Simplify CPU vs GPU benchmark: remove GPU_MIN_SIZE threshold logic

Author: Sébastien Loisel

tools/benchmark_cpu_vs_gpu.jl

@@ -1,16 +1,12 @@
 #!/usr/bin/env julia
 #
-# Benchmark: CPU vs Auto (GPU with size threshold) for fem2d_mpi_solve
+# Benchmark: CPU vs GPU for fem2d_mpi_solve
 #
 # Run with:
-#   mpiexec -n 1 julia --project=MultiGridBarrierMPI.jl MultiGridBarrierMPI.jl/tools/benchmark_cpu_vs_gpu.jl
+#   mpiexec -n 1 julia --project=. tools/benchmark_cpu_vs_gpu.jl
 #
 # Note: Metal only supports Float32, so we use Float32 for both CPU and GPU
 #       to ensure a fair comparison.
-#
-# Two modes:
-#   - CPU: Pure CPU (no backend parameter)
-#   - Auto: Automatic GPU/CPU selection based on GPU_MIN_SIZE threshold
 
 using MPI
 MPI.Init()
@@ -23,85 +19,72 @@ using Metal
 using MultiGridBarrierMPI
 using MultiGridBarrier
 using LinearAlgebraMPI
-using LinearAlgebraMPI: GPU_MIN_SIZE
 using BenchmarkTools
 using Printf
 
-# Configurable threshold for "Auto" mode
-const AUTO_THRESHOLD = 1000
-
 println("\n" * "="^70)
-println("Benchmark: fem2d_mpi_solve - CPU vs Auto")
+println("Benchmark: fem2d_mpi_solve - CPU vs GPU")
 println("  MPI ranks: $(MPI.Comm_size(comm))")
 println("  Element type: Float32 (Metal requirement)")
-println("  Auto threshold: GPU_MIN_SIZE = $AUTO_THRESHOLD")
-println("  Running L = 1:6")
+println("  Running L = 1:7")
 println("="^70)
 
 # Store results
 results = Vector{NamedTuple}()
 
-for L in 1:6
+for L in 1:7
     # Get grid size
     g = fem2d(Float32; L=L)
     n = size(g.x, 1)
 
     println("\n--- L = $L (n = $n) ---")
 
-    # Benchmark CPU (pure CPU, no backend)
+    # Benchmark CPU
     println("  Benchmarking CPU...")
     LinearAlgebraMPI.clear_plan_cache!()
     b_cpu = @benchmark fem2d_mpi_solve(Float32; L=$L, verbose=false) samples=1 evals=1
     cpu_time = median(b_cpu.times) / 1e9
 
-    # Benchmark Auto (GPU_MIN_SIZE threshold)
-    println("  Benchmarking Auto (threshold=$AUTO_THRESHOLD)...")
+    # Benchmark GPU
+    println("  Benchmarking GPU...")
     LinearAlgebraMPI.clear_plan_cache!()
-    GPU_MIN_SIZE[] = AUTO_THRESHOLD
-    b_auto = @benchmark fem2d_mpi_solve(Float32; L=$L, backend=LinearAlgebraMPI.mtl, verbose=false) samples=1 evals=1
-    auto_time = median(b_auto.times) / 1e9
-
-    # Determine which arrays went to GPU in auto mode
-    GPU_MIN_SIZE[] = AUTO_THRESHOLD
-    g_test = fem2d_mpi(Float32; L=L, backend=LinearAlgebraMPI.mtl)
-    auto_is_gpu = !(g_test.x.A isa Matrix)
+    b_gpu = @benchmark fem2d_mpi_solve(Float32; L=$L, backend=LinearAlgebraMPI.mtl, verbose=false) samples=1 evals=1
+    gpu_time = median(b_gpu.times) / 1e9
 
-    push!(results, (L=L, n=n, cpu=cpu_time, auto=auto_time, auto_gpu=auto_is_gpu))
+    push!(results, (L=L, n=n, cpu=cpu_time, gpu=gpu_time))
 
     # Print results
-    speedup = cpu_time / auto_time
+    speedup = cpu_time / gpu_time
     println("  CPU:  $(round(cpu_time, digits=3))s")
-    println("  Auto: $(round(auto_time, digits=3))s [$(auto_is_gpu ? "GPU" : "CPU")]")
+    println("  GPU:  $(round(gpu_time, digits=3))s")
     if speedup > 1
-        println("  Speedup: $(round(speedup, digits=2))x (Auto faster)")
+        println("  Speedup: $(round(speedup, digits=2))x (GPU faster)")
     else
-        println("  Speedup: $(round(1/speedup, digits=2))x (CPU faster)")
+        println("  Slowdown: $(round(1/speedup, digits=2))x (CPU faster)")
     end
 end
 
 # Summary table
 println("\n" * "="^70)
 println("Summary")
 println("="^70)
-println("\n  L       n         CPU        Auto    Speedup   Auto backend")
-println("  -       -         ---        ----    -------   ------------")
+println("\n  L       n         CPU        GPU      Speedup")
+println("  -       -         ---        ---      -------")
 for r in results
     n_str = lpad(r.n, 7)
     cpu_str = @sprintf("%6.3fs", r.cpu)
-    auto_str = @sprintf("%6.3fs", r.auto)
+    gpu_str = @sprintf("%6.3fs", r.gpu)
 
-    speedup = r.cpu / r.auto
+    speedup = r.cpu / r.gpu
     if speedup > 1
-        speedup_str = @sprintf("%.2fx Auto", speedup)
+        speedup_str = @sprintf("%.2fx GPU", speedup)
     else
         speedup_str = @sprintf("%.2fx CPU", 1/speedup)
     end
     speedup_str = lpad(speedup_str, 10)
 
-    auto_backend = r.auto_gpu ? "GPU" : "CPU"
-    println("  $(r.L)    $n_str    $cpu_str    $auto_str    $speedup_str       $auto_backend")
+    println("  $(r.L)    $n_str    $cpu_str    $gpu_str    $speedup_str")
 end
 
-println("\n  Auto threshold: GPU_MIN_SIZE = $AUTO_THRESHOLD")
-println("  Speedup = CPU time / Auto time (>1 means Auto is faster)")
+println("\n  Speedup = CPU time / GPU time (>1 means GPU is faster)")
 println("="^70)