1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
|
;=========================== begin_copyright_notice ============================
;
; Copyright (C) 2025 Intel Corporation
;
; SPDX-License-Identifier: MIT
;
;============================ end_copyright_notice =============================
;
; REQUIRES: llvm-16-plus
; RUN: igc_opt --typed-pointers %s -S -o - -igc-joint-matrix-resolution --platformpvc 2>&1 | FileCheck %s
; ------------------------------------------------
; JointMatrixFuncsResolutionPass
; ------------------------------------------------
; CHECK-LABEL: @mad_builtin_half_float(
; CHECK-NEXT: [[TMP1:%.*]] = alloca <16 x float>
; CHECK-NEXT: [[TMP2:%.*]] = alloca <32 x i16>
; CHECK-NEXT: [[TMP3:%.*]] = alloca <8 x i32>
; CHECK-NEXT: [[TMP4:%.*]] = alloca <16 x float>
; CHECK-NEXT: [[TMP5:%.*]] = alloca <16 x float>
; CHECK-NEXT: store <32 x i16>{{.*}} <32 x i16>* [[TMP2]]
; CHECK-NEXT: store <8 x i32>{{.*}} <8 x i32>* [[TMP3]]
; CHECK-NEXT: store <16 x float> zeroinitializer, <16 x float>* [[TMP4]]
; CHECK-NEXT: [[TMP6:%.*]] = bitcast <32 x i16>* [[TMP2]] to i8*
; CHECK-NEXT: [[TMP7:%.*]] = bitcast <8 x i32>* [[TMP3]] to i8*
; CHECK-NEXT: [[TMP8:%.*]] = bitcast <16 x float>* [[TMP4]] to i8*
; CHECK-NEXT: [[TMP9:%.*]] = bitcast <16 x float>* [[TMP5]] to i8*
; CHECK-NEXT: call void @__builtin_spriv_OpJointMatrixMadINTEL_16x16x16_i32_fp16_fp32(i8* [[TMP6]], i8* [[TMP7]], i8* [[TMP8]], i8* [[TMP9]])
; CHECK-NEXT: [[TMP10:%.*]] = load <16 x float>, <16 x float>* [[TMP5]]
; CHECK-NEXT: store <16 x float> [[TMP10]], <16 x float>* [[TMP1]]
; CHECK-NEXT: [[TMP11:%.*]] = bitcast <16 x float>* [[TMP1]] to i8*
; CHECK-NEXT: call void @__builtin_spriv_OpJointMatrixStoreINTEL_Accumulator_RowMajor_SG16_16x16_i32_16_global_pi64_v8i8(float addrspace(1)* [[DST:%.*]], i8* [[TMP11]], i64 [[STRIDE:%.*]])
; CHECK-NEXT: ret void
; CHECK-NOT: error:
define spir_kernel void @mad_builtin_half_float(i8 addrspace(1)* %src, i64 %stride, float addrspace(1)* %dst) {
%1 = call spir_func target("spirv.CooperativeMatrixKHR", float, 3, 16, 16, 2) @_Z26__spirv_CompositeConstructf(float 0.0)
%2 = call spir_func target("spirv.CooperativeMatrixKHR", float, 3, 16, 16, 0) @_Z76__spirv_CompositeConstructf(float 1.0)
%3 = call spir_func target("spirv.CooperativeMatrixKHR", half, 3, 16, 16, 1) @_Z80__spirv_CompositeConstructh(half -1.0)
%4 = call spir_func target("spirv.CooperativeMatrixKHR", float, 3, 16, 16, 2) @_Z34__spirv_CooperativeMatrixMulAddKHRPU3AS141__spirv.CooperativeMatrixKHR._float_3_16_16_0PU3AS142__spirv_CooperativeMatrixKHR._half_3_16_16_1PU3AS140__spirv_CooperativeMatrixKHR__float_3_16_16_2(target("spirv.CooperativeMatrixKHR", float, 3, 16, 16, 0) %2, target("spirv.CooperativeMatrixKHR", half, 3, 16, 16, 1) %3, target("spirv.CooperativeMatrixKHR", float, 3, 16, 16, 2) %1)
call spir_func void @_Z33__spirv_CooperativeMatrixStoreKHRPU3AS1iPU3AS140__spirv_CooperativeMatrixKHR__float_3_16_16_2ili(float addrspace(1)* %dst, target("spirv.CooperativeMatrixKHR", float, 3, 16, 16, 2) %4, i32 0, i64 %stride, i32 3)
ret void
}
declare spir_func target("spirv.CooperativeMatrixKHR", float, 3, 16, 16, 2) @_Z26__spirv_CompositeConstructf(float)
declare spir_func target("spirv.CooperativeMatrixKHR", float, 3, 16, 16, 0) @_Z76__spirv_CompositeConstructf(float)
declare spir_func target("spirv.CooperativeMatrixKHR", half, 3, 16, 16, 1) @_Z80__spirv_CompositeConstructh(half)
declare spir_func target("spirv.CooperativeMatrixKHR", float, 3, 16, 16, 2) @_Z34__spirv_CooperativeMatrixMulAddKHRPU3AS141__spirv.CooperativeMatrixKHR._float_3_16_16_0PU3AS142__spirv_CooperativeMatrixKHR._half_3_16_16_1PU3AS140__spirv_CooperativeMatrixKHR__float_3_16_16_2(target("spirv.CooperativeMatrixKHR", float, 3, 16, 16, 0), target("spirv.CooperativeMatrixKHR", half, 3, 16, 16, 1), target("spirv.CooperativeMatrixKHR", float, 3, 16, 16, 2))
declare spir_func void @_Z33__spirv_CooperativeMatrixStoreKHRPU3AS1iPU3AS140__spirv_CooperativeMatrixKHR__float_3_16_16_2ili(float addrspace(1)*, target("spirv.CooperativeMatrixKHR", float, 3, 16, 16, 2), i32, i64, i32)
!igc.functions = !{!0}
!0 = !{void (i8 addrspace(1)*, i64, float addrspace(1)*)* @mad_builtin_half_float, !1}
!1 = !{!2, !3}
!2 = !{!"function_type", i32 0}
!3 = !{!"sub_group_size", i32 16}
|
