daily update on 20241226 by nitload youmu

This commit is contained in:
Nitload
2024-12-26 03:09:46 +08:00
parent 7c8d462ca5
commit 592ad55d57
94 changed files with 3767 additions and 1484 deletions
+82 -20
View File
@@ -2,14 +2,42 @@
#include "ssc_intrinsic.h"
#include <stdio.h>
#ifdef _MSC_VER
#ifdef _MSC_VER
#include <intrin.h>
#include <isa_availability.h>
#else
#else
#include <cpuid.h>
#endif // _MSC_VER
PRIVATE void load_intrinsic_sse(ssc_intrinsic* context);
PRIVATE void load_intrinsic_sse(ssc_intrinsic *context);
PRIVATE void load_intrinsic_avx(ssc_intrinsic *context);
PRIVATE void load_intrinsic_512(ssc_intrinsic *context);
PRIVATE int check_isa(most_advanced_instruction instruction_family)
{
if (instruction_family != NEON) {
#if _MSC_VER
return __check_isa_support(instruction_family, 0);
#else
switch (instruction_family) {
case SSE:
return __builtin_cpu_supports("sse4.2");
case AVX256:
return __builtin_cpu_supports("avx2");
case AVX512:
return __builtin_cpu_supports("avx512f");
default:
return 0;
}
#endif // _MSC_VER
}
#if defined(__ARM_NEON) || defined(__ARM_NEON__)
return 1;
#else
return 0;
#endif
}
SSCAPI int load_intrinsic_function(void *intrinsic_context)
{
@@ -17,31 +45,65 @@ SSCAPI int load_intrinsic_function(void *intrinsic_context)
ssc_intrinsic *context = (ssc_intrinsic *)intrinsic_context;
if (__check_isa_support(__IA_SUPPORT_VECTOR512,0))
{
}
else if (__check_isa_support(__IA_SUPPORT_VECTOR256, 0))
{
}
else if (__check_isa_support(__IA_SUPPORT_SSE42, 0))
{
if (check_isa(AVX512)) {
load_intrinsic_512(context);
return 3;
} else if (check_isa(AVX256)) {
load_intrinsic_avx(context);
return 2;
} else if (check_isa(SSE)) {
load_intrinsic_sse(context);
}
else
{
return 1;
} else if (check_isa(NEON)) {
return -1;
}
return 0;
}
void load_intrinsic_sse(ssc_intrinsic* context)
void load_intrinsic_sse(ssc_intrinsic *context)
{
context->add_m2 = addmatrix2_sse;
context->add_m3 = ;
context->add_m4 = ;
context->add_m3 = addmatrix3_sse;
context->add_m4 = addmatrix4_sse;
context->transpose_23 = transpose23;
context->transpose_24 = transpose24;
context->transpose_32 = transpose32;
context->transpose_33 = transpose33;
context->transpose_34 = transpose34;
context->transpose_42 = transpose42;
context->transpose_43 = transpose43;
context->transpose_44 = transpose44;
context->normalize_3 = normalize;
context->det_matrix33 = det_mat3;
context->det_matrix44 = det_mat4;
context->string_quick_hash = quick_string_hash;
}
void load_intrinsic_avx(ssc_intrinsic *context)
{
context->add_m2 = addmatrix2_avx;
context->add_m3 = addmatrix3_avx;
context->add_m4 = addmatrix4_avx;
context->transpose_23 = transpose23;
context->transpose_24 = transpose24;
context->transpose_32 = transpose32;
context->transpose_33 = transpose33;
context->transpose_34 = transpose34;
context->transpose_42 = transpose42;
context->transpose_43 = transpose43;
context->transpose_44 = transpose44;
context->normalize_3 = normalize;
context->det_matrix33 = det_mat3;
context->det_matrix44 = det_mat4;
context->string_quick_hash = quick_string_hash;
}
void load_intrinsic_512(ssc_intrinsic *context)
{
context->add_m2 = addmatrix2_avx;
context->add_m3 = addmatrix3_512;
context->add_m4 = addmatrix4_512;
context->transpose_23 = transpose23;
context->transpose_24 = transpose24;
context->transpose_32 = transpose32;