[gentoo-commits] proj/sci:master commit in: dev-util/Tensile/files/, dev-util/Tensile/

["Andrew Ammerlaan" <andrewammerlaan@gentoo.org>]

commit:     f1b7b457296505b40a3ff204fed5d605086e39e6
Author:     Andrew Ammerlaan <andrewammerlaan <AT> gentoo <DOT> org>
AuthorDate: Sat Mar  5 14:41:24 2022 +0000
Commit:     Andrew Ammerlaan <andrewammerlaan <AT> gentoo <DOT> org>
CommitDate: Sat Mar  5 14:42:12 2022 +0000
URL:        https://gitweb.gentoo.org/proj/sci.git/commit/?id=f1b7b457

dev-util/Tensile: add version 4.3.0

https://github.com/gentoo/gentoo/pull/23450

Package-Manager: Portage-3.0.30, Repoman-3.0.3
Signed-off-by: Andrew Ammerlaan <andrewammerlaan <AT> gentoo.org>

 dev-util/Tensile/Tensile-4.3.0.ebuild              | 74 +++++++++++++++++
 ...ensile-4.3.0-hsaco-compile-specified-arch.patch | 96 ++++++++++++++++++++++
 .../files/Tensile-4.3.0-output-commands.patch      | 23 ++++++
 3 files changed, 193 insertions(+)

diff --git a/dev-util/Tensile/Tensile-4.3.0.ebuild b/dev-util/Tensile/Tensile-4.3.0.ebuild
new file mode 100644
index 000000000..bfd86ecea
--- /dev/null
+++ b/dev-util/Tensile/Tensile-4.3.0.ebuild
@@ -0,0 +1,74 @@
+# Copyright 1999-2022 Gentoo Authors
+# Distributed under the terms of the GNU General Public License v2
+
+EAPI=8
+
+PYTHON_COMPAT=( python3_{8..10} )
+DISTUTILS_USE_PEP517=setuptools
+inherit distutils-r1
+
+DESCRIPTION="Stretching GPU performance for GEMMs and tensor contractions"
+HOMEPAGE="https://github.com/ROCmSoftwarePlatform/Tensile"
+SRC_URI="https://github.com/ROCmSoftwarePlatform/Tensile/archive/rocm-${PV}.tar.gz -> rocm-Tensile-${PV}.tar.gz"
+S="${WORKDIR}/${PN}-rocm-${PV}"
+
+LICENSE="MIT"
+KEYWORDS="~amd64"
+SLOT="0"
+
+# Not compatible with recent versions of pytest
+RESTRICT="test"
+
+RDEPEND="${PYTHON_DEPS}
+	dev-python/pyyaml[${PYTHON_USEDEP}]
+	dev-python/msgpack[${PYTHON_USEDEP}]
+"
+DEPEND="${RDEPEND}
+	dev-util/hip:${SLOT}
+"
+BDEPEND="test? (
+	dev-util/rocminfo:${SLOT}
+)"
+
+PATCHES=(
+	"${FILESDIR}/Tensile-${PV}-hsaco-compile-specified-arch.patch" # backported from upstream, should remove after 4.3.0
+	"${FILESDIR}/Tensile-4.3.0-output-commands.patch"
+)
+
+CMAKE_USE_DIR="${WORKDIR}/Source"
+
+distutils_enable_tests pytest
+
+src_prepare() {
+	distutils-r1_src_prepare
+
+	mv ${PN}/Source "${WORKDIR}"/ || die
+	sed -e "/ROCM_SMI_ROOT/s,lib,$(get_libdir)," \
+		-i "${WORKDIR}"/Source/cmake/FindROCmSMI.cmake || die
+	sed -r -e "/TENSILE_USE_LLVM/s/ON/OFF/" \
+		-i "${WORKDIR}"/Source/CMakeLists.txt || die
+
+	sed -e "/HipClangVersion/s/0,0,0/$(ver_rs 1-3 ,)/" \
+		-e "/SourcePath/s,os\.path\.join.*$,\"${EPREFIX}/usr/share/${PN}\"," \
+		-i ${PN}/Common.py || die
+
+	sed -e "s|os\.path\.dirname.*$|\"${EPREFIX}/usr/share/Tensile\", end='')|" \
+		-i ${PN}/__init__.py || die
+}
+
+src_test() {
+	ROCM_PATH="${EPREFIX}/usr/" distutils-r1_src_test
+}
+
+src_install() {
+	distutils-r1_src_install
+
+	# Move the cmake files to the correct directory
+	mkdir -p "${ED}/usr/$(get_libdir)/cmake/${PN}" || die
+	mv "${ED}/usr/cmake/"* "${ED}/usr/$(get_libdir)/cmake/${PN}" || die
+	rm -r "${ED}/usr/cmake" || die
+
+	insinto /usr/share/${PN}
+	doins -r "${WORKDIR}"/Source/*
+	dosym . /usr/share/${PN}/Source
+}

diff --git a/dev-util/Tensile/files/Tensile-4.3.0-hsaco-compile-specified-arch.patch b/dev-util/Tensile/files/Tensile-4.3.0-hsaco-compile-specified-arch.patch
new file mode 100644
index 000000000..8e6753781
--- /dev/null
+++ b/dev-util/Tensile/files/Tensile-4.3.0-hsaco-compile-specified-arch.patch
@@ -0,0 +1,96 @@
+https://github.com/ROCmSoftwarePlatform/Tensile/issues/1395
+https://github.com/ROCmSoftwarePlatform/Tensile/pull/1398
+
+--- a/Tensile/TensileCreateLibrary.py
++++ b/Tensile/TensileCreateLibrary.py
+@@ -136,6 +136,35 @@ def which(p):
+                 return candidate
+     return None
+ 
++def splitArchs():
++  # Helper for architecture
++  def isSupported(arch):
++    return globalParameters["AsmCaps"][arch]["SupportedISA"] and \
++           globalParameters["AsmCaps"][arch]["SupportedSource"]
++
++  if ";" in globalParameters["Architecture"]:
++    wantedArchs = globalParameters["Architecture"].split(";")
++  else:
++    wantedArchs = globalParameters["Architecture"].split("_")
++  archs = []
++  cmdlineArchs = []
++  if "all" in wantedArchs:
++    for arch in globalParameters['SupportedISA']:
++      if isSupported(arch):
++        if (arch == (9,0,6) or arch == (9,0,8) or arch == (9,0,10)):
++          if (arch == (9,0,10)):
++            archs += [gfxName(arch) + '-xnack+']
++            cmdlineArchs += [gfxName(arch) + ':xnack+']
++          archs += [gfxName(arch) + '-xnack-']
++          cmdlineArchs += [gfxName(arch) + ':xnack-']
++        else:
++          archs += [gfxName(arch)]
++          cmdlineArchs += [gfxName(arch)]
++  else:
++    for arch in wantedArchs:
++      archs += [re.sub(":", "-", arch)]
++      cmdlineArchs += [arch]
++  return archs, cmdlineArchs
+ 
+ def buildSourceCodeObjectFile(CxxCompiler, outputPath, kernelFile):
+     buildPath = ensurePath(os.path.join(globalParameters['WorkingPath'], 'code_object_tmp'))
+@@ -149,24 +178,8 @@ def buildSourceCodeObjectFile(CxxCompiler, outputPath, kernelFile):
+     objectFilename = base + '.o'
+     soFilename = base + '.so'
+ 
+-    def isSupported(arch):
+-        return globalParameters["AsmCaps"][arch]["SupportedISA"] and \
+-               globalParameters["AsmCaps"][arch]["SupportedSource"]
+-
+     if (CxxCompiler == "hipcc"):
+-      archs = []
+-      cmdlineArchs = []
+-      for arch in globalParameters['SupportedISA']:
+-        if isSupported(arch):
+-          if (arch == (9,0,6) or arch == (9,0,8) or arch == (9,0,10)):
+-            if (arch == (9,0,10)):
+-              archs += [gfxName(arch) + '-xnack+']
+-              cmdlineArchs += [gfxName(arch) + ':xnack+']
+-            archs += [gfxName(arch) + '-xnack-']
+-            cmdlineArchs += [gfxName(arch) + ':xnack-']
+-          else:
+-            archs += [gfxName(arch)]
+-            cmdlineArchs += [gfxName(arch)]
++      archs, cmdlineArchs = splitArchs()
+ 
+       archFlags = ['--offload-arch=' + arch for arch in cmdlineArchs]
+ 
+@@ -1063,11 +1076,6 @@ def buildObjectFileNames(solutionWriter, kernelWriterSource, kernelWriterAssembl
+   sourceKernels = list([k for k in kernels if k['KernelLanguage'] == 'Source'])
+   asmKernels = list([k for k in kernels if k['KernelLanguage'] == 'Assembly'])
+ 
+-  # Helper for architecture
+-  def isSupported(arch):
+-        return globalParameters["AsmCaps"][arch]["SupportedISA"] and \
+-               globalParameters["AsmCaps"][arch]["SupportedSource"]
+-
+   # Build a list of kernel object names.
+   for kernel in sourceKernels:
+     sourceKernelNames += [kernelWriterSource.getKernelFileBase(kernel)]
+@@ -1081,15 +1089,7 @@ def buildObjectFileNames(solutionWriter, kernelWriterSource, kernelWriterAssembl
+ 
+   # Source based kernels are built for all supported architectures
+   if (cxxCompiler == 'hipcc'):
+-    sourceArchs = []
+-    for arch in globalParameters['SupportedISA']:
+-      if isSupported(arch):
+-        if (arch == (9,0,6) or arch == (9,0,8) or arch == (9,0,10)):
+-          if (arch == (9,0,10)):
+-            sourceArchs += [gfxName(arch) + '-xnack+']
+-          sourceArchs += [gfxName(arch) + '-xnack-']
+-        else:
+-          sourceArchs += [gfxName(arch)]
++    sourceArchs, _ = splitArchs()
+   else:
+     raise RuntimeError("Unknown compiler %s" % cxxCompiler)
+ 

diff --git a/dev-util/Tensile/files/Tensile-4.3.0-output-commands.patch b/dev-util/Tensile/files/Tensile-4.3.0-output-commands.patch
new file mode 100644
index 000000000..be5a4db21
--- /dev/null
+++ b/dev-util/Tensile/files/Tensile-4.3.0-output-commands.patch
@@ -0,0 +1,23 @@
+diff --color -uprN orig/Tensile/cmake/TensileConfig.cmake Tensile-rocm-4.3.0/Tensile/cmake/TensileConfig.cmake
+--- orig/Tensile/cmake/TensileConfig.cmake	2021-08-18 17:48:00.115478470 +0800
++++ Tensile-rocm-4.3.0/Tensile/cmake/TensileConfig.cmake	2021-08-18 17:48:49.963478801 +0800
+@@ -234,6 +234,7 @@ function(TensileCreateLibraryFiles
+         COMMENT "Generating Tensile Libraries"
+         OUTPUT ${Tensile_EMBED_LIBRARY_SOURCE};${Tensile_MANIFEST_CONTENTS}
+         COMMAND ${CommandLine}
++		USES_TERMINAL
+       )
+ 
+       set("${Tensile_VAR_PREFIX}_ALL_FILES" ${Tensile_MANIFEST_CONTENTS} PARENT_SCOPE)
+diff --color -uprN orig/Tensile/Common.py Tensile-rocm-4.3.0/Tensile/Common.py
+--- orig/Tensile/Common.py	2021-08-18 17:48:00.075478470 +0800
++++ Tensile-rocm-4.3.0/Tensile/Common.py	2021-08-18 17:48:23.287478624 +0800
+@@ -179,7 +179,7 @@ globalParameters["PrintTensorD"] = 0
+ globalParameters["PrintTensorRef"] = 0          # Print reference tensor.  0x1=after init; 0x2=after copy-back; 0x3=both
+ globalParameters["PrintIndexAssignments"] = 0      # Print the tensor index assignment info
+ globalParameters["PrintWinnersOnly"] = False      # Only print the solutions which become the fastest
+-globalParameters["PrintCodeCommands"] = False  # print the commands used to generate the code objects (asm,link,hip-clang, etc)
++globalParameters["PrintCodeCommands"] = True  # print the commands used to generate the code objects (asm,link,hip-clang, etc)
+ globalParameters["DumpTensors"] = False        # If True, dump tensors to binary files instead of printing them.
+ 
+ # TODO - remove this when NewClient is mainstream