Skip to content

GitLab

C
compass-github-pull
Commit 9fc4aea1 authored by Peter Steinbach's avatar Peter Steinbach
Browse files
Options

fixed off-by-1 error in intel cacheline detection

parent f0b01553
Hide whitespace changes
Inline Side-by-side
Showing with 117 additions and 117 deletions
include/detail/rt/x86_sizes.hpp
View file @ 9fc4aea1
......@@ -33,76 +33,76 @@ namespace compass {
std::vector<std::uint32_t> sizes_in_bytes_;
void on_intel(){
void on_intel(){
std::uint32_t maxlevel = 8;//maximum - 1 that can be mapped to 3 bits in eax[7:5]
std::uint32_t eax = 0;
sizes_in_bytes_.reserve(maxlevel);
for(std::uint32_t l = 0;l<maxlevel;++l)
{
auto regs = cpuid(0x04,0,l);
std::uint32_t maxlevel = 8;//maximum - 1 that can be mapped to 3 bits in eax[7:5]
std::uint32_t eax = 0;
sizes_in_bytes_.reserve(maxlevel);
eax = regs[ct::eax];
auto bv = bitview(eax);
for(std::uint32_t l = 0;l<maxlevel;++l)
{
auto regs = cpuid(0x04,0,l);
if(!bv.test(1))//this is not a data cache
continue;
eax = regs[ct::eax];
auto bv = bitview(eax);
auto truelevel = bv.range(5,8);
if(truelevel != l)//this is the wrong level
continue;
if(!bv.test(1))//this is not a data cache
continue;
std::uint32_t value = bitview(regs[ct::ebx]).range(0,11);
sizes_in_bytes_.push_back(value);
auto truelevel = bv.range(5,8);
if(truelevel != l)//this is the wrong level
continue;
}
}
std::uint32_t value = bitview(regs[ct::ebx]).range(0,11);
sizes_in_bytes_.push_back(value+1);
}
}
void on_amd(){
sizes_in_bytes_.reserve(3);
auto regs = cpuid(0x80000005);
void on_amd(){
std::uint32_t ecx = regs[ct::ecx];
auto bv = bitview(ecx);//L1data cache
std::uint32_t linesize = bv.range(0,7);
if(!linesize)//this is not a data cache, as the L1 cacheline size is 0
return;
sizes_in_bytes_.reserve(3);
auto regs = cpuid(0x80000005);
sizes_in_bytes_.push_back(linesize);
std::uint32_t ecx = regs[ct::ecx];
auto bv = bitview(ecx);//L1data cache
std::uint32_t linesize = bv.range(0,7);
if(!linesize)//this is not a data cache, as the L1 cacheline size is 0
return;
auto l23regs = cpuid(0x80000006);
ecx = l23regs[ct::ecx];
auto bv2 = bitview(ecx);//L2 cache
linesize = bv2.range(0,7);
sizes_in_bytes_.push_back(linesize);
sizes_in_bytes_.push_back(linesize);
auto l23regs = cpuid(0x80000006);
ecx = l23regs[ct::ecx];
auto bv2 = bitview(ecx);//L2 cache
linesize = bv2.range(0,7);
sizes_in_bytes_.push_back(linesize);
auto bv3 = bitview(l23regs[ct::edx]);//L3 cache
linesize = bv3.range(0,7);
auto bv3 = bitview(l23regs[ct::edx]);//L3 cache
linesize = bv3.range(0,7);
sizes_in_bytes_.push_back(linesize);
sizes_in_bytes_.push_back(linesize);
}
}
cacheline():
sizes_in_bytes_()
{
auto brand = compass::runtime::detail::vendor( current_arch_t() );
if(brand.find("AMD") != std::string::npos){
on_amd();
}
auto brand = compass::runtime::detail::vendor( current_arch_t() );
if(brand.find("Intel") != std::string::npos){
on_intel();
}
if(brand.find("AMD") != std::string::npos){
on_amd();
}
if(brand.find("Intel") != std::string::npos){
on_intel();
}
}
......@@ -128,90 +128,90 @@ namespace compass {
class cache
{
std::vector<std::uint32_t> sizes_in_bytes_;
//TODO: refactor this sooner than later
void on_intel() {
std::uint32_t eax = 0;
std::uint32_t maxlevel = 8;//maximum - 1 that can be mapped to 3 bits in eax[7:5]
sizes_in_bytes_.reserve(8);
for(std::uint32_t l = 0;l<maxlevel;++l)
{
auto regs = cpuid(0x04,0,l);
eax = regs[ct::eax];
auto bv = bitview(eax);
if(!bv.test(1))//this is not a data cache
continue;
auto truelevel = bv.range(5,8);
if(truelevel != l)//this is the wrong level
continue;
std::uint32_t ebx = regs[ct::ebx];
const bitview bv_ebx = bitview(ebx);
const std::uint32_t ecx = regs[ct::ecx];
std::uint32_t ways = 1 + bv_ebx.range(22,31);
std::uint32_t partitions = 1 + bv_ebx.range(12,21);
std::uint32_t line_size = 1 + bv_ebx.range(0,11);
std::uint32_t sets = 1 + ecx;
std::uint32_t value = ways*partitions*line_size*sets;
sizes_in_bytes_.push_back(value);
}
std::vector<std::uint32_t> sizes_in_bytes_;
}
//TODO: refactor this sooner than later
void on_intel() {
void on_amd(){
std::uint32_t eax = 0;
std::uint32_t maxlevel = 8;//maximum - 1 that can be mapped to 3 bits in eax[7:5]
sizes_in_bytes_.reserve(8);
sizes_in_bytes_.reserve(3);
auto regs = cpuid(0x80000005);
for(std::uint32_t l = 0;l<maxlevel;++l)
{
auto regs = cpuid(0x04,0,l);
std::uint32_t ecx = regs[ct::ecx];
auto bv = bitview(ecx);//L1data cache
std::uint32_t test_linesize = bv.range(0,7);
if(!test_linesize)//this is not a data cache, as the L1 cacheline size is 0
return;
eax = regs[ct::eax];
auto bv = bitview(eax);
sizes_in_bytes_.push_back(bv.range(24,31)*1024);//AMD puts the numbers in kB
if(!bv.test(1))//this is not a data cache
continue;
auto l23regs = cpuid(0x80000006);
ecx = l23regs[ct::ecx];
auto bv2 = bitview(ecx);//L2 cache
auto l2size = bv2.range(16,31);
l2size &= 0xffff;
auto truelevel = bv.range(5,8);
if(truelevel != l)//this is the wrong level
continue;
sizes_in_bytes_.push_back(l2size*1024);//AMD puts the numbers in kB
std::uint32_t ebx = regs[ct::ebx];
const bitview bv_ebx = bitview(ebx);
const std::uint32_t ecx = regs[ct::ecx];
auto bv3 = bitview(l23regs[ct::edx]);
auto l3size = bv3.range(19,31);//AMD manual says bits [18,31], experiments on a Ryzen Threadripper 1900X showed that [19,31] gives the right result
l3size *= 512*1024;
sizes_in_bytes_.push_back(l3size);//AMD puts the numbers in kB
std::uint32_t ways = 1 + bv_ebx.range(22,31);
std::uint32_t partitions = 1 + bv_ebx.range(12,21);
std::uint32_t line_size = 1 + bv_ebx.range(0,11);
std::uint32_t sets = 1 + ecx;
}
std::uint32_t value = ways*partitions*line_size*sets;
sizes_in_bytes_.push_back(value);
}
}
void on_amd(){
sizes_in_bytes_.reserve(3);
auto regs = cpuid(0x80000005);
std::uint32_t ecx = regs[ct::ecx];
auto bv = bitview(ecx);//L1data cache
std::uint32_t test_linesize = bv.range(0,7);
if(!test_linesize)//this is not a data cache, as the L1 cacheline size is 0
return;
sizes_in_bytes_.push_back(bv.range(24,31)*1024);//AMD puts the numbers in kB
auto l23regs = cpuid(0x80000006);
ecx = l23regs[ct::ecx];
auto bv2 = bitview(ecx);//L2 cache
auto l2size = bv2.range(16,31);
l2size &= 0xffff;
sizes_in_bytes_.push_back(l2size*1024);//AMD puts the numbers in kB
auto bv3 = bitview(l23regs[ct::edx]);
auto l3size = bv3.range(19,31);//AMD manual says bits [18,31], experiments on a Ryzen Threadripper 1900X showed that [19,31] gives the right result
l3size *= 512*1024;
sizes_in_bytes_.push_back(l3size);//AMD puts the numbers in kB
}
cache():
sizes_in_bytes_()
sizes_in_bytes_()
{
auto brand = compass::runtime::detail::vendor( current_arch_t() );
if(brand.find("AMD") != std::string::npos){
on_amd();
}
auto brand = compass::runtime::detail::vendor( current_arch_t() );
if(brand.find("AMD") != std::string::npos){
on_amd();
}
if(brand.find("Intel") != std::string::npos){
on_intel();
}
if(brand.find("Intel") != std::string::npos){
on_intel();
}
}
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment