Use the LLVM SPIR-V back-end on the LTS stack

src/compiler/compilation.jl

@@ -212,24 +212,19 @@ end
             throw(ArgumentError("Sub-group size $sub_group_size is not supported by this device, which supports $(join(sizes, ", "))"))
     end
 
-    # SPIR-V codegen path. The Aurora LTS NEO/IGC runtime only accepts SPIR-V from the
-    # Khronos translator; the rolling stack uses the LLVM SPIR-V back-end. GPUCompiler picks
-    # the tool from the target's `backend` field and loads the JLL lazily, so both can be
-    # listed as deps and the choice is made here at compile time. Either way the extensions
-    # below have to be declared explicitly; the ones that atomics need are added by
-    # GPUCompiler based on `atomics`.
+    # The extensions below have to be declared explicitly; the ones that atomics need are
+    # added by GPUCompiler based on `atomics`.
     # TODO: emit printf format strings in constant memory
     extensions = String[
         "SPV_EXT_relaxed_printf_string_address_space",
     ]
+    backend = :llvm
     if oneL0.LTS[]
-        backend = :khronos
         # The LTS SPIR-V stack cannot codegen native bfloat in generic kernels (clamp! ->
         # InvalidIRError, and SPV_KHR_bfloat16 segfaults NEO), so keep bf16 off entirely on
         # this path; the test suite gates bf16 eltypes on `!oneL0.LTS[]` to match.
         supports_bfloat16 = false
     else
-        backend = :llvm
         supports_bfloat16 = _driver_supports_bfloat16_spirv(dev)
         if supports_bfloat16
             push!(extensions, "SPV_KHR_bfloat16")

src/compiler/precompile.jl

@@ -3,16 +3,9 @@ using PrecompileTools: @compile_workload
 # Warm up the GPUCompiler -> SPIR-V pipeline during precompilation so the first real
 # `zefunction` call is cheap. This doesn't need a GPU: SPIR-V codegen is device-independent
 # (the target only carries version/extension/capability knobs), and the workload never
-# launches anything. It is gated on the SPIR-V tool for the selected codegen path being
-# available so the package still precompiles on platforms/toolchains without it.
-#
-# The workload mirrors the runtime backend choice (see `_compiler_config`): the LTS stack
-# compiles through the Khronos translator, the rolling stack through the LLVM SPIR-V
-# back-end. `oneL0.__init__` has not run at precompile time, so resolve ONEAPI_LTS
-# directly from the environment with the same parser.
-const _precompile_lts = oneL0.parse_env_bool("ONEAPI_LTS", false)
-if _precompile_lts ? SPIRV_LLVM_Translator_jll.is_available() :
-        SPIRV_LLVM_Backend_jll.is_available()
+# launches anything. It is gated on the SPIR-V back-end being available so the package
+# still precompiles on platforms/toolchains without it.
+if SPIRV_LLVM_Backend_jll.is_available()
     @compile_workload begin
         let
             function _precompile_kernel(a)
@@ -24,7 +17,7 @@ if _precompile_lts ? SPIRV_LLVM_Translator_jll.is_available() :
             # these knobs from the device; here we use conservative, portable defaults (the
             # workload only exercises the pipeline, it does not target a specific device).
             target = SPIRVCompilerTarget(;
-                backend = (_precompile_lts ? :khronos : :llvm),
+                backend = :llvm,
                 extensions = "", supports_fp16 = true,
                 supports_fp64 = true, supports_bfloat16 = false, driver = :intel,
             )