remove xclbin (and XRT) dep

nod-ai · Oct 12, 2024 · c5aeb01 · c5aeb01
1 parent 793714d
commit c5aeb01
Show file tree

Hide file tree

Showing 26 changed files with 456 additions and 432 deletions.
diff --git a/.github/workflows/ci-linux.yml b/.github/workflows/ci-linux.yml
@@ -51,7 +51,7 @@ jobs:
           git reset --hard FETCH_HEAD
           git -c submodule."third_party/torch-mlir".update=none \
               -c submodule."third_party/stablehlo".update=none \
-              -c submodule."src/runtime_src/core/common/aiebu".update=none \
+              -c submodule."third_party/XRT".update=none \
               submodule update --init --recursive --depth 1 --single-branch -j 10
           
       - name: Install deps
@@ -64,11 +64,6 @@ jobs:
         run: |
           pip install "numpy<2" pyyaml "pybind11[global]==2.10.3" nanobind pytest
 
-      - name: Peano dep
-        run: |
-          bash build_tools/download_peano.sh
-          echo "PEANO_INSTALL_DIR=$PWD/llvm-aie" >> $GITHUB_ENV
-
       - name: Run Pytest
         run: |
           pytest build_tools/ci
@@ -80,6 +75,11 @@ jobs:
           key:  ${{ env.CACHE_KEY }}
           restore-keys: linux-build-test-cpp-
 
+      - name: Peano dep
+        run: |
+          bash build_tools/download_peano.sh
+          echo "PEANO_INSTALL_DIR=$PWD/llvm-aie" >> $GITHUB_ENV
+
       - name: Build packages
         run: |
           export cache_dir="${{ env.CACHE_DIR }}"
@@ -161,64 +161,36 @@ jobs:
       - name : E2E comparison of AIE to llvm-cpu
         run: |
           source .venv/bin/activate
-          source /opt/xilinx/xrt/setup.sh
           python build_tools/ci/cpu_comparison/run.py \
             test_aie_vs_cpu \
             $PWD/iree-install \
             $PWD/llvm-aie \
-            --xrt-dir /opt/xilinx/xrt \
             --vitis-dir /opt/Xilinx/Vitis/2024.2 \
             --reset-npu-between-runs -v
 
       - name: E2E correctness matmul test
         run: |
-          # Without this additional line an error like
-          #
-          # [XRT] ERROR: Failed to allocate host memory buffer (mmap(len=10616832, prot=3, flags=8193, offset=4294967296)
-          # failed (err=11): Resource temporarily unavailable), make sure host bank is enabled (see xbutil configure --host-mem)
-          # iree-amd-aie/runtime/src/iree-amd-aie/driver/xrt/direct_allocator.cc:179: RESOURCE_EXHAUSTED; could not allocate
-          # memory for buffer; while invoking C++ function matmul_test.generate_random_matrix; while calling import;
-          #
-          # might be observed when too much memory is allocated. This
-          # error was seen when running a bf16->f32 matmul with m=n=k=2304.
-          #
-          # This line was suggested at https://github.com/Xilinx/mlir-air/issues/566
-          #
-          # Note that this is only half of the fix. It is also necessary that
-          # the machine that CI is running on has permission to run this line.
-          #
-          # This permission can be adding by adding the line
-          # ```
-          # %github ALL=(ALL) NOPASSWD: /usr/bin/prlimit *
-          # ```
-          #
-          # to the file /etc/sudoers.d/github, which can be done by running
-          # ```
-          # sudo visudo -f /etc/sudoers.d/github
-          # ```
-          # on the github CI machine.
+          # https://stackoverflow.com/a/17567422
+          # shim_xdna::bo::map_drm_bo does an mmap with MAP_LOCKED
+          # which can fail if limit is to low
           sudo prlimit -lunlimited --pid $$
           source .venv/bin/activate
-          source /opt/xilinx/xrt/setup.sh
           bash build_tools/ci/run_matmul_test.sh \
             test_matmuls \
             iree-install \
             $PWD/llvm-aie \
-            /opt/xilinx/xrt \
             /opt/Xilinx/Vitis/2024.2
 
       - name: Python tests
         run: |
           source .venv/bin/activate
-          source /opt/xilinx/xrt/setup.sh
           pytest -v tests \
             --capture=tee-sys \
             --iree-install-dir=$PWD/iree-install \
             --peano-install-dir=$PWD/llvm-aie
 
       - name: XRT-LITE tests
         run: |
-          source /opt/xilinx/xrt/setup.sh
           DEVICE_TEST_DIR="$PWD/iree-install/device_tests"
           for t in $(ls $DEVICE_TEST_DIR); do
             $DEVICE_TEST_DIR/$t

diff --git a/build_tools/ci/run_matmul_test.sh b/build_tools/ci/run_matmul_test.sh
@@ -102,21 +102,11 @@ if [ ! -d "${PEANO}" ]; then
   exit 1
 fi
 
-# Parameter 4) <xrt-dir>
+# Parameter 4) <vitis-install-dir>
 if [ -z "${4-}" ]; then
-  XRT_DIR=/opt/xilinx/xrt
-else
-  XRT_DIR=`realpath "$4"`
-fi
-if [ -d "$XRT_DIR" ]; then
-  source $XRT_DIR/setup.sh
-fi
-
-# Parameter 5) <vitis-install-dir>
-if [ -z "${5-}" ]; then
   VITIS=/opt/Xilinx/Vitis/2024.2
 else
-  VITIS=`realpath "$5"`
+  VITIS=`realpath "$4"`
 fi
 
 THIS_DIR="$(cd $(dirname $0) && pwd)"
@@ -139,9 +129,6 @@ fi
 
 GITHUB_ACTIONS="${GITHUB_ACTIONS:-false}"
 
-# Circumvent xclbin security (no longer needed as of April 2024 XDNA driver)
-export XRT_HACK_UNSECURE_LOADING_XCLBIN=1
-
 cd ${OUTPUT_DIR}
 
 export MATMUL_TESTS_RUN=0
@@ -405,6 +392,7 @@ function run_matmul_test() {
                       --iree-amd-aie-enable-chess=${use_chess} \
                       --iree-amdaie-enable-packet-flow=${enable_packet_flow} \
                       --iree-hal-dump-executable-files-to=$PWD \
+                      --iree-amdaie-device-hal=xrt-lite \
                       --iree-hal-memoization=false \
                       --iree-hal-indirect-command-buffers=false \
                       --mlir-elide-resource-strings-if-larger=10 \

diff --git a/compiler/plugins/target/AMD-AIE/iree-amd-aie/CMakeLists.txt b/compiler/plugins/target/AMD-AIE/iree-amd-aie/CMakeLists.txt
@@ -26,6 +26,7 @@ iree_cc_library(
     iree::base::core_headers
     iree::base::internal::flatcc::building
     iree-amd-aie::schemas::xrt_executable_def_c_fbs
+    iree-amd-aie::schemas::pdi_executable_def_c_fbs
   PUBLIC
 )
 

diff --git a/compiler/plugins/target/AMD-AIE/iree-amd-aie/PluginRegistration.cpp b/compiler/plugins/target/AMD-AIE/iree-amd-aie/PluginRegistration.cpp
@@ -46,9 +46,16 @@ struct AMDAIESession
   }
 
   void populateHALTargetDevices(IREE::HAL::TargetDeviceList &targets) override {
+    // #hal.device.target<"xrt", ...
+    targets.add("xrt", [=] {
+      options.deviceHal = AMDAIE::AMDAIEOptions::DeviceHAL::XRT;
+      return AMDAIE::createTarget(options);
+    });
     // #hal.device.target<"xrt-lite", ...
-    // #hal.executable.target<"amd-aie", ...
-    targets.add("xrt-lite", [=]() { return AMDAIE::createTarget(options); });
+    targets.add("xrt-lite", [=] {
+      options.deviceHal = AMDAIE::AMDAIEOptions::DeviceHAL::XRT_LITE;
+      return AMDAIE::createTarget(options);
+    });
   }
 
   void populateHALTargetBackends(