shylie/llvm-project
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
llvm-project/llvm/test/CodeGen/SPIRV/transcoding/group_ops.ll
Vyacheslav Levytskyy 54cc4141e4
[SPIR-V] Rework duplicate tracker and tracking of IR entities and types to improve compile-time performance (#130605) 
This PR is to thoroughly rework duplicate tracker implementation and
tracking of IR entities and types. These are legacy parts of the project
resulting in an extremely bloated intermediate representation and
computational delays due to inefficient data flow and structure choices.

Main results of the rework:

1) Improved compile-time performance. The reference binary LLVM IR used
to measure speed gains in
https://github.com/llvm/llvm-project/pull/120415 shows ~x5 speed up also
after this PR. The timing before this PR is ~42s and after this PR it's
~7.5s. In total this PR and the previous overhaul of the module analysis
in https://github.com/llvm/llvm-project/pull/120415 results in ~x25
speed improvement.
```
$ time llc -O0 -mtriple=spirv64v1.6-unknown-unknown _group_barrier_phi.bc -o 1 --filetype=obj

real    0m7.545s
user    0m6.685s
sys     0m0.859s
```

2) Less bloated intermediate representation of internal translation
steps. Elimination of `spv_track_constant` intrinsic usage for scalar
constants, rework of `spv_assign_name`, removal of the gMIR `GET_XXX`
pseudo code and a smaller number of generated `ASSIGN_TYPE` pseudo codes
substantially decrease volume of data generated during translation.

3) Simpler code and easier maintenance. The duplicate tracker
implementation is simplified, as well as other features.

4) Numerous fixes of issues and logical flaws in different passes. The
main achievement is rework of the duplicate tracker itself that had
never guaranteed a correct caching of LLVM IR entities, rarely and
randomly returning stale/incorrect records (like, remove an instruction
from gMIR but still refer to it). Other fixes comprise consistent
generation of OpConstantNull, assigning types to newly created
registers, creation of integer/bool types, and other minor fixes.

5) Numerous fixes of LIT tests: mainly CHECK-DAG to properly reflect
SPIR-V spec guarantees, `{{$}}` at the end of constants to avoid
matching of substrings, and XFAILS for `SPV_INTEL_long_composites` test
cases, because the feature is not completed in full yet and doesn't
generate a requested by the extension sequence of instructions.

6) New test cases are added.
2025-03-26 17:58:10 +01:00

316 lines
14 KiB
LLVM
Raw Blame History

; RUN: llc -verify-machineinstrs -O0 -mtriple=spirv64-unknown-unknown %s -o - | FileCheck %s --check-prefix=CHECK-SPIRV
; RUN: %if spirv-tools %{ llc -O0 -mtriple=spirv64-unknown-unknown %s -o - -filetype=obj | spirv-val %}
; RUN: llc -verify-machineinstrs -O0 -mtriple=spirv32-unknown-unknown %s -o - | FileCheck %s --check-prefix=CHECK-SPIRV
; RUN: %if spirv-tools %{ llc -O0 -mtriple=spirv32-unknown-unknown %s -o - -filetype=obj | spirv-val %}
; CHECK-SPIRV-DAG: %[[#int:]] = OpTypeInt 32 0
; CHECK-SPIRV-DAG: %[[#intv2:]] = OpTypeVector %[[#int]] 2
; CHECK-SPIRV-DAG: %[[#intv3:]] = OpTypeVector %[[#int]] 3
; CHECK-SPIRV-DAG: %[[#float:]] = OpTypeFloat 32
; CHECK-SPIRV-DAG: %[[#ScopeCrossWorkgroup:]] = OpConstantNull %[[#int]]
; CHECK-SPIRV-DAG: %[[#ScopeWorkgroup:]] = OpConstant %[[#int]] 2{{$}}
; CHECK-SPIRV-DAG: %[[#ScopeSubgroup:]] = OpConstant %[[#int]] 3{{$}}
; CHECK-SPIRV: OpFunction
; CHECK-SPIRV: %[[#]] = OpGroupFMax %[[#float]] %[[#ScopeWorkgroup]] Reduce
; CHECK-SPIRV: OpFunctionEnd
;; kernel void testWorkGroupFMax(float a, global float *res) {
;; res[0] = work_group_reduce_max(a);
;; }
define dso_local spir_kernel void @testWorkGroupFMax(float noundef %a, float addrspace(1)* nocapture noundef writeonly %res) local_unnamed_addr {
entry:
%call = call spir_func float @_Z21work_group_reduce_maxf(float noundef %a)
store float %call, float addrspace(1)* %res, align 4
ret void
}
declare spir_func float @_Z21work_group_reduce_maxf(float noundef) local_unnamed_addr
; CHECK-SPIRV: OpFunction
; CHECK-SPIRV: %[[#]] = OpGroupFMin %[[#float]] %[[#ScopeWorkgroup]] Reduce
; CHECK-SPIRV: OpFunctionEnd
;; kernel void testWorkGroupFMin(float a, global float *res) {
;; res[0] = work_group_reduce_min(a);
;; }
define dso_local spir_kernel void @testWorkGroupFMin(float noundef %a, float addrspace(1)* nocapture noundef writeonly %res) local_unnamed_addr {
entry:
%call = call spir_func float @_Z21work_group_reduce_minf(float noundef %a)
store float %call, float addrspace(1)* %res, align 4
ret void
}
declare spir_func float @_Z21work_group_reduce_minf(float noundef) local_unnamed_addr
; CHECK-SPIRV: OpFunction
; CHECK-SPIRV: %[[#]] = OpGroupFAdd %[[#float]] %[[#ScopeWorkgroup]] Reduce
; CHECK-SPIRV: OpFunctionEnd
;; kernel void testWorkGroupFAdd(float a, global float *res) {
;; res[0] = work_group_reduce_add(a);
;; }
define dso_local spir_kernel void @testWorkGroupFAdd(float noundef %a, float addrspace(1)* nocapture noundef writeonly %res) local_unnamed_addr {
entry:
%call = call spir_func float @_Z21work_group_reduce_addf(float noundef %a)
store float %call, float addrspace(1)* %res, align 4
ret void
}
declare spir_func float @_Z21work_group_reduce_addf(float noundef) local_unnamed_addr
; CHECK-SPIRV: OpFunction
; CHECK-SPIRV: %[[#]] = OpGroupFMax %[[#float]] %[[#ScopeWorkgroup]] InclusiveScan
; CHECK-SPIRV: OpFunctionEnd
;; kernel void testWorkGroupScanInclusiveFMax(float a, global float *res) {
;; res[0] = work_group_scan_inclusive_max(a);
;; }
define dso_local spir_kernel void @testWorkGroupScanInclusiveFMax(float noundef %a, float addrspace(1)* nocapture noundef writeonly %res) local_unnamed_addr {
entry:
%call = call spir_func float @_Z29work_group_scan_inclusive_maxf(float noundef %a)
store float %call, float addrspace(1)* %res, align 4
ret void
}
declare spir_func float @_Z29work_group_scan_inclusive_maxf(float noundef) local_unnamed_addr
; CHECK-SPIRV: OpFunction
; CHECK-SPIRV: %[[#]] = OpGroupFMax %[[#float]] %[[#ScopeWorkgroup]] ExclusiveScan
; CHECK-SPIRV: OpFunctionEnd
;; kernel void testWorkGroupScanExclusiveFMax(float a, global float *res) {
;; res[0] = work_group_scan_exclusive_max(a);
;; }
define dso_local spir_kernel void @testWorkGroupScanExclusiveFMax(float noundef %a, float addrspace(1)* nocapture noundef writeonly %res) local_unnamed_addr {
entry:
%call = call spir_func float @_Z29work_group_scan_exclusive_maxf(float noundef %a)
store float %call, float addrspace(1)* %res, align 4
ret void
}
declare spir_func float @_Z29work_group_scan_exclusive_maxf(float noundef) local_unnamed_addr
; CHECK-SPIRV: OpFunction
; CHECK-SPIRV: %[[#]] = OpGroupSMax %[[#int]] %[[#ScopeWorkgroup]] Reduce
; CHECK-SPIRV: OpFunctionEnd
;; kernel void testWorkGroupSMax(int a, global int *res) {
;; res[0] = work_group_reduce_max(a);
;; }
define dso_local spir_kernel void @testWorkGroupSMax(i32 noundef %a, i32 addrspace(1)* nocapture noundef writeonly %res) local_unnamed_addr {
entry:
%call = call spir_func i32 @_Z21work_group_reduce_maxi(i32 noundef %a)
store i32 %call, i32 addrspace(1)* %res, align 4
ret void
}
declare spir_func i32 @_Z21work_group_reduce_maxi(i32 noundef) local_unnamed_addr
; CHECK-SPIRV: OpFunction
; CHECK-SPIRV: %[[#]] = OpGroupSMin %[[#int]] %[[#ScopeWorkgroup]] Reduce
; CHECK-SPIRV: OpFunctionEnd
;; kernel void testWorkGroupSMin(int a, global int *res) {
;; res[0] = work_group_reduce_min(a);
;; }
define dso_local spir_kernel void @testWorkGroupSMin(i32 noundef %a, i32 addrspace(1)* nocapture noundef writeonly %res) local_unnamed_addr {
entry:
%call = call spir_func i32 @_Z21work_group_reduce_mini(i32 noundef %a)
store i32 %call, i32 addrspace(1)* %res, align 4
ret void
}
declare spir_func i32 @_Z21work_group_reduce_mini(i32 noundef) local_unnamed_addr
; CHECK-SPIRV: OpFunction
; CHECK-SPIRV: %[[#]] = OpGroupIAdd %[[#int]] %[[#ScopeWorkgroup]] Reduce
; CHECK-SPIRV: OpFunctionEnd
;; kernel void testWorkGroupIAddSigned(int a, global int *res) {
;; res[0] = work_group_reduce_add(a);
;; }
define dso_local spir_kernel void @testWorkGroupIAddSigned(i32 noundef %a, i32 addrspace(1)* nocapture noundef writeonly %res) local_unnamed_addr {
entry:
%call = call spir_func i32 @_Z21work_group_reduce_addi(i32 noundef %a)
store i32 %call, i32 addrspace(1)* %res, align 4
ret void
}
declare spir_func i32 @_Z21work_group_reduce_addi(i32 noundef) local_unnamed_addr
; CHECK-SPIRV: OpFunction
; CHECK-SPIRV: %[[#]] = OpGroupIAdd %[[#int]] %[[#ScopeWorkgroup]] Reduce
; CHECK-SPIRV: OpFunctionEnd
;; kernel void testWorkGroupIAddUnsigned(uint a, global uint *res) {
;; res[0] = work_group_reduce_add(a);
;; }
define dso_local spir_kernel void @testWorkGroupIAddUnsigned(i32 noundef %a, i32 addrspace(1)* nocapture noundef writeonly %res) local_unnamed_addr {
entry:
%call = call spir_func i32 @_Z21work_group_reduce_addj(i32 noundef %a)
store i32 %call, i32 addrspace(1)* %res, align 4
ret void
}
declare spir_func i32 @_Z21work_group_reduce_addj(i32 noundef) local_unnamed_addr
; CHECK-SPIRV: OpFunction
; CHECK-SPIRV: %[[#]] = OpGroupUMax %[[#int]] %[[#ScopeWorkgroup]] Reduce
; CHECK-SPIRV: OpFunctionEnd
;; kernel void testWorkGroupUMax(uint a, global uint *res) {
;; res[0] = work_group_reduce_max(a);
;; }
define dso_local spir_kernel void @testWorkGroupUMax(i32 noundef %a, i32 addrspace(1)* nocapture noundef writeonly %res) local_unnamed_addr {
entry:
%call = call spir_func i32 @_Z21work_group_reduce_maxj(i32 noundef %a)
store i32 %call, i32 addrspace(1)* %res, align 4
ret void
}
declare spir_func i32 @_Z21work_group_reduce_maxj(i32 noundef) local_unnamed_addr
; CHECK-SPIRV: OpFunction
; CHECK-SPIRV: %[[#]] = OpGroupUMax %[[#int]] %[[#ScopeSubgroup]] Reduce
; CHECK-SPIRV: OpFunctionEnd
;; #pragma OPENCL EXTENSION cl_khr_subgroups: enable
;; kernel void testSubGroupUMax(uint a, global uint *res) {
;; res[0] = sub_group_reduce_max(a);
;; }
;; #pragma OPENCL EXTENSION cl_khr_subgroups: disable
define dso_local spir_kernel void @testSubGroupUMax(i32 noundef %a, i32 addrspace(1)* nocapture noundef writeonly %res) local_unnamed_addr {
entry:
%call = call spir_func i32 @_Z20sub_group_reduce_maxj(i32 noundef %a)
store i32 %call, i32 addrspace(1)* %res, align 4
ret void
}
declare spir_func i32 @_Z20sub_group_reduce_maxj(i32 noundef) local_unnamed_addr
; CHECK-SPIRV: OpFunction
; CHECK-SPIRV: %[[#]] = OpGroupUMax %[[#int]] %[[#ScopeWorkgroup]] InclusiveScan
; CHECK-SPIRV: OpFunctionEnd
;; kernel void testWorkGroupScanInclusiveUMax(uint a, global uint *res) {
;; res[0] = work_group_scan_inclusive_max(a);
;; }
define dso_local spir_kernel void @testWorkGroupScanInclusiveUMax(i32 noundef %a, i32 addrspace(1)* nocapture noundef writeonly %res) local_unnamed_addr {
entry:
%call = call spir_func i32 @_Z29work_group_scan_inclusive_maxj(i32 noundef %a)
store i32 %call, i32 addrspace(1)* %res, align 4
ret void
}
declare spir_func i32 @_Z29work_group_scan_inclusive_maxj(i32 noundef) local_unnamed_addr
; CHECK-SPIRV: OpFunction
; CHECK-SPIRV: %[[#]] = OpGroupUMax %[[#int]] %[[#ScopeWorkgroup]] ExclusiveScan
; CHECK-SPIRV: OpFunctionEnd
;; kernel void testWorkGroupScanExclusiveUMax(uint a, global uint *res) {
;; res[0] = work_group_scan_exclusive_max(a);
;; }
define dso_local spir_kernel void @testWorkGroupScanExclusiveUMax(i32 noundef %a, i32 addrspace(1)* nocapture noundef writeonly %res) local_unnamed_addr {
entry:
%call = call spir_func i32 @_Z29work_group_scan_exclusive_maxj(i32 noundef %a)
store i32 %call, i32 addrspace(1)* %res, align 4
ret void
}
declare spir_func i32 @_Z29work_group_scan_exclusive_maxj(i32 noundef) local_unnamed_addr
; CHECK-SPIRV: OpFunction
; CHECK-SPIRV: %[[#]] = OpGroupUMin %[[#int]] %[[#ScopeWorkgroup]] Reduce
; CHECK-SPIRV: OpFunctionEnd
;; kernel void testWorkGroupUMin(uint a, global uint *res) {
;; res[0] = work_group_reduce_min(a);
;; }
define dso_local spir_kernel void @testWorkGroupUMin(i32 noundef %a, i32 addrspace(1)* nocapture noundef writeonly %res) local_unnamed_addr {
entry:
%call = call spir_func i32 @_Z21work_group_reduce_minj(i32 noundef %a)
store i32 %call, i32 addrspace(1)* %res, align 4
ret void
}
declare spir_func i32 @_Z21work_group_reduce_minj(i32 noundef) local_unnamed_addr
; CHECK-SPIRV: OpFunction
; CHECK-SPIRV: %[[#]] = OpGroupBroadcast %[[#int]] %[[#ScopeWorkgroup]] %[[#BroadcastValue:]] %[[#BroadcastLocalId:]]
; CHECK-SPIRV: %[[#BroadcastVec2:]] = OpCompositeConstruct %[[#intv2]] %[[#BroadcastLocalId]] %[[#BroadcastLocalId]]
; CHECK-SPIRV: %[[#]] = OpGroupBroadcast %[[#int]] %[[#ScopeWorkgroup]] %[[#BroadcastValue]] %[[#BroadcastVec2]]
; CHECK-SPIRV: %[[#BroadcastVec3:]] = OpCompositeConstruct %[[#intv3]] %[[#BroadcastLocalId]] %[[#BroadcastLocalId]] %[[#BroadcastLocalId]]
; CHECK-SPIRV: %[[#]] = OpGroupBroadcast %[[#int]] %[[#ScopeWorkgroup]] %[[#BroadcastValue]] %[[#BroadcastVec3]]
; CHECK-SPIRV: %[[#]] = OpGroupBroadcast %[[#int]] %[[#ScopeCrossWorkgroup]] %[[#BroadcastValue]] %[[#BroadcastLocalId]]
; CHECK-SPIRV: OpFunctionEnd
;; kernel void testWorkGroupBroadcast(uint a, global size_t *id, global int *res) {
;; res[0] = work_group_broadcast(a, *id);
;; }
define dso_local spir_kernel void @testWorkGroupBroadcast(i32 noundef %a, i32 addrspace(1)* nocapture noundef readonly %id, i32 addrspace(1)* nocapture noundef writeonly %res) local_unnamed_addr {
entry:
%0 = load i32, i32 addrspace(1)* %id, align 4
%call = call spir_func i32 @_Z20work_group_broadcastjj(i32 noundef %a, i32 noundef %0)
%call_v2 = call spir_func i32 @_Z20work_group_broadcastjj(i32 noundef %a, i32 noundef %0, i32 noundef %0)
%call_v3 = call spir_func i32 @_Z20work_group_broadcastjj(i32 noundef %a, i32 noundef %0, i32 noundef %0, i32 noundef %0)
store i32 %call, i32 addrspace(1)* %res, align 4
%call1 = call spir_func i32 @__spirv_GroupBroadcast(i32 0, i32 noundef %a, i32 noundef %0)
ret void
}
declare spir_func i32 @_Z20work_group_broadcastjj(i32 noundef, i32 noundef) local_unnamed_addr
declare spir_func i32 @_Z20work_group_broadcastjjj(i32 noundef, i32 noundef, i32 noundef) local_unnamed_addr
declare spir_func i32 @_Z20work_group_broadcastjjjj(i32 noundef, i32 noundef, i32 noundef, i32 noundef) local_unnamed_addr
declare spir_func i32 @__spirv_GroupBroadcast(i32 noundef, i32 noundef, i32 noundef) local_unnamed_addr
; CHECK-SPIRV: OpFunction
; CHECK-SPIRV: %[[#]] = OpGroupFAdd %[[#float]] %[[#ScopeCrossWorkgroup]] Reduce %[[#FValue:]]
; CHECK-SPIRV: %[[#]] = OpGroupFMin %[[#float]] %[[#ScopeWorkgroup]] InclusiveScan %[[#FValue]]
; CHECK-SPIRV: %[[#]] = OpGroupFMax %[[#float]] %[[#ScopeSubgroup]] ExclusiveScan %[[#FValue]]
; CHECK-SPIRV: %[[#]] = OpGroupIAdd %[[#int]] %[[#ScopeCrossWorkgroup]] Reduce %[[#IValue:]]
; CHECK-SPIRV: %[[#]] = OpGroupUMin %[[#int]] %[[#ScopeWorkgroup]] InclusiveScan %[[#IValue]]
; CHECK-SPIRV: %[[#]] = OpGroupSMin %[[#int]] %[[#ScopeSubgroup]] ExclusiveScan %[[#IValue]]
; CHECK-SPIRV: %[[#]] = OpGroupUMax %[[#int]] %[[#ScopeCrossWorkgroup]] Reduce %[[#IValue]]
; CHECK-SPIRV: %[[#]] = OpGroupSMax %[[#int]] %[[#ScopeWorkgroup]] InclusiveScan %[[#IValue]]
; CHECK-SPIRV: OpFunctionEnd
define spir_kernel void @foo(float %a, i32 %b) {
entry:
%f1 = call spir_func float @__spirv_GroupFAdd(i32 0, i32 0, float %a)
%f2 = call spir_func float @__spirv_GroupFMin(i32 2, i32 1, float %a)
%f3 = call spir_func float @__spirv_GroupFMax(i32 3, i32 2, float %a)
%i1 = call spir_func i32 @__spirv_GroupIAdd(i32 0, i32 0, i32 %b)
%i2 = call spir_func i32 @__spirv_GroupUMin(i32 2, i32 1, i32 %b)
%i3 = call spir_func i32 @__spirv_GroupSMin(i32 3, i32 2, i32 %b)
%i4 = call spir_func i32 @__spirv_GroupUMax(i32 0, i32 0, i32 %b)
%i5 = call spir_func i32 @__spirv_GroupSMax(i32 2, i32 1, i32 %b)
ret void
}
declare spir_func float @__spirv_GroupFAdd(i32, i32, float)
declare spir_func float @__spirv_GroupFMin(i32, i32, float)
declare spir_func float @__spirv_GroupFMax(i32, i32, float)
declare spir_func i32 @__spirv_GroupIAdd(i32, i32, i32)
declare spir_func i32 @__spirv_GroupUMin(i32, i32, i32)
declare spir_func i32 @__spirv_GroupSMin(i32, i32, i32)
declare spir_func i32 @__spirv_GroupUMax(i32, i32, i32)
declare spir_func i32 @__spirv_GroupSMax(i32, i32, i32)
Reference in New Issue View Git Blame Copy Permalink