Misaki/Misaki.HighPerformance
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Refactor SPMD to HPC; add SIMD source generators

Browse Source 
Major namespace migration from SPMD to HPC across all code, templates, and projects. Introduced Misaki.HighPerformance.HPC.Generator with Roslyn-based source generators for SIMD code (e.g., AVX2), including attribute and method generators. Renamed MultipleAdd to MultiplyAdd in all lanes and updated usages. Added AVX2 utility methods via codegen. Updated tests, benchmarks, and project references to use the new framework. Improved SIMD memory utilities and modernized project files. Removed legacy SPMD project from the solution.
This commit is contained in:
Misaki
2026-05-06 13:43:58 +09:00
parent d3e497c7d8
commit c8f78f9d02
36 changed files with 895 additions and 130 deletions
Download Patch File Download Diff File Expand all files Collapse all files

90
Misaki.HighPerformance.LowLevel/Utilities/MemoryUtility.Byte.cs
View File

@@ -405,20 +405,54 @@ public static unsafe partial class MemoryUtility
}
var i = 0;
if (Vector.IsHardwareAccelerated && a.Length >= Vector<byte>.Count)
if (Vector512.IsHardwareAccelerated && a.Length >= Vector512<byte>.Count)
{
ref var ptrA = ref MemoryMarshal.GetReference(a);
ref var ptrB = ref MemoryMarshal.GetReference(b);
var limit = a.Length - Vector<byte>.Count;
for (; i <= limit; i += Vector<byte>.Count)
var limit = a.Length - Vector512<byte>.Count;
for (; i <= limit; i += Vector512<byte>.Count)
{
var vecA = Vector.LoadUnsafe(ref ptrA, (nuint)i);
var vecB = Vector.LoadUnsafe(ref ptrB, (nuint)i);
var vecA = Vector512.LoadUnsafe(ref ptrA, (nuint)i);
var vecB = Vector512.LoadUnsafe(ref ptrB, (nuint)i);
var mask = Vector.Equals(vecA, Vector<byte>.Zero);
var mask = Vector512.Equals(vecA, Vector512<byte>.Zero);
var result = Vector.ConditionalSelect(mask, vecB, vecA);
var result = Vector512.ConditionalSelect(mask, vecB, vecA);
result.StoreUnsafe(ref ptrA, (nuint)i);
}
}
else if (Vector256.IsHardwareAccelerated && a.Length >= Vector256<byte>.Count)
{
ref var ptrA = ref MemoryMarshal.GetReference(a);
ref var ptrB = ref MemoryMarshal.GetReference(b);
var limit = a.Length - Vector256<byte>.Count;
for (; i <= limit; i += Vector256<byte>.Count)
{
var vecA = Vector256.LoadUnsafe(ref ptrA, (nuint)i);
var vecB = Vector256.LoadUnsafe(ref ptrB, (nuint)i);
var mask = Vector256.Equals(vecA, Vector256<byte>.Zero);
var result = Vector256.ConditionalSelect(mask, vecB, vecA);
result.StoreUnsafe(ref ptrA, (nuint)i);
}
}
else if (Vector128.IsHardwareAccelerated && a.Length >= Vector128<byte>.Count)
{
ref var ptrA = ref MemoryMarshal.GetReference(a);
ref var ptrB = ref MemoryMarshal.GetReference(b);
var limit = a.Length - Vector128<byte>.Count;
for (; i <= limit; i += Vector128<byte>.Count)
{
var vecA = Vector128.LoadUnsafe(ref ptrA, (nuint)i);
var vecB = Vector128.LoadUnsafe(ref ptrB, (nuint)i);
var mask = Vector128.Equals(vecA, Vector128<byte>.Zero);
var result = Vector128.ConditionalSelect(mask, vecB, vecA);
result.StoreUnsafe(ref ptrA, (nuint)i);
}
}
@@ -440,18 +474,48 @@ public static unsafe partial class MemoryUtility
nuint i = 0u;
if (Vector.IsHardwareAccelerated && length >= (nuint)Vector<byte>.Count)
if (Vector512.IsHardwareAccelerated && length >= (nuint)Vector512<byte>.Count)
{
var vectorSize = (nuint)Vector<byte>.Count;
var vectorSize = (nuint)Vector512<byte>.Count;
var limit = length - vectorSize;
for (; i <= limit; i += vectorSize)
{
var vecA = Vector.Load(ptrA + i);
var vecB = Vector.Load(ptrB + i);
var vecA = Vector512.Load(ptrA + i);
var vecB = Vector512.Load(ptrB + i);
var mask = Vector.Equals(vecA, Vector<byte>.Zero);
var mask = Vector512.Equals(vecA, Vector512<byte>.Zero);
var result = Vector.ConditionalSelect(mask, vecB, vecA);
var result = Vector512.ConditionalSelect(mask, vecB, vecA);
result.Store(ptrA + i);
}
}
else if (Vector256.IsHardwareAccelerated && length >= (nuint)Vector256<byte>.Count)
{
var vectorSize = (nuint)Vector256<byte>.Count;
var limit = length - vectorSize;
for (; i <= limit; i += vectorSize)
{
var vecA = Vector256.Load(ptrA + i);
var vecB = Vector256.Load(ptrB + i);
var mask = Vector256.Equals(vecA, Vector256<byte>.Zero);
var result = Vector256.ConditionalSelect(mask, vecB, vecA);
result.Store(ptrA + i);
}
}
else if (Vector128.IsHardwareAccelerated && length >= (nuint)Vector128<byte>.Count)
{
var vectorSize = (nuint)Vector128<byte>.Count;
var limit = length - vectorSize;
for (; i <= limit; i += vectorSize)
{
var vecA = Vector128.Load(ptrA + i);
var vecB = Vector128.Load(ptrB + i);
var mask = Vector128.Equals(vecA, Vector128<byte>.Zero);
var result = Vector128.ConditionalSelect(mask, vecB, vecA);
result.Store(ptrA + i);
}
}