This is an archive of the discontinued LLVM Phabricator instance.

Differential D105858

opencl-c.h: add 3.0 optional extension support for a few more bits
ClosedPublic

Authored by airlied on Jul 12 2021, 4:35 PM.

Details

Reviewers
Anastasia
Commits
rG1854db74c549: opencl-c.h: add 3.0 optional extension support for a few more bits
Summary

These 3 are fairly simple, pipes, workgroups and subgroups.

Diff Detail

Event Timeline

airlied created this revision.Jul 12 2021, 4:35 PM
Herald added subscribers: ldrumm, Anastasia, yaxunl. · View Herald TranscriptJul 12 2021, 4:35 PM
airlied requested review of this revision.Jul 12 2021, 4:35 PM
Herald added a subscriber: cfe-commits. · View Herald TranscriptJul 12 2021, 4:35 PM
Harbormaster completed remote builds in B113614: Diff 358098.Jul 12 2021, 5:17 PM
Anastasia added a subscriber: azabaznov.Jul 13 2021, 9:06 AM
Anastasia added inline comments.
clang/lib/Headers/opencl-c-base.h
336–342

We had a discussion with @azabaznov around features that are aliasing each other and we have discussed to use one feature macro for those. Clang should already ensure that both are set/unset simultaneously? And for those that are not set in clang we can set them correctly here in the header directly.

clang/lib/Headers/opencl-c.h
16159

I think that here guarding the OpenCL version is correct?

azabaznov added inline comments.Jul 13 2021, 9:13 AM
clang/lib/Headers/opencl-c-base.h
336–342

Yeah, I we did. Note that this is applicable to fp64 and 3d image writes, while __openc_c_subgroups and cl_khr_subgroups are not equivalent as extension requires subgroup-independent forward progress but subgroup-independent forward progress is optional in OpenCL C 3.0. I'll try submit a patch for 3d image writes feature macro support this week.

airlied updated this revision to Diff 359114.Jul 15 2021, 1:39 PM

fixed up missing cl_khr_subgroups checks.

Harbormaster completed remote builds in B114342: Diff 359114.Jul 15 2021, 3:36 PM
airlied updated this revision to Diff 359166.Jul 15 2021, 4:41 PM

leave version checks alone since they are inside the outer extension check.

airlied marked an inline comment as done.Jul 15 2021, 4:42 PM

yeah I agree I'll drop those two changes.

Harbormaster completed remote builds in B114381: Diff 359166.Jul 15 2021, 5:54 PM
Anastasia added inline comments.Jul 19 2021, 3:03 PM
clang/lib/Headers/opencl-c-base.h
336–342

Ok, I see so while the functions are identical they are not entirely equivalent extensions so vendors might support one but not the other? In this case I think we should keep checking both but it would be good to add a comment explaining why we are checking both macros here.

Btw do you happen to have spec reference? I can't find anything relevant.

airlied updated this revision to Diff 364372.Aug 5 2021, 12:31 AM

Added a comment on the subgroups difference.

Harbormaster completed remote builds in B118092: Diff 364372.Aug 5 2021, 1:11 AM
Anastasia accepted this revision.Aug 6 2021, 10:24 AM

LGTM! Thanks

This revision is now accepted and ready to land.Aug 6 2021, 10:24 AM
This revision was landed with ongoing or failed builds.Aug 6 2021, 4:25 PM
Closed by commit rG1854db74c549: opencl-c.h: add 3.0 optional extension support for a few more bits (authored by airlied). · Explain Why
This revision was automatically updated to reflect the committed changes.
airlied added a commit: rG1854db74c549: opencl-c.h: add 3.0 optional extension support for a few more bits.

Revision Contents

PathSize
clang/
lib/
Headers/
8 lines
12 lines

Diff 364909

clang/lib/Headers/opencl-c-base.h

Show First 20 Lines • Show All 327 Lines • ▼ Show 20 Linestypedef enum memory_scope {
memory_scope_work_group = __OPENCL_MEMORY_SCOPE_WORK_GROUP, memory_scope_work_group = __OPENCL_MEMORY_SCOPE_WORK_GROUP,
memory_scope_device = __OPENCL_MEMORY_SCOPE_DEVICE, memory_scope_device = __OPENCL_MEMORY_SCOPE_DEVICE,
#if defined(__opencl_c_atomic_scope_all_devices) #if defined(__opencl_c_atomic_scope_all_devices)
memory_scope_all_svm_devices = __OPENCL_MEMORY_SCOPE_ALL_SVM_DEVICES, memory_scope_all_svm_devices = __OPENCL_MEMORY_SCOPE_ALL_SVM_DEVICES,
#if (__OPENCL_C_VERSION__ >= CL_VERSION_3_0) #if (__OPENCL_C_VERSION__ >= CL_VERSION_3_0)
memory_scope_all_devices = memory_scope_all_svm_devices, memory_scope_all_devices = memory_scope_all_svm_devices,
#endif // __OPENCL_C_VERSION__ >= CL_VERSION_3_0 #endif // __OPENCL_C_VERSION__ >= CL_VERSION_3_0
#endif // defined(__opencl_c_atomic_scope_all_devices) #endif // defined(__opencl_c_atomic_scope_all_devices)
#if defined(cl_intel_subgroups) || defined(cl_khr_subgroups) /**
* Subgroups have different requirements on forward progress, so just test
* all the relevant macros.
* CL 3.0 sub-groups "they are not guaranteed to make independent forward progress"
* KHR subgroups "Subgroups within a workgroup are independent, make forward progress with respect to each other"
*/
#if defined(cl_intel_subgroups) || defined(cl_khr_subgroups) || defined(__opencl_c_subgroups)
AnastasiaUnsubmitted
Not Done
ReplyInline Actions

We had a discussion with @azabaznov around features that are aliasing each other and we have discussed to use one feature macro for those. Clang should already ensure that both are set/unset simultaneously? And for those that are not set in clang we can set them correctly here in the header directly.

Anastasia: We had a discussion with @azabaznov around features that are aliasing each other and we have…
azabaznovUnsubmitted
Not Done
ReplyInline Actions

Yeah, I we did. Note that this is applicable to fp64 and 3d image writes, while __openc_c_subgroups and cl_khr_subgroups are not equivalent as extension requires subgroup-independent forward progress but subgroup-independent forward progress is optional in OpenCL C 3.0. I'll try submit a patch for 3d image writes feature macro support this week.

azabaznov: Yeah, I we did. Note that this is applicable to fp64 and 3d image writes, while…
AnastasiaUnsubmitted
Not Done
ReplyInline Actions

Ok, I see so while the functions are identical they are not entirely equivalent extensions so vendors might support one but not the other? In this case I think we should keep checking both but it would be good to add a comment explaining why we are checking both macros here.

Btw do you happen to have spec reference? I can't find anything relevant.

Anastasia: Ok, I see so while the functions are identical they are not entirely equivalent extensions so…
memory_scope_sub_group = __OPENCL_MEMORY_SCOPE_SUB_GROUP memory_scope_sub_group = __OPENCL_MEMORY_SCOPE_SUB_GROUP
#endif #endif
} memory_scope; } memory_scope;
/** /**
* Queue a memory fence to ensure correct ordering of memory * Queue a memory fence to ensure correct ordering of memory
* operations between work-items of a work-group to * operations between work-items of a work-group to
* image memory. * image memory.
▲ Show 20 Lines • Show All 406 Lines • Show Last 20 Lines

clang/lib/Headers/opencl-c.h

  • This file is larger than 256 KB, so syntax highlighting is disabled by default.
Show First 20 Lines • Show All 16,005 Lines • ▼ Show 20 Lines
int __ovld get_image_num_samples(read_write image2d_msaa_depth_t image); int __ovld get_image_num_samples(read_write image2d_msaa_depth_t image);
int __ovld get_image_num_samples(read_write image2d_array_msaa_t image); int __ovld get_image_num_samples(read_write image2d_array_msaa_t image);
int __ovld get_image_num_samples(read_write image2d_array_msaa_depth_t image); int __ovld get_image_num_samples(read_write image2d_array_msaa_depth_t image);
#endif //defined(__OPENCL_CPP_VERSION__) || (__OPENCL_C_VERSION__ >= CL_VERSION_2_0) #endif //defined(__OPENCL_CPP_VERSION__) || (__OPENCL_C_VERSION__ >= CL_VERSION_2_0)
#endif #endif
// OpenCL v2.0 s6.13.15 - Work-group Functions // OpenCL v2.0 s6.13.15 - Work-group Functions
#if defined(__OPENCL_CPP_VERSION__) || (__OPENCL_C_VERSION__ >= CL_VERSION_2_0) #if defined(__opencl_c_work_group_collective_functions)
int __ovld __conv work_group_all(int predicate); int __ovld __conv work_group_all(int predicate);
int __ovld __conv work_group_any(int predicate); int __ovld __conv work_group_any(int predicate);
#ifdef cl_khr_fp16 #ifdef cl_khr_fp16
half __ovld __conv work_group_broadcast(half a, size_t local_id); half __ovld __conv work_group_broadcast(half a, size_t local_id);
half __ovld __conv work_group_broadcast(half a, size_t x, size_t y); half __ovld __conv work_group_broadcast(half a, size_t x, size_t y);
half __ovld __conv work_group_broadcast(half a, size_t x, size_t y, size_t z); half __ovld __conv work_group_broadcast(half a, size_t x, size_t y, size_t z);
#endif #endif
▲ Show 20 Lines • Show All 81 Lines • ▼ Show 20 Lines
double __ovld __conv work_group_scan_exclusive_add(double x); double __ovld __conv work_group_scan_exclusive_add(double x);
double __ovld __conv work_group_scan_exclusive_min(double x); double __ovld __conv work_group_scan_exclusive_min(double x);
double __ovld __conv work_group_scan_exclusive_max(double x); double __ovld __conv work_group_scan_exclusive_max(double x);
double __ovld __conv work_group_scan_inclusive_add(double x); double __ovld __conv work_group_scan_inclusive_add(double x);
double __ovld __conv work_group_scan_inclusive_min(double x); double __ovld __conv work_group_scan_inclusive_min(double x);
double __ovld __conv work_group_scan_inclusive_max(double x); double __ovld __conv work_group_scan_inclusive_max(double x);
#endif //cl_khr_fp64 #endif //cl_khr_fp64
#endif //defined(__OPENCL_CPP_VERSION__) || (__OPENCL_C_VERSION__ >= CL_VERSION_2_0) #endif //defined(__opencl_c_work_group_collective_functions)
// OpenCL v2.0 s6.13.16 - Pipe Functions // OpenCL v2.0 s6.13.16 - Pipe Functions
#if defined(__OPENCL_CPP_VERSION__) || (__OPENCL_C_VERSION__ >= CL_VERSION_2_0) #if defined(__opencl_c_pipes)
bool __ovld is_valid_reserve_id(reserve_id_t reserve_id); bool __ovld is_valid_reserve_id(reserve_id_t reserve_id);
#endif //defined(__OPENCL_CPP_VERSION__) || (__OPENCL_C_VERSION__ >= CL_VERSION_2_0) #endif //defined(__opencl_c_pipes)
// OpenCL v2.0 s6.13.17 - Enqueue Kernels // OpenCL v2.0 s6.13.17 - Enqueue Kernels
#if defined(__OPENCL_CPP_VERSION__) || (__OPENCL_C_VERSION__ >= CL_VERSION_2_0) #if defined(__OPENCL_CPP_VERSION__) || (__OPENCL_C_VERSION__ >= CL_VERSION_2_0)
ndrange_t __ovld ndrange_1D(size_t); ndrange_t __ovld ndrange_1D(size_t);
ndrange_t __ovld ndrange_1D(size_t, size_t); ndrange_t __ovld ndrange_1D(size_t, size_t);
ndrange_t __ovld ndrange_1D(size_t, size_t, size_t); ndrange_t __ovld ndrange_1D(size_t, size_t, size_t);
Show All 20 Lines
void __ovld capture_event_profiling_info(clk_event_t, clk_profiling_info, __global void* value); void __ovld capture_event_profiling_info(clk_event_t, clk_profiling_info, __global void* value);
queue_t __ovld get_default_queue(void); queue_t __ovld get_default_queue(void);
#endif //defined(__OPENCL_CPP_VERSION__) || (__OPENCL_C_VERSION__ >= CL_VERSION_2_0) #endif //defined(__OPENCL_CPP_VERSION__) || (__OPENCL_C_VERSION__ >= CL_VERSION_2_0)
// OpenCL Extension v2.0 s9.17 - Sub-groups // OpenCL Extension v2.0 s9.17 - Sub-groups
#if defined(cl_intel_subgroups) || defined(cl_khr_subgroups) #if defined(cl_intel_subgroups) || defined(cl_khr_subgroups) || defined(__opencl_c_subgroups)
// Shared Sub Group Functions // Shared Sub Group Functions
uint __ovld get_sub_group_size(void); uint __ovld get_sub_group_size(void);
uint __ovld get_max_sub_group_size(void); uint __ovld get_max_sub_group_size(void);
uint __ovld get_num_sub_groups(void); uint __ovld get_num_sub_groups(void);
#if defined(__OPENCL_CPP_VERSION__) || (__OPENCL_C_VERSION__ >= CL_VERSION_2_0) #if defined(__OPENCL_CPP_VERSION__) || (__OPENCL_C_VERSION__ >= CL_VERSION_2_0)
AnastasiaUnsubmitted
Done
ReplyInline Actions

I think that here guarding the OpenCL version is correct?

Anastasia: I think that here guarding the OpenCL version is correct?
uint __ovld get_enqueued_num_sub_groups(void); uint __ovld get_enqueued_num_sub_groups(void);
#endif //defined(__OPENCL_CPP_VERSION__) || (__OPENCL_C_VERSION__ >= CL_VERSION_2_0) #endif //defined(__OPENCL_CPP_VERSION__) || (__OPENCL_C_VERSION__ >= CL_VERSION_2_0)
uint __ovld get_sub_group_id(void); uint __ovld get_sub_group_id(void);
uint __ovld get_sub_group_local_id(void); uint __ovld get_sub_group_local_id(void);
void __ovld __conv sub_group_barrier(cl_mem_fence_flags flags); void __ovld __conv sub_group_barrier(cl_mem_fence_flags flags);
#if defined(__OPENCL_CPP_VERSION__) || (__OPENCL_C_VERSION__ >= CL_VERSION_2_0) #if defined(__OPENCL_CPP_VERSION__) || (__OPENCL_C_VERSION__ >= CL_VERSION_2_0)
void __ovld __conv sub_group_barrier(cl_mem_fence_flags flags, memory_scope scope); void __ovld __conv sub_group_barrier(cl_mem_fence_flags flags, memory_scope scope);
▲ Show 20 Lines • Show All 77 Lines • ▼ Show 20 Lines
double __ovld __conv sub_group_scan_exclusive_add(double x); double __ovld __conv sub_group_scan_exclusive_add(double x);
double __ovld __conv sub_group_scan_exclusive_min(double x); double __ovld __conv sub_group_scan_exclusive_min(double x);
double __ovld __conv sub_group_scan_exclusive_max(double x); double __ovld __conv sub_group_scan_exclusive_max(double x);
double __ovld __conv sub_group_scan_inclusive_add(double x); double __ovld __conv sub_group_scan_inclusive_add(double x);
double __ovld __conv sub_group_scan_inclusive_min(double x); double __ovld __conv sub_group_scan_inclusive_min(double x);
double __ovld __conv sub_group_scan_inclusive_max(double x); double __ovld __conv sub_group_scan_inclusive_max(double x);
#endif //cl_khr_fp64 #endif //cl_khr_fp64
#endif //cl_khr_subgroups cl_intel_subgroups #endif //cl_khr_subgroups cl_intel_subgroups __opencl_c_subgroups
#if defined(cl_khr_subgroup_extended_types) #if defined(cl_khr_subgroup_extended_types)
char __ovld __conv sub_group_broadcast( char value, uint index ); char __ovld __conv sub_group_broadcast( char value, uint index );
char2 __ovld __conv sub_group_broadcast( char2 value, uint index ); char2 __ovld __conv sub_group_broadcast( char2 value, uint index );
char3 __ovld __conv sub_group_broadcast( char3 value, uint index ); char3 __ovld __conv sub_group_broadcast( char3 value, uint index );
char4 __ovld __conv sub_group_broadcast( char4 value, uint index ); char4 __ovld __conv sub_group_broadcast( char4 value, uint index );
char8 __ovld __conv sub_group_broadcast( char8 value, uint index ); char8 __ovld __conv sub_group_broadcast( char8 value, uint index );
char16 __ovld __conv sub_group_broadcast( char16 value, uint index ); char16 __ovld __conv sub_group_broadcast( char16 value, uint index );
▲ Show 20 Lines • Show All 1,907 Lines • Show Last 20 Lines