This is an archive of the discontinued LLVM Phabricator instance.

Differential D18284

Implement get_local_size for AMDGPU (allows us to use get_global_id)
ClosedPublic

Authored by • kzhuravl-AMD on Mar 18 2016, 2:53 PM.

Details

Reviewers
• tstellarAMD
arsenm
Commits
rGf8a81f869fb9: [AMDGPU] Implement get_local_size for amdgcn--amdhsa triple
rL265713: [AMDGPU] Implement get_local_size for amdgcn--amdhsa triple

Diff Detail

Event Timeline

• kzhuravl-AMD updated this revision to Diff 51071.Mar 18 2016, 2:53 PM
• kzhuravl-AMD retitled this revision from to Implement get_local_size for AMDGPU (allows us to use get_global_id).
• kzhuravl-AMD updated this object.
• kzhuravl-AMD added reviewers: arsenm, • tstellarAMD.
arsenm edited edge metadata.Mar 18 2016, 4:25 PM

We need to add a separate HSA ABI triple to the libclc build for this

• kzhuravl-AMD updated this revision to Diff 51826.Mar 28 2016, 12:44 PM
• kzhuravl-AMD edited edge metadata.

Separated into separate HSA ABI triple

arsenm added inline comments.Mar 29 2016, 12:20 AM
configure.py
106

The SI devices should be removed here

• kzhuravl-AMD updated this revision to Diff 51946.Mar 29 2016, 10:29 AM

Removed SI devices

• kzhuravl-AMD marked an inline comment as done.Mar 29 2016, 10:29 AM
• tstellarAMD accepted this revision.Apr 1 2016, 5:09 PM
• tstellarAMD edited edge metadata.

LGTM.

This revision is now accepted and ready to land.Apr 1 2016, 5:09 PM
• kzhuravl-AMD added a comment.Apr 5 2016, 9:32 AM

Ping - Matt

Closed by commit rL265713: [AMDGPU] Implement get_local_size for amdgcn--amdhsa triple (authored by kzhuravl). · Explain WhyApr 7 2016, 12:59 PM
This revision was automatically updated to reflect the committed changes.

Revision Contents

PathSize
1 line
amdgcn-amdhsa/
lib/
1 line
workitem/
35 lines
5 lines

Diff 51946

.gitignore

Makefile Makefile
amdgcn-- amdgcn--
amdgcn--amdhsa
build/*.pyc build/*.pyc
built_libs/ built_libs/
generic-- generic--
generic/lib/convert.cl generic/lib/convert.cl
libclc.pc libclc.pc
nvptx--nvidiacl nvptx--nvidiacl
nvptx64--nvidiacl nvptx64--nvidiacl
r600-- r600--
utils/prepare-builtins utils/prepare-builtins
utils/prepare-builtins.o utils/prepare-builtins.o
utils/prepare-builtins.o.d utils/prepare-builtins.o.d

amdgcn-amdhsa/lib/OVERRIDES

The contents of this file were not changed.

amdgcn-amdhsa/lib/SOURCES

workitem/get_local_size.ll

amdgcn-amdhsa/lib/workitem/get_local_size.ll

declare i8 addrspace(2)* @llvm.amdgcn.dispatch.ptr() #0
define i32 @get_local_size(i32 %dim) #1 {
%dispatch_ptr = call noalias nonnull dereferenceable(64) i8 addrspace(2)* @llvm.amdgcn.dispatch.ptr()
%dispatch_ptr_i32 = bitcast i8 addrspace(2)* %dispatch_ptr to i32 addrspace(2)*
%xy_size_ptr = getelementptr inbounds i32, i32 addrspace(2)* %dispatch_ptr_i32, i64 1
%xy_size = load i32, i32 addrspace(2)* %xy_size_ptr, align 4, !invariant.load !0
switch i32 %dim, label %default [
i32 0, label %x_dim
i32 1, label %y_dim
i32 2, label %z_dim
]
x_dim:
%x_size = and i32 %xy_size, 65535
ret i32 %x_size
y_dim:
%y_size = lshr i32 %xy_size, 16
ret i32 %y_size
z_dim:
%z_size_ptr = getelementptr inbounds i32, i32 addrspace(2)* %dispatch_ptr_i32, i64 2
%z_size = load i32, i32 addrspace(2)* %z_size_ptr, align 4, !invariant.load !0, !range !1
ret i32 %z_size
default:
ret i32 1
}
attributes #0 = { nounwind readnone }
attributes #1 = { alwaysinline norecurse nounwind readonly }
!0 = !{}
!1 = !{ i32 0, i32 257 }

configure.py

Show First 20 Lines • Show All 96 Lines • ▼ Show 20 Lines 'r600--' : { 'devices' :
'defines' : {}}, 'defines' : {}},
{'gpu' : 'barts', 'aliases' : ['turks', 'caicos'], {'gpu' : 'barts', 'aliases' : ['turks', 'caicos'],
'defines' : {}}, 'defines' : {}},
{'gpu' : 'cayman', 'aliases' : ['aruba'], {'gpu' : 'cayman', 'aliases' : ['aruba'],
'defines' : {}} ]}, 'defines' : {}} ]},
'amdgcn--': { 'devices' : 'amdgcn--': { 'devices' :
[{'gpu' : 'tahiti', 'aliases' : ['pitcairn', 'verde', 'oland', 'hainan', 'bonaire', 'kabini', 'kaveri', 'hawaii','mullins','tonga','carrizo','iceland','fiji','stoney'], [{'gpu' : 'tahiti', 'aliases' : ['pitcairn', 'verde', 'oland', 'hainan', 'bonaire', 'kabini', 'kaveri', 'hawaii','mullins','tonga','carrizo','iceland','fiji','stoney'],
'defines' : {}} ]}, 'defines' : {}} ]},
'amdgcn--amdhsa': { 'devices' :
[{'gpu' : '', 'aliases' : ['bonaire', 'hawaii', 'kabini', 'kaveri', 'mullins', 'carrizo', 'stoney', 'fiji', 'iceland', 'tonga'],
arsenmUnsubmitted
Done
ReplyInline Actions

The SI devices should be removed here

arsenm: The SI devices should be removed here
'defines' : {}} ]},
'nvptx--' : { 'devices' : [{'gpu' : '', 'aliases' : [], 'nvptx--' : { 'devices' : [{'gpu' : '', 'aliases' : [],
'defines' : {'all' : ['cl_khr_fp64']}}]}, 'defines' : {'all' : ['cl_khr_fp64']}}]},
'nvptx64--' : { 'devices' : [{'gpu' : '', 'aliases' : [], 'nvptx64--' : { 'devices' : [{'gpu' : '', 'aliases' : [],
'defines' : {'all' : ['cl_khr_fp64']}}]}, 'defines' : {'all' : ['cl_khr_fp64']}}]},
'nvptx--nvidiacl' : { 'devices' : [{'gpu' : '', 'aliases' : [], 'nvptx--nvidiacl' : { 'devices' : [{'gpu' : '', 'aliases' : [],
'defines' : {'all' : ['cl_khr_fp64']}}]}, 'defines' : {'all' : ['cl_khr_fp64']}}]},
'nvptx64--nvidiacl' : { 'devices' : [{'gpu' : '', 'aliases' : [], 'nvptx64--nvidiacl' : { 'devices' : [{'gpu' : '', 'aliases' : [],
'defines' : {'all' : ['cl_khr_fp64']}}]}, 'defines' : {'all' : ['cl_khr_fp64']}}]},
} }
default_targets = ['nvptx--nvidiacl', 'nvptx64--nvidiacl', 'r600--', 'amdgcn--'] default_targets = ['nvptx--nvidiacl', 'nvptx64--nvidiacl', 'r600--', 'amdgcn--', 'amdgcn--amdhsa']
targets = args targets = args
if not targets: if not targets:
targets = default_targets targets = default_targets
b = metabuild.from_name(options.g) b = metabuild.from_name(options.g)
b.rule("LLVM_AS", "%s -o $out $in" % os.path.join(llvm_bindir, "llvm-as"), b.rule("LLVM_AS", "%s -o $out $in" % os.path.join(llvm_bindir, "llvm-as"),
▲ Show 20 Lines • Show All 173 Lines • Show Last 20 Lines