CUTLASS 3.8 Release (#2059)

* CUTLASS 3.8 Release

* update

* Update README.md

* Revert "Update README.md"

This reverts commit b353e36fe83e0815f99b44e46c0c95494c44726b.

* update

* update

---------

Co-authored-by: Haicheng Wu <57973641+hwu36@users.noreply.github.com>
Co-authored-by: Haicheng Wu <haichengw@nvidia.com>
This commit is contained in:
mihir-awatramani
2025-01-25 02:44:06 -05:00
committed by GitHub
co-authored by Haicheng Wu Haicheng Wu
parent 9eb01fa0b0
commit 389e493055
290 changed files with 91222 additions and 291 deletions
@@ -0,0 +1,290 @@
/***************************************************************************************************
* Copyright (c) 2017 - 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.
*
**************************************************************************************************/
/* \file
\brief Blockscale Gemm Profiler
*/
#pragma once
#include <vector>
#include <string>
#include <memory>
#include <algorithm>
#include <unordered_map>
// CUTLASS Library includes
#include "cutlass/library/library.h"
#include "cutlass/library/util.h"
#include "cutlass/library/manifest.h"
// Profiler includes
#include "options.h"
#include "device_context.h"
#include "operation_profiler.h"
#include "performance_result.h"
#include "problem_space.h"
#include "reduction_operation_profiler.h"
/////////////////////////////////////////////////////////////////////////////////////////////////
namespace cutlass {
namespace profiler {
/////////////////////////////////////////////////////////////////////////////////////////////////
/// Abstract base class for each math function
class BlockScaledGemmOperationProfiler : public OperationProfiler {
public:
/// Problem structure obtained from problem space
struct GemmProblem {
cutlass::library::GemmUniversalMode mode{library::GemmUniversalMode::kGemm};
int64_t m{16};
int64_t n{16};
int64_t k{16};
int cluster_m{1};
int cluster_n{1};
int cluster_k{1};
int cluster_m_fallback{1};
int cluster_n_fallback{1};
int cluster_k_fallback{1};
int64_t lda{0};
int64_t ldb{0};
int64_t ldc{0};
std::vector<uint8_t> alpha;
std::vector<uint8_t> beta;
cutlass::library::SplitKMode split_k_mode{library::SplitKMode::kNone};
int split_k_slices{1};
int batch_count{1};
cutlass::library::RasterOrder raster_order{cutlass::library::RasterOrder::kHeuristic};
int swizzle_size{1};
cutlass::library::RuntimeDatatype runtime_input_datatype_a{};
cutlass::library::RuntimeDatatype runtime_input_datatype_b{};
// gemm with parallel interleaved reduction
// gemm epilogue (alpha, beta) = (1.0, 0.0)
// reduction epilogue (alpha, beta) = (GemmProblem::alpha, GemmProblem::beta)
std::vector<uint8_t> alpha_one;
std::vector<uint8_t> beta_zero;
bool use_pdl{false};
//
// Methods
//
/// Parses the problem
Status parse(
library::BlockScaledGemmDescription const &operation_desc,
ProblemSpace const &problem_space,
ProblemSpace::Problem const &problem);
/// Total number of bytes loaded
int64_t bytes(library::BlockScaledGemmDescription const &operation_desc) const;
/// Total number of flops computed
int64_t flops(library::BlockScaledGemmDescription const &operation_desc) const;
/// Initializes a performance result
void initialize_result(
PerformanceResult &result,
library::BlockScaledGemmDescription const &operation_desc,
ProblemSpace const &problem_space);
};
/// Workspace used
struct GemmWorkspace {
DeviceAllocation *A{nullptr};
DeviceAllocation *SFA{nullptr};
DeviceAllocation *B{nullptr};
DeviceAllocation *SFB{nullptr};
DeviceAllocation *C{nullptr};
DeviceAllocation *Computed{nullptr};
DeviceAllocation *Reference{nullptr};
DeviceAllocation *Computed_SFD{nullptr};
DeviceAllocation *Reference_SFD{nullptr};
DeviceAllocation *Norm_constant{nullptr};
/// Number of copies of the problem workspace which are visited sequentially during
/// profiling to avoid camping in the last level cache.
int problem_count{1};
library::GemmUniversalConfiguration configuration;
library::BlockScaledGemmArguments arguments;
/// Buffer used for the operation's host workspace
std::vector<uint8_t> host_workspace;
/// Buffer used for the operations' device workspace
DeviceAllocation device_workspace;
/// Library configuration and arguments for reduction operator
library::ReductionConfiguration reduction_configuration;
library::ReductionArguments reduction_arguments;
/// Buffer used for the cutlass reduction operations' host workspace
std::vector<uint8_t> reduction_host_workspace;
};
protected:
//
// Data members
//
/// GEMM problem obtained from problem space
GemmProblem problem_;
/// Device memory allocations
GemmWorkspace gemm_workspace_;
/// CUTLASS parallel reduction operation to follow this* gemm operation
library::Operation const *reduction_op_;
public:
//
// Methods
//
/// Ctor
BlockScaledGemmOperationProfiler(Options const &options);
/// Destructor
virtual ~BlockScaledGemmOperationProfiler();
GemmProblem const& problem() const { return problem_; }
/// Prints usage statement for the math function
virtual void print_usage(std::ostream &out) const;
/// Prints examples
virtual void print_examples(std::ostream &out) const;
/// Extracts the problem dimensions
virtual Status initialize_configuration(
Options const &options,
PerformanceReport &report,
DeviceContext &device_context,
library::Operation const *operation,
ProblemSpace const &problem_space,
ProblemSpace::Problem const &problem);
/// Initializes workspace
virtual Status initialize_workspace(
Options const &options,
PerformanceReport &report,
DeviceContext &device_context,
library::Operation const *operation,
ProblemSpace const &problem_space,
ProblemSpace::Problem const &problem);
/// Verifies CUTLASS against references
virtual bool verify_cutlass(
Options const &options,
PerformanceReport &report,
DeviceContext &device_context,
library::Operation const *operation,
ProblemSpace const &problem_space,
ProblemSpace::Problem const &problem);
/// Measures performance results
virtual bool profile(
Options const &options,
PerformanceReport &report,
DeviceContext &device_context,
library::Operation const *operation,
ProblemSpace const &problem_space,
ProblemSpace::Problem const &problem);
protected:
/// Initializes the performance result
void initialize_result_(
PerformanceResult &result,
Options const &options,
library::BlockScaledGemmDescription const &operation_desc,
ProblemSpace const &problem_space);
/// Verifies CUTLASS against references
bool verify_with_cublas_(
Options const &options,
PerformanceReport &report,
DeviceContext &device_context,
library::Operation const *operation,
ProblemSpace const &problem_space,
ProblemSpace::Problem const &problem);
/// Verifies CUTLASS against host and device references
bool verify_with_reference_(
Options const &options,
PerformanceReport &report,
DeviceContext &device_context,
library::Operation const *operation,
ProblemSpace const &problem_space,
ProblemSpace::Problem const &problem,
cutlass::library::NumericTypeID element_A,
cutlass::library::NumericTypeID element_B);
/// Method to profile a CUTLASS Operation
Status profile_cutlass_(
PerformanceResult &result,
Options const &options,
library::Operation const *operation,
void *arguments,
void *host_workspace,
void *device_workspace);
/// Initialize reduction problem dimensions and library::Operation
bool initialize_reduction_configuration_(
library::Operation const *operation,
ProblemSpace::Problem const &problem);
};
/////////////////////////////////////////////////////////////////////////////////////////////////
} // namespace profiler
} // namespace cutlass
/////////////////////////////////////////////////////////////////////////////////////////////////
@@ -73,6 +73,15 @@ public:
int64_t n{16};
int64_t k{16};
int cluster_m{1};
int cluster_n{1};
int cluster_k{1};
int cluster_m_fallback{1};
int cluster_n_fallback{1};
int cluster_k_fallback{1};
int64_t lda{0};
int64_t ldb{0};
int64_t ldc{0};
@@ -86,6 +95,11 @@ public:
cutlass::library::RasterOrder raster_order{cutlass::library::RasterOrder::kHeuristic};
int swizzle_size{1};
cutlass::library::RuntimeDatatype runtime_input_datatype_a{};
cutlass::library::RuntimeDatatype runtime_input_datatype_b{};
// gemm with parallel interleaved reduction
// gemm epilogue (alpha, beta) = (1.0, 0.0)
// reduction epilogue (alpha, beta) = (GemmProblem::alpha, GemmProblem::beta)
@@ -942,6 +942,18 @@ bool arg_as_IteratorAlgorithmID(
ProblemSpace const &problem_space,
ProblemSpace::Problem const &problem);
/// Lexically casts an argument to an int64 if it is defined. Returns true if not null.
bool arg_as_RuntimeDatatype(library::RuntimeDatatype &runtime_datatype, KernelArgument::Value const *value_ptr);
/// Lexically casts an argument to an int64 if it is defined. Returns true if not null.
bool arg_as_RuntimeDatatype(
library::RuntimeDatatype &runtime_datatype,
char const *name,
ProblemSpace const &problem_space,
ProblemSpace::Problem const &problem);
/// Lexically casts an argument to an int64 if it is defined. Returns true if not null.
bool arg_as_RasterOrder(library::RasterOrder &raster_order, KernelArgument::Value const *value_ptr);