
} else {
return nullptr;
}
}
${includes}
${declaration}
${impl}
"""
_PYTORCH_GEMM_CPP_TEMPLATE = common._CSTYLE_AUTOGEN_COMMENT + """
#include <torch/extension.h>
#include <ATen/ATen.h>
#include <pybind11/stl.h>
// CUDA forward declarations
at::Tensor ${name}_kernel(const at::Tensor& A, const at::Tensor& B, at::optional<const at::Tensor> C=at::nullopt, float alpha=1.f, float beta=0.f);
// C++ interface
at::Tensor ${name}(const at::Tensor& A, const at::Tensor& B, at::optional<const at::Tensor> C=at::nullopt, float alpha=1.f, float beta=0.f) {
return ${name}_kernel(A, B, C, alpha, beta);
}
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.def("run", py::overload_cast<const at::Tensor&, const at::Tensor&, at::optional<const at::Tensor>, float, float>(&${name}), py::arg("A"), py::arg("B"), py::arg("C") = nullptr, py::arg("alpha") = 1.f, py::arg("beta") = 0.f);
}
"""
_PYTORCH_GROUPED_GEMM_CPP_TEMPLATE = common._CSTYLE_AUTOGEN_COMMENT + """
#include <torch/extension.h>
#include <ATen/ATen.h>
#include <pybind11/stl.h>
// CUDA forward declarations
std::vector<at::Tensor> ${name}_kernel(const std::vector<at::Tensor>& A, const std::vector<at::Tensor>& B, at::optional<const std::vector<at::Tensor>> C=at::nullopt, float alpha=1.f, float beta=0.f);
// C++ interface
std::vector<at::Tensor> ${name}(const std::vector<at::Tensor>& A, const std::vector<at::Tensor>& B, at::optional<const std::vector<at::Tensor>> C=at::nullopt, float alpha=1.f, float beta=0.f) {
return ${name}_kernel(A, B, C, alpha, beta);
}
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.def("run", py::overload_cast<const std::vector<at::Tensor>&, const std::vector<at::Tensor>&, at::optional<const std::vector<at::Tensor>>, float, float>(&${name}),
py::arg("A"), py::arg("B"), py::arg("C") = nullptr, py::arg("alpha") = 1.f, py::arg("beta") = 0.f);
}
"""
_PYTORCH_CONV2D_FPROP_CPP_TEMPLATE = common._CSTYLE_AUTOGEN_COMMENT + """
#include <torch/extension.h>
#include <ATen/ATen.h>
#include <pybind11/stl.h>
// CUDA forward declarations
at::Tensor ${name}_kernel(
const at::Tensor& A, const at::Tensor& B, at::optional<const at::Tensor> C=at::nullopt,
std::tuple<int, int> stride={1, 1}, std::tuple<int, int> padding={0, 0}, std::tuple<int, int> dilation={1, 1},
float alpha=1.f, float beta=0.f,
std::string split_k_mode="serial", int split_k_slices=1);
// C++ interface
at::Tensor ${name}(
const at::Tensor& A, const at::Tensor& B, at::optional<const at::Tensor> C=at::nullopt,
std::tuple<int, int> stride={1, 1}, std::tuple<int, int> padding={0, 0}, std::tuple<int, int> dilation={1, 1},
float alpha=1.f, float beta=0.f,
std::string split_k_mode="serial", int split_k_slices=1) {
return ${name}_kernel(A, B, C, stride, padding, dilation, alpha, beta, split_k_mode, split_k_slices);
}
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.def("run",
py::overload_cast<
const at::Tensor&, const at::Tensor&, at::optional<const at::Tensor>,
std::tuple<int, int>, std::tuple<int, int>, std::tuple<int, int>, float, float, std::string, int>(
&${name}), py::arg("A"), py::arg("B"), py::arg("C") = nullptr,
py::arg("stride") = std::make_tuple(1, 1), py::arg("padding") = std::make_tuple(1, 1), py::arg("dilation") = std::make_tuple(1, 1),
py::arg("alpha") = 1.f, py::arg("beta") = 0.f,
py::arg("split_k_mode") = "serial", py::arg("split_k_slices") = 1);
}
"""
_PYTORCH_CONV2D_GRAD_CPP_TEMPLATE = common._CSTYLE_AUTOGEN_COMMENT + """
#include <torch/extension.h>
#include <ATen/ATen.h>
#include <pybind11/stl.h>
// CUDA forward declarations
at::Tensor ${name}_kernel(
std::tuple<int, int, int, int> result_size, const at::Tensor& A, const at::Tensor& B, at::optional<const at::Tensor> C=at::nullopt,
std::tuple<int, int> stride={1, 1}, std::tuple<int, int> padding={0, 0}, std::tuple<int, int> dilation={1, 1},
float alpha=1.f, float beta=0.f,
std::string split_k_mode="serial", int split_k_slices=1);
// C++ interface
at::Tensor ${name}(
std::tuple<int, int, int, int> result_size, const at::Tensor& A, const at::Tensor& B, at::optional<const at::Tensor> C=at::nullopt,
std::tuple<int, int> stride={1, 1}, std::tuple<int, int> padding={0, 0}, std::tuple<int, int> dilation={1, 1},
float alpha=1.f, float beta=0.f,
std::string split_k_mode="serial", int split_k_slices=1) {
return ${name}_kernel(result_size, A, B, C, stride, padding, dilation, alpha, beta, split_k_mode, split_k_slices);
}
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.def("run",
py::overload_cast<
std::tuple<int, int, int, int>, const at::Tensor&, const at::Tensor&, at::optional<const at::Tensor>,
std::tuple<int, int>, std::tuple<int, int>, std::tuple<int, int>, float, float, std::string, int>(
&${name}), py::arg("result_size"), py::arg("A"), py::arg("B"), py::arg("C") = nullptr,
py::arg("stride") = std::make_tuple(1, 1), py::arg("padding") = std::make_tuple(1, 1), py::arg("dilation") = std::make_tuple(1, 1),
py::arg("alpha") = 1.f, py::arg("beta") = 0.f,
py::arg("split_k_mode") = "serial", py::arg("split_k_slices") = 1);
}
"""
_PYTORCH_GEMM_INCLUDES = {
ApiVersion.v2x: """
#include "cutlass/gemm/device/gemm_universal.h"
""",
ApiVersion.v3x: """
#include "cutlass/gemm/device/gemm_universal_adapter.h"
#include "cutlass/gemm/collective/collective_builder.hpp"
#include "cutlass/gemm/device/gemm_universal_adapter.h"
#include "cutlass/gemm/kernel/gemm_universal.hpp"
#include "cutlass/epilogue/collective/collective_builder.hpp"
#include "cutlass/util/packed_stride.hpp"
""",
}
_PYTORCH_GROUPED_GEMM_INCLUDES = """
#include "cutlass/gemm/kernel/default_gemm_grouped.h"
#include "cutlass/gemm/device/gemm_grouped.h"
"""
_PYTORCH_CONV2D_INCLUDES = """
#include "cutlass/conv/kernel/default_conv2d_fprop.h"
#include "cutlass/conv/kernel/default_conv2d_dgrad.h"
#include "cutlass/conv/kernel/default_conv2d_wgrad.h"
#include "cutlass/conv/device/implicit_gemm_convolution.h"
"""
_CUTLASS_TYPE_TO_TORCH_TYPE = {
DataType.f16: "torch::kF16",
DataType.f32: "torch::kF32",
DataType.f64: "torch::kF64",
DataType.s8: "torch::kI8",
DataType.s32: "torch::kI32",
DataType.bf16: "torch::kBFloat16",
}
_PYTORCH_GEMM_IMPL_TEMPLATE_2x = (
common._CUTLASS_KERNEL_RUN_GEMM_2x
+ """
at::Tensor ${name}_kernel(const at::Tensor& A, const at::Tensor& B, at::optional<const at::Tensor> C, float alpha, float beta) {
int M = A.size(0);
int N = B.size(1);
int K = A.size(1);
typename DeviceKernel::ElementC* ptrC = (C == at::nullopt) ?
nullptr :
reinterpret_cast<typename DeviceKernel::ElementC*>(C->contiguous().data_ptr());
at::Tensor D = B.new_empty({M, N}, ${torch_type_C});
cutlass::Status status = ${name}_kernel_run(M, N, K,
reinterpret_cast<typename DeviceKernel::ElementA*>(A.contiguous().data_ptr()),
reinterpret_cast<typename DeviceKernel::ElementB*>(B.contiguous().data_ptr()),
ptrC,
reinterpret_cast<typename DeviceKernel::ElementC*>(D.contiguous().data_ptr()),
ElementCompute(alpha), ElementCompute(beta));
TORCH_CHECK(status == cutlass::Status::kSuccess, "CUTLASS kernel failed");
return D;
}
"""
)
_PYTORCH_GEMM_IMPL_TEMPLATE_3x = (
common._CUTLASS_KERNEL_RUN_GEMM_3x
+ """
bool hw_info_queried = false;
cutlass::KernelHardwareInfo hw_info;
at::Tensor ${name}_kernel(const at::Tensor& A, const at::Tensor& B, at::optional<const at::Tensor> C, float alpha, float beta) {
int M = A.size(0);
int N = B.size(1);
int K = A.size(1);
int L = 1;
// Query hardware info if we haven't already
if (!hw_info_queried) {
hw_info.device_id = 0;
hw_info.sm_count = cutlass::KernelHardwareInfo::query_device_multiprocessor_count(hw_info.device_id);
}
typename DeviceKernel::ElementC* ptrC = (C == at::nullopt) ?
nullptr :
reinterpret_cast<typename DeviceKernel::ElementC*>(C->contiguous().data_ptr());
at::Tensor D = B.new_empty({M, N}, ${torch_type_C});
cutlass::Status status = ${name}_kernel_run(M, N, K, L,
reinterpret_cast<typename DeviceKernel::ElementA*>(A.contiguous().data_ptr()),
reinterpret_cast<typename DeviceKernel::ElementB*>(B.contiguous().data_ptr()),
ptrC,
reinterpret_cast<typename DeviceKernel::ElementC*>(D.contiguous().data_ptr()),
ElementCompute(alpha), ElementCompute(beta),
hw_info);
TORCH_CHECK(status == cutlass::Status::kSuccess, "CUTLASS kernel failed");
return D;
}
"""
)
_PYTORCH_GROUPED_GEMM_IMPL_TEMPLATE = (
common._CUTLASS_KERNEL_RUN_GROUPED_GEMM_2x
+ """
std::vector<at::Tensor> ${name}_kernel(const std::vector<at::Tensor>& A, const std::vector<at::Tensor>& B, at::optional<const std::vector<at::Tensor>> C, float alpha, float beta) {
size_t num = A.size();
// To avoid performing many small cudaMallocs and host-to-device copies,
// we serialize the grouped GEMM arguments on the host, allocate one
// large chunk of device memory, and perform a single cudaMemcpy to
// copy the host data to the device. Allocation overheads could be
// avoided by using a memory pool.
// Calculate the total size of the data to be copied from host to device
size_t total_size = sizeof(cutlass::gemm::GemmCoord) +
sizeof(DeviceKernel::ElementA*) +
sizeof(DeviceKernel::ElementB*) +
sizeof(DeviceKernel::ElementC*) +
sizeof(DeviceKernel::ElementC*) +
sizeof(int64_t) +
sizeof(int64_t) +
sizeof(int64_t);
total_size *= num;
// num * sizeof(cutlass::gemm::GemmCoord) may leave one at a non-multiple
// of sizeof(DeviceKernel::ElementA*) (which will be 64 on a 64-bit system).
// To ensure that we don't end up having misaligned loads in the kernel,
// we pad to the nearest multiple of 8.
//
// Note that, even on a 32-bit system (for which sizeof(X*) will not equal
// sizeof(int64_t)), only padding between the list of GemmCoords and the
// list of ptr_As is sufficient because the set of four equal-length lists of pointers
// (A*, B*, C*, D*) will ensure that the first list of int64_ts will always
// start on a multiple of 8.
int64_t padding = 8 - (total_size % 8);
total_size += padding;
uint8_t* host_data = new uint8_t[total_size];
cutlass::DeviceAllocation<uint8_t> device_data(total_size);
uint8_t* start = host_data;
cutlass::gemm::GemmCoord* problem_sizes_host = reinterpret_cast<cutlass::gemm::GemmCoord*>(start);
// Apply the padding after the list of GemmCoords
start += num * sizeof(cutlass::gemm::GemmCoord) + padding;
int64_t ptr_A_offset = start - host_data;
DeviceKernel::ElementA** ptr_A_host = reinterpret_cast<DeviceKernel::ElementA**>(start);
start += num * sizeof(DeviceKernel::ElementA*);
int64_t ptr_B_offset = start - host_data;
DeviceKernel::ElementB** ptr_B_host = reinterpret_cast<DeviceKernel::ElementB**>(start);
start += num * sizeof(DeviceKernel::ElementB*);
int64_t ptr_C_offset = start - host_data;
DeviceKernel::ElementC** ptr_C_host = reinterpret_cast<DeviceKernel::ElementC**>(start);
start += num * sizeof(DeviceKernel::ElementC*);
int64_t ptr_D_offset = start - host_data;
DeviceKernel::ElementC** ptr_D_host = reinterpret_cast<DeviceKernel::ElementC**>(start);
start += num * sizeof(DeviceKernel::ElementC*);
int64_t lda_offset = start - host_data;
int64_t* lda_host = reinterpret_cast<int64_t*>(start);
start += num * sizeof(int64_t);
int64_t ldb_offset = start - host_data;
int64_t* ldb_host = reinterpret_cast<int64_t*>(start);
start += num * sizeof(int64_t);
int64_t ldc_offset = start - host_data;
int64_t* ldc_host = reinterpret_cast<int64_t*>(start);
start += num * sizeof(int64_t);
std::vector<at::Tensor> D(num);
bool need_C = (C != at::nullopt) && (beta != 0.f);
for (size_t i = 0; i < num; ++i) {
int M = A[i].size(0);
int N = B[i].size(1);
int K = A[i].size(1);
*(problem_sizes_host + i) = {M, N, K};
*(ptr_A_host + i) = reinterpret_cast<typename DeviceKernel::ElementA*>(A[i].contiguous().data_ptr());
*(ptr_B_host + i) = reinterpret_cast<typename DeviceKernel::ElementB*>(B[i].contiguous().data_ptr());
if (need_C) {
*(ptr_C_host + i) = reinterpret_cast<typename DeviceKernel::ElementC*>(C->at(i).contiguous().data_ptr());
}
else {
*(ptr_C_host + i) = nullptr;
}
D[i] = B[i].new_empty({M, N}, ${torch_type_C});
*(ptr_D_host + i) = reinterpret_cast<typename DeviceKernel::ElementC*>(D[i].contiguous().data_ptr());
*(lda_host + i) = DeviceKernel::LayoutA::packed({M, K}).stride(0);
*(ldb_host + i) = DeviceKernel::LayoutB::packed({K, N}).stride(0);
*(ldc_host + i) = DeviceKernel::LayoutC::packed({M, N}).stride(0);
}
device_data.copy_from_host(host_data);
cutlass::Status status = ${name}_kernel_run(
num,
reinterpret_cast<cutlass::gemm::GemmCoord*>(device_data.get()),
reinterpret_cast<DeviceKernel::ElementA**>(device_data.get() + ptr_A_offset),
reinterpret_cast<DeviceKernel::ElementB**>(device_data.get() + ptr_B_offset),
reinterpret_cast<DeviceKernel::ElementC**>(device_data.get() + ptr_C_offset),
reinterpret_cast<DeviceKernel::ElementC**>(device_data.get() + ptr_D_offset),
reinterpret_cast<int64_t*>(device_data.get() + lda_offset),
reinterpret_cast<int64_t*>(device_data.get() + ldb_offset),
reinterpret_cast<int64_t*>(device_data.get() + ldc_offset),
reinterpret_cast<int64_t*>(device_data.get() + ldc_offset),
ElementCompute(alpha), ElementCompute(beta));
delete[] host_data;
TORCH_CHECK(status == cutlass::Status::kSuccess, "CUTLASS kernel failed");
return D;
}
"""
)
_PYTORCH_CONV2D_IMPL_TEMPLATE_2x = """
cudaStream_t stream = at::cuda::getCurrentCUDAStream();
cutlass::Status status = ${name}_kernel_run(
&problem_size,
reinterpret_cast<typename UnderlyingKernel::ElementA*>(A.data_ptr()),
reinterpret_cast<typename UnderlyingKernel::ElementB*>(B.data_ptr()),
ptrC,
reinterpret_cast<typename UnderlyingKernel::ElementC*>(D.data_ptr()),
alpha, beta,
split_k_mode, stream, B.device().index());
TORCH_CHECK(status == cutlass::Status::kSuccess, "CUTLASS kernel failed");
return D;
}
"""
_PYTORCH_CONV2D_FPROP_IMPL_TEMPLATE_2x = (
common._CUTLASS_KERNEL_RUN_CONV2D_2x
+ """
at::Tensor ${name}_kernel(const at::Tensor& A, const at::Tensor& B, at::optional<const at::Tensor> C=at::nullopt,
std::tuple<int, int> stride={1, 1}, std::tuple<int, int> padding={0, 0}, std::tuple<int, int> dilation={1, 1},
float alpha=1.f, float beta=0.f, std::string split_k_mode="serial", int split_k_slices=1) {
int N, H, W, C_, K, R, S, P, Q;
N = A.size(0);
C_ = A.size(1);
H = A.size(2);
W = A.size(3);
K = B.size(0);
R = B.size(2);
S = B.size(3);
cutlass::conv::Conv2dProblemSize problem_size(
cutlass::Tensor4DCoord(N, H, W, C_),
cutlass::Tensor4DCoord(K, R, S, C_),
cutlass::Tensor4DCoord(std::get<0>(padding), std::get<0>(padding), std::get<1>(padding), std::get<1>(padding)),
cutlass::MatrixCoord(std::get<0>(stride), std::get<1>(stride)),
cutlass::MatrixCoord(std::get<0>(dilation), std::get<1>(dilation)),
cutlass::conv::Mode::kCrossCorrelation,
split_k_slices
);
P = problem_size.P;
Q = problem_size.Q;
typename UnderlyingKernel::ElementC* ptrC = (C == at::nullopt) ?
nullptr :
reinterpret_cast<typename UnderlyingKernel::ElementC*>(C->data_ptr());
torch::TensorOptions options = torch::TensorOptions().dtype(${torch_type_C}).device(B.device()).memory_format(at::MemoryFormat::ChannelsLast);
at::Tensor D = torch::zeros({N, K, P, Q}, options);
""" + _PYTORCH_CONV2D_IMPL_TEMPLATE_2x
)
_PYTORCH_CONV2D_DGRAD_IMPL_TEMPLATE_2x = (
common._CUTLASS_KERNEL_RUN_CONV2D_2x
+ """
at::Tensor ${name}_kernel(std::tuple<int, int, int, int> input_size, const at::Tensor& A, const at::Tensor& B, at::optional<const at::Tensor> C=at::nullopt,
std::tuple<int, int> stride={1, 1}, std::tuple<int, int> padding={0, 0}, std::tuple<int, int> dilation={1, 1}, float alpha=1.f, float beta=0.f,
std::string split_k_mode="serial", int split_k_slices=1) {
int N, H, W, C_, K, R, S;
N = std::get<0>(input_size);
C_ = std::get<1>(input_size);
H = std::get<2>(input_size);
W = std::get<3>(input_size);
K = B.size(0);
R = B.size(2);
S = B.size(3);
cutlass::conv::Conv2dProblemSize problem_size(
cutlass::Tensor4DCoord(N, H, W, C_),
cutlass::Tensor4DCoord(K, R, S, C_),
cutlass::Tensor4DCoord(std::get<0>(padding), std::get<0>(padding), std::get<1>(padding), std::get<1>(padding)),
cutlass::MatrixCoord(std::get<0>(stride), std::get<1>(stride)),
cutlass::MatrixCoord(std::get<0>(dilation), std::get<1>(dilation)),
cutlass::conv::Mode::kCrossCorrelation,
split_k_slices
);
typename UnderlyingKernel::ElementC* ptrC = (C == at::nullopt) ?
nullptr :
reinterpret_cast<typename UnderlyingKernel::ElementC*>(C->data_ptr());
torch::TensorOptions options = torch::TensorOptions().dtype(${torch_type_C}).device(B.device()).memory_format(at::MemoryFormat::ChannelsLast);
at::Tensor D = torch::empty({N, C_, H, W}, options);
""" + _PYTORCH_CONV2D_IMPL_TEMPLATE_2x
)
_PYTORCH_CONV2D_WGRAD_IMPL_TEMPLATE_2x = (
common._CUTLASS_KERNEL_RUN_CONV2D_2x
+ """
at::Tensor ${name}_kernel(std::tuple<int, int, int, int> weight_size, const at::Tensor& A, const at::Tensor& B, at::optional<const at::Tensor> C=at::nullopt,
std::tuple<int, int> stride={1, 1}, std::tuple<int, int> padding={0, 0}, std::tuple<int, int> dilation={1, 1}, float alpha=1.f, float beta=0.f,
std::string split_k_mode="serial", int split_k_slices=1) {
int N, H, W, C_, K, R, S;
K = std::get<0>(weight_size);
C_ = std::get<1>(weight_size);
R = std::get<2>(weight_size);
S = std::get<3>(weight_size);
N = B.size(0);
H = B.size(2);
W = B.size(3);
cutlass::conv::Conv2dProblemSize problem_size(
cutlass::Tensor4DCoord(N, H, W, C_),
cutlass::Tensor4DCoord(K, R, S, C_),
cutlass::Tensor4DCoord(std::get<0>(padding), std::get<0>(padding), std::get<1>(padding), std::get<1>(padding)),
cutlass::MatrixCoord(std::get<0>(stride), std::get<1>(stride)),
cutlass::MatrixCoord(std::get<0>(dilation), std::get<1>(dilation)),
cutlass::conv::Mode::kCrossCorrelation,
split_k_slices
);
typename UnderlyingKernel::ElementC* ptrC = (C == at::nullopt) ?
nullptr :
reinterpret_cast<typename UnderlyingKernel::ElementC*>(C->data_ptr());
torch::TensorOptions options = torch::TensorOptions().dtype(${torch_type_C}).device(B.device()).memory_format(at::MemoryFormat::ChannelsLast);
at::Tensor D = torch::empty({K, C_, R, S}, options);
""" + _PYTORCH_CONV2D_IMPL_TEMPLATE_2x
)
_PYTORCH_SETUP_PY = common._PYSTYLE_AUTOGEN_COMMENT + """
from setuptools import setup
from torch.utils.cpp_extension import BuildExtension, CUDAExtension
setup(
name='${name}',
ext_modules=[
CUDAExtension('${name}', [
'${name}.cpp',
'${name}_kernel.cu',
],
include_dirs=['${cutlass_path}/include', '${cutlass_path}/tools/util/include'],
extra_compile_args={
'cxx': ['-std=c++17'],
'nvcc': ['-std=c++17', ${extra_compile_args}],
},
libraries=['cuda']
),
],
cmdclass={
'build_ext': BuildExtension
})
"""
def _generate_setup(name: str, sourcedir: str, extra_compile_args: str=""):
"""
Generates a setup.py file for the extension
:param name: name of the module to generate
:type name: str
:param sourcedir: directory to which generated source files should be written
:type sourcedir: str
:param extra_compile_args: additional arguments to pass to setup.py
:type extra_args: str
"""
setup_py_file = os.path.join(sourcedir, "setup.py")
setup_source = SubstituteTemplate(
_PYTORCH_SETUP_PY, {"name": name, "cutlass_path": CUTLASS_PATH, "extra_compile_args": extra_compile_args}
)
with open(setup_py_file, "w") as outfile:
outfile.write(setup_source)
class _ArchListSetter:
"""
Utility context manager for temporarily setting the value of the ``TORCH_CUDA_ARCH_LIST``
environment variable when building a PyTorch CUDA module.
``TORCH_CUDA_ARCH_LIST`` is a space-delmited list of compute capabilites for which a PyTorch
CUDA module should be compiled.
For example, ``TORCH_CUDA_ARCH_LIST="7.0 8.0"`` would result in the inclusion of
``-gencode=arch=compute_70,code=sm_70`` and ``-gencode=arch=compute_80,code=sm_80`` in the
compilation of the module.
This utility wraps the building of a PyTorch CUDA module with a setting of this environment
variable according to the current compute capability being targetted.
Example usage:
.. highlight:: python
.. code-block:: python
# Temporarily set TORCH_CUDA_ARCH_LIST="8.0"
with _ArchListSetter(80):
# Perform JIT compilation and loading of the module
mod = torch.utils.cpp_extension.load(...)
:param cc: compute capability
:type cc: int
"""
_TORCH_CUDA_ARCH_LIST = "TORCH_CUDA_ARCH_LIST"
def __init__(self, cc: int):
self.cc_str = ".".join(list(str(cc)))
def __enter__(self):
"""
Saves the old value of TORCH_CUDA_ARCH_LIST and reset it to the new value based on ``cc``
"""
self.old_arch_list = os.getenv(_ArchListSetter._TORCH_CUDA_ARCH_LIST)
os.environ[_ArchListSetter._TORCH_CUDA_ARCH_LIST] = self.cc_str
return self
def __exit__(self, exc_type, exc_val, traceback):
"""
Restores the old value of TORCH_CUDA_ARCH_LIST
"""
if self.old_arch_list is None:
del os.environ[_ArchListSetter._TORCH_CUDA_ARCH_LIST]
else:
os.environ[_ArchListSetter._TORCH_CUDA_ARCH_LIST] = self.old_arch_list
def _jit(name: str, cc: int, cpp_file: str, cuda_file: str):
"""
JIT compiles and loads a PyTorch CUDA extension.
:param name: name of the module to generate
:type name: str
:param cc: compute capability of the device the module should target
:type cc: int
:param cpp_file: path to file containing extension's C++ interface
:type cpp_file: str
:param cuda_file: path to file containing extension's CUDA interface
:type cuda_file: str
:return: loaded PyTorch module
"""
from torch.utils.cpp_extension import load
extra_cuda_cflags = ["-std=c++17"]
if cc in [90, 100, 101, 103]:
# PyTorch does not currently add the sm_90a target when compute capability
# 9.0 is set within TORCH_CUDA_ARCH_LIST. Thus, we manually add the sm_90a target.
extra_cuda_cflags.append(f"-gencode=arch=compute_{cc}a,code=sm_{cc}a")
with _ArchListSetter(cc):
jitmodule = load(
name,
[cpp_file, cuda_file],
extra_cuda_cflags=extra_cuda_cflags,
extra_include_paths=[
os.path.join(CUTLASS_PATH, "include"),
os.path.join(CUTLASS_PATH, "tools/util/include"),
],
extra_ldflags=["-lcuda"],
verbose=(logger.level == logging.DEBUG)
)
return jitmodule
def _pytorch_gemm(op, name: str, cc: int, jit: bool = False, sourcedir: str = ""):
"""
Generates source for building a PyTorch CUDA module that leverages the CUTLASS GEMM
specified by ``op``. If the ``jit`` parameter is set to true, the module is just-in-time
compiled, loaded, and returned.
:param op: operation to emit in the module
:param name: name of the module to generate
:type name: str
:param cc: compute capability of the device the module should target
:type cc: int
:param jit: whether the module should be just-in-time compiled
:type jit: bool
:param sourcedir: directory to which generated source files should be written
:type sourcedir: str
:return: loaded PyTorch module if ``jit=True`` or ``None`` otherwise
"""
if sourcedir != "" and not os.path.isdir(sourcedir):
os.makedirs(sourcedir)
cuda_file = os.path.join(sourcedir, name + "_kernel.cu")
extra_kw = {}
if op.api == ApiVersion.v3x:
impl_template = _PYTORCH_GEMM_IMPL_TEMPLATE_3x
else:
impl_template = _PYTORCH_GEMM_IMPL_TEMPLATE_2x
if op.swizzling_functor == swizzle.ThreadblockSwizzleStreamK:
extra_kw["args"] = common._CUTLASS_KERNEL_ARGS_2x_STREAM_K
else:
extra_kw["args"] = common._CUTLASS_KERNEL_ARGS_2x
impl_template = (
_PYTORCH_GEMM_IMPL_TEMPLATE_3x
if op.api == ApiVersion.v3x
else _PYTORCH_GEMM_IMPL_TEMPLATE_2x
)
cuda_impl = SubstituteTemplate(impl_template, {"name": name, **extra_kw})
cuda_source = SubstituteTemplate(
_PYTORCH_CUDA_TEMPLATE,
{
"includes": _PYTORCH_GEMM_INCLUDES[op.api],
"declaration": op.rt_module.emit(),
"procedural_name": op.procedural_name(),
"impl": cuda_impl,
"torch_type_C": _CUTLASS_TYPE_TO_TORCH_TYPE[op.C.element],
},
)
with open(cuda_file, "w") as outfile:
outfile.write(cuda_source)
cpp_file = os.path.join(sourcedir, name + ".cpp")
cpp_source = SubstituteTemplate(
_PYTORCH_GEMM_CPP_TEMPLATE,
{"name": name, "description": f"CUTLASS {op.procedural_name()} GEMM"},
)
with open(cpp_file, "w") as outfile:
outfile.write(cpp_source)
extra_compile_args = ""
if cc in [90, 100, 101, 103]:
extra_compile_args = f"'--generate-code=arch=compute_{cc}a,code=[sm_{cc}a]'"
_generate_setup(name, sourcedir, extra_compile_args)
if jit:
return _jit(name, cc, cpp_file, cuda_file)
return None
def _pytorch_grouped_gemm(
op, name: str, cc: int, jit: bool = False, sourcedir: str = ""
):
"""
Generates source for building a PyTorch CUDA module that leverages the CUTLASS grouped GEMM
specified by ``op``. If the ``jit`` parameter is set to true, the module is just-in-time
compiled, loaded, and returned.
:param op: operation to emit in the module
:param name: name of the module to generate
:type name: str
:param cc: compute capability of the device the module should target
:type cc: int
:param jit: whether the module should be just-in-time compiled
:type jit: bool
:param sourcedir: directory to which generated source files should be written
:type sourcedir: str
:return: loaded PyTorch module if ``jit=True`` or ``None`` otherwise
"""
if op.api != ApiVersion.v2x:
raise Exception("Grouped GEMM is currently only supported for CUTLASS 2.x")
if sourcedir != "" and not os.path.isdir(sourcedir):
os.makedirs(sourcedir)
cuda_file = os.path.join(sourcedir, name + "_kernel.cu")
cuda_impl = SubstituteTemplate(_PYTORCH_GROUPED_GEMM_IMPL_TEMPLATE, {"name": name})
cuda_source = SubstituteTemplate(
_PYTORCH_CUDA_TEMPLATE,
{
"includes": _PYTORCH_GROUPED_GEMM_INCLUDES,
"declaration": op.rt_module.emit(),
"procedural_name": op.procedural_name(),
"impl": cuda_impl,
"torch_type_C": _CUTLASS_TYPE_TO_TORCH_TYPE[op.C.element],
},
)
with open(cuda_file, "w") as outfile:
outfile.write(cuda_source)
cpp_file = os.path.join(sourcedir, name + ".cpp")
cpp_source = SubstituteTemplate(
_PYTORCH_GROUPED_GEMM_CPP_TEMPLATE,
{"name": name, "description": f"CUTLASS {op.procedural_name()} grouped GEMM"},
)
with open(cpp_file, "w") as outfile:
outfile.write(cpp_source)
_generate_setup(name, sourcedir)
if jit:
return _jit(name, cc, cpp_file, cuda_file)
return None
def _pytorch_conv2d(op, name: str, cc: int, jit: bool = False, sourcedir: str = ""):
"""
Generates source for building a PyTorch CUDA module that leverages the CUTLASS Conv2d
specified by ``op``. If the ``jit`` parameter is set to true, the module is just-in-time
compiled, loaded, and returned.
:param op: operation to emit in the module
:param name: name of the module to generate
:type name: str
:param cc: compute capability of the device the module should target
:type cc: int
:param jit: whether the module should be just-in-time compiled
:type jit: bool
:param sourcedir: directory to which generated source files should be written
:type sourcedir: str
Note that the when conv kind is `dgrad` or `wgrad`, the size of the input `(N, C, H, W)` or
weight `(K, C, R, S)` should be provided. This is because there are multiple valid solutions
for H/W/R/S given the same P/Q.
:return: loaded PyTorch module if ``jit=True`` or ``None`` otherwise
"""
if sourcedir != "" and not os.path.isdir(sourcedir):
os.makedirs(sourcedir)
cuda_file = os.path.join(sourcedir, name + "_kernel.cu")
extra_kw = {}
if op.conv_kind == ConvKind.Fprop:
impl_template = _PYTORCH_CONV2D_FPROP_IMPL_TEMPLATE_2x
cpp_template = _PYTORCH_CONV2D_FPROP_CPP_TEMPLATE
elif op.conv_kind == ConvKind.Dgrad:
impl_template = _PYTORCH_CONV2D_DGRAD_IMPL_TEMPLATE_2x
cpp_template = _PYTORCH_CONV2D_GRAD_CPP_TEMPLATE
elif op.conv_kind == ConvKind.Wgrad:
impl_template = _PYTORCH_CONV2D_WGRAD_IMPL_TEMPLATE_2x
cpp_template = _PYTORCH_CONV2D_GRAD_CPP_TEMPLATE
extra_kw["conv_kind_name"] = ConvKindNames[op.conv_kind].capitalize()
extra_kw["torch_type_C"] = _CUTLASS_TYPE_TO_TORCH_TYPE[op.C.element]
cuda_impl = SubstituteTemplate(impl_template, {"name": name, **extra_kw})
cuda_source = SubstituteTemplate(
_PYTORCH_CUDA_TEMPLATE,
{
"includes": _PYTORCH_CONV2D_INCLUDES,
"declaration": op.rt_module.emit(),
"procedural_name": op.procedural_name(),
"impl": cuda_impl,
"torch_type_C": _CUTLASS_TYPE_TO_TORCH_TYPE[op.C.element],
},
)
with open(cuda_file, "w") as outfile:
outfile.write(cuda_source)
cpp_file = os.path.join(sourcedir, name + ".cpp")
cpp_source = SubstituteTemplate(
cpp_template,
{"name": name, "description": f"CUTLASS {op.procedural_name()} Conv2d"},
)
with open(cpp_file, "w") as outfile:
outfile.write(cpp_source)
_generate_setup(name, sourcedir)
if jit:
return _jit(name, cc, cpp_file, cuda_file)
return None
def pytorch(op, name: str, cc: int, jit: bool = False, sourcedir: str = ""):
"""
Generates source for building a PyTorch CUDA module that leverages the CUTLASS kernel
specified by ``op``. If the ``jit`` parameter is set to true, the module is just-in-time
compiled, loaded, and returned.
The result of this method is files within ``sourcedir`` that can be used for building
a PyTorch module.
:param op: operation to emit in the module
:param name: name of the module to generate
:type name: str
:param cc: compute capability of the device the module should target
:type cc: int
:param jit: whether the module should be just-in-time compiled
:type jit: bool
:param sourcedir: directory to which generated source files should be written
:type sourcedir: str
:return: loaded PyTorch module (if ``jit=True``) or None
"""
device_op = op.device_op()
if isinstance(op, GemmOperationUniversal):
return _pytorch_gemm(device_op, name, cc, jit, sourcedir)
elif isinstance(op, GemmOperationGrouped):
return _pytorch_grouped_gemm(device_op, name, cc, jit, sourcedir)
elif isinstance(op, Conv2dOperation):
return _pytorch_conv2d(device_op, name, cc, jit, sourcedir)
else:
raise Exception(
f"Operation type {type(op)} is not currently supported for PyTorch emission."
)
Reference in New Issue View Git Blame Copy Permalink