* add rocblas

rocblas/PKGBUILD

@@ -0,0 +1,60 @@
+# POWER Maintainer: Alexander Baldeck <alex.bldck@gmail.com>
+# Maintainer: Torsten Keßler <tpkessler at archlinux dot org>
+# Contributor: Markus Näther <naether.markus@gmail.com>
+pkgname=rocblas
+pkgver=6.0.2
+pkgrel=1
+pkgdesc='Next generation BLAS implementation for ROCm platform'
+arch=(x86_64 powerpc64le powerpc64 riscv64)
+url='https://rocblas.readthedocs.io/en/latest'
+license=('MIT')
+depends=('rocm-core' 'hip' 'glibc' 'gcc-libs' 'openmp')
+makedepends=('rocm-cmake' 'python' 'python-virtualenv' 'python-pyaml' 'python-wheel'
+             'python-msgpack' 'python-joblib' 'perl-file-which' 'msgpack-cxx' 'gcc-fortran')
+_rocblas='https://github.com/ROCmSoftwarePlatform/rocBLAS'
+_tensile='https://github.com/ROCmSoftwarePlatform/Tensile'
+source=("$pkgname-$pkgver.tar.gz::$_rocblas/archive/rocm-$pkgver.tar.gz"
+        "$pkgname-tensile-$pkgver.tar.gz::$_tensile/archive/refs/tags/rocm-$pkgver.tar.gz"
+        "find-msgpack-5.patch")
+sha256sums=('d1bf31063a2d349797b88c994c91d05f94e681bafb5550ad9b53529703d89dbb'
+            '1d8a92422560c1e908fa25fd97a4aa07a96659528a543f77618408ffcfe1f307'
+            '3f91bf087e4ea72eaef5acd500e16b61aa69c029cfcca14666799a7c42a0c5aa')
+options=(!lto)
+_dirname="$(basename "$_rocblas")-$(basename "${source[0]}" ".tar.gz")"
+_tensile_dir="$(basename "$_tensile")-$(basename "${source[1]}" ".tar.gz")"
+
+prepare() {
+    cd "$_tensile_dir"
+    patch -Np1 -i "$srcdir/find-msgpack-5.patch"
+}
+
+build() {
+  # Compile source code for supported GPU archs in parallel
+  export HIPCC_COMPILE_FLAGS_APPEND="-parallel-jobs=$(nproc)"
+  export HIPCC_LINK_FLAGS_APPEND="-parallel-jobs=$(nproc)"
+  # -fcf-protection is not supported by HIP, see
+  # https://rocm.docs.amd.com/en/latest/reference/rocmcc.html#support-status-of-other-clang-options
+  local cmake_args=(
+    -Wno-dev
+    -S "$_dirname"
+    -B build
+    -D CMAKE_BUILD_TYPE=None
+    -D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc
+    -D CMAKE_TOOLCHAIN_FILE=toolchain-linux.cmake
+    -D CMAKE_CXX_FLAGS="${CXXFLAGS} -fcf-protection=none"
+    -D CMAKE_INSTALL_PREFIX=/opt/rocm
+    -D CMAKE_PREFIX_PATH=/opt/rocm/llvm/lib/cmake/llvm
+    -D amd_comgr_DIR=/opt/rocm/lib/cmake/amd_comgr
+    -D BUILD_WITH_TENSILE=ON
+    -D Tensile_LIBRARY_FORMAT=msgpack
+    -D Tensile_TEST_LOCAL_PATH="$srcdir/$_tensile_dir"
+  )
+  cmake "${cmake_args[@]}"
+  cmake --build build
+}
+
+package() {
+  DESTDIR="$pkgdir" cmake --install build
+
+  install -Dm644 "$_dirname/LICENSE.md" "$pkgdir/usr/share/licenses/$pkgname/LICENSE"
+}

rocblas/find-msgpack-5.patch

@@ -0,0 +1,11 @@
+--- Tensile-rocm-5.4.2/Tensile/Source/lib/CMakeLists.txt.bak	2023-01-27 08:30:16.374451318 +0100
++++ Tensile-rocm-5.4.2/Tensile/Source/lib/CMakeLists.txt	2023-01-27 08:30:33.194515443 +0100
+@@ -103,7 +103,7 @@
+ endif()
+ 
+ if(TENSILE_USE_MSGPACK)
+-    find_package(msgpack REQUIRED)
++    find_package(msgpackc-cxx REQUIRED)
+     target_compile_definitions(TensileHost PUBLIC -DTENSILE_MSGPACK=1)
+ 
+     if(TARGET msgpackc-cxx)

rocblas/test.cpp

@@ -0,0 +1,70 @@
+#include <rocblas/rocblas.h>
+#include <hip/hip_runtime.h>
+#include <vector>
+#include <random>
+#include <algorithm>
+#include <cmath>
+#include <iostream>
+
+int main()
+{
+    size_t n = 128;
+    size_t size = n * n;
+
+    std::random_device rd;
+    std::mt19937 gen(rd());
+    std::uniform_real_distribution<float> dist(-1.0, 1.0);
+    auto myrand = [&](){return dist(gen);};
+
+    float *x;
+    float *y;
+    float *z;
+    hipMalloc((void**)&x, sizeof *x * size);
+    hipMalloc((void**)&y, sizeof *y * size);
+    hipMalloc((void**)&z, sizeof *z * size);
+
+    std::vector<float> xin(size);
+    std::vector<float> yin(size);
+
+    std::generate(xin.begin(), xin.end(), myrand);
+    std::generate(yin.begin(), yin.end(), myrand);
+
+    hipMemcpy(x, xin.data(), sizeof *x * size, hipMemcpyHostToDevice);
+    hipMemcpy(y, yin.data(), sizeof *x * size, hipMemcpyHostToDevice);
+
+    rocblas_handle handle;
+    rocblas_create_handle(&handle);
+
+    float alpha = 15.412f;
+    float beta = 0.0f;
+    rocblas_sgemm(handle, rocblas_operation_none, rocblas_operation_none,
+        n, n, n, &alpha, x, n, y, n, &beta, z, n);
+
+    std::vector<float> zout(size);
+    hipMemcpy(zout.data(), z, sizeof *z * size, hipMemcpyDeviceToHost);
+
+    for(size_t j = 0; j < n; j++){
+        for(size_t i = 0; i < n; i++){
+            for(size_t k = 0; k < n; k++){
+                zout[i + j * n] -= alpha * xin[i + k * n] * yin[k + j * n];
+            }
+        }
+    }
+
+    float tol = 0.001f;
+    for(size_t i = 0; i < size; i++){
+        if(std::abs(zout[i]) > tol){
+            std::cout << "Element mismatch at index " << i << "\n";
+            std::cout << "Expected: 0\n";
+            std::cout << "Actual  : " << zout[i] << "\n";
+            return 1;
+        }
+    }
+
+    std::cout << "TESTS PASSED!" << std::endl;
+
+    hipFree(x);
+    hipFree(y);
+    hipFree(z);
+    rocblas_destroy_handle(handle);
+}

rocblas/test.sh

@@ -0,0 +1,5 @@
+#! /usr/bin/env sh
+
+OUT=$(mktemp -d)
+/opt/rocm/bin/hipcc -o "$OUT"/test test.cpp -lrocblas
+"$OUT"/test