potok/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2026-07-30 15:53:03 +04:00
Code Packages Releases Wiki Activity

added detection & dispatching of some modern NEON instructions (NEON_FP16, NEON_BF16) (#24420)

Browse Source 
* added more or less cross-platform (based on POSIX signal() semantics) method to detect various NEON extensions, such as FP16 SIMD arithmetics, BF16 SIMD arithmetics, SIMD dotprod etc. It could be propagated to other instruction sets if necessary.

* hopefully fixed compile errors

* continue to fix CI

* another attempt to fix build on Linux aarch64

* * reverted to the original method to detect special arm neon instructions without signal()
* renamed FP16_SIMD & BF16_SIMD to NEON_FP16 and NEON_BF16, respectively

* removed extra whitespaces
This commit is contained in:
Vadim Pisarevsky
2023-10-18 22:06:20 +03:00
committed by GitHub
parent 2f1d529a71
commit ba4d6c859d
6 changed files with 130 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/core/src/system.cpp
+25 -11
View File
@@ -408,6 +408,8 @@ struct HWFeatures
g_hwFeatureNames[CPU_NEON] = "NEON";
g_hwFeatureNames[CPU_NEON_DOTPROD] = "NEON_DOTPROD";
g_hwFeatureNames[CPU_NEON_FP16] = "NEON_FP16";
g_hwFeatureNames[CPU_NEON_BF16] = "NEON_BF16";
g_hwFeatureNames[CPU_VSX] = "VSX";
g_hwFeatureNames[CPU_VSX3] = "VSX3";
@@ -566,10 +568,15 @@ struct HWFeatures
while ((size_t)read(cpufile, &auxv, size_auxv_t) == size_auxv_t)
{
// see https://elixir.bootlin.com/linux/latest/source/arch/arm64/include/uapi/asm/hwcap.h
if (auxv.a_type == AT_HWCAP)
{
have[CV_CPU_NEON_DOTPROD] = (auxv.a_un.a_val & (1 << 20)) != 0;
break;
have[CV_CPU_NEON_DOTPROD] = (auxv.a_un.a_val & (1 << 20)) != 0; // HWCAP_ASIMDDP
have[CV_CPU_NEON_FP16] = (auxv.a_un.a_val & (1 << 10)) != 0; // HWCAP_ASIMDHP
}
else if (auxv.a_type == AT_HWCAP2)
{
have[CV_CPU_NEON_BF16] = (auxv.a_un.a_val & (1 << 14)) != 0; // HWCAP2_BF16
}
}
@@ -623,16 +630,23 @@ struct HWFeatures
have[CV_CPU_NEON] = true;
#endif
#if (defined __ARM_FP && (((__ARM_FP & 0x2) != 0) && defined __ARM_NEON__))
have[CV_CPU_FP16] = true;
#endif
#if (defined __ARM_FEATURE_DOTPROD)
int has_feat_dotprod = 0;
size_t has_feat_dotprod_size = sizeof(has_feat_dotprod);
sysctlbyname("hw.optional.arm.FEAT_DotProd", &has_feat_dotprod, &has_feat_dotprod_size, NULL, 0);
if (has_feat_dotprod) {
have[CV_CPU_NEON_DOTPROD] = true;
}
have[CV_CPU_FP16] = have[CV_CPU_NEON_FP16] = true;
#endif
// system.cpp may be compiled w/o special -march=armv8...+dotprod, -march=armv8...+bf16 etc.,
// so we check for the features in any case, no mater what are the compile flags.
// We check the real hardware capabilities here.
int has_feat_dotprod = 0;
size_t has_feat_dotprod_size = sizeof(has_feat_dotprod);
sysctlbyname("hw.optional.arm.FEAT_DotProd", &has_feat_dotprod, &has_feat_dotprod_size, NULL, 0);
if (has_feat_dotprod) {
have[CV_CPU_NEON_DOTPROD] = true;
}
int has_feat_bf16 = 0;
size_t has_feat_bf16_size = sizeof(has_feat_bf16);
sysctlbyname("hw.optional.arm.FEAT_BF16", &has_feat_bf16, &has_feat_bf16_size, NULL, 0);
if (has_feat_bf16) {
have[CV_CPU_NEON_BF16] = true;
}
#elif (defined __clang__)
#if (defined __ARM_NEON__ || (defined __ARM_NEON && defined __aarch64__))
have[CV_CPU_NEON] = true;