Implement MessagePack backend (ROCm#904)

* Implement MessagePack backend

* Install msgpack on Docker by adding Debian repo and installing from there

* Revert "Install msgpack on Docker by adding Debian repo and installing from there"

This reverts commit 94ae9de.

HostLibraryTests/ContractionLibraryLoading_test.cpp

@@ -37,7 +37,7 @@ using namespace Tensile;
 TEST(ContractionLibraryLoadingTest, MultipleKernels)
 {
     auto library = LoadLibraryFile<ContractionProblem>(
-        TestData::Instance().file("SampleTensileKernels.yaml").native());
+        TestData::Instance().file("SampleTensileKernels").native());
     ASSERT_NE(library, nullptr);
 
     AMDGPU hardware;
@@ -87,7 +87,7 @@ TEST(ContractionLibraryLoadingTest, MultipleKernels)
 
 TEST(ContractionLibraryLoadingTest, SGEMM_Kernels_Lite)
 {
-    auto library = LoadLibraryFile<ContractionProblem>(
-        TestData::Instance().file("KernelsLite.yaml").native());
+    auto library
+        = LoadLibraryFile<ContractionProblem>(TestData::Instance().file("KernelsLite").native());
     ASSERT_NE(library, nullptr);
 }

HostLibraryTests/TestData_test.cpp

@@ -37,10 +37,10 @@ TEST(TestData, Simple)
     auto is_regular_file
         = static_cast<bool (*)(boost::filesystem::path const&)>(boost::filesystem::is_regular_file);
 
-    EXPECT_PRED1(is_regular_file, data.file("KernelsLite.yaml"));
+    EXPECT_PRED1(is_regular_file, data.file("KernelsLite"));
     EXPECT_FALSE(is_regular_file(data.file("fjdlksljfjldskj")));
 
-    auto files = data.glob("*.yaml");
+    auto files = data.glob(std::string("*.") + TestData::defaultExtension);
     EXPECT_EQ(files.size(), 6);
     for(auto file : files)
     {

HostLibraryTests/configs/CMakeLists.txt

@@ -1,11 +1,21 @@
-
-set(SOLUTION_LIBRARY_FILES
-    ${SOLUTION_LIBRARY_FILES}
-    "${CMAKE_CURRENT_SOURCE_DIR}/SolutionLibraries/KernelsLite.yaml"
-    "${CMAKE_CURRENT_SOURCE_DIR}/SolutionLibraries/KernelsLiteMixed.yaml"
-    "${CMAKE_CURRENT_SOURCE_DIR}/SolutionLibraries/KernelsLiteNavi.yaml"
-    "${CMAKE_CURRENT_SOURCE_DIR}/SolutionLibraries/KernelsTileLite.yaml"
-    "${CMAKE_CURRENT_SOURCE_DIR}/SolutionLibraries/SampleTensileKernels.yaml"
-    "${CMAKE_CURRENT_SOURCE_DIR}/SolutionLibraries/rocBLAS_Full.yaml"
-    PARENT_SCOPE)
-
+if(Tensile_YAML)
+    set(SOLUTION_LIBRARY_FILES
+        ${SOLUTION_LIBRARY_FILES}
+        "${CMAKE_CURRENT_SOURCE_DIR}/SolutionLibraries/KernelsLite.yaml"
+        "${CMAKE_CURRENT_SOURCE_DIR}/SolutionLibraries/KernelsLiteMixed.yaml"
+        "${CMAKE_CURRENT_SOURCE_DIR}/SolutionLibraries/KernelsLiteNavi.yaml"
+        "${CMAKE_CURRENT_SOURCE_DIR}/SolutionLibraries/KernelsTileLite.yaml"
+        "${CMAKE_CURRENT_SOURCE_DIR}/SolutionLibraries/SampleTensileKernels.yaml"
+        "${CMAKE_CURRENT_SOURCE_DIR}/SolutionLibraries/rocBLAS_Full.yaml"
+        PARENT_SCOPE)
+else()
+    set(SOLUTION_LIBRARY_FILES
+        ${SOLUTION_LIBRARY_FILES}
+        "${CMAKE_CURRENT_SOURCE_DIR}/SolutionLibraries/KernelsLite.dat"
+        "${CMAKE_CURRENT_SOURCE_DIR}/SolutionLibraries/KernelsLiteMixed.dat"
+        "${CMAKE_CURRENT_SOURCE_DIR}/SolutionLibraries/KernelsLiteNavi.dat"
+        "${CMAKE_CURRENT_SOURCE_DIR}/SolutionLibraries/KernelsTileLite.dat"
+        "${CMAKE_CURRENT_SOURCE_DIR}/SolutionLibraries/SampleTensileKernels.dat"
+        "${CMAKE_CURRENT_SOURCE_DIR}/SolutionLibraries/rocBLAS_Full.dat"
+        PARENT_SCOPE)
+endif()

HostLibraryTests/hip/RunGEMMKernelTileSelection_test.cpp

@@ -176,17 +176,20 @@ struct RunGEMMKernelSolutionSelectionTest : public ::testing::TestWithParam<Cont
 #endif
     }
 
-    void TearDown() override
-    {
-        hipFree(a_d);
-        hipFree(b_d);
-        hipFree(c_d);
-        hipFree(d_d);
-        hipFree(d_ref_d);
+}
 
-        hipDeviceReset();
-    }
-};
+void TearDown() override
+{
+    hipFree(a_d);
+    hipFree(b_d);
+    hipFree(c_d);
+    hipFree(d_d);
+    hipFree(d_ref_d);
+
+    hipDeviceReset();
+}
+}
+;
 
 TEST_P(RunGEMMKernelSolutionSelectionTest, KernelsTileSelection)
 {
@@ -195,11 +198,11 @@ TEST_P(RunGEMMKernelSolutionSelectionTest, KernelsTileSelection)
     // std::cout << problem << std::endl;
 
     auto library = LoadLibraryFile<ContractionProblem>(
-        TestData::File("kernels_tile_selection/TensileLibrary.yaml").native());
+        TestData::Instance().file("kernels_tile_selection/TensileLibrary").native());
 
     hip::SolutionAdapter adapter(false);
     adapter.loadCodeObjectFile(
-        TestData::File("kernels_tile_selection/TensileLibrary_gfx906.co").native());
+        TestData::Instance().file("kernels_tile_selection/TensileLibrary_gfx906", "co").native());
 
     ASSERT_NE(library, nullptr);
 

HostLibraryTests/hip/RunGEMMKernel_test.cpp

@@ -534,7 +534,7 @@ std::vector<std::tuple<std::shared_ptr<SolutionLibrary<ContractionProblem>>,
 
     {
         auto library = LoadLibraryFile<ContractionProblem>(
-            TestData::Instance().file("kernels_lite/TensileLibrary.yaml").native());
+            TestData::Instance().file("kernels_lite/TensileLibrary").native());
         auto adapter = std::make_shared<hip::SolutionAdapter>(debug, "kernels_lite (file)");
         for(auto file : TestData::Instance().glob("kernels_lite/*.*co"))
             adapter->loadCodeObjectFile(file.native());
@@ -544,7 +544,7 @@ std::vector<std::tuple<std::shared_ptr<SolutionLibrary<ContractionProblem>>,
 
     {
         auto library = LoadLibraryFile<ContractionProblem>(
-            TestData::Instance().file("kernels_lite_mixed/TensileLibrary.yaml").native());
+            TestData::Instance().file("kernels_lite_mixed/TensileLibrary").native());
         auto adapter = std::make_shared<hip::SolutionAdapter>(debug, "kernels_lite_mixed (file)");
         for(auto file : TestData::Instance().glob("kernels_lite_mixed/*.*co"))
             adapter->loadCodeObjectFile(file.native());
@@ -554,7 +554,7 @@ std::vector<std::tuple<std::shared_ptr<SolutionLibrary<ContractionProblem>>,
 
     {
         auto library = LoadLibraryFile<ContractionProblem>(
-            TestData::Instance().file("tile_aware_selection/library/TensileLibrary.yaml").native());
+            TestData::Instance().file("tile_aware_selection/library/TensileLibrary").native());
 
         auto adapter = std::make_shared<hip::SolutionAdapter>(debug, "tile_aware_selection");
         for(auto file : TestData::Instance().glob("tile_aware_selection/library/*.*co"))
@@ -569,8 +569,7 @@ std::vector<std::tuple<std::shared_ptr<SolutionLibrary<ContractionProblem>>,
     auto envDir = TestData::Env("TENSILE_TEST_LIBRARY");
     if(envDir)
     {
-        auto library
-            = LoadLibraryFile<ContractionProblem>(envDir.file("TensileLibrary.yaml").native());
+        auto library = LoadLibraryFile<ContractionProblem>(envDir.file("TensileLibrary").native());
         auto adapter = std::make_shared<hip::SolutionAdapter>(debug, "TENSILE_TEST_LIBRARY");
 
         for(auto file : envDir.glob("*.co"))

HostLibraryTests/llvm/LibraryPerformance_test.cpp

@@ -37,16 +37,16 @@ using namespace Tensile;
 /**
  * LibraryPerformanceTest:
  *
- * This suite contains micro-benchmarks for pieces of the runtime library.  It
- * does not exercise any of the Hip-specific code.
+ * This suite contains micro-benchmarks for pieces of the runtime library.  It does not
+ * exercise any of the Hip-specific code.
  *
- * There are no performance-based assertions or checks.  The timing results are
- * provided by googletest.
+ * There are no performance-based assertions or checks.  The timing results are provided by
+ * googletest.
  *
- * Most of these tests depend on a library being loaded from a YAML file.  The
- * library objects are cached so that the deserialization time is not a part of
- * the actual test (outside of the LoadLibrary test). PopulateCache is an empty
- * test whose purpose is to ensure the cache is populated for the actual tests.
+ * Most of these tests depend on a library being loaded from a DAT/YAML file.  The library objects
+ * are cached so that the deserialization time is not a part of the actual test (outside of the
+ * LoadLibrary test). PopulateCache is an empty test whose purpose is to ensure the cache is
+ * populated for the actual tests.
  */
 struct LibraryPerformanceTest
     : public ::testing::TestWithParam<std::tuple<AMDGPU, std::string, bool, bool>>
@@ -223,17 +223,17 @@ std::vector<LibraryPerformanceTest::ParamType> GetParams()
 
     for(auto const& gpu : gpus)
     {
-        rv.push_back(std::make_tuple(gpu, "KernelsLite.yaml", false, false));
-        rv.push_back(std::make_tuple(gpu, "KernelsLiteMixed.yaml", false, true));
-        rv.push_back(std::make_tuple(gpu, "KernelsLiteNavi.yaml", true, false));
-        rv.push_back(std::make_tuple(gpu, "KernelsTileLite.yaml", false, false));
-        rv.push_back(std::make_tuple(gpu, "rocBLAS_Full.yaml", false, true));
+        rv.push_back(std::make_tuple(gpu, "KernelsLite", false, false));
+        rv.push_back(std::make_tuple(gpu, "KernelsLiteMixed", false, true));
+        rv.push_back(std::make_tuple(gpu, "KernelsLiteNavi", true, false));
+        rv.push_back(std::make_tuple(gpu, "KernelsTileLite", false, false));
+        rv.push_back(std::make_tuple(gpu, "rocBLAS_Full", false, true));
     }
 
     rv.push_back(std::make_tuple(
-        AMDGPU(AMDGPU::Processor::gfx908, 64, "Arcturus"), "rocBLAS_Full.yaml", false, true));
+        AMDGPU(AMDGPU::Processor::gfx908, 64, "Arcturus"), "rocBLAS_Full", false, true));
     rv.push_back(std::make_tuple(
-        AMDGPU(AMDGPU::Processor::gfx1010, 40, "Navi"), "KernelsLiteNavi.yaml", true, true));
+        AMDGPU(AMDGPU::Processor::gfx1010, 40, "Navi"), "KernelsLiteNavi", true, true));
 
     return rv;
 }

HostLibraryTests/testlib/include/TestData.hpp

@@ -41,7 +41,9 @@ struct TestData : public Tensile::LazySingleton<TestData>
 
     boost::filesystem::path dataDir() const;
 
-    boost::filesystem::path file(std::string const& filename) const;
+    static const std::string defaultExtension;
+    boost::filesystem::path  file(std::string const& filename,
+                                  std::string const& extension = defaultExtension) const;
 
     std::vector<boost::filesystem::path> glob(std::string const& pattern) const;
 

HostLibraryTests/testlib/source/TestData.cpp

@@ -39,6 +39,12 @@
 
 #include <Tensile/Utils.hpp>
 
+#ifdef TENSILE_YAML
+const std::string TestData::defaultExtension = "yaml";
+#else
+const std::string TestData::defaultExtension = "dat";
+#endif
+
 TestData::operator bool() const
 {
     return boost::filesystem::is_directory(dataDir());
@@ -64,9 +70,10 @@ boost::filesystem::path TestData::dataDir() const
     return m_dataDir;
 }
 
-boost::filesystem::path TestData::file(std::string const& filename) const
+boost::filesystem::path TestData::file(std::string const& filename,
+                                       std::string const& extension) const
 {
-    return dataDir() / filename;
+    return dataDir() / (filename + "." + extension);
 }
 
 std::vector<boost::filesystem::path> TestData::glob(std::string const& pattern) const

Tensile/BenchmarkProblems.py

@@ -32,7 +32,7 @@
 
 from . import ClientExecutable
 from . import SolutionLibrary
-from . import YAMLIO
+from . import LibraryIO
 from . import Utils
 from .BenchmarkStructs import BenchmarkProcess
 from .ClientWriter import runClient, writeClientParameters, writeClientConfig
@@ -362,7 +362,7 @@ def benchmarkProblemType( problemTypeConfig, problemSizeGroupConfig, \
     ############################################################################
     # Write Solutions YAML
     ############################################################################
-    YAMLIO.writeSolutions(solutionsFileName, benchmarkStep.problemSizes, \
+    LibraryIO.writeSolutions(solutionsFileName, benchmarkStep.problemSizes, \
         solutions )
 
     # End Iteration
@@ -520,11 +520,12 @@ def writeBenchmarkFiles(stepBaseDir, solutions, problemSizes, stepName, filesToC
       globalParameters["WorkingPath"], globalParameters["CxxCompiler"], [problemType], solutions, kernels, kernelsBetaOnly, \
       solutionWriter, kernelWriterSource, kernelWriterAssembly, errorTolerant=True )
 
+  newLibraryFilename = "TensileLibrary.yaml" if globalParameters["YAML"] else "TensileLibrary.dat"
   newLibraryDir = ensurePath(os.path.join(globalParameters["WorkingPath"], 'library'))
-  newLibraryFile = os.path.join(newLibraryDir, "TensileLibrary.yaml")
+  newLibraryFile = os.path.join(newLibraryDir, newLibraryFilename)
   newLibrary = SolutionLibrary.MasterSolutionLibrary.BenchmarkingLibrary(solutions)
   newLibrary.applyNaming(kernelMinNaming)
-  YAMLIO.write(newLibraryFile, Utils.state(newLibrary))
+  LibraryIO.configWriter(globalParameters["YAML"]).write(newLibraryFile, Utils.state(newLibrary))
 
   codeObjectFiles = [os.path.relpath(f, globalParameters["WorkingPath"]) for f in codeObjectFiles]
 

Tensile/ClientWriter.py

@@ -22,7 +22,7 @@
 from .Common import globalParameters, HR, pushWorkingPath, popWorkingPath, print1, CHeader, printWarning, listToInitializer, ClientExecutionLock
 from . import ClientExecutable
 from . import Common
-from . import YAMLIO
+from . import LibraryIO
 
 import os
 import subprocess
@@ -98,7 +98,7 @@ def main( config ):
   for logicFileName in logicFiles:
     (scheduleName, deviceNames, problemType, solutionsForType, \
         indexOrder, exactLogic, rangeLogic, newLibrary, architectureName) \
-        = YAMLIO.readLibraryLogicForSchedule(logicFileName)
+        = LibraryIO.readLibraryLogicForSchedule(logicFileName)
     if problemType["DataType"].isHalf():
         enableHalf = True
     functions.append((scheduleName, problemType))
@@ -559,7 +559,8 @@ def param(key, value):
             f.write("{}={}\n".format(key, value))
 
         sourceDir = os.path.join(stepBaseDir, "source")
-        libraryFile = os.path.join(sourceDir, "library", "TensileLibrary.yaml")
+        libraryFilename = "TensileLibrary.yaml" if globalParameters["YAML"] else "TensileLibrary.dat"
+        libraryFile = os.path.join(sourceDir, "library", libraryFilename)
         param("library-file", libraryFile)
 
         currentGFXName = Common.gfxName(globalParameters["CurrentISA"])

Tensile/Common.py

@@ -108,6 +108,7 @@
 ########################################
 globalParameters["CMakeBuildType"] = "Release"            # whether benchmark clients and library client should be release or debug
 globalParameters["PrintSolutionRejectionReason"] = False  # when a solution is marked as invalid, print why
+globalParameters["YAML"] = False                          # whether to use the YAML backend (as opposed to the MessagePack backend)
 
 # how to initialize tensor data
 # serial-in-u will use a sequence that increments in the K dimension

Tensile/Component.py

@@ -1,5 +1,5 @@
 ################################################################################
-# Copyright (C) 2020 Advanced Micro Devices, Inc. All rights reserved.
+# Copyright 2020 Advanced Micro Devices, Inc. All rights reserved.
 #
 # Permission is hereby granted, free of charge, to any person obtaining a copy
 # of this software and associated documentation files (the "Software"), to deal
@@ -72,7 +72,7 @@ def PartialMatch(pattern, obj):
     for key, value in pattern.items():
         if key not in obj:
             return False
-        
+
         objValue = obj[key]
         if isinstance(value, collections.abc.Mapping) and \
            isinstance(objValue, collections.abc.Mapping):

Tensile/Components/__init__.py

@@ -1,4 +1,4 @@
 
 def use(): pass
 
-__all__ = ["MAC_F16"]
+__all__ = ["MAC_F16"]

Tensile/KernelWriterAssembly.py

@@ -6027,7 +6027,7 @@ def endSummation(self, kernel):
       #         lastVgprForReads ^  ^         ^
       #              startVgprReuse ^         ^
       #                             lastValuC ^
-      # if valuC does not include all of lastVgprForReads, we can reuse the 
+      # if valuC does not include all of lastVgprForReads, we can reuse the
       # non-overlapped part of lastVgprForReads
       # |<-------------- valuC -------------->|
       # |xxxxxxxxxxxxxxxxxxxxx|xxxxxxxxxxxxxxx|oooooo|xx|
@@ -6036,7 +6036,7 @@ def endSummation(self, kernel):
       #                                  startVgprReuse ^
       vbegin = self.numVgprValuC
       vsize = max(0, self.lastVgprForReads-self.numVgprValuC)
-    else: 
+    else:
       vbegin = self.startVgprValuA
       vsize = self.lastVgprForReads - self.startVgprValuA