Add autotuning

Project.toml

@@ -15,6 +15,7 @@ CUDA_Tile_jll = "2068806d-a867-5dbd-af0e-42c2eb5d895d"
 CompilerCaching = "9db33cc3-5358-4881-8759-fa4194144afd"
 EnumX = "4e289a0a-7415-4d19-859d-a7e5c4648b56"
 GPUArrays = "0c68f7d7-f131-5f86-a1c3-88cf8149b2d7"
+GPUCompiler = "61eb1bfa-7361-4325-ad38-22787b887f55"
 GPUToolbox = "096a3bc2-3ced-46d0-87f4-dd12716f4bfc"
 IRStructurizer = "93e32bba-5bb8-402b-805d-ffb066edee93"
 LMDB = "11f193de-5e89-5f17-923a-7d207d56daf9"
@@ -43,6 +44,7 @@ CompilerCaching = "0.3.1, 0.4"
 DLFP8Types = "0.1.0"
 EnumX = "1.0"
 GPUArrays = "11"
+GPUCompiler = "2.4"
 GPUToolbox = "3"
 IRStructurizer = "0.6.3"
 LMDB = "3"
@@ -53,3 +55,7 @@ REPL = "1"
 SHA = "0.7.0"
 Scratch = "1.2"
 julia = "1.11"
+
+[sources]
+CompilerCaching = {url = "https://github.com/AntonOresten/CompilerCaching.jl", rev = "concurrent-const-seed"}
+GPUCompiler = {url = "https://github.com/AntonOresten/GPUCompiler.jl", rev = "inference_batch"}

src/compiler/driver.jl

@@ -307,7 +307,7 @@ function emit_tile!(cache::CacheView, mi::Core.MethodInstance,
 
     # Compute bytecode via driver
     sci, rettype, kernel_meta = ir_result
-    key = cache.owner::TileCacheKey
+    key = tile_cache_key(cache.owner)
     opts = CGOpts((sm_arch=unpack_version(key.sm_arch),
                    opt_level=unpack_hint(key.opt_level),
                    num_ctas=unpack_hint(key.num_ctas),

src/compiler/interpreter.jl

@@ -1,6 +1,5 @@
 # Integration with Julia's abstract interpreter
 
-
 Base.Experimental.@MethodTable cuTileMethodTable
 
 function get_method_table_view(world::UInt)
@@ -21,11 +20,10 @@ end
 
 function cuTileInterpreter(cache::CacheView; always_inline::Bool=true)
     method_table = get_method_table_view(cache.world)
-    @static if isdefined(CC, :InferenceCache)
-        inf_cache = CC.InferenceCache()
-    else
-        inf_cache = Vector{CC.InferenceResult}()
-    end
+    # Fresh per interpreter, unless inside `GPUCompiler.inference_batch`, which
+    # shares one per (task, owner, world) so const-prop results carry across
+    # the batch (e.g. autotuning over many const-seeded variants of a kernel).
+    inf_cache = GPUCompiler.inference_cache(cache.owner, cache.world)
     inf_params = CC.InferenceParams()
     opt_params = if always_inline
         CC.OptimizationParams(; inline_cost_threshold=typemax(Int))

src/cuTile.jl

@@ -22,6 +22,7 @@ using EnumX
 public BFloat16
 
 import CompilerCaching
+import GPUCompiler
 using CompilerCaching: CacheView, method_instance, match_method_instance, typeinf!, results, lookup, get_source
 
 # Shared definitions
@@ -127,4 +128,6 @@ end
 
 include("precompile.jl")
 
+include("experimental/Experimental.jl")
+
 end # module cuTile

src/experimental/Experimental.jl

@@ -0,0 +1,15 @@
+module Experimental
+
+using ..cuTile
+using ..cuTile: cuTileconvert, default_sm_arch, temporary_cufunction
+
+using CUDACore: CUDACore
+using GPUCompiler: inference_batch
+
+using Random
+
+include("search_space.jl")
+include("autotune.jl")
+include("autotune_macro.jl")
+
+end