[rocmlibs] Use tensile while building rocblas

bin/rocmlibs/build_hipblas.sh

@@ -25,8 +25,9 @@ for _arch in $GFXLIST ; do
  fi
  _sep=";"
 done
-#export CC=$AOMP_INSTALL_DIR/bin/clang 
-#export CXX=$AOMP_INSTALL_DIR/bin/clang++
+export CC=$LLVM_INSTALL_LOC/bin/clang
+export CXX=$LLVM_INSTALL_LOC/bin/clang++
+export FC=$LLVM_INSTALL_LOC/bin/flang
 export ROCM_DIR=$AOMP_INSTALL_DIR
 export ROCM_PATH=$AOMP_INSTALL_DIR
 export PATH=$AOMP_SUPP/cmake/bin:$AOMP_INSTALL_DIR/bin:$PATH

bin/rocmlibs/build_powerinfer.sh

@@ -36,10 +36,10 @@ export CXXFLAGS="-I$AOMP_INSTALL_DIR/include -D__HIP_PLATFORM_AMD__=1"
 
 MYCMAKEOPTS="-DCMAKE_BUILD_TYPE=$BUILD_TYPE \
 -DLLAMA_HIPBLAS=ON \
--DCMAKE_PREFIX_PATH=$AOMP_INSTALL_DIR/lib/cmake  \
+-DCMAKE_PREFIX_PATH=$LLVM_INSTALL_LOC/lib/cmake  \
 -DCMAKE_INSTALL_PREFIX=$AOMP_INSTALL_DIR/PowerInfer \
--DCMAKE_C_COMPILER=$AOMP/bin/clang \
--DCMAKE_CXX_COMPILER=$AOMP/bin/clang++ \
+-DCMAKE_C_COMPILER=$LLVM_INSTALL_LOC/bin/clang \
+-DCMAKE_CXX_COMPILER=$LLVM_INSTALL_LOC/bin/clang++ \
 -DHIP_PLATFORM=amd \
 -DAMDGPU_TARGETS="\'$GFXSEMICOLONS\'" \
 "
@@ -115,6 +115,22 @@ if [ "$1" == "install" ] ; then
    echo
    echo "SUCCESSFUL INSTALL to $AOMP_INSTALL_DIR/PowerInfer"
    echo
+   pushd $_repo_dir
+   cd gguf-py
+   echo "Installing gguf python package"
+   pip install .
+   if [ $? != 0 ] ; then
+      echo "ERROR pip install failed for PowerInfer/gguf-py package"
+      exit 1
+   fi
+   cd ../powerinfer-py
+   echo "Installing powerinfer python package"
+   pip install .
+   if [ $? != 0 ] ; then
+      echo "ERROR pip install failed for PowerInfer/powerinfer-py package"
+      exit 1
+   fi
+   popd
    removepatch $_repo_dir
 else 
    echo 

bin/rocmlibs/build_rocblas.sh

@@ -14,24 +14,17 @@ thisdir=`dirname $realpath`
 _repo_dir=$AOMP_REPOS/rocmlibs/rocBLAS
 _build_dir=$_repo_dir/build
 
-AOMP_BUILD_TENSILE=${AOMP_BUILD_TENSILE:-0}
-
-if [ $AOMP_BUILD_TENSILE == 0 ] ; then 
-   echo 
-   echo "WARNING: Building rocblas without Tensile"
-   _local_tensile_opt=""
-else
-   _tensile_repo_dir=$AOMP_REPOS/rocmlibs/Tensile
-   _cwd=$PWD
-   cd $_tensile_repo_dir
-   git checkout release/rocm-rel-6.2
-   git pull
-   # FIXME:  We should get the Tensile hash from rocBLAS/tensile_tag.txt
-   #git checkout 97e0cfc2c8cb87a1e38901d99c39090dc4181652
-   git checkout 66ab453c3fcfc3f3816e3383cb0eccb528a1b5a9
-   cd $_cwd
-   _local_tensile_opt="--test_local_path=$_tensile_repo_dir"
-fi
+_tensile_repo_dir=$AOMP_REPOS/rocmlibs/Tensile
+_cwd=$PWD
+cd $_tensile_repo_dir
+git checkout release/rocm-rel-6.2
+git pull
+# Read the commit SHA from the file rocBLAS/tensile_tag.txt
+_tensile_commit_sha=$(cat $_repo_dir/tensile_tag.txt)
+# Checkout the specific commit SHA
+git checkout $_tensile_commit_sha
+cd $_cwd
+_local_tensile_opt="--test_local_path=$_tensile_repo_dir"
 
 patchrepo $_repo_dir
 
@@ -100,11 +93,9 @@ if [ "$1" != "install" ] ; then
    echo rm -rf $_build_dir
    rm -rf $_build_dir
    mkdir -p $_build_dir
-   if [ $AOMP_BUILD_TENSILE != 0 ] ; then 
-      # Cleanup possible old tensile build area
-      echo rm -rf $_tensile_repo_dir/build
-      rm -rf $_tensile_repo_dir/build
-   fi
+   # Cleanup possible old tensile build area
+   echo rm -rf $_tensile_repo_dir/build
+   rm -rf $_tensile_repo_dir/build
 else
    if [ ! -d $_build_dir ] ; then 
       echo "ERROR: The build directory $_build_dir"
@@ -126,7 +117,6 @@ $_local_tensile_opt \
 --install_invoked \
 --build_dir $_build_dir \
 --src_path=$_repo_dir \
---no_tensile \
 --jobs=$AOMP_JOB_THREADS \
 --architecture="""$_gfxlist""" \
 "

bin/rocmlibs/build_rocprim.sh

@@ -32,14 +32,14 @@ export PATH=$AOMP_SUPP/cmake/bin:$AOMP/bin:$PATH
 export USE_PERL_SCRIPTS=1
 export NUM_PROC=$AOMP_JOB_THREADS
 export AMDGPU_TARGETS="$GFXSEMICOLONS"
-export CXXFLAGS="-I$AOMP_INSTALL_DIR/include -D__HIP_PLATFORM_AMD__=1"
+export CXXFLAGS="-I$LLVM_INSTALL_LOC/include -D__HIP_PLATFORM_AMD__=1"
 export LDFLAGS="-fPIC"
-export CMAKE_PREFIX_PATH="$AOMP_INSTALL_DIR/lib/cmake"
+export CMAKE_PREFIX_PATH="$LLVM_INSTALL_LOC/lib/cmake"
 MYCMAKEOPTS="-DCMAKE_BUILD_TYPE=$BUILD_TYPE \
--DCMAKE_CXX_COMPILER=$AOMP_INSTALL_DIR/bin/clang++ \
--DHIP_COMPILER=$AOMP_INSTALL_DIR/bin/clang \
+-DCMAKE_CXX_COMPILER=$LLVM_INSTALL_LOC/bin/clang++ \
+-DHIP_COMPILER=$LLVM_INSTALL_LOC/bin/clang \
 -DHIP_CXX_COMPILER=$AOMP_INSTALL_DIR/bin/hipcc \
--DCMAKE_PREFIX_PATH=$AOMP_INSTALL_DIR/lib/cmake \
+-DCMAKE_PREFIX_PATH=$LLVM_INSTALL_LOC/lib/cmake \
 -DCMAKE_INSTALL_PREFIX=$AOMP_INSTALL_DIR \
 -DAMDGPU_TARGETS="\'$GFXSEMICOLONS\'"
 -DROCM_DIR=$ROCM_DIR \

bin/rocmlibs/build_rocsolver.sh

@@ -25,8 +25,8 @@ for _arch in $GFXLIST ; do
  fi
  _sep=";"
 done
-export CC=$AOMP_INSTALL_DIR/bin/clang 
-export CXX=$AOMP_INSTALL_DIR/bin/clang++
+export CC=$LLVM_INSTALL_LOC/bin/clang 
+export CXX=$LLVM_INSTALL_LOC/bin/clang++
 export ROCM_DIR=$AOMP_INSTALL_DIR
 export ROCM_PATH=$AOMP_INSTALL_DIR
 export PATH=$AOMP_SUPP/cmake/bin:$AOMP_INSTALL_DIR/bin:$PATH

bin/rocmlibs/build_rocsparse.sh

@@ -25,15 +25,15 @@ for _arch in $GFXLIST ; do
  fi
  _sep=";"
 done
-export CC=$AOMP_INSTALL_DIR/bin/clang 
-export CXX=$AOMP_INSTALL_DIR/bin/clang++
+export CC=$LLVM_INSTALL_LOC/bin/clang 
+export CXX=$LLVM_INSTALL_LOC/bin/clang++
 export ROCM_DIR=$AOMP_INSTALL_DIR
 export ROCM_PATH=$AOMP_INSTALL_DIR
 export PATH=$AOMP_SUPP/cmake/bin:$AOMP_INSTALL_DIR/bin:$PATH
 export HIP_USE_PERL_SCRIPTS=1
 export USE_PERL_SCRIPTS=1
 export NUM_PROC=$AOMP_JOB_THREADS
-export CXXFLAGS="-I$AOMP_INSTALL_DIR/include -D__HIP_PLATFORM_AMD__=1"
+export CXXFLAGS="-I$LLVM_INSTALL_LOC/include -D__HIP_PLATFORM_AMD__=1"
 export LDFLAGS="-fPIC"
 if [ "$AOMP_USE_CCACHE" != 0 ] ; then
    _ccache_bin=`which ccache`

bin/rocmlibs/patches/rocblas.patch

@@ -1,60 +1,67 @@
 diff --git a/CMakeLists.txt b/CMakeLists.txt
-index 9424c65b..e454bc07 100644
+index f70de1cf..5eedfcc3 100644
 --- a/CMakeLists.txt
 +++ b/CMakeLists.txt
-@@ -111,7 +111,7 @@ list( APPEND CMAKE_PREFIX_PATH ${ROCM_PATH}/llvm ${ROCM_PATH} ${ROCM_PATH}/hip /
- set( TARGET_LIST_ROCM_5.6 "gfx803;gfx900;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102")
- set( TARGET_LIST_ROCM_5.7 "gfx803;gfx900;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102")
- set( TARGET_LIST_ROCM_6.0 "gfx900;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102")
--set( TARGET_LIST_ROCM_6.3 "gfx900;gfx906:xnack-;gfx908:xnack-;gfx90a;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201")
-+set( TARGET_LIST_ROCM_6.3 "gfx900;gfx906:xnack-;gfx908:xnack-;gfx90a;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201;gfx1103;gfx90c")
-
- if(ROCM_PLATFORM_VERSION)
-   if(${ROCM_PLATFORM_VERSION} VERSION_LESS 5.7.0)
+@@ -112,7 +112,7 @@ if (NOT BUILD_ADDRESS_SANITIZER)
+   set( TARGET_LIST_ROCM_5.6 "gfx803;gfx900;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102")
+   set( TARGET_LIST_ROCM_5.7 "gfx803;gfx900;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102")
+   set( TARGET_LIST_ROCM_6.0 "gfx900;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102")
+-  set( TARGET_LIST_ROCM_6.3 "gfx900;gfx906:xnack-;gfx908:xnack-;gfx90a;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102;gfx1151;gfx1200;gfx1201")
++  set( TARGET_LIST_ROCM_6.3 "gfx900;gfx906:xnack-;gfx908:xnack-;gfx90a;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102;gfx1151;gfx1200;gfx1201;gfx1103;gfx90c")
+ else()
+   set( TARGET_LIST_ROCM_5.6 "gfx908:xnack+;gfx90a:xnack+")
+   set( TARGET_LIST_ROCM_5.7 "gfx908:xnack+;gfx90a:xnack+;gfx942:xnack+")
 diff --git a/library/src/handle.cpp b/library/src/handle.cpp
-index 44be1ee1..8f4af5c5 100644
+index ea1fa1a7..cb14874c 100644
 --- a/library/src/handle.cpp
 +++ b/library/src/handle.cpp
-@@ -101,6 +101,14 @@ static Processor getActiveArch(int deviceId)
+@@ -106,6 +106,10 @@ static Processor getActiveArch(int deviceId)
      {
          return Processor::gfx90a;
      }
 +    else if(deviceString.find("gfx90c") != std::string::npos)
 +    {
 +        return Processor::gfx90c;
 +    }
+     else if(deviceString.find("gfx940") != std::string::npos)
+     {
+         return Processor::gfx940;
+@@ -146,6 +150,10 @@ static Processor getActiveArch(int deviceId)
+     {
+         return Processor::gfx1102;
+     }
 +    else if(deviceString.find("gfx1103") != std::string::npos)
 +    {
 +        return Processor::gfx1103;
 +    }
-     else if(deviceString.find("gfx940") != std::string::npos)
+     else if(deviceString.find("gfx1151") != std::string::npos)
      {
-         return Processor::gfx940;
+         return Processor::gfx1151;
 diff --git a/library/src/include/handle.hpp b/library/src/include/handle.hpp
-index b0717b5d..8df7f0f5 100644
+index 94d18c7b..c47cefe8 100644
 --- a/library/src/include/handle.hpp
 +++ b/library/src/include/handle.hpp
-@@ -79,6 +79,7 @@ enum class Processor : int
+@@ -82,6 +82,7 @@ enum class Processor : int
      gfx906  = 906,
      gfx908  = 908,
      gfx90a  = 910,
 +    gfx90c  = 912,
      gfx940  = 940,
      gfx941  = 941,
      gfx942  = 942,
-@@ -93,6 +94,7 @@ enum class Processor : int
+@@ -96,6 +97,7 @@ enum class Processor : int
      gfx1100 = 1100,
      gfx1101 = 1101,
      gfx1102 = 1102,
 +    gfx1103 = 1103,
+     gfx1151 = 1151,
      gfx1200 = 1200,
      gfx1201 = 1201
- };
 diff --git a/library/src/tensile_host.cpp b/library/src/tensile_host.cpp
-index 53bebba1..f1367429 100644
+index f7d4bdb1..47d69c0d 100644
 --- a/library/src/tensile_host.cpp
 +++ b/library/src/tensile_host.cpp
-@@ -266,6 +266,10 @@ namespace
+@@ -271,6 +271,10 @@ namespace
          {
              return Tensile::LazyLoadingInit::gfx90a;
          }
@@ -65,7 +72,7 @@ index 53bebba1..f1367429 100644
          else if(deviceString.find("gfx940") != std::string::npos)
          {
              return Tensile::LazyLoadingInit::gfx940;
-@@ -306,6 +310,10 @@ namespace
+@@ -311,6 +315,10 @@ namespace
          {
              return Tensile::LazyLoadingInit::gfx1102;
          }

bin/rocmlibs/patches/rocprim.patch

@@ -1,13 +1,13 @@
 diff --git a/CMakeLists.txt b/CMakeLists.txt
-index bdff5831..55c05947 100644
+index 68e9b7bf..049ed365 100644
 --- a/CMakeLists.txt
 +++ b/CMakeLists.txt
 @@ -91,7 +91,7 @@ if(NOT USE_HIP_CPU)
 
    if(GPU_TARGETS STREQUAL "all")
      rocm_check_target_ids(DEFAULT_AMDGPU_TARGETS
--      TARGETS "gfx803;gfx900:xnack-;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack-;gfx90a:xnack+;gfx940;gfx941;gfx942;gfx1030;gfx1100;gfx1101;gfx1102"
-+      TARGETS "gfx803;gfx900:xnack-;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack-;gfx90a:xnack+;gfx90c;gfx940;gfx941;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1103"
+-      TARGETS "gfx803;gfx900:xnack-;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack-;gfx90a:xnack+;gfx940;gfx941;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1151"
++      TARGETS "gfx803;gfx900:xnack-;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack-;gfx90a:xnack+;gfx90c;gfx940;gfx941;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1103;gfx1151"
      )
      set(GPU_TARGETS "${DEFAULT_AMDGPU_TARGETS}" CACHE STRING "GPU architectures to compile for" FORCE)
    endif()

bin/rocmlibs/patches/rocsolver.patch

@@ -1,5 +1,5 @@
 diff --git a/CMakeLists.txt b/CMakeLists.txt
-index f857f1e..de6b11e 100644
+index 0f88788..95d5ac6 100644
 --- a/CMakeLists.txt
 +++ b/CMakeLists.txt
 @@ -159,12 +159,14 @@ if(NOT DEFINED AMDGPU_TARGETS)
@@ -14,6 +14,6 @@ index f857f1e..de6b11e 100644
        gfx1101
        gfx1102
 +      gfx1103
+       gfx1151
    )
    set(AMDGPU_TARGETS_INIT
-     gfx900

bin/rocmlibs/patches/rocsparse.patch

@@ -1,22 +1,22 @@
 diff --git a/CMakeLists.txt b/CMakeLists.txt
-index afcba556..7dc9bf14 100644
+index 6a4f9d21..ff9d3635 100644
 --- a/CMakeLists.txt
 +++ b/CMakeLists.txt
-@@ -148,14 +148,14 @@ if(COMMAND rocm_check_target_ids)
+@@ -162,14 +162,14 @@ if(COMMAND rocm_check_target_ids)
          TARGETS "gfx908:xnack+;gfx90a:xnack+;gfx940:xnack+;gfx941:xnack+;gfx942:xnack+")
      else()
        rocm_check_target_ids(DEFAULT_AMDGPU_TARGETS
--          TARGETS "gfx803;gfx900:xnack-;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack-;gfx90a:xnack+;gfx940;gfx941;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201")
-+          TARGETS "gfx803;gfx900:xnack-;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack-;gfx90a:xnack+;gfx940;gfx941;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201;gfx1103;gfx90c")
+-          TARGETS "gfx803;gfx900:xnack-;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack-;gfx90a:xnack+;gfx940;gfx941;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1151;gfx1200;gfx1201")
++          TARGETS "gfx803;gfx900:xnack-;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack-;gfx90a:xnack+;gfx90c;gfx940;gfx941;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1103;gfx1151;gfx1200;gfx1201")
      endif()
  else()
      # Use target ID syntax if supported for AMDGPU_TARGETS
      if(TARGET_ID_SUPPORT)
--        set(DEFAULT_AMDGPU_TARGETS "gfx803;gfx900:xnack-;gfx906:xnack-;gfx908:xnack-;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201")
-+        set(DEFAULT_AMDGPU_TARGETS "gfx803;gfx900:xnack-;gfx906:xnack-;gfx908:xnack-;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201;gfx1103;gfx90c")
+-        set(DEFAULT_AMDGPU_TARGETS "gfx803;gfx900:xnack-;gfx906:xnack-;gfx908:xnack-;gfx1030;gfx1100;gfx1101;gfx1102;gfx1151;gfx1200;gfx1201")
++        set(DEFAULT_AMDGPU_TARGETS "gfx803;gfx900:xnack-;gfx906:xnack-;gfx908:xnack-;gfx90c;gfx1030;gfx1100;gfx1101;gfx1102;gfx1103;gfx1151;gfx1200;gfx1201")
      else()
 -        set(DEFAULT_AMDGPU_TARGETS "gfx803;gfx900;gfx906;gfx908")
-+        set(DEFAULT_AMDGPU_TARGETS "gfx803;gfx900;gfx906;gfx908;gfx1103;gfx90c")
++        set(DEFAULT_AMDGPU_TARGETS "gfx803;gfx900;gfx906;gfx908;gfx90c;gfx1103")
      endif()
  endif()
 

bin/rocmlibs/test_powerinfer.sh

@@ -23,6 +23,8 @@ if [ ! -d $MODEL_DIR ]; then
 fi
 pushd $MODEL_DIR
 
+pushd $MODEL_DIR
+
 # Install huggingface-cli to download the PowerInfer GGUF models
 pip install -U "huggingface_hub[cli]"
 export PATH=$PATH:$HOME/.local/bin
@@ -64,7 +66,7 @@ popd
 PREDICTION_LENGTH=128
 THREAD_COUNT=8
 PROMPT="Once upon a time"
-VRAM_BUDGET=4
+VRAM_BUDGET=16
 
 MODEL_NAME_BASE=$(basename $MODEL_NAME)
 
@@ -83,7 +85,6 @@ fi
 $POWERINFER_INSTALL_DIR/bin/main -m $MODEL_DIR/$MODEL_NAME -n $PREDICTION_LENGTH -t $THREAD_COUNT -p "$PROMPT" --vram-budget $VRAM_BUDGET 2>&1 | tee $LOG_FILE_NAME.log
 
 set +x
-
 # Print the log file name and parameters with which the model was run
 echo "Inference via PowerInfer completed for the following:"
 echo "Model: $MODEL_NAME"