Bump IREE to iree-org/iree@c37c680ae6e71

Abhishek-Varma · Abhishek-Varma · commit bb545533ed2e · 2025-09-02T07:04:53.000Z
Signed-off-by: Abhishek Varma &lt;abhvarma@amd.com&gt;
diff --git a/build_tools/ci/cpu_comparison/run.py b/build_tools/ci/cpu_comparison/run.py
@@ -2493,29 +2493,27 @@ def __init__(self):
         # Note: The error tolerance for npu4 is higher than that for npu1_4col.
         # npu1_4col uses a lookup table to compute exponentials,
         # whereas npu4 uses a native exp2 instruction, which is less accurate.
-        # TODO: Disable till iree-org/iree issue https://github.com/iree-org/iree/issues/21633
-        #       gets fixed.
-        # for target, rtol in [["npu1_4col", 4e-2], ["npu4", 8e-2]]:
-        #     for run_benchmark in [False, True]:
-        #         self.register(
-        #             Softmax(
-        #                 8192,
-        #                 1024,
-        #                 "bf16",
-        #                 test_params=TestParams(
-        #                     run_on_target=target,
-        #                     name_suffix=target,
-        #                     use_chess=True,
-        #                     use_chess_for_ukernel=True,
-        #                     use_ukernel=True,
-        #                     tile_pipeline="general-copy",
-        #                     run_benchmark=run_benchmark,
-        #                     n_repeats=2,
-        #                     n_kernel_runs=100,
-        #                     rtol=rtol,
-        #                 ),
-        #             )
-        #         )
+        for target, rtol in [["npu1_4col", 4e-2], ["npu4", 8e-2]]:
+            for run_benchmark in [False, True]:
+                self.register(
+                    Softmax(
+                        8192,
+                        1024,
+                        "bf16",
+                        test_params=TestParams(
+                            run_on_target=target,
+                            name_suffix=target,
+                            use_chess=True,
+                            use_chess_for_ukernel=True,
+                            use_ukernel=True,
+                            tile_pipeline="general-copy",
+                            run_benchmark=run_benchmark,
+                            n_repeats=2,
+                            n_kernel_runs=100,
+                            rtol=rtol,
+                        ),
+                    )
+                )
 
         # Reduction op tests:
         self.register(
diff --git a/runtime/src/iree-amd-aie/driver/xrt-lite/CMakeLists.txt b/runtime/src/iree-amd-aie/driver/xrt-lite/CMakeLists.txt
@@ -40,6 +40,7 @@ iree_cc_library(
     iree::base
     iree::base::core_headers
     iree::hal::utils::deferred_command_buffer
+    iree::hal::utils::queue_emulation
     iree::hal::utils::semaphore_base
     iree::base::internal::flatcc::parsing
     iree-amd-aie::schemas::pdi_executable_def_c_fbs
diff --git a/runtime/src/iree-amd-aie/driver/xrt-lite/device.cc b/runtime/src/iree-amd-aie/driver/xrt-lite/device.cc
@@ -15,6 +15,7 @@
 #include "iree-amd-aie/driver/xrt-lite/util.h"
 #include "iree/hal/utils/deferred_command_buffer.h"
 #include "iree/hal/utils/deferred_work_queue.h"
+#include "iree/hal/utils/queue_emulation.h"
 
 #define ARENA_BLOCK_SIZE (32 * 1024)
 
@@ -98,8 +99,9 @@ static iree_status_t iree_hal_xrt_lite_device_create_command_buffer(
 }
 
 static iree_status_t iree_hal_xrt_lite_device_create_semaphore(
-    iree_hal_device_t* base_device, uint64_t initial_value,
-    iree_hal_semaphore_flags_t flags, iree_hal_semaphore_t** out_semaphore) {
+    iree_hal_device_t* base_device, iree_hal_queue_affinity_t queue_affinity,
+    uint64_t initial_value, iree_hal_semaphore_flags_t flags,
+    iree_hal_semaphore_t** out_semaphore) {
   IREE_TRACE_ZONE_BEGIN(z0);
 
   iree_hal_xrt_lite_device* device = IREE_HAL_XRT_LITE_CHECKED_VTABLE_CAST(
@@ -190,8 +192,9 @@ static iree_status_t iree_hal_xrt_lite_device_queue_alloca(
   iree_hal_xrt_lite_device* device = IREE_HAL_XRT_LITE_CHECKED_VTABLE_CAST(
       base_device, iree_hal_xrt_lite_device_vtable, iree_hal_xrt_lite_device);
   IREE_RETURN_AND_END_ZONE_IF_ERROR(
-      z0, iree_hal_semaphore_list_wait(wait_semaphore_list,
-                                       iree_infinite_timeout()));
+      z0,
+      iree_hal_semaphore_list_wait(wait_semaphore_list, iree_infinite_timeout(),
+                                   IREE_HAL_WAIT_FLAG_DEFAULT));
   IREE_RETURN_AND_END_ZONE_IF_ERROR(
       z0, iree_hal_allocator_allocate_buffer(device->device_allocator, params,
                                              allocation_size, out_buffer));
@@ -207,8 +210,9 @@ static iree_status_t iree_hal_xrt_lite_device_queue_dealloca(
     const iree_hal_semaphore_list_t wait_semaphore_list,
     const iree_hal_semaphore_list_t signal_semaphore_list,
     iree_hal_buffer_t* buffer, iree_hal_alloca_flags_t flags) {
-  IREE_RETURN_IF_ERROR(iree_hal_semaphore_list_wait(wait_semaphore_list,
-                                                    iree_infinite_timeout()));
+  IREE_RETURN_IF_ERROR(
+      iree_hal_semaphore_list_wait(wait_semaphore_list, iree_infinite_timeout(),
+                                   IREE_HAL_WAIT_FLAG_DEFAULT));
   iree_status_t status = iree_hal_semaphore_list_signal(signal_semaphore_list);
   return status;
 }
diff --git a/runtime/src/iree-amd-aie/driver/xrt-lite/direct_command_buffer.cc b/runtime/src/iree-amd-aie/driver/xrt-lite/direct_command_buffer.cc
@@ -64,8 +64,11 @@ iree_status_t iree_hal_xrt_lite_direct_command_buffer_create(
   command_buffer->host_allocator = host_allocator;
   command_buffer->device = device;
   iree_arena_initialize(block_pool, &command_buffer->arena);
-  iree_status_t status =
-      iree_hal_resource_set_allocate(block_pool, &command_buffer->resource_set);
+  iree_status_t status = iree_ok_status();
+  if (!iree_all_bits_set(mode, IREE_HAL_COMMAND_BUFFER_MODE_UNRETAINED)) {
+    status = iree_hal_resource_set_allocate(block_pool,
+                                            &command_buffer->resource_set);
+  }
   if (iree_status_is_ok(status)) {
     *out_command_buffer = &command_buffer->base;
   } else {
diff --git a/runtime/src/iree-amd-aie/driver/xrt/direct_command_buffer.cc b/runtime/src/iree-amd-aie/driver/xrt/direct_command_buffer.cc
@@ -82,8 +82,11 @@ iree_status_t iree_hal_xrt_direct_command_buffer_create(
       &iree_hal_xrt_direct_command_buffer_vtable, &command_buffer->base);
   command_buffer->host_allocator = host_allocator;
   iree_arena_initialize(block_pool, &command_buffer->arena);
-  iree_status_t status =
-      iree_hal_resource_set_allocate(block_pool, &command_buffer->resource_set);
+  iree_status_t status = iree_ok_status();
+  if (!iree_all_bits_set(mode, IREE_HAL_COMMAND_BUFFER_MODE_UNRETAINED)) {
+    status = iree_hal_resource_set_allocate(block_pool,
+                                            &command_buffer->resource_set);
+  }
   if (iree_status_is_ok(status)) {
     *out_command_buffer = &command_buffer->base;
   } else {
diff --git a/runtime/src/iree-amd-aie/driver/xrt/xrt_device.cc b/runtime/src/iree-amd-aie/driver/xrt/xrt_device.cc
@@ -230,8 +230,9 @@ static iree_status_t iree_hal_xrt_device_import_file(
 }
 
 static iree_status_t iree_hal_xrt_device_create_semaphore(
-    iree_hal_device_t* base_device, uint64_t initial_value,
-    iree_hal_semaphore_flags_t flags, iree_hal_semaphore_t** out_semaphore) {
+    iree_hal_device_t* base_device, iree_hal_queue_affinity_t queue_affinity,
+    uint64_t initial_value, iree_hal_semaphore_flags_t flags,
+    iree_hal_semaphore_t** out_semaphore) {
   iree_hal_xrt_device_t* device = iree_hal_xrt_device_cast(base_device);
   return iree_hal_xrt_semaphore_create(device->host_allocator, initial_value,
                                        out_semaphore);
@@ -251,8 +252,9 @@ static iree_status_t iree_hal_xrt_device_queue_alloca(
     iree_device_size_t allocation_size, iree_hal_execute_flags_t flags,
     iree_hal_buffer_t** IREE_RESTRICT out_buffer) {
   // TODO: queue-ordered allocations.
-  IREE_RETURN_IF_ERROR(iree_hal_semaphore_list_wait(wait_semaphore_list,
-                                                    iree_infinite_timeout()));
+  IREE_RETURN_IF_ERROR(
+      iree_hal_semaphore_list_wait(wait_semaphore_list, iree_infinite_timeout(),
+                                   IREE_HAL_WAIT_FLAG_DEFAULT));
   IREE_RETURN_IF_ERROR(
       iree_hal_allocator_allocate_buffer(iree_hal_device_allocator(base_device),
                                          params, allocation_size, out_buffer));
@@ -265,8 +267,9 @@ static iree_status_t iree_hal_xrt_device_queue_dealloca(
     const iree_hal_semaphore_list_t wait_semaphore_list,
     const iree_hal_semaphore_list_t signal_semaphore_list,
     iree_hal_buffer_t* buffer, iree_hal_alloca_flags_t flags) {
-  IREE_RETURN_IF_ERROR(iree_hal_semaphore_list_wait(wait_semaphore_list,
-                                                    iree_infinite_timeout()));
+  IREE_RETURN_IF_ERROR(
+      iree_hal_semaphore_list_wait(wait_semaphore_list, iree_infinite_timeout(),
+                                   IREE_HAL_WAIT_FLAG_DEFAULT));
   iree_status_t status = iree_hal_semaphore_list_signal(signal_semaphore_list);
   return status;
 }
@@ -352,7 +355,8 @@ static iree_status_t iree_hal_xrt_device_queue_flush(
 
 static iree_status_t iree_hal_xrt_device_wait_semaphores(
     iree_hal_device_t* base_device, iree_hal_wait_mode_t wait_mode,
-    const iree_hal_semaphore_list_t semaphore_list, iree_timeout_t timeout) {
+    const iree_hal_semaphore_list_t semaphore_list, iree_timeout_t timeout,
+    iree_hal_wait_flags_t flags) {
   return iree_make_status(IREE_STATUS_UNIMPLEMENTED,
                           "Unimplemented semaphore wait");
 }
diff --git a/third_party/iree b/third_party/iree
@@ -1 +1 @@
-Subproject commit 337c8aaf544053ac512bb7c023c572f8e2d51e7e
+Subproject commit c37c680ae6e71f715bd7c540909155061bc44491
