Even more aie4 shim test enablement by hlaccabu · Pull Request #1025 · amd/xdna-driver

hlaccabu · 2026-01-23T01:19:30Z

42 passing shim tests with these changes

Copilot

Pull request overview

This PR enables additional shim tests for AIE4/NPU3 devices, bringing the total to 42 passing tests. The changes add support for a new kernel type (KERNEL_TYPE_TXN_FULL_ELF_PREEMPT_AIE4), implement NPU3-specific device filtering and configuration, and handle NPU3's lack of telemetry support for preemption verification.

Changes:

Added NPU3 device filter and virtual context configuration (128 contexts for NPU3)
Introduced new AIE4 preempt test infrastructure with elf_preempt_aie4_io_test_bo_set class
Updated test parameter handling to support device-specific xclbin selection

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
test/shim_test/shim_test.cpp	Added NPU3 device filter, updated virtual context limits, modified test parameters for multi-context tests, and added new AIE4 preemption test
test/shim_test/io_test.cpp	Added AIE4 kernel type support, implemented telemetry availability check for NPU3, and added conditional xclbin selection
test/shim_test/io.h	Declared new `elf_preempt_aie4_io_test_bo_set` class for AIE4 preemption testing
test/shim_test/io.cpp	Implemented AIE4 preemption test class with weight offset handling and result verification
test/shim_test/hwctx.h	Updated full ELF check to include new AIE4 preempt kernel type
test/shim_test/exec_buf.h	Added method for patched 64-bit arguments and updated copyright year
test/shim_test/dev_info.h	Added new kernel type enum value for AIE4 full ELF preemption
test/shim_test/dev_info.cpp	Added xclbin configuration entries for resnet50.elf on NPU3 devices

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

xdavidz

@hlaccabu have you verified this on board 10+ times? make sure it is stable enough please.

hlaccabu · 2026-01-26T20:03:29Z

@hlaccabu have you verified this on board 10+ times? make sure it is stable enough please.

yes just went 10/10 on silicon, all 42 tests passed each time

Copilot

Pull request overview

Copilot reviewed 11 out of 11 changed files in this pull request and generated 2 comments.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Copilot · 2026-01-27T23:00:53Z

+  // XDNA driver by default supports 6 virtual context on npu1, 128 on npu3, and 32 virtual context on npu4
  if (device_id == npu1_device_id)
    num_virt_ctx = 6;
+  else if (device_id == npu3_device_id)


The condition only checks for npu3_device_id but should also check for npu3_device_id1 to match the new dev_filter_is_npu3 function behavior that handles both device IDs.

Suggested change

else if (device_id == npu3_device_id)

else if (device_id == npu3_device_id || device_id == npu3_device_id1)

@xdavidz Can VF do same amount of ctxs?

can you test if you can get some number? I think VF can.

Sure, I can test that!

Added implementation based on Max's suggestion, works well

xdavidz · 2026-01-29T19:04:26Z

 		goto free_buf;
 	}

-	print_hex_dump_debug("telemetry: ", DUMP_PREFIX_OFFSET, 16, 4, buff,


removing this will make dmesg disabled for debugging. can we have this line unchanged or it is flooding?

It is flooding, since telemetry info is large it completely fills the dmesg. We can figure something else out in between, maybe just print first couple of lines?

hlaccabu · 2026-01-30T02:00:03Z

@xdavidz addressed several of your comments and made some enhancements, let me know what you think. All tests passing on classic, 1 preemption test fails on VF due to issues we spoke about.

Copilot

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.

maxzhen · 2026-02-06T00:25:10Z

retest this please

Signed-off-by: Hayden Laccabue <hlaccabu@amd.com>

Copilot

Pull request overview

Copilot reviewed 11 out of 11 changed files in this pull request and generated 13 comments.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Copilot · 2026-02-08T04:15:49Z

      { "DPU:dpu", {0xffffffff} },
    },
-    .workspace = "local_shim_test_data/elf_bad_ctrl_npu3",
+    .workspace = "local_shim_test_data/npu3a/bad_ctrl",


workspace paths previously had a trailing /, and get_xclbin_path() concatenates wrk + xclbin_name directly. Removing the trailing slash will produce invalid paths (e.g., .../vaddfoo.xclbin). Restore the trailing / here, or change get_xclbin_workspace/get_xclbin_path to join paths safely (ensuring exactly one separator).

Copilot · 2026-02-08T04:15:49Z

+    .ip_name2idx = {
+      { "DPU:dpu", {0xffffffff} },
+    },
+    .workspace = "local_shim_test_data/npu3a/resnet50",


workspace paths previously had a trailing /, and get_xclbin_path() concatenates wrk + xclbin_name directly. Removing the trailing slash will produce invalid paths (e.g., .../vaddfoo.xclbin). Restore the trailing / here, or change get_xclbin_workspace/get_xclbin_path to join paths safely (ensuring exactly one separator).

Copilot · 2026-02-08T04:15:49Z

      { "DPU:dpu", {0xffffffff} },
    },
-    .workspace = "local_shim_test_data/npu3a/nop/",
+    .workspace = "local_shim_test_data/npu3a/nop",


workspace paths previously had a trailing /, and get_xclbin_path() concatenates wrk + xclbin_name directly. Removing the trailing slash will produce invalid paths (e.g., .../vaddfoo.xclbin). Restore the trailing / here, or change get_xclbin_workspace/get_xclbin_path to join paths safely (ensuring exactly one separator).

Copilot · 2026-02-08T04:15:50Z

+  struct aie4_fw_telemetry {
+    uint8_t enabled;
+    uint8_t reserved[3];
+    struct aie4_clk_deep_slp deep_slp;
+    uint64_t l1_interrupt;
+    uint64_t context_starting[AIE4_MAX_NUM_SUPERVISORS + 1];
+    uint64_t scheduler_scheduled[AIE4_MAX_NUM_SUPERVISORS + 1];
+    uint64_t did_dma;
+    uint64_t resource_acquired[AIE4_MAX_NUM_SUPERVISORS];
+    struct aie4_telemetry_opcodes opcodes;
+    uint64_t preemption_frame_boundary_counter[AIE4_TOTAL_NUM_UC];
+    uint64_t preemption_checkpoint_event_counter[AIE4_TOTAL_NUM_UC];
+  };


These firmware-facing telemetry structs are parsed via reinterpret_cast<aie4_fw_telemetry*> from a raw byte buffer, but the structs are not explicitly packed/aligned to the firmware ABI. Default C++ layout/padding can differ from the firmware layout and will yield incorrect field decoding. Make the structs explicitly packed (and ideally add static_assert checks for expected offsets/sizes), or parse using explicit offsets with memcpy.

Copilot · 2026-02-08T04:15:50Z

+    virtual xrt_core::query::rtos_telemetry::result_type
+    query_rtos_telemetry(const shim_xdna::pdev& pci_dev) const {
+      xrt_core::query::rtos_telemetry::result_type output;


The previous implementation gated RTOS telemetry to NPU4 only; that guard has been removed in the base (AIE2) handler. This will now attempt RTOS telemetry queries for any non-AIE4 device using the legacy amdxdna_drm_query_telemetry layout, which can return garbage or fail on devices/firmware that don’t implement those fields. Reintroduce appropriate device/feature gating (either in get() before dispatching, or inside the base query_rtos_telemetry() implementation).

Copilot · 2026-02-08T04:15:50Z

      auto cmd_hdl = std::get<0>(cmd).get()->get();
      auto cmd_pkt = std::get<1>(cmd);

-      if (cmd_pkt->opcode == ERT_CMD_CHAIN) {
-        size_t start_idx = cmd_idx * cmds_per_list;
-        size_t end_idx = std::min(start_idx + cmds_per_list, bo_set.size());
-        for (size_t i = start_idx; i < end_idx; i++) {
-          auto subcmd_bo = bo_set[i]->get_bos()[IO_TEST_BO_CMD].tbo.get();
-          auto subcmd_pkt = reinterpret_cast<ert_start_kernel_cmd *>(subcmd_bo->map());
-          bo_set[i]->restore_cmd_header(subcmd_bo->get(), subcmd_pkt);
-        }
-      } else {
-        bo_set[cmd_idx]->restore_cmd_header(cmd_hdl, cmd_pkt);
-        bo_set[cmd_idx]->cache_cmd_header(cmd_hdl, cmd_pkt);
-      }
+      restore_cmd_headers_before_submit(bo_set, cmd_idx, cmds_per_list, cmd_hdl, cmd_pkt);

      hwq->submit_command(cmd_hdl);


The refactor removed cache_cmd_header() calls that previously occurred for non-chain command packets (and also removed initial caching in other paths later in the file). If any io_test_bo_set_base relies on caching the original header before patching/mutation, restore_cmd_header() may restore from an uninitialized/empty cache, causing malformed submits or flaky behavior across iterations. Ensure command headers are cached at least once per BO-set (including non-chain and chain sub-commands) before the first restore/submit loop, or restore the removed caching in the appropriate setup location.

Copilot · 2026-02-08T04:15:50Z

+void
+elf_preempt_aie4_io_test_bo_set::
+init_cmd(hw_ctx& hwctx, bool dump)
+{
+  std::vector<uint32_t> wts_offsets;
+  uint32_t offset;
+
+  exec_buf ebuf(*m_bo_array[IO_TEST_BO_CMD].tbo.get(), ERT_START_DPU);


elf_preempt_io_test_bo_set::init_cmd() now caches the command header after building the packet, but the new AIE4 variant does not. With the caching removed/refactored elsewhere, this AIE4 path is especially likely to submit packets that can’t be reliably restored between iterations. Add equivalent header caching for IO_TEST_BO_CMD after the packet is constructed in this init_cmd().

hlaccabu · 2026-02-09T19:51:41Z

Closing to create several smaller PRs with these changes

hlaccabu requested review from Copilot and xdavidz January 23, 2026 01:19

Copilot AI reviewed Jan 23, 2026

View reviewed changes

Comment thread test/shim_test/io.cpp

xdavidz previously approved these changes Jan 26, 2026

View reviewed changes

hlaccabu added the do-not-merge label Jan 26, 2026

hlaccabu dismissed xdavidz’s stale review via de8802f January 26, 2026 23:53

hlaccabu force-pushed the shim branch from d9faec3 to de8802f Compare January 26, 2026 23:53

hlaccabu removed the do-not-merge label Jan 26, 2026

hlaccabu force-pushed the shim branch from de8802f to f1e9770 Compare January 27, 2026 23:00

Copilot AI review requested due to automatic review settings January 27, 2026 23:00

Copilot AI reviewed Jan 27, 2026

View reviewed changes

hlaccabu requested a review from xdavidz January 28, 2026 23:42

xdavidz reviewed Jan 29, 2026

View reviewed changes

Comment thread src/driver/amdxdna/aie4_message.c

xdavidz reviewed Jan 29, 2026

View reviewed changes

Comment thread src/driver/amdxdna/aie4_pci.c

xdavidz reviewed Jan 29, 2026

View reviewed changes

Comment thread src/shim/device.cpp Outdated

xdavidz reviewed Jan 29, 2026

View reviewed changes

Comment thread src/shim/device.cpp Outdated

hlaccabu force-pushed the shim branch from f1e9770 to e0a86d6 Compare January 30, 2026 01:58

Copilot AI review requested due to automatic review settings January 30, 2026 17:59

hlaccabu force-pushed the shim branch from e0a86d6 to f59602f Compare January 30, 2026 17:59

Copilot AI reviewed Jan 30, 2026

View reviewed changes

root and others added 2 commits February 6, 2026 15:13

Even more aie4 shim test enablement

3a21893

Signed-off-by: Hayden Laccabue <hlaccabu@amd.com>

Support telemetry query for npu3

01bcdd1

Signed-off-by: Hayden Laccabue <hlaccabu@amd.com>

hlaccabu force-pushed the shim branch from f59602f to 01bcdd1 Compare February 6, 2026 23:17

Clean up caching logic for shim tests

19532e2

Signed-off-by: Hayden Laccabue <hlaccabu@amd.com>

Copilot AI review requested due to automatic review settings February 8, 2026 04:13

	else if (device_id == npu3_device_id)
	else if (device_id == npu3_device_id \|\| device_id == npu3_device_id1)

Conversation

hlaccabu commented Jan 23, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

xdavidz left a comment

Choose a reason for hiding this comment

Uh oh!

hlaccabu commented Jan 26, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Copilot AI Jan 27, 2026

Choose a reason for hiding this comment

Uh oh!

hlaccabu Jan 27, 2026

Choose a reason for hiding this comment

Uh oh!

xdavidz Jan 29, 2026

Choose a reason for hiding this comment

Uh oh!

hlaccabu Jan 29, 2026

Choose a reason for hiding this comment

Uh oh!

hlaccabu Jan 30, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

xdavidz Jan 29, 2026

Choose a reason for hiding this comment

Uh oh!

hlaccabu Jan 29, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

hlaccabu commented Jan 30, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Uh oh!

maxzhen commented Feb 6, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Copilot AI Feb 8, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 8, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 8, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 8, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 8, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 8, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 8, 2026

Choose a reason for hiding this comment

Uh oh!

hlaccabu commented Feb 9, 2026

Uh oh!

Reviewers

Assignees