rename FUSION_LOOKAHEAD

Makefile

@@ -16,7 +16,7 @@ JIT_PIPELINE_FALLBACK ?= 0
 # how many unique kernels to batch before triggering a JIT compile
 JIT_BATCH_SIZE ?= 16
 # how many pending queue events to scan ahead when looking for fusion candidates
-FUSION_LOOKAHEAD ?= 32
+QUEUE_ABSORB_LIMIT ?= 32
 # horizontal fusion: max independent parallel chains per kernel pass
 MAX_HFUSE_CHAINS ?= 9
 # vertical fusion: max RPN opcodes per chain (caps how deep chains can be fused)
@@ -176,7 +176,7 @@ reconfigure:
 	@echo "PERFETTO_HOME=$(PERFETTO_HOME)" >> $(CONFIG_STAMP)
 	@echo "DEBUG_KEEP_JIT_DIR=$(DEBUG_KEEP_JIT_DIR)" >> $(CONFIG_STAMP)
 	@echo "ENABLE_PROMOTION_REDUCTIONS=$(ENABLE_PROMOTION_REDUCTIONS)" >> $(CONFIG_STAMP)
-	@echo "FUSION_LOOKAHEAD=$(FUSION_LOOKAHEAD)" >> $(CONFIG_STAMP)
+	@echo "QUEUE_ABSORB_LIMIT=$(QUEUE_ABSORB_LIMIT)" >> $(CONFIG_STAMP)
 	@echo "ENABLE_OOM_RECOVERY=1" >> $(CONFIG_STAMP)
 	@echo "MAX_HFUSE_CHAINS=$(MAX_HFUSE_CHAINS)" >> $(CONFIG_STAMP)
 	@echo "MAX_VFUSE_OPS=$(MAX_VFUSE_OPS)" >> $(CONFIG_STAMP)

benchmarks/src/config.jl

@@ -112,7 +112,7 @@ struct FusionProfile
 end
 
 const FUSION_BUILD_KNOBS = (
-    "FUSION_LOOKAHEAD",
+    "QUEUE_ABSORB_LIMIT",
     "MAX_HFUSE_CHAINS",
     "JIT_BATCH_SIZE",
     "MAX_VFUSE_OPS",

benchmarks/src/tuning.jl

@@ -1,5 +1,5 @@
 const DEFAULT_FUSION_BUILD = Dict(
-    "FUSION_LOOKAHEAD" => 128,
+    "QUEUE_ABSORB_LIMIT" => 32,
     "MAX_HFUSE_CHAINS" => 9,
     "JIT_BATCH_SIZE" => 16,
     "MAX_VFUSE_OPS" => 128,
@@ -8,15 +8,14 @@ const DEFAULT_FUSION_BUILD = Dict(
 )
 
 const DEFAULT_FUSION_SEARCH = Dict(
-    "FUSION_LOOKAHEAD" => [0, 1, 2, 4, 8, 16, 32, 64, 128],
     # 9 is the ceiling: a tenth chain overflows WRAM at stack depth 3.
     "MAX_HFUSE_CHAINS" => [1, 2, 4, 6, 8, 9],
     "JIT_BATCH_SIZE" => [0, 1, 2, 4, 8, 16, 32],
     "MAX_VFUSE_OPS" => [1, 8, 16, 32, 64, 96, 128, 192],
 )
 
 const CORE_FUSION_KNOBS = (
-    "FUSION_LOOKAHEAD",
+    "QUEUE_ABSORB_LIMIT",
     "MAX_HFUSE_CHAINS",
     "JIT_BATCH_SIZE",
     "MAX_VFUSE_OPS",
@@ -72,6 +71,10 @@ function coordinate_descent(seed::Dict{String,Int}, search, passes::Int,
     for pass in 1:passes
         changed = false
         for knob in CORE_FUSION_KNOBS
+            # A knob with no candidate list is not swept this run: the absorb
+            # limit is out of the default search but --absorb-limit puts it
+            # back, so the order here stays canonical either way.
+            haskey(search, knob) || continue
             local_build, local_result = best, best_result
             for candidate in unique([best[knob]; search[knob]])
                 trial = copy(best)
@@ -101,7 +104,7 @@ function tune_usage(io::IO = stdout)
       --warmup N                  Override warmup iterations
       --iterations N              Override measured iterations
       --passes N                  Maximum coordinate-descent passes
-      --lookahead N[,N...]        FUSION_LOOKAHEAD candidates
+      --absorb-limit N[,N...]        QUEUE_ABSORB_LIMIT candidates
       --hfuse-chains N[,N...]     MAX_HFUSE_CHAINS candidates
       --jit-batch N[,N...]        JIT_BATCH_SIZE candidates
       --vfuse-ops N[,N...]        MAX_VFUSE_OPS candidates
@@ -151,7 +154,7 @@ function parse_tune_args(args)
             options.reset = true
             options.resume = false
         elseif arg in ("--dpus", "--elements-per-dpu", "--warmup", "--iterations",
-                       "--passes", "--lookahead", "--hfuse-chains", "--jit-batch",
+                       "--passes", "--absorb-limit", "--hfuse-chains", "--jit-batch",
                        "--vfuse-ops", "--workspace", "--profiles", "--checkpoints",
                        "--timeout", "--build-timeout", "--config")
             value = option_value(args, index, arg)
@@ -167,8 +170,8 @@ function parse_tune_args(args)
             elseif arg == "--passes"
                 options.passes = parse(Int, value)
                 options.passes > 0 || error("--passes must be positive")
-            elseif arg == "--lookahead"
-                options.search["FUSION_LOOKAHEAD"] = nonnegative_list(value)
+            elseif arg == "--absorb-limit"
+                options.search["QUEUE_ABSORB_LIMIT"] = nonnegative_list(value)
             elseif arg == "--hfuse-chains"
                 options.search["MAX_HFUSE_CHAINS"] = int_list(value)
             elseif arg == "--jit-batch"

benchmarks/test/tests/timings.jl

@@ -43,7 +43,7 @@ const SAMPLE_APP_TIME =
         @test run["min"] == "3.1"
         @test run["max"] == "3.4"
         @test [run[knob] for knob in BenchmarkRunner.FUSION_BUILD_KNOBS] ==
-              ["128", "9", "16", "128", "11", "4"]
+              ["32", "9", "16", "128", "11", "4"]
 
         sections = csv_records(BenchmarkRunner.sections_csv(path))
         @test length(sections) == 4

benchmarks/test/tests/tuning.jl

@@ -4,17 +4,17 @@
           [1, 2, 4, 6, 8, 9]
     seed = copy(BenchmarkRunner.DEFAULT_FUSION_BUILD)
     search = Dict(knob => [value] for (knob, value) in seed)
-    search["FUSION_LOOKAHEAD"] = [1, 2, 3]
+    search["QUEUE_ABSORB_LIMIT"] = [1, 2, 3]
     evaluations = Ref(0)
     evaluate = build -> begin
         evaluations[] += 1
-        objective = Float64((build["FUSION_LOOKAHEAD"] - 2)^2 + 1)
+        objective = Float64((build["QUEUE_ABSORB_LIMIT"] - 2)^2 + 1)
         BenchmarkRunner.TuneResult("ok", objective, [objective], [(1, 1)])
     end
 
     best, result = BenchmarkRunner.coordinate_descent(seed, search, 3, evaluate)
 
-    @test best["FUSION_LOOKAHEAD"] == 2
+    @test best["QUEUE_ABSORB_LIMIT"] == 2
     @test result.objective == 1.0
     @test evaluations[] < 3 * sum(length, values(search))
 end
@@ -32,7 +32,7 @@ end
             TOML.print(io, profile)
         end
         loaded = BenchmarkRunner.load_fusion_profile(directory, "elementwise")
-        @test loaded.build["FUSION_LOOKAHEAD"] == 1
+        @test loaded.build["QUEUE_ABSORB_LIMIT"] == 1
 
         text = captured_output() do
             run_cli([
@@ -42,7 +42,7 @@ end
             ])
         end
         @test occursin("Using fusion profile", text)
-        @test occursin("FUSION_LOOKAHEAD=1", text)
+        @test occursin("QUEUE_ABSORB_LIMIT=1", text)
     end
 end
 
@@ -65,7 +65,7 @@ end
         result = BenchmarkRunner.TuneResult("ok", 4.5, [4.5], [(2, 64)])
         checkpoint = BenchmarkRunner.load_checkpoint(config, spec, options)
         @test BenchmarkRunner.record!(
-            checkpoint, build, result, 1, "FUSION_LOOKAHEAD", 128)
+            checkpoint, build, result, 1, "QUEUE_ABSORB_LIMIT", 128)
 
         resumed = BenchmarkRunner.load_checkpoint(
             config, spec, BenchmarkRunner.TuneOptions(
@@ -103,7 +103,7 @@ end
         failed = BenchmarkRunner.TuneResult(
             "runtime_failed", Inf, Float64[], Tuple{Int,Int}[])
         BenchmarkRunner.record!(checkpoint, copy(BenchmarkRunner.DEFAULT_FUSION_BUILD),
-                                failed, 1, "FUSION_LOOKAHEAD", 128)
+                                failed, 1, "QUEUE_ABSORB_LIMIT", 128)
         resumed = BenchmarkRunner.load_checkpoint(config, spec, options)
         @test isempty(resumed.cache)
         @test length(resumed.trials) == 1

host/detail/queue.cc

@@ -474,7 +474,7 @@ void EventQueue::grow_fusion_batch(const std::shared_ptr<Event>& e) {
 #if PIPELINE
   if (e->op == Event::OperationType::COMPUTE) {
     size_t absorbed = 0;
-    while (absorbed < FUSION_LOOKAHEAD && !operations_.empty()) {
+    while (absorbed < QUEUE_ABSORB_LIMIT && !operations_.empty()) {
       auto next = operations_.front();
       if (next->op != Event::OperationType::COMPUTE) break;
       if (!try_fuse(e, next)) break;

julia/src/display.jl

@@ -77,7 +77,7 @@ function versioninfo(io::IO = stdout)
                                       " batch=", cfg("JIT_BATCH_SIZE")))
     _field(io, "Build", string(inst("BUILD_TYPE"), ", ", cfg("CXX_STANDARD"),
                                ", ", inst("CXX")))
-    _field(io, "Fusion", string("lookahead=", cfg("FUSION_LOOKAHEAD"),
+    _field(io, "Fusion", string("absorb=", cfg("QUEUE_ABSORB_LIMIT"),
                                 " chains=", cfg("MAX_HFUSE_CHAINS"),
                                 " ops=", cfg("MAX_VFUSE_OPS")))
     _field(io, "Slots", string(MAX_VFUSE_INPUTS, " operands, ",

test/framework.cc

@@ -169,7 +169,7 @@ void print_build_config() {
             << " MAX_VFUSE_OPS=" << MAX_VFUSE_OPS
             << " MAX_VFUSE_INPUTS=" << MAX_VFUSE_INPUTS
             << "\n       MAX_PIPELINE_STACK_DEPTH=" << MAX_PIPELINE_STACK_DEPTH
-            << " FUSION_LOOKAHEAD=" << FUSION_LOOKAHEAD
+            << " QUEUE_ABSORB_LIMIT=" << QUEUE_ABSORB_LIMIT
             << " BLOCK_SIZE=" << BLOCK_SIZE << " NR_TASKLETS=" << NR_TASKLETS
             << "\n";
 }

test/framework.h

@@ -217,7 +217,7 @@ constexpr size_t max_reduction_chains() {
 constexpr size_t max_hfuse_chains() { return (size_t)MAX_HFUSE_CHAINS; }
 constexpr size_t max_combined_inputs() { return (size_t)MAX_COMBINED_INPUTS; }
 constexpr size_t max_vfuse_ops() { return (size_t)MAX_VFUSE_OPS; }
-constexpr size_t fusion_lookahead() { return (size_t)FUSION_LOOKAHEAD; }
+constexpr size_t queue_absorb_limit() { return (size_t)QUEUE_ABSORB_LIMIT; }
 
 // Turns a limit into an expected kernel count.
 constexpr size_t ceil_div(size_t a, size_t b) { return (a + b - 1) / b; }