hack tests for xrt-lite

.github/workflows/ci-windows.yml

@@ -102,6 +102,8 @@ jobs:
           # Remove-Item -Path "$pwd\llvm-build" -Force
           $env:llvm_install_dir = "$pwd\llvm-install"
           echo $env:llvm_install_dir
+          .\build_tools\download_peano.ps1
+          $env:peano_install_dir = "$pwd\llvm-aie"
           .\build_tools.\build_test_cpp.ps1
 
       - name: Create artifacts

build_tools/build_test_cpp.sh

@@ -155,5 +155,7 @@ if [ -d "$llvm_install_dir" ]; then
 fi
 
 cp "$build_dir/tools/testing/e2e/iree-e2e-matmul-test" "$install_dir/bin"
-mkdir -p "$install_dir/device_tests"
-cp "$build_dir"/runtime/plugins/AMD-AIE/iree-amd-aie/driver/xrt-lite/cts/*test "$install_dir/device_tests"
+if [[ "$OSTYPE" == "linux"* ]]; then
+  mkdir -p "$install_dir/device_tests"
+  cp "$build_dir"/runtime/plugins/AMD-AIE/iree-amd-aie/driver/xrt-lite/cts/*test "$install_dir/device_tests"
+fi

build_tools/ci/cpu_comparison/run.py

@@ -191,7 +191,7 @@ def generate_aie_output(config, aie_vmfb, input_args, function_name, name, outpu
         config.iree_run_exe,
         f"--module={aie_vmfb}",
         *input_args,
-        "--device=xrt",
+        "--device=xrt-lite",
         f"--output=@{aie_bin}",
     ]
     if function_name:

build_tools/ci/run_matmul_test.sh

@@ -176,7 +176,7 @@ function run_matmul_test() {
 
   local target_device="npu1_4col"
 
-  local device="xrt"
+  local device="xrt-lite"
 
   local peano_install_path="${PEANO}"
 
@@ -530,7 +530,7 @@ run_matmul_test \
     --acc_type "f32" \
     --target_backend "amd-aie" \
     --target_device "npu1_4col" \
-    --device "xrt" \
+    --device "xrt-lite" \
     --peano_install_path "${PEANO}" \
     --amd_aie_install_path "${IREE_INSTALL_DIR}" \
     --vitis_path  "${VITIS}" \

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

@@ -46,9 +46,9 @@ struct AMDAIESession
   }
 
   void populateHALTargetDevices(IREE::HAL::TargetDeviceList &targets) override {
-    // #hal.device.target<"xrt", ...
+    // #hal.device.target<"xrt-lite", ...
     // #hal.executable.target<"amd-aie", ...
-    targets.add("xrt", [=]() { return AMDAIE::createTarget(options); });
+    targets.add("xrt-lite", [=]() { return AMDAIE::createTarget(options); });
   }
 
   void populateHALTargetBackends(

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

@@ -101,7 +101,7 @@ class AIETargetDevice final : public IREE::HAL::TargetDevice {
     targetRegistry.getTargetBackend("amd-aie")->getDefaultExecutableTargets(
         context, "amd-aie", configAttr, executableTargetAttrs);
 
-    return IREE::HAL::DeviceTargetAttr::get(context, b.getStringAttr("xrt"),
+    return IREE::HAL::DeviceTargetAttr::get(context, b.getStringAttr("xrt-lite"),
                                             configAttr, executableTargetAttrs);
   }
 
@@ -113,7 +113,7 @@ class AIETargetBackend final : public IREE::HAL::TargetBackend {
  public:
   explicit AIETargetBackend(const AMDAIEOptions &options) : options(options) {}
 
-  std::string getLegacyDefaultDeviceID() const override { return "xrt"; }
+  std::string getLegacyDefaultDeviceID() const override { return "xrt-lite"; }
 
   void getDefaultExecutableTargets(
       MLIRContext *context, StringRef deviceID, DictionaryAttr deviceConfigAttr,

runtime/src/iree-amd-aie/driver/xrt-lite/allocator.cc

@@ -24,11 +24,11 @@ extern const iree_hal_allocator_vtable_t iree_hal_xrt_lite_allocator_vtable;
 struct iree_hal_xrt_lite_allocator {
   iree_hal_resource_t resource;
   iree_allocator_t host_allocator;
-  std::shared_ptr<shim_xdna::device> shim_device;
+  shim_xdna::device* shim_device;
   IREE_STATISTICS(iree_hal_allocator_statistics_t statistics;)
 
   iree_hal_xrt_lite_allocator(iree_allocator_t host_allocator,
-                              std::shared_ptr<shim_xdna::device> shim_device)
+                              shim_xdna::device* shim_device)
       : host_allocator(host_allocator), shim_device(shim_device) {
     IREE_TRACE_ZONE_BEGIN(z0);
     iree_hal_resource_initialize(&iree_hal_xrt_lite_allocator_vtable,
@@ -266,7 +266,7 @@ static iree_hal_xrt_lite_allocator* iree_hal_xrt_lite_allocator_cast(
 }
 
 iree_status_t iree_hal_xrt_lite_allocator_create(
-    iree_allocator_t host_allocator, std::shared_ptr<shim_xdna::device> device,
+    iree_allocator_t host_allocator, shim_xdna::device* device,
     iree_hal_allocator_t** out_allocator) {
   IREE_ASSERT_ARGUMENT(out_allocator);
   IREE_TRACE_ZONE_BEGIN(z0);
@@ -296,11 +296,7 @@ static void iree_hal_xrt_lite_allocator_destroy(
       iree_hal_xrt_lite_allocator_cast(base_allocator);
   IREE_TRACE_ZONE_BEGIN(z0);
 
-  // TODO(max): shouldn't this be happening automatically via the refcounting
-  // (or just the dtor of device?)
-  allocator->shim_device.reset();
   iree_hal_resource_release(&allocator->resource);
-  // something's not happening here?
   iree_allocator_free(allocator->host_allocator, allocator);
 
   IREE_TRACE_ZONE_END(z0);

runtime/src/iree-amd-aie/driver/xrt-lite/allocator.h

@@ -13,7 +13,7 @@
 
 // Creates a buffer allocator used for persistent allocations.
 iree_status_t iree_hal_xrt_lite_allocator_create(
-    iree_allocator_t host_allocator, std::shared_ptr<shim_xdna::device> device,
+    iree_allocator_t host_allocator, shim_xdna::device* device,
     iree_hal_allocator_t** out_allocator);
 
 #endif  // IREE_HAL_DRIVERS_XRT_LITE_ALLOCATOR_H_

runtime/src/iree-amd-aie/driver/xrt-lite/device.cc

@@ -26,7 +26,7 @@ struct iree_hal_xrt_lite_device {
   // Block pool used for command buffers with a larger block size (as command
   // buffers can contain inlined data uploads).
   iree_arena_block_pool_t block_pool;
-  std::shared_ptr<shim_xdna::device> shim_device;
+  shim_xdna::device* shim_device;
 
   iree_status_t create_executable_cache(
       iree_string_view_t identifier, iree_loop_t loop,
@@ -162,7 +162,7 @@ iree_status_t iree_hal_xrt_lite_device_create(
       identifier, &device->identifier,
       reinterpret_cast<char*>(device) + total_size - identifier.size);
   device->host_allocator = host_allocator;
-  device->shim_device = std::make_shared<shim_xdna::device>();
+  device->shim_device = new shim_xdna::device;
 
   // TODO(null): pass device handles and pool configuration to the allocator.
   // Some implementations may share allocators across multiple devices created
@@ -205,7 +205,7 @@ static void iree_hal_xrt_lite_device_destroy(iree_hal_device_t* base_device) {
   // and joined first.
 
   iree_hal_allocator_release(device->device_allocator);
-  device->shim_device.reset();
+  delete device->shim_device;
   iree_allocator_free(host_allocator, device);
 
   IREE_TRACE_ZONE_END(z0);

runtime/src/iree-amd-aie/driver/xrt-lite/direct_command_buffer.cc

@@ -9,6 +9,7 @@
 #include "iree-amd-aie/driver/xrt-lite/buffer.h"
 #include "iree-amd-aie/driver/xrt-lite/executable.h"
 #include "iree-amd-aie/driver/xrt-lite/shim/linux/kmq/hwq.h"
+#include "iree-amd-aie/driver/xrt-lite/shim/linux/kmq/kernel.h"
 #include "iree/hal/utils/resource_set.h"
 
 // The max number of bindings per descriptor set allowed in the XRT HAL
@@ -29,7 +30,7 @@ struct iree_hal_xrt_lite_direct_command_buffer {
   // Staging arena used for host->device transfers.
   iree_arena_allocator_t arena;
 
-  std::shared_ptr<shim_xdna::device> shim_device;
+  shim_xdna::device* shim_device;
 
   struct {
     shim_xdna::bo* bindings[IREE_HAL_XRT_LITE_MAX_DESCRIPTOR_SET_BINDING_COUNT];
@@ -52,12 +53,12 @@ iree_hal_xrt_lite_direct_command_buffer_cast(
     iree_hal_command_buffer_t* base_value) {
   IREE_HAL_ASSERT_TYPE(base_value,
                        &iree_hal_xrt_lite_direct_command_buffer_vtable);
-  return (iree_hal_xrt_lite_direct_command_buffer*)base_value;
+  return reinterpret_cast<iree_hal_xrt_lite_direct_command_buffer*>(base_value);
 }
 
 iree_status_t iree_hal_xrt_lite_direct_command_buffer_create(
-    std::shared_ptr<shim_xdna::device> shim_device,
-    iree_hal_allocator_t* device_allocator, iree_hal_command_buffer_mode_t mode,
+    shim_xdna::device* shim_device, iree_hal_allocator_t* device_allocator,
+    iree_hal_command_buffer_mode_t mode,
     iree_hal_command_category_t command_categories,
     iree_host_size_t binding_capacity, iree_arena_block_pool_t* block_pool,
     iree_allocator_t host_allocator,
@@ -106,7 +107,7 @@ static void iree_hal_xrt_lite_direct_command_buffer_destroy(
       iree_hal_xrt_lite_direct_command_buffer_cast(base_command_buffer);
   iree_allocator_t host_allocator = command_buffer->host_allocator;
   IREE_TRACE_ZONE_BEGIN(z0);
-  command_buffer->shim_device.reset();
+
   iree_hal_resource_set_free(command_buffer->resource_set);
   iree_arena_deinitialize(&command_buffer->arena);
   iree_allocator_free(host_allocator, command_buffer);
@@ -135,18 +136,6 @@ static iree_status_t iree_hal_xrt_lite_direct_command_buffer_end(
   return iree_ok_status();
 }
 
-static void iree_hal_xrt_lite_direct_command_buffer_begin_debug_group(
-    iree_hal_command_buffer_t* base_command_buffer, iree_string_view_t label,
-    iree_hal_label_color_t label_color,
-    const iree_hal_label_location_t* location) {
-  (void)iree_status_from_code(IREE_STATUS_UNIMPLEMENTED);
-}
-
-static void iree_hal_xrt_lite_direct_command_buffer_end_debug_group(
-    iree_hal_command_buffer_t* base_command_buffer) {
-  (void)iree_status_from_code(IREE_STATUS_UNIMPLEMENTED);
-}
-
 static iree_status_t iree_hal_xrt_lite_direct_command_buffer_execution_barrier(
     iree_hal_command_buffer_t* base_command_buffer,
     iree_hal_execution_stage_t source_stage_mask,
@@ -300,8 +289,7 @@ static iree_status_t iree_hal_xrt_lite_direct_command_buffer_dispatch(
   cu_name += ":IREE";
   shim_xdna::cuidx_t cu_idx = kernel_params.context->open_cu_context(cu_name);
 
-  shim_xdna::exec_buf ebuf(command_buffer->shim_device->get_pdev(),
-                           ERT_START_CU);
+  shim_xdna::kernel ebuf(command_buffer->shim_device->get_pdev(), ERT_START_CU);
   ebuf.set_cu_idx(cu_idx);
   unsigned int opcode = 3;
   ebuf.add_arg_64(opcode);
@@ -311,13 +299,9 @@ static iree_status_t iree_hal_xrt_lite_direct_command_buffer_dispatch(
     shim_xdna::bo* bo = iree_hal_xrt_lite_buffer_handle(
         iree_hal_buffer_allocated_buffer(bindings.values[j].buffer));
     ebuf.add_arg_bo(*bo);
-  }
-
-  for (iree_host_size_t j = 0; j < bindings.count; ++j) {
-    shim_xdna::bo* bo = iree_hal_xrt_lite_buffer_handle(
-        iree_hal_buffer_allocated_buffer(bindings.values[j].buffer));
     bo->sync(shim_xdna::direction::host2device);
   }
+
   shim_xdna::hw_q* hwq = kernel_params.context->get_hw_queue();
   hwq->issue_command(ebuf.get_exec_buf_bo());
   hwq->wait_command(ebuf.get_exec_buf_bo(), 0);
@@ -328,11 +312,6 @@ static iree_status_t iree_hal_xrt_lite_direct_command_buffer_dispatch(
     bo->sync(shim_xdna::direction::device2host);
   }
 
-  for (iree_host_size_t j = 0; j < bindings.count; ++j) {
-    shim_xdna::bo* bo = iree_hal_xrt_lite_buffer_handle(
-        iree_hal_buffer_allocated_buffer(bindings.values[j].buffer));
-  }
-
   IREE_TRACE_ZONE_END(z0);
 
   return iree_ok_status();

runtime/src/iree-amd-aie/driver/xrt-lite/direct_command_buffer.h

@@ -18,7 +18,7 @@ extern "C" {
 // |out_command_buffer| must be released by the caller (see
 // iree_hal_command_buffer_release).
 iree_status_t iree_hal_xrt_lite_direct_command_buffer_create(
-    std::shared_ptr<shim_xdna::device> shim_device,
+    shim_xdna::device* shim_device,
     iree_hal_allocator_t* device_allocator, iree_hal_command_buffer_mode_t mode,
     iree_hal_command_category_t command_categories,
     iree_host_size_t binding_capacity, iree_arena_block_pool_t* block_pool,

runtime/src/iree-amd-aie/driver/xrt-lite/executable.cc

@@ -100,7 +100,7 @@ iree_amd_aie_hal_xrt_lite_native_executable_flatbuffer_verify(
 }
 
 iree_status_t iree_hal_xrt_lite_native_executable_create(
-    std::shared_ptr<shim_xdna::device> shim_device,
+    shim_xdna::device* shim_device,
     const iree_hal_executable_params_t* executable_params,
     iree_allocator_t host_allocator, iree_hal_executable_t** out_executable) {
   IREE_ASSERT_ARGUMENT(executable_params);
@@ -176,8 +176,8 @@ iree_status_t iree_hal_xrt_lite_native_executable_create(
     std::vector<char> xclbinVector(
         xclbin_fb, xclbin_fb + flatbuffers_string_len(xclbin_fb));
     params->xclbinVector = xclbinVector;
-//    xrt::xclbin xclbin = xrt::xclbin(xclbinVector);
-//    params->context = shim_device->create_hw_context(xclbin);
+    //    xrt::xclbin xclbin = xrt::xclbin(xclbinVector);
+    //    params->context = shim_device->create_hw_context(xclbin);
 
     uint32_t asm_instr_index =
         flatbuffers_uint32_vec_at(asm_instr_indices_vec, entry_ordinal);
@@ -186,13 +186,13 @@ iree_status_t iree_hal_xrt_lite_native_executable_create(
     params->asm_inst =
         iree_amd_aie_hal_xrt_AsmInstDef_asm_inst_get(asminst_def);
 
-//    uint32_t num_instr = flatbuffers_uint32_vec_len(asm_inst);
-//    size_t ctrl_code_size = num_instr * sizeof(uint32_t);
-//    params->bo_ctrl_code =
-//        shim_device->alloc_bo(ctrl_code_size, XCL_BO_FLAGS_CACHEABLE);
-//    uint32_t* instr_buffer =
-//        static_cast<uint32_t*>(params->bo_ctrl_code->map());
-//    memcpy(instr_buffer, asm_inst, ctrl_code_size);
+    //    uint32_t num_instr = flatbuffers_uint32_vec_len(asm_inst);
+    //    size_t ctrl_code_size = num_instr * sizeof(uint32_t);
+    //    params->bo_ctrl_code =
+    //        shim_device->alloc_bo(ctrl_code_size, XCL_BO_FLAGS_CACHEABLE);
+    //    uint32_t* instr_buffer =
+    //        static_cast<uint32_t*>(params->bo_ctrl_code->map());
+    //    memcpy(instr_buffer, asm_inst, ctrl_code_size);
 
     // Stash the entry point name in the string table for use when tracing.
     IREE_TRACE({

runtime/src/iree-amd-aie/driver/xrt-lite/executable.h

@@ -36,7 +36,7 @@ struct iree_hal_xrt_lite_kernel_params_t {
 // |out_executable| must be released by the caller (see
 // iree_hal_executable_release).
 iree_status_t iree_hal_xrt_lite_native_executable_create(
-    std::shared_ptr<shim_xdna::device> shim_device,
+    shim_xdna::device* shim_device,
     const iree_hal_executable_params_t* executable_params,
     iree_allocator_t host_allocator, iree_hal_executable_t** out_executable);
 

runtime/src/iree-amd-aie/driver/xrt-lite/nop_executable_cache.cc

@@ -17,7 +17,7 @@ struct iree_hal_xrt_lite_nop_executable_cache_t {
   // Abstract resource used for injecting reference counting and vtable; must be
   // at offset 0.
   iree_hal_resource_t resource;
-  std::shared_ptr<shim_xdna::device> shim_device;
+  shim_xdna::device* shim_device;
   iree_allocator_t host_allocator;
 };
 
@@ -35,8 +35,8 @@ iree_hal_xrt_lite_nop_executable_cache_cast(
 }
 
 iree_status_t iree_hal_xrt_lite_nop_executable_cache_create(
-    std::shared_ptr<shim_xdna::device> shim_device,
-    iree_string_view_t identifier, iree_allocator_t host_allocator,
+    shim_xdna::device* shim_device, iree_string_view_t identifier,
+    iree_allocator_t host_allocator,
     iree_hal_executable_cache_t** out_executable_cache) {
   IREE_ASSERT_ARGUMENT(out_executable_cache);
   *out_executable_cache = nullptr;
@@ -62,7 +62,6 @@ static void iree_hal_xrt_lite_nop_executable_cache_destroy(
       iree_hal_xrt_lite_nop_executable_cache_cast(base_executable_cache);
   IREE_TRACE_ZONE_BEGIN(z0);
 
-  executable_cache->shim_device.reset();
   iree_allocator_free(executable_cache->host_allocator, executable_cache);
 
   IREE_TRACE_ZONE_END(z0);

runtime/src/iree-amd-aie/driver/xrt-lite/nop_executable_cache.h

@@ -22,7 +22,7 @@ extern "C" {
 // |out_executable_cache| must be released by the caller (see
 // iree_hal_executable_cache_release).
 iree_status_t iree_hal_xrt_lite_nop_executable_cache_create(
-    std::shared_ptr<shim_xdna::device> shim_device,
+    shim_xdna::device* shim_device,
     iree_string_view_t identifier, iree_allocator_t host_allocator,
     iree_hal_executable_cache_t** out_executable_cache);
 

runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/CMakeLists.txt

@@ -21,6 +21,8 @@ iree_cc_library(
     hwctx.h
     hwq.cpp
     hwq.h
+    kernel.cpp
+    kernel.h
     shim_debug.cpp
     shim_debug.h
   DEPS