static void extraInteger(this ExtraWriter w, string i, bool avx)
{
w.writeExtra(eExtra.int_common, i, "all");
w.writeExtra(eExtra.bitwise, i, "all");
w.line();
w.line(" // ==== Couple missing integer comparison intrinsics ====");
w.writeExtra(eExtra.int_math, i, "epi8");
w.writeExtra(eExtra.int_math, i, "epi16");
w.writeExtra(eExtra.int_math, i, "epi32");
if (avx)
{
w.writeExtra(eExtra.int_math, i, "epi64");
}
w.line();
w.line("#if INTRINSICS_SUPPORT_OPERATORS");
w.writeExtra(eExtra.int_math_ops, i, "epi8", "Int8");
w.writeExtra(eExtra.int_math_ops, i, "epi16", "Int16");
w.writeExtra(eExtra.int_math_ops, i, "epi32", "Int32");
if (avx)
{
w.writeExtra(eExtra.int_math_ops, i, "epi64", "Int64");
}
w.line("#endif // INTRINSICS_SUPPORT_OPERATORS");
}
static void extraFloat(this ExtraWriter w, string f, string suffix, bool avx)
{
if (avx)
{
w.writeExtra(eExtra.float_avx_cmp, f, suffix);
}
w.writeExtra(eExtra.float_math, f, suffix);
w.writeExtra(eExtra.float_cmp_ops, f, suffix);
w.writeExtra(eExtra.bitwise, f, suffix);
}
public override void Write(char value)
{
this.BaseWriter.Write(value);
foreach (TextWriter ExtraWriter in this.ExtraWriters)
{
ExtraWriter.Write(value);
}
base.Write(value);
}
public override void Flush()
{
this.BaseWriter.Flush();
foreach (TextWriter ExtraWriter in this.ExtraWriters)
{
ExtraWriter.Flush();
}
base.Flush();
}
public override void Close()
{
this.BaseWriter.Close();
foreach (TextWriter ExtraWriter in this.ExtraWriters)
{
ExtraWriter.Close();
}
base.Close();
}
protected override void Dispose(bool disposing)
{
if (disposing)
{
this.BaseWriter.Dispose();
foreach (TextWriter ExtraWriter in this.ExtraWriters)
{
ExtraWriter.Dispose();
}
}
base.Dispose(disposing);
}
public static void generateExtras(string destFolder)
{
ExtraCodeParser.copyStatic("fp16c.hpp", destFolder);
string extra = Path.Combine(destFolder, "Extra");
if (!Directory.Exists(extra))
{
Directory.CreateDirectory(extra);
}
ExtraCodeParser.copyStatic("common.hpp", extra);
string f = "__m128";
string d = "__m128d";
string i = "__m128i";
// SSE1
using (var w = new ExtraWriter(extra, "sse-extra", "Sse"))
{
w.writeExtra(eExtra.sse1, f, "ps");
w.extraFloat(f, "ps", false);
}
// SSE2
using (var w = new ExtraWriter(extra, "sse2-extra", "Sse", @"""../sse.hpp"""))
{
// Double stuff
w.writeExtra(eExtra.sse2, d, "pd");
w.extraFloat(d, "pd", false);
w.line();
w.extraInteger(i, false);
}
// SSE 4.2
using (var w = new ExtraWriter(extra, "sse4.2-extra", "Sse", @"""../sse2.hpp""", @"""../sse4.1.hpp""")) // cmpeq_epi64 is SSE 4.1, while cmpgt_epi64 is SSE 4.2. Looks like Intel forgot about that, and fixed later :-)
{
w.writeExtra(eExtra.sse42, null, null);
w.writeExtra(eExtra.int_math, i, "epi64");
w.line();
w.line("#if INTRINSICS_SUPPORT_OPERATORS");
w.writeExtra(eExtra.int_math_ops, i, "epi64", "Int64");
w.line("#endif // INTRINSICS_SUPPORT_OPERATORS");
}
// AVX1
f = "__m256";
d = "__m256d";
i = "__m256i";
using (var w = new ExtraWriter(extra, "avx-extra", "Avx"))
{
w.writeExtra(eExtra.avx1, f, "ps");
w.extraFloat(f, "ps", true);
w.extraFloat(d, "pd", true);
}
// AVX2
using (var w = new ExtraWriter(extra, "avx2-extra", "Avx", @"""../avx.hpp"""))
{
w.writeExtra(eExtra.avx2, f, "ps");
w.extraInteger(i, true);
}
}