[WIP] use peano for ukernels

build_tools/ci/run_matmul_test.sh

@@ -182,7 +182,7 @@ function run_matmul_test() {
 
   local amd_aie_install_path="${IREE_INSTALL_DIR}"
 
-  local vitis_path="${VITIS}"
+  local vitis_path=""
 
   local use_chess="false"
 
@@ -540,16 +540,15 @@ run_matmul_test \
 # MLIR-AIR Matmul tests
 ###################################################################
 
-if [ -d "$VITIS" ]; then
-  run_matmul_test \
-      --name_prefix "ukern" \
-      --lower_to_aie_pipeline "air" \
-      --tile_pipeline "pad-pack" \
-      --lhs_rhs_type "bf16" \
-      --acc_type "f32" \
-      --m "256"  --k "256" --n "256" \
-      --use_ukernel "1"
-fi
+run_matmul_test \
+    --name_prefix "ukern" \
+    --lower_to_aie_pipeline "air" \
+    --tile_pipeline "pad-pack" \
+    --lhs_rhs_type "bf16" \
+    --acc_type "f32" \
+    --m "256"  --k "256" --n "256" \
+    --vitis_path  "${VITIS}" \
+    --use_ukernel "1"
 
 # Example of a run with a group of 2+ matmuls. Currently this test is passed
 # the flag '--num_repeat_runs 0" as there is currently an issue with the runtime if
@@ -720,6 +719,7 @@ if [ -d "$VITIS" ]; then
       --lhs_rhs_type "bf16" \
       --acc_type "f32" \
       --num_repeat_runs "2" \
+      --vitis_path  "${VITIS}" \
       --use_ukernel "1"
 
   run_matmul_test_on_shapes ${bf16_ukernel_shapes_medium[@]} \
@@ -729,6 +729,7 @@ if [ -d "$VITIS" ]; then
       --lhs_rhs_type "bf16" \
       --acc_type "f32" \
       --num_repeat_runs "2" \
+      --vitis_path  "${VITIS}" \
       --use_ukernel "1"
 fi
 
@@ -746,6 +747,7 @@ if [ -d "$VITIS" ]; then
       --n "32" \
       --k "32" \
       --use_chess "1" \
+      --vitis_path  "${VITIS}" \
       --num_repeat_runs "10"
 
   run_matmul_test \
@@ -757,6 +759,7 @@ if [ -d "$VITIS" ]; then
       --k "64" \
       --use_chess "1" \
       --num_repeat_runs "10" \
+      --vitis_path  "${VITIS}" \
       --use_ukernel "1"
 
   run_matmul_test \
@@ -769,6 +772,7 @@ if [ -d "$VITIS" ]; then
       --n "32" \
       --k "32" \
       --use_chess "1" \
+      --vitis_path  "${VITIS}" \
       --num_repeat_runs "10"
 
 fi

build_tools/download_peano.sh

@@ -1,5 +1,5 @@
 #!/bin/bash
 
-RELEASE=19.0.0.2024082221+90abe71b
+RELEASE=19.0.0.2024083101+42158757
 pip download llvm_aie==$RELEASE -f https://github.com/Xilinx/llvm-aie/releases/expanded_assets/nightly
 unzip llvm_aie*whl

compiler/plugins/target/AMD-AIE/iree-amd-aie/Target/CMakeLists.txt

@@ -10,10 +10,11 @@ iree_cc_library(
   NAME
     AIETargets
   SRCS
-    "AMDAIETargetBCF.cpp"
-    "AMDAIETargetCDODirect.cpp"
-    "AMDAIETargetLdScript.cpp"
-    "XCLBinGen.cpp"
+    AMDAIETargetBCF.cpp
+    AMDAIETargetCDODirect.cpp
+    AMDAIETargetLdScript.cpp
+    PeanoDriver.cpp
+    XCLBinGen.cpp
   DEPS
     iree-amd-aie::aie_runtime::iree_aie_runtime_static
     iree::target::amd-aie::Transforms
@@ -28,9 +29,9 @@ iree_cc_library(
   NAME
     Target
   HDRS
-    "AIETarget.h"
+    AIETarget.h
   SRCS
-    "AIETarget.cpp"
+    AIETarget.cpp
   DEPS
     ::AIETargets
     iree-amd-aie::schemas::xrt_executable_def_c_fbs

compiler/plugins/target/AMD-AIE/iree-amd-aie/Target/PeanoDriver.cpp

@@ -0,0 +1,106 @@
+// Copyright 2024 The IREE Authors
+//
+// Licensed under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+
+#include "PeanoDriver.h"
+
+#include <filesystem>
+#include <string>
+#include <vector>
+
+#include "llvm/Support/Error.h"
+
+using Path = std::filesystem::path;
+
+void addExternCSystemInclude(std::vector<std::string> &CC1Args,
+                             const std::string &Path) {
+  CC1Args.push_back("-internal-externc-isystem");
+  CC1Args.push_back(Path);
+}
+
+void addSystemInclude(std::vector<std::string> &CC1Args,
+                      const std::string &Path) {
+  CC1Args.push_back("-internal-isystem");
+  CC1Args.push_back(Path);
+}
+
+void AddClangSystemIncludeArgs(std::vector<std::string> &CC1Args,
+                               const Path &peanoDir, const std::string &target,
+                               bool novitisheaders, bool nostdlibinc) {
+  // Always include our instrinsics, for compatibility with existing toolchain.
+  if (!novitisheaders) {
+    std::string path;
+    if (target.rfind("aie2-", 0) == 0) {
+      path = peanoDir / "lib" / "clang" / "19" / "include" / "aiev2intrin.h";
+    } else {
+      llvm::report_fatal_error(("unsupported target: " + target).c_str());
+    }
+    CC1Args.push_back("-include");
+    CC1Args.push_back(path);
+  }
+
+  CC1Args.push_back("-D__AIENGINE__");
+  if (target.rfind("aie2-", 0) == 0) {
+    CC1Args.push_back("-D__AIEARCH__=20");
+  } else {
+    llvm::report_fatal_error(("unsupported target: " + target).c_str());
+  }
+
+  // Don't pull in system headers from /usr/include or /usr/local/include.
+  // All the basic headers that we need come from the compiler.
+  CC1Args.push_back("-nostdsysteminc");
+
+  if (nostdlibinc) return;
+  addExternCSystemInclude(CC1Args, peanoDir / "include" / target);
+}
+
+void addLibCxxIncludePaths(std::vector<std::string> &CC1Args,
+                           const Path &peanoDir, const std::string &target,
+                           bool nostdinc, bool nostdlibinc, bool nostdincxx) {
+  if (nostdinc || nostdlibinc || nostdincxx) return;
+  addSystemInclude(CC1Args, peanoDir / "include" / target / "c++" / "v1");
+  // Second add the generic one.
+  addSystemInclude(CC1Args, peanoDir / "include" / "c++" / "v1");
+}
+
+void addOptTargetOptions(std::vector<std::string> &CC1Args) {
+  // For now, we disable the auto-vectorizers by default, as the backend cannot
+  // handle many vector types. For experimentation the vectorizers can still be
+  // enabled explicitly by the user
+  CC1Args.push_back("-vectorize-loops=false");
+  CC1Args.push_back("-vectorize-slp=false");
+  // An if-then-else cascade requires at least 5 delay slots for evaluating the
+  // condition and 5 delay slots for one of the branches, thus speculating 10
+  // instructions should be fine
+  CC1Args.push_back("--two-entry-phi-node-folding-threshold=10");
+  // Make sure to perform most optimizations before mandatory inlinings,
+  // otherwise noalias attributes can get lost and hurt AA results.
+  CC1Args.push_back("-mandatory-inlining-before-opt=false");
+  // Perform complete AA analysis on phi nodes.
+  CC1Args.push_back("-basic-aa-full-phi-analysis=true");
+  // Extend the max limit of the search depth in BasicAA
+  CC1Args.push_back("-basic-aa-max-lookup-search-depth=10");
+}
+
+void addClangTargetOptions(std::vector<std::string> &CC1Args,
+                           const std::string &target) {
+  CC1Args.push_back("-triple");
+  CC1Args.push_back(target);
+  CC1Args.push_back("-fno-use-init-array");
+  // Pass -fno-threadsafe-statics to prevent dependence on lock acquire/release
+  // handling for static local variables.
+  CC1Args.push_back("-fno-threadsafe-statics");
+  std::vector<std::string> peanoArgs;
+  addOptTargetOptions(peanoArgs);
+  CC1Args.reserve(CC1Args.size() + 2 * peanoArgs.size());
+  for (const std::string &item : peanoArgs) {
+    CC1Args.emplace_back("-mllvm");
+    CC1Args.emplace_back(item);
+  }
+}
+
+// Avoid using newer dwarf versions, as the simulator doesn't understand newer
+// dwarf.
+unsigned getMaxDwarfVersion() { return 4; }

compiler/plugins/target/AMD-AIE/iree-amd-aie/Target/PeanoDriver.h

@@ -0,0 +1,28 @@
+// Copyright 2024 The IREE Authors
+//
+// Licensed under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+
+#include <filesystem>
+#include <string>
+#include <vector>
+
+#include "llvm/Support/Error.h"
+
+void AddClangSystemIncludeArgs(std::vector<std::string> &CC1Args,
+                               const std::filesystem::path &peanoDir,
+                               const std::string &target,
+                               bool novitisheaders = false,
+                               bool nostdlibinc = false);
+
+void addLibCxxIncludePaths(std::vector<std::string> &CC1Args,
+                           const std::filesystem::path &peanoDir,
+                           const std::string &target, bool nostdinc = false,
+                           bool nostdlibinc = false, bool nostdincxx = false);
+
+void addOptTargetOptions(std::vector<std::string> &CC1Args);
+void addClangTargetOptions(std::vector<std::string> &CC1Args,
+                           const std::string &target);
+
+unsigned getMaxDwarfVersion();