[#5153] Update the build flag metadata when loading SPIRV kernel.

[chengjunlu]

This is workaround for feature requirement in #5153. The IGC build flag is updated when the large GRF mode is used in loading SPIRV kernel when register spill size > 1000.

[Copilot]

Pull Request Overview

This PR implements a workaround to update build flag metadata when loading SPIRV kernels, specifically addressing feature requirement #5153. The change enables IGC build flag updates when large GRF mode is used during SPIRV kernel loading when register spill size exceeds 1000.

Key changes:

• Modified C driver to return build flags as an additional return value from load_binary function
• Updated Python compiler to capture and apply updated build flags to kernel metadata when they differ from original flags

Reviewed Changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
third_party/intel/backend/driver.c	Modified load_binary return value to include build flags string
python/triton/compiler/compiler.py	Added logic to capture and update metadata with new build flags when changed

---

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

[Copilot]

The addition of build_flags().data() to the return tuple creates a breaking change in the API. Consider versioning this function or providing a backward-compatible wrapper to avoid breaking existing callers that expect a 5-tuple instead of a 6-tuple.

[chengjunlu]

This is known breaking. It is a short term solution.

[chengjunlu]

@etaf Please help to confirm the change works as expected.

[etaf]

@etaf Please help to confirm the change works as expected.

Thanks.

[anmyachev]

@chengjunlu we have the following file in cache: ~/.triton/cache/QTPP7LDAKIPHLZY2LCDGJTPLKOKKXNI5HP4H7MHFPULVU5UT3MBA/kernel.json

That looks like:
{"hash": "84deffac60521e75e71a588664cdeb5394abb51d3bf87fb0e57d175a7693db02", "target": {"backend": "xpu", "arch": {"architecture": 13136561920, "device_id": 3034, "driver_version": "1.6.33578+15", "gpu_eu_count": 448, "gpu_subslice_count": 56, "has_atomic64": true, "has_bfloat16_conversions": true, "has_fp16": true, "has_fp64": true, "has_subgroup_2d_block_io": true, "has_subgroup_matrix_multiply_accumulate": true, "has_subgroup_matrix_multiply_accumulate_tensor_float32": false, "max_compute_units": 448, "max_num_sub_groups": 64, "max_work_group_size": 1024, "name": "Intel(R) Data Center GPU Max 1100", "platform_name": "Intel(R) oneAPI Unified Runtime over Level-Zero", "sub_group_sizes": [16, 32], "total_memory": 51522830336, "type": "gpu", "vendor": "Intel(R) Corporation", "version": "12.60.7"}, "warp_size": 32}, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "cluster_dims": [1, 1, 1], "warp_size": 32, "optimize_epilogue": false, "enable_fp_fusion": true, "launch_cooperative_grid": false, "reduce_variable_liveness": true, "supported_fp8_dtypes": ["fp8e5", "fp8e4nv", "fp8e4b15"], "deprecated_fp8_dot_operand_dtypes": [], "default_dot_input_precision": "tf32", "allowed_dot_input_precisions": ["tf32", "tf32x3", "ieee"], "allow_fp8e4nv": true, "allow_fp8e4b15": true, "grf_mode": ["small", "large", "auto", "default"], "split_barriers_scope": "None", "max_num_imprecise_acc_default": 0, "extern_libs": [["libdevice", "/home/jovyan/intel-xpu-backend-for-triton/python/triton/backends/intel/lib/libsycl-spir64-unknown-unknown.bc"]], "debug": false, "backend_name": "intel", "sanitize_overflow": false, "generate_native_code": false, "arch": null, "instrumentation_mode": "", "cache_dir": "/home/jovyan/.triton/cache/QTPP7LDAKIPHLZY2LCDGJTPLKOKKXNI5HP4H7MHFPULVU5UT3MBA", "triton_version": "3.5.0", "threads_per_warp": 32, "shared": 0, "global_scratch_size": 0, "global_scratch_align": 1, "profile_scratch_size": 0, "profile_scratch_align": 1, "name": "kernel", "build_flags": ""}

We need to make sure build_flags is updated here as well.

[chengjunlu]

@chengjunlu we have the following file in cache: ~/.triton/cache/QTPP7LDAKIPHLZY2LCDGJTPLKOKKXNI5HP4H7MHFPULVU5UT3MBA/kernel.json

We need to make sure build_flags is updated here as well.

This is a speific workaround only works for #5153. The build flag in json is only up to the make SPIRV stage not the load binary phase.
Do you have any concern that this may cause issue?

[anmyachev]

Do you have any concern that this may cause issue?

On second thought, it seems there shouldn't be a problem here. What about #5402 (comment)?

[anmyachev]

LGTM!