Add SYCL config options

Signed-off-by: Lukas Sommer <lukas.sommer@codeplay.com>

Author: sommerlukas

torch/_inductor/codecache.py

@@ -3200,22 +3200,29 @@ def _sycl_lib_options() -> list[str]:
 
 
 def _dpcpp_compiler_options() -> list[str]:
-    # TODO Select device target architecture based on environment.
+    # TODO: Automatically detect device architecture.
+    arch = f"intel_gpu_{config.sycl.arch}" if config.sycl.arch is not None else "spir64"
     options = [
         "-fsycl",
         "-std=c++17",
         "-fPIC",
         "-Xspirv-translator", "-spirv-ext=+SPV_INTEL_split_barrier",
         "-fsycl-range-rounding=disable",
-        # TODO: Replace this with device-specific architecture.
-        "-fsycl-targets=spir64",
+        f"-fsycl-targets={arch}",
+        config.sycl.compile_opt_level,
         "-DCUTLASS_ENABLE_SYCL",
         "-DSYCL_INTEL_TARGET",
-        # TODO: Add optimization level.
     ]
     # TODO: Add special case for FB?
-    # TODO: Add debug info handling.
-    # TODO: Add fast-math handling.
+    if config.sycl.enable_debug_info:
+        options.extend(["-g", "-DCUTLASS_DEBUG_TRACE_LEVEL=1"])
+    if config.sycl.use_fast_math:
+        options.extend(
+            [
+                "-ffast-math",
+                "-DCUTLASS_USE_TANH_FOR_SIGMOID=1",
+            ]
+        )
 
     return options
 

torch/_inductor/config.py

@@ -1394,6 +1394,22 @@ class rocm:
     split_k_threshold: int = 16
 
 
+class sycl:
+    # Intel GPU arch to use for SYCL template kernel compilation.
+    # e.g. "pvc", "bmg", etc.
+    # When arch is None, generates SPIR-V that is finalized at runtime.
+    arch: Optional[str] = None
+
+    # Optimization level for the host compiler.
+    compile_opt_level: Literal["-O0", "-O1", "-O2", "-O3", "-OS"] = "-O1"
+
+    # Whether to enable debug info, e.g. line number, cutlass debug info.
+    enable_debug_info = False
+
+    # Whether to use fast math.
+    use_fast_math = False
+
+
 # Backend to use for CPU codegen either "cpp" or "triton" (experimental) or "halide" (experimental)
 cpu_backend: Literal["cpp", "triton", "halide"] = "cpp"