Merge pull request #261 from Evovest/debug-gpu

Debug gpu

Project.toml

@@ -1,7 +1,7 @@
 name = "EvoTrees"
 uuid = "f6006082-12f8-11e9-0c9c-0d5d367ab1e5"
 authors = ["jeremiedb <jeremie.db@evovest.com>"]
-version = "0.16.3"
+version = "0.16.4"
 
 [deps]
 BSON = "fbb218c0-5317-5bc6-957e-2ee96dd4b1f0"

experiments/hist/perf-gpu.jl

@@ -0,0 +1,202 @@
+using Revise
+using CUDA
+using StatsBase: sample
+using BenchmarkTools
+using Base.Threads: @threads
+using Random: seed!
+
+"""
+    hist_kernel!
+"""
+function hist_kernel!(h∇::CuDeviceArray{T,3}, ∇::CuDeviceMatrix{S}, x_bin, is, js) where {T,S}
+    tix, tiy, k = threadIdx().z, threadIdx().y, threadIdx().x
+    bdx, bdy = blockDim().z, blockDim().y
+    bix, biy = blockIdx().z, blockIdx().y
+    gdx = gridDim().z
+
+    j = tiy + bdy * (biy - 1)
+    if j <= length(js)
+        jdx = js[j]
+        i_max = length(is)
+        niter = cld(i_max, bdx * gdx)
+        @inbounds for iter = 1:niter
+            i = tix + bdx * (bix - 1) + bdx * gdx * (iter - 1)
+            if i <= i_max
+                @inbounds idx = is[i]
+                @inbounds bin = x_bin[idx, jdx]
+                hid = Base._to_linear_index(h∇, k, bin, jdx)
+                CUDA.atomic_add!(pointer(h∇, hid), T(∇[k, idx]))
+            end
+        end
+    end
+    sync_threads()
+    return nothing
+end
+
+function update_hist_gpu!(h, h∇, ∇, x_bin, is, js, jsc)
+    kernel = @cuda launch = false hist_kernel!(h∇, ∇, x_bin, is, js)
+    config = launch_configuration(kernel.fun)
+    max_threads = config.threads ÷ 4
+    max_blocks = config.blocks * 4
+    k = size(h∇, 1)
+    ty = max(1, min(length(js), fld(max_threads, k)))
+    tx = min(64, max(1, min(length(is), fld(max_threads, k * ty))))
+    threads = (k, ty, tx)
+    by = cld(length(js), ty)
+    bx = min(cld(max_blocks, by), cld(length(is), tx))
+    blocks = (1, by, bx)
+    h∇ .= 0
+    kernel(h∇, ∇, x_bin, is, js; threads, blocks)
+    CUDA.synchronize()
+    CUDA.@sync for j in jsc
+        nbins = size(h[j], 2)
+        copyto!(h[j], view(h∇, :, 1:nbins, j))
+    end
+    return nothing
+end
+
+function update_hist_gpu1!(h, h∇, ∇, x_bin, is, js, jsc)
+    kernel = @cuda launch = false hist_kernel!(h∇, ∇, x_bin, is, js)
+    config = launch_configuration(kernel.fun)
+    max_threads = config.threads ÷ 4
+    max_blocks = config.blocks * 4
+    k = size(h∇, 1)
+    ty = max(1, min(length(js), fld(max_threads, k)))
+    tx = min(64, max(1, min(length(is), fld(max_threads, k * ty))))
+    threads = (k, ty, tx)
+    by = cld(length(js), ty)
+    bx = min(cld(max_blocks, by), cld(length(is), tx))
+    blocks = (1, by, bx)
+    h∇ .= 0
+    kernel(h∇, ∇, x_bin, is, js; threads, blocks)
+    CUDA.synchronize()
+    return nothing
+end
+
+function update_hist_gpu2!(h, h∇_cpu, h∇, ∇, x_bin, is, js, jsc)
+    kernel = @cuda launch = false hist_kernel!(h∇, ∇, x_bin, is, js)
+    config = launch_configuration(kernel.fun)
+    max_threads = config.threads ÷ 4
+    max_blocks = config.blocks * 4
+    k = size(h∇, 1)
+    ty = max(1, min(length(js), fld(max_threads, k)))
+    tx = min(64, max(1, min(length(is), fld(max_threads, k * ty))))
+    threads = (k, ty, tx)
+    by = cld(length(js), ty)
+    bx = min(cld(max_blocks, by), cld(length(is), tx))
+    blocks = (1, by, bx)
+    h∇ .= 0
+    kernel(h∇, ∇, x_bin, is, js; threads, blocks)
+    copyto!(h∇_cpu, h∇)
+    CUDA.synchronize()
+    return nothing
+end
+
+
+function update_hist_gpu3!(h, h∇_cpu, h∇, ∇, x_bin, is, js, jsc)
+    kernel = @cuda launch = false hist_kernel!(h∇, ∇, x_bin, is, js)
+    config = launch_configuration(kernel.fun)
+    max_threads = config.threads ÷ 4
+    max_blocks = config.blocks * 4
+    k = size(h∇, 1)
+    ty = max(1, min(length(js), fld(max_threads, k)))
+    tx = min(64, max(1, min(length(is), fld(max_threads, k * ty))))
+    threads = (k, ty, tx)
+    by = cld(length(js), ty)
+    bx = min(cld(max_blocks, by), cld(length(is), tx))
+    blocks = (1, by, bx)
+    h∇ .= 0
+    kernel(h∇, ∇, x_bin, is, js; threads, blocks)
+    # CUDA.synchronize()
+    copyto!(h∇_cpu, h∇)
+    # CUDA.synchronize()
+    @threads for j in jsc
+        nbins = size(h[j], 2)
+        @views h[j] .= h∇_cpu[:, 1:nbins, j]
+        # h[j] .= h∇_cpu[:, 1:nbins, j]
+    end
+    return nothing
+end
+
+
+seed!(123)
+nbins = 32
+nfeats = 100
+nobs = Int(1e6)
+x_bin = UInt8.(rand(1:nbins, nobs, nfeats));
+∇ = rand(Float32, 3, nobs);
+h∇ = [zeros(Float32, 3, nbins) for n in 1:nfeats]
+rowsample = 0.5
+colsample = 0.5
+is = sample(1:nobs, Int(round(rowsample * nobs)), replace=false, ordered=true)
+js = sample(1:nfeats, Int(round(rowsample * nfeats)), replace=false, ordered=true)
+
+∇_gpu = CuArray(∇)
+x_bin_gpu = CuArray(x_bin)
+h∇_cpu = zeros(Float32, 3, nbins, nfeats)
+h∇_gpu = CuArray(h∇_cpu)
+is_gpu = CuArray(is)
+js_gpu = CuArray(js)
+
+CUDA.allowscalar(false)
+CUDA.@time update_hist_gpu!(h∇, h∇_gpu, ∇_gpu, x_bin_gpu, is_gpu, js_gpu, js)
+# ref without copy to cpu: ~same
+# ref 10K: 875.100 μs (168 allocations: 7.08 KiB)
+# ref 100K: 1.236 ms (215 allocations: 9.91 KiB)
+# ref 1M:  6.138 ms (227 allocations: 12.00 KiB)
+# ref 10M: 67.075 ms (235 allocations: 13.38 KiB)
+
+# with copy
+# CUDA v4 1M: 2.903 ms (124 allocations: 6.98 KiB)
+# CUDA v5 1M: 3.542 ms (848 allocations: 37.14 KiB)
+@btime update_hist_gpu!(h∇, h∇_gpu, ∇_gpu, x_bin_gpu, is_gpu, js_gpu, js)
+
+# without copy
+# CUDA v4 1M: 2.599 ms (74 allocations: 4.64 KiB)
+# CUDA v5 1M: 2.274 ms (48 allocations: 2.77 KiB)
+@btime update_hist_gpu1!(h∇, h∇_gpu, ∇_gpu, x_bin_gpu, is_gpu, js_gpu, js)
+
+# without single array copy
+# CUDA v4 1M: 
+# CUDA v5 1M: 2.447 ms (48 allocations: 2.77 KiB)
+@btime update_hist_gpu2!(h∇, h∇_cpu, h∇_gpu, ∇_gpu, x_bin_gpu, is_gpu, js_gpu, js)
+
+# without single array copy
+# CUDA v4 1M: 
+# CUDA v5 1M: 2.442 ms (48 allocations: 2.77 KiB)
+@btime update_hist_gpu3!(h∇, h∇_cpu, h∇_gpu, ∇_gpu, x_bin_gpu, is_gpu, js_gpu, js)
+
+
+using CUDA, BenchmarkTools
+function gpu_copy!(h, h∇, jsc)
+    CUDA.@sync for j in jsc
+        nbins = size(h[j], 2)
+        copyto!(h[j], view(h∇, :, 1:nbins, j))
+    end
+    return nothing
+end
+
+h∇ = [zeros(Float32, 3, 32) for n in 1:100];
+h∇_gpu = CUDA.zeros(Float32, 3, 32, 100);
+js = 1:100
+
+# CUDA v4: 534.480 μs (100 allocations: 4.69 KiB)
+# CUDA v5: 1.203 ms (1600 allocations: 68.75 KiB)
+@btime gpu_copy!(h∇, h∇_gpu, js)
+
+
+function gpu_copy2!(h, h∇, jsc)
+    for j in jsc
+        nbins = size(h[j], 2)
+        @async copyto!(h[j], view(h∇, :, 1:nbins, j))
+    end
+    return nothing
+end
+
+h∇ = [zeros(Float32, 3, 32) for n in 1:100];
+h∇_gpu = CUDA.zeros(Float32, 3, 32, 100);
+js = 1:100
+
+# CUDA v4: 534.480 μs (100 allocations: 4.69 KiB)
+# CUDA v5: 1.203 ms (1600 allocations: 68.75 KiB)
+@btime gpu_copy2!(h∇, h∇_gpu, js)