Benchmark trying to emulate cscmv with csrmv with T flag

cscmv-with-T/benchmark.jl

@@ -0,0 +1,38 @@
+using FactCheck
+include("funk.jl")
+function benchmark(k)
+
+ println("-------------- k = $k ------------") 
+
+ a = sprand(10^7, 10^7, k*1e-7)
+ rowptr = map(x -> Int32(x), a.colptr)
+ colval = map(x -> Int32(x), a.rowval)
+ m = a.n
+ n = a.m
+
+ println("Time taken to transfer matrix to device:")
+ @time ad = CudaSparseMatrixCSR(Float64, CudaArray(rowptr), CudaArray(colval), CudaArray(a.nzval), (m, n))
+
+ b = rand(10^7)
+ println("Time taken to transfer vector: ")
+ @time bd = CudaArray(b)
+
+ println("Time taken for multiplication: ")
+ @time cd = ad * bd
+
+ println("Time taken to transfer back to host: ")
+ @time c = to_host(cd)
+
+ println("Time taken for multiplication on CPU")
+ @time ch = a * b
+
+ facts("Checking for correctness for k = $k") do
+ @fact c => roughly(ch)
+ end
+end
+
+function run_benchmark()
+ for k = 1:9
+ benchmark(k)
+ end
+end

cscmv-with-T/funk.jl

@@ -0,0 +1,6 @@
+using CUSPARSE, CUDArt
+
+import Base.*
+function *(a::CudaSparseMatrixCSR, b::CudaArray)
+ CUSPARSE.csrmv('T', a, b, 'O')
+end