Add: enable_dump_args — orchestrator args dump by vegetabledoww · Pull Request #792 · hw-native-sys/simpler

vegetabledoww · 2026-05-18T01:34:06Z

…lars)

New diagnostic flag parallel to enable_dump_tensor. When enabled, dumps orchestrator-level tensor shapes/dtype/addresses and scalar values to args_dump.json via the existing tensor dump collector SHM infrastructure.

Supports both a2a3 slim、 a5 slim platforms and a2a3 onboard, Python and C++ APIs.

Summary

Add --dump-args diagnostic flag that captures orchestrator-level tensor metadata (shape, dtype, address) and scalar values into args_dump.json via the existing tensor dump collector infrastructure.

Design

Reuse existing tensor dump channel — no new data transfer path; shares the existing dump SHM, collector, and host export pipeline.
New TensorDumpStage::ARGS_DUMP — stage distinguishes args dump from regular tensor dumps.
Independent flag — --dump-args is independent of --dump-tensor; can be enabled separately.
Dual-platform — supports both a2a3 and a5, sim and onboard.

Scope

~40 files across Python CLI/bindings, common config, a2a3 and a5 device/host/runtime layers.

Verification

a2a3sim: 1 passed
a2a3 onboard: 1 passed
a5sim: 1 passed
All three produce valid args_dump.json (json.load() validated)
UT: 244 passed (1 pre-existing failure, 1 skipped HW test)

gemini-code-assist

Code Review

This pull request introduces a feature to dump orchestrator-level arguments (tensors and scalars) into an args_dump.json file for the a2a3 and a5 platforms. Key changes include updating the CallConfig structure, Python bindings, and the TensorDumpCollector to handle a new ARGS_DUMP stage. Feedback identifies several critical issues: the collector currently records arena offsets instead of actual scalar values, and the JSON export is missing essential tensor metadata such as shapes and data pointers. Additionally, several leftover debug logs need to be removed, an indentation error in the a5 collector should be corrected, and the flush logic in the a2a3 executor should be updated to a loop for consistency across platforms.

gemini-code-assist · 2026-05-18T01:38:29Z

+            e.ndims = rec.ndims;
+            memcpy(e.shapes.data(), rec.shapes, sizeof(e.shapes));
+            e.type = (rec.ndims == 0) ? "scalar" : "tensor";
+            e.scalar_value = (rec.ndims == 0 && rec.payload_size > 0 && rec.payload_offset > 0) ? rec.payload_offset : 0;


The scalar_value is being assigned the payload_offset, which is the offset into the circular arena where the value was written, not the actual 8-byte scalar value. The collector needs to read the uint64_t value from the arena using this offset (handling circular wrap if necessary) to correctly capture the argument value. Additionally, per repository rules, ensure this 64-bit unsigned integer is represented as a string in the JSON output to prevent precision loss in JavaScript-based parsers.

References

When serializing 64-bit unsigned integers to JSON, represent them as strings to prevent precision loss in JavaScript-based parsers.

gemini-code-assist · 2026-05-18T01:38:29Z

+            e.ndims = rec.ndims;
+            memcpy(e.shapes.data(), rec.shapes, sizeof(e.shapes));
+            e.type = (rec.ndims == 0) ? "scalar" : "tensor";
+            e.scalar_value = (rec.ndims == 0 && rec.payload_size > 0) ? rec.payload_offset : 0;


Similar to the a2a3 implementation, scalar_value is being set to the arena offset (payload_offset) instead of the actual value stored in the arena. This will result in incorrect values in the exported JSON. Ensure the resulting 64-bit value is serialized as a string.

References

When serializing 64-bit unsigned integers to JSON, represent them as strings to prevent precision loss.

gemini-code-assist · 2026-05-18T01:38:29Z

+        json << "    {\"arg_index\": " << e.arg_index << ", \"type\": \"" << e.type << "\", \"dtype\": \"" << dtype_name << "\", \"ndims\": " << static_cast<uint32_t>(e.ndims);
+        json << ", \"scalar_value\": " << e.scalar_value;
+        json << "}";


The JSON export for orchestrator arguments is missing critical metadata for tensors. The PR description states that it should dump shapes and addresses, but the current implementation only exports arg_index, type, dtype, ndims, and scalar_value. For tensors, the shapes array and the data_ptr (address) should also be included in the JSON object. Note that the data_ptr (a 64-bit address) must be represented as a string to comply with repository standards.

References

When serializing 64-bit unsigned integers (such as hardware addresses) to JSON, represent them as strings to prevent precision loss.

gemini-code-assist · 2026-05-18T01:38:29Z

 #endif
+            {
+                bool en = is_dump_args_enabled();
+                LOG_ERROR("ARGS_DEBUG: Thread %d is_dump_args_enabled=%d args_addr=%p", thread_idx, en, (void*)orch_args_cached_);


This LOG_ERROR with the ARGS_DEBUG prefix appears to be leftover debug code and should be removed before merging.

gemini-code-assist · 2026-05-18T01:38:29Z

+            // Flush the args dump buffer so the host collector can pick it up.
+            {
+                bool en = is_dump_args_enabled();
+                LOG_ERROR("ARGS_FLUSH: Thread %d enabled=%d", thread_idx, en);


This LOG_ERROR with the ARGS_FLUSH prefix appears to be leftover debug code and should be removed.

gemini-code-assist · 2026-05-18T01:38:29Z

+        int orch_thread = thread_num_ - 1;
+        if (orch_thread >= sched_thread_num_ && orch_thread < thread_num_) {
+            dump_tensor_flush(orch_thread);
+        }


The logic here only flushes the last thread (thread_num_ - 1), but the comment above (line 757) says it should flush all threads from sched_thread_num_ to thread_num_ - 1. If there are multiple orchestrator threads, some buffers will not be flushed. This should be a loop to maintain consistency with the implementation in other runtimes like a5.

for (int t = sched_thread_num_; t < thread_num_; t++) { dump_tensor_flush(t); }

References

For files shared across different runtimes, maintain identical code for consistency.

gemini-code-assist · 2026-05-18T01:38:30Z

+        LOG_ERROR("A5_ARGS_FLUSH2: thread=%d count=%u", thread_idx, buf->count);
        uint64_t buf_addr = reinterpret_cast<uint64_t>(buf);
        uint32_t seq = s_dump_states[thread_idx]->current_buf_seq;
-        if (enqueue_dump_ready_buffer(thread_idx, buf_addr, seq) != 0) {
+        int enq_rc = enqueue_dump_ready_buffer(thread_idx, buf_addr, seq);
+        LOG_ERROR("A5_ARGS_ENQ: thread=%d rc=%d seq=%u", thread_idx, enq_rc, seq);


These LOG_ERROR statements with A5_ARGS_FLUSH2 and A5_ARGS_ENQ prefixes are clearly debug logs and should be removed.

gemini-code-assist · 2026-05-18T01:38:30Z

-        return 0;
-    }
+    } else {
    auto export_start = std::chrono::steady_clock::now();


The code inside the else block is not properly indented.

gemini-code-assist · 2026-05-18T01:38:30Z

+        json << "    {\"arg_index\": " << e.arg_index << ", \"type\": \"" << e.type << "\", \"dtype\": \""
+             << dtype_name << "\", \"ndims\": " << static_cast<uint32_t>(e.ndims);
+        json << ", \"scalar_value\": " << e.scalar_value;
+        json << "}";


The JSON export for a5 is also missing tensor shapes and addresses (data_ptr), which are required for the feature to be fully functional. Ensure data_ptr is serialized as a string per repository rules for 64-bit integers.

References

When serializing 64-bit unsigned integers (such as hardware addresses) to JSON, represent them as strings to prevent precision loss.

…lars) New diagnostic flag parallel to enable_dump_tensor. When enabled, dumps orchestrator-level tensor shapes/dtype/addresses and scalar values to args_dump.json via the existing tensor dump collector SHM infrastructure. Supports both a2a3 and a5 platforms, sim and onboard, Python and C++ APIs.

…s/data_ptr, remove debug logs, fix indent

…atch

gemini-code-assist Bot reviewed May 18, 2026

View reviewed changes

vegetabledoww force-pushed the main branch from b62c769 to 1674138 Compare May 18, 2026 07:32

vegetabledoww added 2 commits May 18, 2026 16:50

Fix: address PR review feedback — scalar_value arena read, JSON shape…

f1f1849

…s/data_ptr, remove debug logs, fix indent

feat: add a2a5 tensor dump platform support and refine scheduler disp…

ba4a214

…atch

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add: enable_dump_args — orchestrator args dump #792

Add: enable_dump_args — orchestrator args dump #792
vegetabledoww wants to merge 3 commits into
hw-native-sys:mainfrom
vegetabledoww:main

vegetabledoww commented May 18, 2026 •

edited

Loading

Uh oh!

gemini-code-assist Bot left a comment

Uh oh!

gemini-code-assist Bot May 18, 2026

Uh oh!

gemini-code-assist Bot May 18, 2026

Uh oh!

gemini-code-assist Bot May 18, 2026

Uh oh!

gemini-code-assist Bot May 18, 2026

Uh oh!

gemini-code-assist Bot May 18, 2026

Uh oh!

gemini-code-assist Bot May 18, 2026

Uh oh!

gemini-code-assist Bot May 18, 2026

Uh oh!

gemini-code-assist Bot May 18, 2026

Uh oh!

gemini-code-assist Bot May 18, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

vegetabledoww commented May 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Design

Scope

Verification

Uh oh!

gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist Bot May 18, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot May 18, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot May 18, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot May 18, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot May 18, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot May 18, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot May 18, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot May 18, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot May 18, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

vegetabledoww commented May 18, 2026 •

edited

Loading