Skip to content

Commit 34e038c

Browse files
bokrzesiigcbot
bokrzesi
authored and
igcbot
committed
Adding support for TargetExtensionTypes for JointMatrixFuncsResolutionPass
In this patch I've added support for TargetExtensionTypes (TET) which are described here: https://reviews.llvm.org/D135202 https://discourse.llvm.org/t/rfc-adding-opaque-types-to-llvm-ir/65326 For JointMatrixFuncsResolutionPass it means that instead of parsing type names we have mature primitive for representing e.g SPIRV types. Before, we were extracting those values `float 16 16 3 3 2` from typename string e.g: ``` %spirv.JointMatrixINTEL._float_16_16_3_3_2 = type opaque %spirv.JointMatrixINTEL._float_8_16_3_3_2 = type opaque %spirv.JointMatrixINTEL._float_32_64_3_3_2 = type opaque ``` Since TET: it is LLVM IR primitive Which is represented like that (right side) ``` %spirv.JointMatrixINTEL._float_16_16_3_3_2 = type { target("spirv.JointMatrixINTEL", float, 16, 16, 3, 3, 2) } %spirv.JointMatrixINTEL._float_8_16_3_3_2 = type { target("spirv.JointMatrixINTEL", float, 8, 16, 3, 3, 2) } %spirv.JointMatrixINTEL._float_32_64_3_3_2 = type { target("spirv.JointMatrixINTEL", float, 32, 64, 3, 3, 2) } ``` and can be accessed like that (first example): ``` llvm::Type* typeParam = *(targetTy->type_param_begin()); // float a = targetTy->int_params()[0]; // 16 b = targetTy->int_params()[1]; // 16 c = targetTy->int_params()[2]; // 3 d = targetTy->int_params()[3]; // 3 e = targetTy->int_params()[4]; // 2 ``` Additionally, I've extracted some common logic between non-TET and TET handlers e.g for `Use` and `Layout` calculation and GetMatrixTypeName.
1 parent 693cfe5 commit 34e038c

16 files changed

+1613
-97
lines changed

‎IGC/Compiler/Optimizer/OpenCLPasses/JointMatrixFuncsResolutionPass/JointMatrixFuncsResolutionPass.cpp

Lines changed: 311 additions & 97 deletions
Large diffs are not rendered by default.

‎IGC/Compiler/Optimizer/OpenCLPasses/JointMatrixFuncsResolutionPass/JointMatrixFuncsResolutionPass.h

Lines changed: 11 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -82,6 +82,17 @@ namespace IGC
8282

8383
bool parseMatrixTypeNameLegacy(const llvm::Type *opaqueType, JointMatrixTypeDescription *outDescription);
8484
bool ParseMatrixTypeName(llvm::Type *opaqueType, JointMatrixTypeDescription *outDescription);
85+
bool ParseMatrixTypeNameNonExtTypeDetails(llvm::Type* opaqueType,
86+
llvm::StringRef name,
87+
bool IsJointMatrix,
88+
JointMatrixTypeDescription* outDescription);
89+
#if LLVM_VERSION_MAJOR >= 16
90+
bool ParseMatrixTypeNameExtTypeDetails(llvm::Type* opaqueType, bool IsJointMatrix, IGC::JointMatrixTypeDescription* outDescription);
91+
#endif
92+
93+
llvm::StringRef GetMatrixTypeName(llvm::Type* opaqueType);
94+
bool SetLayoutFromUse(unsigned int use, IGC::JointMatrixTypeDescription* outDescription);
95+
unsigned GetUseFromLegacyLayout(unsigned int legacyLayout);
8596

8697
unsigned getNumRowsPerWI(const JointMatrixTypeDescription *desc);
8798
llvm::Type *ResolveType(llvm::Type *opaqueType, JointMatrixTypeDescription *outDesc);

‎IGC/Compiler/tests/JointMatrixFuncsResolutionPass/acc_fill_store-typed-pointers-targetextensiontype.ll

Lines changed: 58 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,58 @@
1+
;=========================== begin_copyright_notice ============================
2+
;
3+
; Copyright (C) 2025 Intel Corporation
4+
;
5+
; SPDX-License-Identifier: MIT
6+
;
7+
;============================ end_copyright_notice =============================
8+
;
9+
; REQUIRES: llvm-16-plus
10+
; RUN: igc_opt --typed-pointers --platformpvc --igc-joint-matrix-resolution -S 2>&1 < %s | FileCheck %s
11+
; ------------------------------------------------
12+
; JointMatrixFuncsResolutionPass
13+
; ------------------------------------------------
14+
15+
; CHECK-LABEL: define spir_kernel void @test_fill_store(
16+
; CHECK-SAME: float addrspace(1)* [[DST0:%.*]], float addrspace(1)* [[DST1:%.*]], float addrspace(1)* [[DST2:%.*]]) {
17+
define spir_kernel void @test_fill_store(float addrspace(1)* %dst0, float addrspace(1)* %dst1, float addrspace(1)* %dst2){
18+
; CHECK-NEXT: [[TMP5:%.*]] = alloca { <64 x float>, <64 x float> }
19+
; CHECK-NEXT: [[TMP3:%.*]] = alloca <8 x float>
20+
; CHECK-NEXT: [[TMP1:%.*]] = alloca <16 x float>
21+
; CHECK-NEXT: store <16 x float> <float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00>, <16 x float>* [[TMP1]]
22+
%1 = call spir_func target("spirv.JointMatrixINTEL", float, 16, 16, 3, 3, 2) @_Z26__spirv_CompositeConstructf(float 5.000000e+00)
23+
24+
; CHECK-NEXT: [[TMP2:%.*]] = bitcast <16 x float>* [[TMP1]] to i8*
25+
; CHECK-NEXT: call void @__builtin_spriv_OpJointMatrixStoreINTEL_Accumulator_RowMajor_SG16_16x16_i32_16_global_pi64_v8i8(float addrspace(1)* [[DST0]], i8* [[TMP2]], i64 16, i32 0)
26+
call spir_func void @_Z29__spirv_JointMatrixStoreINTELPU3AS1fPU3AS143__spirv_JointMatrixINTEL__float_16_16_3_3_2liii(float addrspace(1)* %dst0, target("spirv.JointMatrixINTEL", float, 16, 16, 3, 3, 2) %1, i64 16, i32 0, i32 3, i32 0)
27+
28+
; CHECK-NEXT: store <8 x float> <float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00>, <8 x float>* [[TMP3]]
29+
%2 = call spir_func target("spirv.JointMatrixINTEL", float, 8, 16, 3, 3, 2) @_Z26__spirv_CompositeConstructf.1(float 5.000000e+00)
30+
31+
; CHECK-NEXT: [[TMP4:%.*]] = bitcast <8 x float>* [[TMP3]] to i8*
32+
; CHECK-NEXT: call void @__builtin_spriv_OpJointMatrixStoreINTEL_Accumulator_RowMajor_SG16_8x16_i32_8_global_pi64_v8i8(float addrspace(1)* [[DST1]], i8* [[TMP4]], i64 16, i32 0)
33+
call spir_func void @_Z29__spirv_JointMatrixStoreINTELPU3AS1fPU3AS142__spirv_JointMatrixINTEL__float_8_16_3_3_2liii(float addrspace(1)* %dst1, target("spirv.JointMatrixINTEL", float, 8, 16, 3, 3, 2) %2, i64 16, i32 0, i32 3, i32 0)
34+
35+
; CHECK-NEXT: store { <64 x float>, <64 x float> } { <64 x float> <float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00>, <64 x float> <float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00, float 5.000000e+00> }, { <64 x float>, <64 x float> }* [[TMP5]]
36+
%3 = call spir_func target("spirv.JointMatrixINTEL", float, 32, 64, 3, 3, 2) @_Z26__spirv_CompositeConstructf.2(float 5.000000e+00)
37+
38+
; CHECK-NEXT: [[TMP6:%.*]] = bitcast { <64 x float>, <64 x float> }* [[TMP5]] to i8*
39+
; CHECK-NEXT: call void @__builtin_spriv_OpJointMatrixStoreINTEL_Accumulator_RowMajor_SG16_32x64_i32_128_global_pi64_v8i8(float addrspace(1)* [[DST2]], i8* [[TMP6]], i64 64, i32 0)
40+
call spir_func void @_Z29__spirv_JointMatrixStoreINTELPU3AS1fPU3AS143__spirv_JointMatrixINTEL__float_32_64_3_3_2liii(float addrspace(1)* %dst2, target("spirv.JointMatrixINTEL", float, 32, 64, 3, 3, 2) %3, i64 64, i32 0, i32 3, i32 0)
41+
42+
ret void
43+
}
44+
45+
declare spir_func target("spirv.JointMatrixINTEL", float, 16, 16, 3, 3, 2) @_Z26__spirv_CompositeConstructf(float)
46+
declare spir_func void @_Z29__spirv_JointMatrixStoreINTELPU3AS1fPU3AS143__spirv_JointMatrixINTEL__float_16_16_3_3_2liii(float addrspace(1)*, target("spirv.JointMatrixINTEL", float, 16, 16, 3, 3, 2), i64, i32, i32, i32)
47+
48+
declare spir_func target("spirv.JointMatrixINTEL", float, 8, 16, 3, 3, 2) @_Z26__spirv_CompositeConstructf.1(float)
49+
declare spir_func void @_Z29__spirv_JointMatrixStoreINTELPU3AS1fPU3AS142__spirv_JointMatrixINTEL__float_8_16_3_3_2liii(float addrspace(1)*, target("spirv.JointMatrixINTEL", float, 8, 16, 3, 3, 2), i64, i32, i32, i32)
50+
51+
declare spir_func target("spirv.JointMatrixINTEL", float, 32, 64, 3, 3, 2) @_Z26__spirv_CompositeConstructf.2(float)
52+
declare spir_func void @_Z29__spirv_JointMatrixStoreINTELPU3AS1fPU3AS143__spirv_JointMatrixINTEL__float_32_64_3_3_2liii(float addrspace(1)*, target("spirv.JointMatrixINTEL", float, 32, 64, 3, 3, 2), i64, i32, i32, i32)
53+
54+
!igc.functions = !{!0}
55+
!0 = !{void (float addrspace(1)*, float addrspace(1)*, float addrspace(1)*)* @test_fill_store, !1}
56+
!1 = !{!2, !3}
57+
!2 = !{!"function_type", i32 0}
58+
!3 = !{!"sub_group_size", i32 16}

‎IGC/Compiler/tests/JointMatrixFuncsResolutionPass/access-chain-two-uses-typed-pointers-targetextensiontype.ll

Lines changed: 55 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,55 @@
1+
;=========================== begin_copyright_notice ============================
2+
;
3+
; Copyright (C) 2025 Intel Corporation
4+
;
5+
; SPDX-License-Identifier: MIT
6+
;
7+
;============================ end_copyright_notice =============================
8+
; REQUIRES: llvm-16-plus
9+
; RUN: igc_opt --typed-pointers -platformpvc -igc-joint-matrix-resolution -S 2>&1 < %s | FileCheck %s
10+
; ------------------------------------------------
11+
; JointMatrixFuncsResolutionPass
12+
; ------------------------------------------------
13+
; Checks for multiple uses of __spirv_AccessChain function call - load plus store
14+
; it must result in extract and then insert an element to the matrix's slice
15+
16+
; CHECK: [[SLICE:%.*]] = load <8 x i16>, <8 x i16>* %{{.*}}, align 8
17+
; CHECK: [[ELEMENT:%.*]] = extractelement <8 x i16> [[SLICE]], i64 4, !joint_matrix_apply
18+
; CHECK: [[ADD:%.*]] = add i16 [[ELEMENT]], 1
19+
; CHECK: [[INSERT:%.*]] = insertelement <8 x i16> [[SLICE]], i16 [[ADD]], i64 4
20+
; CHECK: store <8 x i16> [[INSERT]], <8 x i16>* %{{.*}}, align 8
21+
22+
target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v24:32:32-v32:32:32-v48:64:64-v64:64:64-v96:128:128-v128:128:128-v192:256:256-v256:256:256-v512:512:512-v1024:1024:1024-n8:16:32"
23+
target triple = "spir64-unknown-unknown"
24+
25+
26+
; Function Attrs: nounwind
27+
define spir_kernel void @_ZTS5logicILm8ELm16EE(i16 addrspace(1)* %arg) {
28+
entry:
29+
%0 = alloca target("spirv.CooperativeMatrixKHR", i16, 3, 8, 16, 0 )
30+
%1 = call spir_func target("spirv.CooperativeMatrixKHR", i16, 3, 8, 16, 0 ) @_Z86__spirv_CooperativeMatrixLoadKHR_RPU3AS143__spirv_CooperativeMatrixKHR__short_3_8_16_0PU3AS1slii(i16 addrspace(1)* %arg, i32 0, i64 64, i32 0)
31+
store target("spirv.CooperativeMatrixKHR", i16, 3, 8, 16, 0 ) %1, target("spirv.CooperativeMatrixKHR", i16, 3, 8, 16, 0 )* %0
32+
%ptr = call spir_func i16 addrspace(4)* @_Z19__spirv_AccessChainPU3AS4PU3AS143__spirv_CooperativeMatrixKHR._short_3_8_16_0l(target("spirv.CooperativeMatrixKHR", i16, 3, 8, 16, 0 )* %0, i64 4)
33+
%extract = load i16, i16 addrspace(4)* %ptr
34+
%add = add i16 %extract, 1
35+
store i16 %add, i16 addrspace(4)* %ptr
36+
ret void
37+
}
38+
39+
; Function Attrs: nounwind
40+
declare spir_func target("spirv.CooperativeMatrixKHR", i16, 3, 8, 16, 0 ) @_Z86__spirv_CooperativeMatrixLoadKHR_RPU3AS143__spirv_CooperativeMatrixKHR__short_3_8_16_0PU3AS1slii(i16 addrspace(1)* %0, i32 %1, i64 %2, i32 %3)
41+
42+
; Function Attrs: nounwind
43+
declare spir_func i16 addrspace(4)* @_Z19__spirv_AccessChainPU3AS4PU3AS143__spirv_CooperativeMatrixKHR._short_3_8_16_0l(target("spirv.CooperativeMatrixKHR", i16, 3, 8, 16, 0 )* %0, i64 %1)
44+
45+
!spirv.MemoryModel = !{!0}
46+
!spirv.Source = !{!1}
47+
!spirv.Generator = !{!2}
48+
!igc.functions = !{!3}
49+
50+
!0 = !{i32 2, i32 2}
51+
!1 = !{i32 4, i32 100000}
52+
!2 = !{i16 6, i16 14}
53+
!3 = !{void (i16 addrspace(1)*)* @_ZTS5logicILm8ELm16EE, !4}
54+
!4 = !{!5}
55+
!5 = !{!"function_type", i32 0}

0 commit comments

Comments
 (0)