PairedBench vMulPerf(int samples, long cycles)
{
var lhs1 = RVec <double>(samples);
var rhs1 = RVec <double>(samples);
var dst1 = RowVectors.blockalloc <double>(samples);
var sw1 = Time.stopwatch(true);
for (var i = 0; i < cycles; i++)
{
Calcs.mul(lhs1, rhs1, dst1.Unblocked);
}
var time1 = BenchmarkRecord.Define(cycles, Time.snapshot(sw1), "gmath");
var lhs2 = lhs1.Replicate();
var rhs2 = rhs1.Replicate();
var dst2 = dst1.Replicate();
var sw2 = Time.stopwatch(true);
for (var i = 0; i < cycles; i++)
{
mkl.mul(lhs2, rhs2, ref dst2);
}
var time2 = BenchmarkRecord.Define(cycles, Time.snapshot(sw2), "mkl");
return(time1, time2);
}
BenchmarkRecord gemm_direct_check <M, K, N>(IEnumerable <float> src, M m = default, K k = default, N n = default)
where M : unmanaged, ITypeNat
where K : unmanaged, ITypeNat
where N : unmanaged, ITypeNat
{
var A = Matrix.blockalloc <M, K, float>();
var B = Matrix.blockalloc <K, N, float>();
var X = Matrix.blockalloc <M, N, float>();
var E = Matrix.blockalloc <M, N, float>();
var collect = false;
var runtime = Duration.Zero;
for (var i = 0; i < CycleCount; i++)
{
src.StreamTo(A.Unblocked);
src.StreamTo(B.Unblocked);
var sw = stopwatch();
mkl.gemm(A, B, ref X);
runtime += snapshot(sw);
}
var label = $"gemm<{nat32i<M>()},{nat32i<K>()},{nat32i<N>()}>";
BenchmarkRecord timing = Benchmark(CycleCount, runtime, label);
if (collect)
{
ReportBenchmark(label, CycleCount, TimeSpan.FromMilliseconds(runtime.Ms));
}
return(timing);
}
BenchmarkRecord gemm_check <M, K, N, T>(IEnumerable <T> src, T epsilon = default, M m = default, K k = default, N n = default, bool trace = false)
where M : unmanaged, ITypeNat
where K : unmanaged, ITypeNat
where N : unmanaged, ITypeNat
where T : unmanaged
{
var A = Matrix.blockalloc <M, K, T>();
var B = Matrix.blockalloc <K, N, T>();
var X = Matrix.blockalloc <M, N, T>();
var XU = X.Unblocked;
var E = Matrix.blockalloc <M, N, T>();
var EU = E.Unblocked;
var collect = false;
var label = $"gemm<N{nat32i<M>()},N{nat32i<K>()},N{nat32i<N>()},{typeof(T).Name}>";
var runtime = Duration.Zero;
for (var i = 0; i < CycleCount; i++)
{
src.StreamTo(A.Unblocked);
src.StreamTo(B.Unblocked);
var sw = stopwatch();
mkl.gemm(A, B, ref X);
runtime += snapshot(sw);
Matrix.mul(A, B, ref E);
if (trace)
{
var padlen = Int32.MinValue.ToString().Length + 2;
Notify($"X = {X.Format()}");
Notify($"E = {E.Format()}");
}
Claim.close(E.Unblocked, X.Unblocked, epsilon);
}
BenchmarkRecord timing = Benchmark(CycleCount, runtime, label);
if (collect)
{
ReportBenchmark(label, CycleCount, TimeSpan.FromMilliseconds(runtime.Ms));
}
return(timing);
}
void rng_bench <T>(IDataStream <T> stream, [CallerMemberName] string caller = null)
where T : unmanaged
{
var segment = Pow2.T08;
var total = Pow2.T17;
var stats = StatCollector.Create(0.0);
var sw = stopwatch(false);
for (var i = 0; i < total; i += segment)
{
sw.Start();
var sample = stream.TakeArray(segment);
sw.Stop();
for (var j = 0; j < segment; j++)
{
stats.Collect(Numeric.force <T, double>(sample[j]));
}
}
var opname = $"{caller}<{typeof(T).DisplayName()}>";
Deposit(BenchmarkRecord.Define(total, sw.Elapsed, opname));
}
