Add script to tune parameters

.gitignore

@@ -1 +1 @@
-test/Manifest.toml
+Manifest.toml

configs/configs.jl

@@ -3,6 +3,7 @@
 using GemmKernels
 using LinearAlgebra
 using ForwardDiff
+using Octavian
 
 struct Configuration
     name           # Human-readable name of the configuration.
@@ -238,10 +239,10 @@ macro get_wmma_config()
                       CD_type,
                       transpose_a,
                       transpose_b,
-                      mul!,
+                      Octavian.matmul!,
                       Epilogue.Default(),
                       verify_default,
-                      Kernel.matmul_pipelined,
+                      kernel,
                       wmma_baseline)
     end end)
 end
@@ -520,7 +521,8 @@ function get_configs()
             [2, 2, 1],
             [1, 1, 2],
             [2, 2, 2]], [[2048, 2048, 2048]]),
-        zero_c in [false]
+        zero_c in [false],
+        kernel in [Kernel.matmul_pipelined]
 
         push!(rv, @get_wmma_config)
     end

src/kernel.jl

@@ -317,6 +317,7 @@ function matmul_pipelined(conf::GemmKernels.Config, a, b, c, d,
 
             @loopinfo unroll for j = 1 : num_fragments_n
                 b_tile = translate_offset(warp_tile.KN, (K = 0, N = (j-1)*conf.compute_op_shape.N))
+                @cuassert ((b_tile.base.M + b_tile.offset.M) < conf.matmul_shape.M) && ((b_tile.base.K + b_tile.offset.K) < conf.matmul_shape.K)
                 @inbounds @immutable b_frags[nxt_stage, j] = transf_sh2rf_b(Operator.load_b(conf.operator, conf.shared_b_layout, shmem_b, b_tile), b_tile)
             end
 

test/Project.toml

@@ -5,5 +5,6 @@ Distributed = "8ba89e20-285c-5b6f-9357-94700520ee1b"
 ForwardDiff = "f6369f11-7733-5829-9624-2563aa707210"
 InteractiveUtils = "b77e0a4c-d291-57a0-90e8-8db25a27a240"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
+Octavian = "6fd5a793-0b7e-452c-907f-f8bfe9c57db4"
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
 XUnit = "3e3c03f2-1a94-11e9-2981-050a4ca824ab"

tuning/Project.toml

@@ -0,0 +1,5 @@
+[deps]
+Hyperopt = "93e5fe13-2215-51db-baaf-2e9a34fb2712"
+Octavian = "6fd5a793-0b7e-452c-907f-f8bfe9c57db4"
+Plots = "91a5bcdd-55d7-5caf-9e0b-520d859cae80"
+PythonPlot = "274fc56d-3b97-40fa-a1cd-1b4a50311bf9"

tuning/tune-wmma.jl

@@ -0,0 +1,213 @@
+using CUDA, GemmKernels
+using Hyperopt
+using Plots
+pythonplot()
+
+const NUM_SAMPLES = 250
+const NUM_SAMPLES_PLOT = 250
+
+include("../configs/configs.jl")
+
+AB_type = Float16
+CD_type = Float32
+
+zero_c = true
+
+OP_M, OP_N, OP_K = 16, 16, 16
+
+markershapes = Dict(
+                    "NN" => :circle,
+                    "TT" => :cross,
+                    "TN" => :diamond,
+                    "NT" => :dtriangle,
+                   )
+
+function print_counters(counters)
+    count_str(categ) = "$(counters[categ]) ($(round(100*counters[categ]/counters["total"]; digits=1))%)"
+
+    println("Total:                                     $(counters["total"]) configurations")
+    println(repeat("-", 100))
+    println("Skipped due to invalid GemmKernels config: $(count_str("invalid_config"))")
+    println("Produced incorrect result:                 $(count_str("invalid_result"))")
+    println("Threw an error:                            $(count_str("error"))")
+    println("Successful runs:                           $(count_str("success"))")
+end
+
+function optimise(transpose_a, transpose_b)
+    M = N = K = 4096
+
+    BLOCK_M, BLOCK_N, BLOCK_K = 128, 128, 64
+    WARPS_M = 4
+    WARPS_N = 2
+    kernel = Kernel.matmul_singlestage
+    cf = @get_wmma_config
+    c_h, a, b, c, d = generate_inputs(cf)
+
+    @info "Starting hyperopt..."
+
+    counters = Dict(
+        "total" => 0,
+        "invalid_config" => 0,
+        "invalid_result" => 0,
+        "error" => 0,
+        "success" => 0,
+    )
+
+    ho = @hyperopt for i = 1000,
+                BLOCK_M = 2 .^ (1:8),
+                BLOCK_N = 2 .^ (1:8),
+                BLOCK_K = 2 .^ (1:8),
+                WARPS_M = 2 .^ (0:3),
+                WARPS_N = 2 .^ (0:3),
+                kernel in [Kernel.matmul_singlestage, Kernel.matmul_pipelined]
+
+        counters["total"] += 1
+
+        try
+            cf = @get_wmma_config
+        catch err
+            if isa(err, GemmKernels.ConfigError)
+                counters["invalid_config"] += 1
+                return Inf
+            end
+        end
+
+        @info "Trying configuration: $(cf.config)"
+
+        d .= 0
+
+        try
+            run_gemm(cf, a, b, c, d)
+        catch err
+            if isa(err, GemmKernels.ConfigError)
+                counters["invalid_config"] += 1
+                return Inf
+            end
+
+            counters["error"] += 1
+
+            if isa(err, CuError)
+                @error "Configuration failed: $(cf.config)"
+                rethrow()
+            end
+
+            @info "Skipping configuration: $(cf.config)\n" * sprint(Base.showerror, err)
+            return Inf
+        end
+
+        if !verify(cf, c_h, d)
+            @warn "Configuration produced invalid result: $(cf.config)"
+            counters["invalid_result"] += 1
+            return Inf
+        end
+
+        times = []
+
+        try
+            for i in 1:NUM_SAMPLES
+                prof = CUDA.@profile run_gemm(cf, a, b, c, d)
+                push!(times, sum(prof.device[!, "stop"] - prof.device[!, "start"]))
+            end
+        catch err
+            counters["error"] += 1
+
+            if isa(err, CuError)
+                @error "Configuration failed: $(cf.config)"
+                rethrow()
+            end
+
+            @info "Skipping configuration: $(cf.config)\n" * sprint(Base.showerror, err)
+            return Inf
+        end
+
+        counters["success"] += 1
+        return minimum(times)
+    end
+
+    print_counters(counters)
+
+    ho, counters
+end
+
+get_label(transpose_a, transpose_b) = "$(transpose_a ? "T" : "N")$(transpose_b ? "T" : "N")"
+
+function make_plot(BLOCK_M, BLOCK_N, BLOCK_K, WARPS_M, WARPS_N, kernel, transpose_a, transpose_b)
+    label = get_label(transpose_a, transpose_b)
+
+    N_vals = 2 .^ (7:14)
+    gemmkernels = []
+    cublas = []
+
+    for N in N_vals
+        @show N
+        M = K = N
+
+        cf = @get_wmma_config
+        c_h, a, b, c, d = generate_inputs(cf)
+
+        samples = []
+
+        for i in 1:NUM_SAMPLES_PLOT
+            prof = CUDA.@profile run_gemm(cf, a, b, c, d)
+            push!(samples, sum(prof.device[!, "stop"] - prof.device[!, "start"]))
+        end
+
+        push!(gemmkernels, minimum(samples))
+
+        samples = []
+
+        for i in 1:NUM_SAMPLES_PLOT
+            prof = CUDA.@profile run_baseline(cf, a, b, c, d)
+            push!(samples, sum(prof.device[!, "stop"] - prof.device[!, "start"]))
+        end
+
+        push!(cublas, minimum(samples))
+    end
+
+    ratios = 100 .* cublas ./ gemmkernels
+
+    plot!(N_vals, ratios, label=label, markershape=markershapes[label], xscale=:log2, ylims=(0, max(100, ratios...)))
+    title!("$AB_type x $AB_type = $CD_type")
+    xlabel!("Matrix size [-]")
+    ylabel!("Performance relative to cuBLAS [%]")
+end
+
+function main()
+    hos = Dict()
+
+    total_counters = Dict(
+        "total" => 0,
+        "invalid_config" => 0,
+        "invalid_result" => 0,
+        "error" => 0,
+        "success" => 0,
+    )
+
+    for transpose_a in [false, true],
+        transpose_b in [false, true]
+        hos[(transpose_a, transpose_b)], counters = optimise(transpose_a, transpose_b)
+        total_counters = Dict(k => total_counters[k] + counters[k] for k in keys(counters))
+    end
+
+    println(repeat("=", 100))
+    println("Overall configurations:")
+    println(repeat("=", 100))
+
+    print_counters(total_counters)
+
+    println("Optimal parameters:")
+
+    for transpose_a in [false, true],
+        transpose_b in [false, true]
+        println("$(get_label(transpose_a, transpose_b)): $(hos[(transpose_a, transpose_b)].minimizer)")
+    end
+
+    for transpose_a in [false, true],
+        transpose_b in [false, true]
+        make_plot(hos[(transpose_a, transpose_b)].minimizer..., transpose_a, transpose_b)
+    end
+
+    savefig("plot.pdf")
+end
+
+isinteractive() || main()