warp-lang 0.9.0__py3-none-win_amd64.whl → 0.11.0__py3-none-win_amd64.whl

warp/native/cutlass/tools/library/scripts/manifest.py

@@ -0,0 +1,402 @@
+#
+# \file generator.py
+#
+# \brief Generates the CUTLASS Library's instances
+#
+
+import enum
+import os.path
+import shutil
+
+from library import *
+from gemm_operation import *
+from rank_k_operation import *
+from rank_2k_operation import *
+from trmm_operation import *
+from symm_operation import *
+from conv2d_operation import *  
+from conv3d_operation import *  
+
+###################################################################################################
+
+class EmitOperationKindLibrary:
+  def __init__(self, generated_path, kind, args):
+    self.generated_path = generated_path
+    self.kind = kind
+    self.args = args
+    self.emitters = {
+      OperationKind.Gemm: EmitGemmConfigurationLibrary
+      , OperationKind.Conv2d: EmitConv2dConfigurationLibrary  
+      , OperationKind.Conv3d: EmitConv3dConfigurationLibrary  
+      , OperationKind.RankK: EmitRankKConfigurationLibrary
+      , OperationKind.Rank2K: EmitRank2KConfigurationLibrary
+      , OperationKind.Trmm: EmitTrmmConfigurationLibrary
+      , OperationKind.Symm: EmitSymmConfigurationLibrary
+    }
+
+    self.configurations = [];
+
+    self.header_template ="""
+/*
+ Generated by manifest.py - Do not edit.
+*/
+
+#include "cutlass/cutlass.h"
+#include "cutlass/library/library.h"
+#include "cutlass/library/manifest.h"
+
+namespace cutlass {
+namespace library {
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+"""
+    self.entry_template = """
+
+//
+// Entry point to construct operations
+//
+void initialize_all_${operation_name}_operations(Manifest &manifest) {
+"""
+    self.configuration_prototype_template = "void initialize_${configuration_name}(Manifest &manifest);\n"
+    self.configuration_template ="  initialize_${configuration_name}(manifest);\n"
+
+    self.epilogue_template ="""
+
+}
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+} // namespace library
+} // namespace cutlass
+
+"""
+
+  #
+  def __enter__(self):
+    self.operation_path = os.path.join(self.generated_path, OperationKindNames[self.kind])
+    os.mkdir(self.operation_path)
+
+    self.top_level_path = os.path.join(self.operation_path, "all_%s_operations.cu" % OperationKindNames[self.kind])
+
+    self.top_level_file = open(self.top_level_path, "w")
+    self.top_level_file.write(self.header_template)
+
+    self.source_files = [self.top_level_path,]
+
+    return self
+
+  #
+  def emit(self, configuration_name, operations):
+
+    with self.emitters[self.kind](self.operation_path, configuration_name) as configuration_emitter:
+      for operation in operations:
+        configuration_emitter.emit(operation)
+      
+      self.source_files.append(configuration_emitter.configuration_path)
+
+    self.configurations.append(configuration_name)
+    self.top_level_file.write(SubstituteTemplate(self.configuration_prototype_template, {'configuration_name': configuration_name} ))
+
+  #
+  def __exit__(self, exception_type, exception_value, traceback):
+    self.top_level_file.write(SubstituteTemplate(self.entry_template, {'operation_name': OperationKindNames[self.kind]}))
+
+    for configuration_name in self.configurations:
+      self.top_level_file.write(SubstituteTemplate(self.configuration_template, {'configuration_name': configuration_name}))
+
+    self.top_level_file.write(self.epilogue_template)
+    self.top_level_file.close()
+
+class EmitInterfaceLibrary:
+  def __init__(self, generated_path, operation_count, args):
+    self.generated_path = generated_path
+    self.args = args
+
+
+    self.prototypes = []
+    self.fn_calls = []
+    self.operation_count = str(operation_count)
+
+    self.top_level_hdr_template = '''
+/*
+ Generated by manifest.py - Do not edit.
+*/
+'''
+    self.top_level_prologue = '''
+
+#include "cutlass/library/library.h"
+#include "cutlass/library/manifest.h"
+
+namespace cutlass {
+\tnamespace library {
+
+${prototypes}
+
+\t\tvoid initialize_all(Manifest &manifest) {
+\t\t\tmanifest.reserve(${operation_count});\n\n
+${fn_calls}
+\t\t\t}
+
+\t} // namespace library
+} // namespace cutlass
+
+'''
+
+  #
+  def __enter__(self):
+    self.top_level_path = os.path.join(self.generated_path, 'initialize_all.cpp')
+
+    self.top_level_file = open(self.top_level_path, "w")
+    self.top_level_file.write(self.top_level_hdr_template)
+
+    self.source_files = [self.top_level_path,]
+
+    return self
+
+  #
+  def emit(self, operation_name):
+    self.prototypes.append(SubstituteTemplate(
+       "\t\tvoid initialize_all_${operation_kind}_operations(Manifest &manifest);",
+       {'operation_kind': operation_name}))
+    self.fn_calls.append(SubstituteTemplate(
+       "\t\t\tinitialize_all_${operation_kind}_operations(manifest);",
+       {'operation_kind': operation_name}))
+    
+
+
+  #
+  def __exit__(self, exception_type, exception_value, traceback):
+    self.top_level_file.write(SubstituteTemplate(self.top_level_prologue, {'prototypes':"\n".join(self.prototypes),
+                                                                           'fn_calls':"\n".join(self.fn_calls),
+                                                                           'operation_count': self.operation_count}))
+    self.top_level_file.close()
+
+###################################################################################################
+###################################################################################################
+
+class Options:
+  def __init__(self):
+    pass
+
+###################################################################################################
+
+#
+class Manifest:
+
+  #
+  def __init__(self, args = None):
+    self.operations = {}
+    self.args = args
+    self.operation_count = 0
+    self.operations_by_name = {}
+    
+    self.kernel_filter = ''
+    self.kernel_filter_list = []
+    self.kernel_names = []
+    self.operations_enabled = []
+    self.selected_kernels = []
+    self.ignore_kernel_names = []
+    self.compute_capabilities = [50,]
+    self.curr_build_dir = '.'
+    self.filter_by_cc = True
+
+    if self.args:
+      self.kernel_filter = self.args.kernels
+      self.curr_build_dir = args.curr_build_dir
+      architectures = args.architectures.split(';') if len(args.architectures) else ['50',]
+      self.compute_capabilities = [int(x) for x in architectures]
+      
+      if args.filter_by_cc in ['false', 'False', '0']:
+        self.filter_by_cc = False
+
+      if args.operations == 'all':
+        self.operations_enabled = []
+      else:
+        operations_list = [
+          OperationKind.Gemm
+          , OperationKind.Conv2d    
+          , OperationKind.Conv3d    
+          , OperationKind.RankK
+          , OperationKind.Trmm
+          , OperationKind.Symm
+        ] 
+        self.operations_enabled = [x for x in operations_list if OperationKindNames[x] in args.operations.split(',')]
+
+      if args.kernels == 'all':
+        self.kernel_names = []
+      else:
+        self.kernel_names = [x for x in args.kernels.split(',') if x != '']
+
+      self.ignore_kernel_names = [x for x in args.ignore_kernels.split(',') if x != '']
+
+      if args.kernel_filter_file is None:
+          self.kernel_filter_list = []
+      else:
+          self.kernel_filter_list = self.get_kernel_filters(args.kernel_filter_file)
+
+  #
+  def get_kernel_filters (self, kernelListFile):
+    if os.path.isfile(kernelListFile):
+        with open(kernelListFile, 'r') as fileReader:
+            lines = [line.rstrip() for line in fileReader if not line.startswith("#")]
+        
+        lines = [re.compile(line) for line in lines if line]
+        return lines
+    else:
+        return []
+
+  #
+  def filter_out_kernels(self, kernel_name, kernel_filter_list):
+
+    for kernel_filter_re in kernel_filter_list:
+        if kernel_filter_re.search(kernel_name) is not None:
+            return True
+        
+    return False
+
+    
+  #
+  def _filter_string_matches(self, filter_string, haystack):
+    ''' Returns true if all substrings appear in the haystack in order'''
+    substrings = filter_string.split('*')
+    for sub in substrings:
+      idx = haystack.find(sub)
+      if idx < 0:
+        return False
+      haystack = haystack[idx + len(sub):]
+    return True
+
+  #
+  def filter(self, operation):
+    ''' Filtering operations based on various criteria'''
+
+    # filter based on compute capability
+    enabled = not (self.filter_by_cc)
+
+    for cc in self.compute_capabilities:
+      if cc >= operation.tile_description.minimum_compute_capability and \
+         cc <= operation.tile_description.maximum_compute_capability and \
+         (cc not in SharedMemPerCC or SharedMemPerCC[cc] >= CalculateSmemUsage(operation)):
+
+        enabled = True
+        break
+
+    if not enabled:
+      return False
+
+    if len(self.operations_enabled) and not operation.operation_kind in self.operations_enabled:
+      return False
+
+    # eliminate duplicates
+    if operation.procedural_name() in self.operations_by_name.keys():
+      return False
+
+    # Filter based on list of valid substrings
+    if len(self.kernel_names):
+      name = operation.procedural_name()
+      enabled = False
+
+      # compare against the include list
+      for name_substr in self.kernel_names:
+        if self._filter_string_matches(name_substr, name):
+          enabled = True
+          break
+
+      # compare against the exclude list
+      for name_substr in self.ignore_kernel_names:
+        if self._filter_string_matches(name_substr, name):
+          enabled = False
+          break
+          
+    if len(self.kernel_filter_list) > 0:
+        enabled = False
+        if self.filter_out_kernels(operation.procedural_name(), self.kernel_filter_list):
+            enabled = True
+
+    # todo: filter based on compute data type
+    return enabled
+  #
+
+  #
+  def append(self, operation):
+    ''' 
+      Inserts the operation.
+
+      operation_kind -> configuration_name -> []
+    '''
+
+    if self.filter(operation):
+    
+      self.selected_kernels.append(operation.procedural_name())
+
+      self.operations_by_name[operation.procedural_name()] = operation
+
+      # add the configuration
+      configuration_name = operation.configuration_name()
+
+      if operation.operation_kind not in self.operations.keys():
+        self.operations[operation.operation_kind] = {}
+
+      if configuration_name not in self.operations[operation.operation_kind].keys():
+        self.operations[operation.operation_kind][configuration_name] = []
+
+      self.operations[operation.operation_kind][configuration_name].append(operation)
+      self.operation_count += 1
+  #
+
+  #
+  def emit(self, target = GeneratorTarget.Library):
+
+    operation_emitters = {
+      GeneratorTarget.Library: EmitOperationKindLibrary 
+    }
+    interface_emitters = {
+      GeneratorTarget.Library: EmitInterfaceLibrary 
+    }
+
+    generated_path = os.path.join(self.curr_build_dir, 'generated')
+
+    # create generated/
+    if os.path.exists(generated_path):
+      shutil.rmtree(generated_path)
+
+    os.mkdir(generated_path)
+
+    source_files = []
+
+    with interface_emitters[target](generated_path, self.operation_count, self.args) as iface_emitter:
+      for operation_kind, configurations in self.operations.items():
+        iface_emitter.emit(OperationKindNames[operation_kind])
+
+      source_files += iface_emitter.source_files
+
+
+    # for each operation kind, emit initializer for all configurations
+    for operation_kind, configurations in self.operations.items():
+      with operation_emitters[target](generated_path, operation_kind, self.args) as operation_kind_emitter:
+        for configuration_name, operations in configurations.items():
+          operation_kind_emitter.emit(configuration_name, operations)
+
+        source_files += operation_kind_emitter.source_files
+
+    # write the manifest.cmake file containing paths from all targets
+    manifest_path = os.path.join(generated_path, "manifest.cmake")
+    with open(manifest_path, "w") as manifest_file:
+
+      target_name = 'cutlass_library_objs'
+
+      target_text = SubstituteTemplate("""cutlass_target_sources(
+  ${target_name}
+  BATCH_SOURCES ON
+  PRIVATE
+""", { 'target_name': target_name})
+
+      manifest_file.write(target_text)
+
+      for source_file in source_files:
+        manifest_file.write("    %s\n" % str(source_file.replace('\\', '/')))
+      manifest_file.write(")")
+  #
+
+###################################################################################################

warp/native/cutlass/tools/library/scripts/pycutlass/docs/source/conf.py

@@ -0,0 +1,96 @@
+#################################################################################################
+#
+# Copyright (c) 2017 - 2022 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.
+#
+#################################################################################################
+
+# Configuration file for the Sphinx documentation builder.
+#
+# This file only contains a selection of the most common options. For a full
+# list see the documentation:
+# https://www.sphinx-doc.org/en/master/usage/configuration.html
+
+# -- Path setup --------------------------------------------------------------
+
+# If extensions (or modules to document with autodoc) are in another directory,
+# add these directories to sys.path here. If the directory is relative to the
+# documentation root, use os.path.abspath to make it absolute, like shown here.
+#
+# import os
+# import sys
+# sys.path.insert(0, os.path.abspath('.'))
+
+
+# -- Project information -----------------------------------------------------
+
+project = 'PyCutlass'
+copyright = '2022, Zhaodong Chen; Andrew Kerr; Haicheng Wu; Szymon Migacz; Graham Markall'
+author = 'Zhaodong Chen; Andrew Kerr; Haicheng Wu; Szymon Migacz; Graham Markall'
+
+
+# -- General configuration ---------------------------------------------------
+
+# Add any Sphinx extension module names here, as strings. They can be
+# extensions coming with Sphinx (named 'sphinx.ext.*') or your custom
+# ones.
+extensions = [
+    'sphinx.ext.duration',
+    'sphinx.ext.doctest',
+    'sphinx.ext.autodoc',
+    'sphinx.ext.intersphinx',
+    'enum_tools.autoenum',
+    'sphinx.ext.autosummary',
+    'm2r2'
+]
+
+source_suffix = [".rst", ".md"]
+
+autosummary_generate = True
+autosummary_imported_members = True
+
+# Add any paths that contain templates here, relative to this directory.
+templates_path = ['_templates']
+
+# List of patterns, relative to source directory, that match files and
+# directories to ignore when looking for source files.
+# This pattern also affects html_static_path and html_extra_path.
+exclude_patterns = []
+
+
+# -- Options for HTML output -------------------------------------------------
+
+# The theme to use for HTML and HTML Help pages.  See the documentation for
+# a list of builtin themes.
+#
+html_theme = 'bizstyle'
+
+# Add any paths that contain custom static files (such as style sheets) here,
+# relative to this directory. They are copied after the builtin static files,
+# so a file named "default.css" will overwrite the builtin "default.css".
+# html_static_path = ['_static']

warp/native/cutlass/tools/library/scripts/pycutlass/profile/conv/conv2d_f16_sm80.py

@@ -0,0 +1,106 @@
+#################################################################################################
+#
+# Copyright (c) 2017 - 2022 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.
+#
+#################################################################################################
+
+from pycutlass import *
+import pycutlass
+from pycutlass.epilogue import LinearCombination
+from pycutlass.test.conv2d_testbed import Conv2dLauncher
+
+
+if __name__ == "__main__":
+    pycutlass.get_memory_pool(2**33, 2**33)
+    pycutlass.compiler.nvcc()
+
+    math_inst = MathInstruction(
+        instruction_shape=[16, 8, 16],
+        element_a=cutlass.float16, element_b=cutlass.float16,
+        element_accumulator=cutlass.float32, opcode_class=cutlass.OpClass.TensorOp,
+        math_operation=MathOperation.multiply_add
+    )
+
+    A = TensorDescription(
+        element=math_inst.element_a, 
+        layout=cutlass.TensorNHWC,
+        alignment=8)
+    B = TensorDescription(
+        element=math_inst.element_b, 
+        layout=cutlass.TensorNHWC, 
+        alignment=8)
+    C = TensorDescription(
+        element=cutlass.float32,
+        layout=cutlass.TensorNHWC, 
+        alignment=8)
+
+    tile_description = TileDescription(
+        threadblock_shape=[128, 128, 64], stages=4, 
+        warp_count=[2, 2, 1],
+        math_instruction=math_inst
+    )
+
+    epilogue_functor = LinearCombination(cutlass.float32, 4, cutlass.float32, cutlass.float32)
+
+    operation = Conv2dOperation(
+        conv_kind=cutlass.conv.Operator.fprop, iterator_algorithm=cutlass.conv.IteratorAlgorithm.optimized,
+        arch=80, tile_description=tile_description, A=A, B=B, C=C, 
+        element_epilogue=cutlass.float32, stride_support=StrideSupport.Strided,
+        epilogue_functor=epilogue_functor,
+        swizzling_functor=cutlass.IdentitySwizzle1
+    )
+
+    profiler = Conv2dLauncher(operation, verification=False, profiling=True)
+
+    python_runtime = profiler.run(
+        problem_size = cutlass.conv.Conv2dProblemSize(
+            cutlass.Tensor4DCoord(32, 224, 224, 128),
+            cutlass.Tensor4DCoord(128, 3, 3, 128),
+            cutlass.Tensor4DCoord(1, 1, 1, 1),
+            cutlass.MatrixCoord(1, 1),
+            cutlass.MatrixCoord(1, 1),
+            cutlass.conv.Mode.cross_correlation,
+            1, 1
+        ), split_k_mode=cutlass.conv.SplitKMode.Serial
+    )
+
+
+    cpp_runtime = profiler.run_cutlass_profiler(
+        problem_size = cutlass.conv.Conv2dProblemSize(
+            cutlass.Tensor4DCoord(32, 224, 224, 128),
+            cutlass.Tensor4DCoord(128, 3, 3, 128),
+            cutlass.Tensor4DCoord(1, 1, 1, 1),
+            cutlass.MatrixCoord(1, 1),
+            cutlass.MatrixCoord(1, 1),
+            cutlass.conv.Mode.cross_correlation,
+            1, 1
+        ), split_k_mode=cutlass.conv.SplitKMode.Serial
+    )
+
+    print(cpp_runtime / python_runtime)

warp/native/cutlass/tools/library/scripts/pycutlass/profile/gemm/gemm_f32_sm80.py

@@ -0,0 +1,91 @@
+#################################################################################################
+#
+# Copyright (c) 2017 - 2022 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.
+#
+#################################################################################################
+
+import pycutlass
+from pycutlass import *
+from pycutlass.test import *
+from pycutlass.test.gemm_testbed import GemmUniversalLauncher
+
+if __name__ == '__main__':
+    pycutlass.get_memory_pool(2**32, 2**32)
+    pycutlass.compiler.nvcc()
+
+    math_inst = MathInstruction(
+        instruction_shape=[16, 8, 16],
+        element_a=cutlass.float16, element_b=cutlass.float16,
+        element_accumulator=cutlass.float32, opcode_class=cutlass.OpClass.TensorOp,
+        math_operation=MathOperation.multiply_add
+    )
+
+    tile_description = TileDescription(
+        threadblock_shape=[256, 128, 32],
+        stages=3, warp_count=[4, 2, 1],
+        math_instruction=math_inst
+    )
+
+    A = TensorDescription(
+        element=cutlass.float16, layout=cutlass.RowMajor,
+        alignment=4
+    )
+    B = TensorDescription(
+        element=cutlass.float16, layout=cutlass.RowMajor,
+        alignment=4
+    )
+    C = TensorDescription(
+        element=cutlass.float32, layout=cutlass.ColumnMajor,
+        alignment=4
+    )
+
+    element_epilogue = cutlass.float32
+
+    epilogue_functor = LinearCombination(cutlass.float32, 4, cutlass.float32, cutlass.float32)
+    
+    swizzling_functor = cutlass.IdentitySwizzle1
+
+    operation = GemmOperationUniversal(
+        arch=80, tile_description=tile_description,
+        A=A, B=B, C=C, element_epilogue=element_epilogue,
+        epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor
+    )
+
+    profiler = GemmUniversalLauncher(operation, verification=False, profiling=True)
+    python_runtime = profiler.run(
+        mode=cutlass.gemm.Mode.Gemm, 
+        problem_size=cutlass.gemm.GemmCoord(4096, 4096, 4096)
+    )
+
+    cpp_runtime = profiler.run_cutlass_profiler(
+        mode=cutlass.gemm.Mode.Gemm,
+        problem_size=cutlass.gemm.GemmCoord(4096, 4096, 4096),
+    )
+
+    print(cpp_runtime / python_runtime)