newComputePipelineStateWithFunction failed about floor HOT 7 CLOSED

Hi, what exactly are you trying to do? Which OS/version are you using? What is your kernel code?

from floor.

macOSX 10.13.6,kernel code very simple,like kernel void helloWorld(){} , kernel has be compiled to a metallib file . in our projection, llvm and clang linked be a third-party library,not as command line tools.
NSError *err = nullptr;
id<MTLComputePipelineState> kernel_state = nil;
if ([func functionType] == MTLFunctionTypeKernel) {
auto dev = [prog.second.program device];
kernel_state = [dev newComputePipelineStateWithFunction:func error:&err];
if (!kernel_state) {
log_error(
"failed to create kernel state \"%s\" for device \"%s\": %s",
info.name, prog.first->name,
(err != nullptr ? [[err localizedDescription] UTF8String]
: "unknown error"));
continue;
}
}
failed to create kernel state.

from floor.

Can you upload the metallib file? Also, which GPU?
This error usually occurs when the compiler service / backend compiler crashes.

from floor.

temp_file_or_code.metallib.zip
hi,thx for your response .graphic card is gtx750ti.

kernel's source code
#include <common.hh>
#include <flywave/math/color.hh>

void test1(global flywave::color4<float> *cay, global int *size) {
int gid = get_group_id(0);
int glbid = get_global_id(0);
int lid = get_local_id(0);

int gsize = get_group_size(0);
int gblsize = get_global_size(0);
int lsize = get_local_size(0);

int index = glbid + lid;
barrier();

if (index >= *size)
return;

auto &d = cay[index];
auto aver = (d.r * 0.3 + d.g * 0.59 + d.b * 0.11);
d.r = aver;
d.g = aver;
d.b = aver;
}

kernel void helloWorld(global int *size) {
#if !defined(FLYWAVE_COMPUTE_METAL)
printf("hello world! %d \n", *size);
#endif
}

kernel void helloWorld1(global flywave::color4<float> *data, global int *size) {
test1(data, size);
}

from floor.

Hm, probably shouldn't use doubles.
Other than that, there might be other Nvidia backend compiler issues on 10.13.

from floor.

thanks,I got it.cuda do not work on this computer .once invoke cuda's api,then application will be crashed.there might be a Nvidia webdriver's bug.

from floor.

• closing this, since it's outdated

from floor.