Skip to content

Commit 92bcf18

Browse files
ppogotovpszymich
ppogotov
authored and
pszymich
committed
Reform prefetch instructions.
Reform prefetch instructions to maximize width and minimize the number of blocks. (cherry picked from commit fab64a9)
1 parent 1f886a4 commit 92bcf18

File tree

5 files changed

+144
-28
lines changed

5 files changed

+144
-28
lines changed

‎IGC/Compiler/CustomSafeOptPass.cpp‎

Lines changed: 30 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -452,6 +452,30 @@ void CustomSafeOptPass::mergeDotAddToDp4a(llvm::CallInst* I)
452452
}
453453
}
454454

455+
// Reform prefetch instructions to maximize width and minimize the number of blocks. This function assumes that
456+
// it operates on the correct form of the intrinsic (width X array_length matches the requirements).
457+
void CustomSafeOptPass::visitLSC2DBlockPrefetch(CallInst* I) {
458+
const size_t widthInd = 7;
459+
const size_t numBlocksInd = 9;
460+
461+
ConstantInt *constWidth = dyn_cast<ConstantInt>(I->getArgOperand(widthInd));
462+
ConstantInt *constNumBlocks = dyn_cast<ConstantInt>(I->getArgOperand(numBlocksInd));
463+
464+
if (!constWidth || !constNumBlocks) {
465+
return;
466+
}
467+
468+
uint64_t width = constWidth->getZExtValue();
469+
uint64_t numBlocks = constNumBlocks->getZExtValue();
470+
471+
if (numBlocks == 1) {
472+
return;
473+
}
474+
475+
I->setArgOperand(widthInd, ConstantInt::get(Type::getInt32Ty(I->getContext()), width * numBlocks));
476+
I->setArgOperand(numBlocksInd, ConstantInt::get(Type::getInt32Ty(I->getContext()), 1));
477+
}
478+
455479
// Check if Lower 64b to 32b transformation is applicable for binary operator
456480
// i.e. trunc(a op b) == trunc(a) op trunc(b)
457481
static bool isTruncInvariant(unsigned Opcode) {
@@ -878,6 +902,12 @@ void CustomSafeOptPass::visitCallInst(CallInst& C)
878902
break;
879903
}
880904

905+
case GenISAIntrinsic::GenISA_LSC2DBlockPrefetch:
906+
{
907+
visitLSC2DBlockPrefetch(inst);
908+
break;
909+
}
910+
881911
default:
882912
break;
883913
}

‎IGC/Compiler/CustomSafeOptPass.hpp‎

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -72,6 +72,7 @@ namespace IGC
7272
bool isIdentityMatrix(llvm::ExtractElementInst& I);
7373
void visitAnd(llvm::BinaryOperator& I);
7474
void visitXor(llvm::Instruction& XorInstr);
75+
void visitLSC2DBlockPrefetch(llvm::CallInst* I);
7576
void visitShuffleIndex(llvm::CallInst* I);
7677
void visitSelectInst(llvm::SelectInst& S);
7778
void mergeDotAddToDp4a(llvm::CallInst* I);

‎IGC/Compiler/tests/CustomSafeOptPass/2d_block_prefetch.ll‎

Lines changed: 85 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,85 @@
1+
;=========================== begin_copyright_notice ============================
2+
;
3+
; Copyright (C) 2024 Intel Corporation
4+
;
5+
; SPDX-License-Identifier: MIT
6+
;
7+
;============================ end_copyright_notice =============================
8+
;
9+
; REQUIRES: llvm-14-plus, pvc-supported
10+
;
11+
; RUN: igc_opt %s -S -o - -igc-custom-safe-opt -platform pvc | FileCheck %s
12+
13+
define spir_kernel void @block_prefetch_d16.1x16x8nn(i64 %val1, i32 %val2) {
14+
call void @llvm.genx.GenISA.LSC2DBlockPrefetch.isVoid(i64 %val1, i32 8191, i32 4095, i32 8191, i32 0, i32 %val2, i32 16, i32 16, i32 8, i32 1, i1 false, i1 false, i32 4)
15+
;
16+
; CHECK: call void @llvm.genx.GenISA.LSC2DBlockPrefetch.isVoid(i64 %val1, i32 8191, i32 4095, i32 8191, i32 0, i32 %val2, i32 16, i32 16, i32 8, i32 1, i1 false, i1 false, i32 4)
17+
;
18+
ret void
19+
}
20+
21+
define spir_kernel void @block_prefetch_d16.2x16x8nn(i64 %val1, i32 %val2) {
22+
call void @llvm.genx.GenISA.LSC2DBlockPrefetch.isVoid(i64 %val1, i32 8191, i32 4095, i32 8191, i32 0, i32 %val2, i32 16, i32 16, i32 8, i32 2, i1 false, i1 false, i32 4)
23+
;
24+
; CHECK: call void @llvm.genx.GenISA.LSC2DBlockPrefetch.isVoid(i64 %val1, i32 8191, i32 4095, i32 8191, i32 0, i32 %val2, i32 16, i32 32, i32 8, i32 1, i1 false, i1 false, i32 4)
25+
;
26+
ret void
27+
}
28+
29+
define spir_kernel void @block_prefetch_d16.4x8x8nn(i64 %val1, i32 %val2) {
30+
call void @llvm.genx.GenISA.LSC2DBlockPrefetch.isVoid(i64 %val1, i32 8191, i32 4095, i32 8191, i32 0, i32 %val2, i32 16, i32 8, i32 8, i32 4, i1 false, i1 false, i32 4)
31+
;
32+
; CHECK: call void @llvm.genx.GenISA.LSC2DBlockPrefetch.isVoid(i64 %val1, i32 8191, i32 4095, i32 8191, i32 0, i32 %val2, i32 16, i32 32, i32 8, i32 1, i1 false, i1 false, i32 4)
33+
;
34+
ret void
35+
}
36+
37+
define spir_kernel void @block_prefetch_d8.4x64x8nn(i64 %val1, i32 %val2) {
38+
call void @llvm.genx.GenISA.LSC2DBlockPrefetch.isVoid(i64 %val1, i32 8191, i32 4095, i32 8191, i32 0, i32 %val2, i32 8, i32 64, i32 8, i32 4, i1 false, i1 false, i32 4)
39+
;
40+
; CHECK: call void @llvm.genx.GenISA.LSC2DBlockPrefetch.isVoid(i64 %val1, i32 8191, i32 4095, i32 8191, i32 0, i32 %val2, i32 8, i32 256, i32 8, i32 1, i1 false, i1 false, i32 4)
41+
;
42+
ret void
43+
}
44+
45+
define spir_kernel void @block_prefetch_d8.4x16x8nn(i64 %val1, i32 %val2) {
46+
call void @llvm.genx.GenISA.LSC2DBlockPrefetch.isVoid(i64 %val1, i32 8191, i32 4095, i32 8191, i32 0, i32 %val2, i32 8, i32 16, i32 8, i32 4, i1 false, i1 false, i32 4)
47+
;
48+
; CHECK: call void @llvm.genx.GenISA.LSC2DBlockPrefetch.isVoid(i64 %val1, i32 8191, i32 4095, i32 8191, i32 0, i32 %val2, i32 8, i32 64, i32 8, i32 1, i1 false, i1 false, i32 4)
49+
;
50+
ret void
51+
}
52+
53+
define spir_kernel void @block_prefetch_d32.2x16x8nn(i64 %val1, i32 %val2) {
54+
call void @llvm.genx.GenISA.LSC2DBlockPrefetch.isVoid(i64 %val1, i32 8191, i32 4095, i32 8191, i32 0, i32 %val2, i32 32, i32 16, i32 8, i32 2, i1 false, i1 false, i32 4)
55+
;
56+
; CHECK: call void @llvm.genx.GenISA.LSC2DBlockPrefetch.isVoid(i64 %val1, i32 8191, i32 4095, i32 8191, i32 0, i32 %val2, i32 32, i32 32, i32 8, i32 1, i1 false, i1 false, i32 4)
57+
;
58+
ret void
59+
}
60+
61+
define spir_kernel void @block_prefetch_d32.2x32x8nn(i64 %val1, i32 %val2) {
62+
call void @llvm.genx.GenISA.LSC2DBlockPrefetch.isVoid(i64 %val1, i32 8191, i32 4095, i32 8191, i32 0, i32 %val2, i32 32, i32 32, i32 8, i32 2, i1 false, i1 false, i32 4)
63+
;
64+
; CHECK: call void @llvm.genx.GenISA.LSC2DBlockPrefetch.isVoid(i64 %val1, i32 8191, i32 4095, i32 8191, i32 0, i32 %val2, i32 32, i32 64, i32 8, i32 1, i1 false, i1 false, i32 4)
65+
;
66+
ret void
67+
}
68+
69+
define spir_kernel void @block_prefetch_d64.4x8x8nn(i64 %val1, i32 %val2) {
70+
call void @llvm.genx.GenISA.LSC2DBlockPrefetch.isVoid(i64 %val1, i32 8191, i32 4095, i32 8191, i32 0, i32 %val2, i32 64, i32 8, i32 8, i32 4, i1 false, i1 false, i32 4)
71+
;
72+
; CHECK: call void @llvm.genx.GenISA.LSC2DBlockPrefetch.isVoid(i64 %val1, i32 8191, i32 4095, i32 8191, i32 0, i32 %val2, i32 64, i32 8, i32 8, i32 4, i1 false, i1 false, i32 4)
73+
;
74+
ret void
75+
}
76+
77+
define spir_kernel void @block_prefetch_d64.1x8x8nn(i64 %val1, i32 %val2) {
78+
call void @llvm.genx.GenISA.LSC2DBlockPrefetch.isVoid(i64 %val1, i32 8191, i32 4095, i32 8191, i32 0, i32 %val2, i32 64, i32 8, i32 8, i32 1, i1 false, i1 false, i32 4)
79+
;
80+
; CHECK: call void @llvm.genx.GenISA.LSC2DBlockPrefetch.isVoid(i64 %val1, i32 8191, i32 4095, i32 8191, i32 0, i32 %val2, i32 64, i32 8, i32 8, i32 1, i1 false, i1 false, i32 4)
81+
;
82+
ret void
83+
}
84+
85+
declare void @llvm.genx.GenISA.LSC2DBlockPrefetch.isVoid(i64, i32, i32, i32, i32, i32, i32, i32, i32, i32, i1, i1, i32)

‎IGC/ocloc_tests/Builtins/cl_intel_subgroup_2d_block_io/PVC/prefetches.cl‎

Lines changed: 20 additions & 20 deletions
Original file line numberDiff line numberDiff line change
@@ -12,37 +12,37 @@ SPDX-License-Identifier: MIT
1212
// RUN: -DINPUT_TYPE=ushort -DFUNCTION=intel_sub_group_2d_block_prefetch_8b_1r32x2c" \
1313
// RUN: -internal_options "-cl-ext=-all,+cl_intel_subgroup_2d_block_io" | FileCheck %s --check-prefix=CHECK-VISAASM-8B-1R-32X2C
1414

15-
// CHECK-VISAASM-8B-1R-32X2C: lsc_load_block2d.ugm (M1, 1) %null:d8.2x32x1nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
15+
// CHECK-VISAASM-8B-1R-32X2C: lsc_load_block2d.ugm (M1, 1) %null:d8.64x1nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
1616

1717
// RUN: ocloc compile -file %s -device pvc -options "-igc_opts 'DumpVISAASMToConsole=1' \
1818
// RUN: -DINPUT_TYPE=ushort -DFUNCTION=intel_sub_group_2d_block_prefetch_8b_2r32x2c" \
1919
// RUN: -internal_options "-cl-ext=-all,+cl_intel_subgroup_2d_block_io" | FileCheck %s --check-prefix=CHECK-VISAASM-8B-2R-32X2C
2020

21-
// CHECK-VISAASM-8B-2R-32X2C: lsc_load_block2d.ugm (M1, 1) %null:d8.2x32x2nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
21+
// CHECK-VISAASM-8B-2R-32X2C: lsc_load_block2d.ugm (M1, 1) %null:d8.64x2nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
2222

2323
// RUN: ocloc compile -file %s -device pvc -options "-igc_opts 'DumpVISAASMToConsole=1' \
2424
// RUN: -DINPUT_TYPE=ushort -DFUNCTION=intel_sub_group_2d_block_prefetch_8b_4r32x2c" \
2525
// RUN: -internal_options "-cl-ext=-all,+cl_intel_subgroup_2d_block_io" | FileCheck %s --check-prefix=CHECK-VISAASM-8B-4R-32X2C
2626

27-
// CHECK-VISAASM-8B-4R-32X2C: lsc_load_block2d.ugm (M1, 1) %null:d8.2x32x4nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
27+
// CHECK-VISAASM-8B-4R-32X2C: lsc_load_block2d.ugm (M1, 1) %null:d8.64x4nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
2828

2929
// RUN: ocloc compile -file %s -device pvc -options "-igc_opts 'DumpVISAASMToConsole=1' \
3030
// RUN: -DINPUT_TYPE=ushort -DFUNCTION=intel_sub_group_2d_block_prefetch_8b_8r32x2c" \
3131
// RUN: -internal_options "-cl-ext=-all,+cl_intel_subgroup_2d_block_io" | FileCheck %s --check-prefix=CHECK-VISAASM-8B-8R-32X2C
3232

33-
// CHECK-VISAASM-8B-8R-32X2C: lsc_load_block2d.ugm (M1, 1) %null:d8.2x32x8nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
33+
// CHECK-VISAASM-8B-8R-32X2C: lsc_load_block2d.ugm (M1, 1) %null:d8.64x8nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
3434

3535
// RUN: ocloc compile -file %s -device pvc -options "-igc_opts 'DumpVISAASMToConsole=1' \
3636
// RUN: -DINPUT_TYPE=ushort -DFUNCTION=intel_sub_group_2d_block_prefetch_8b_16r32x2c" \
3737
// RUN: -internal_options "-cl-ext=-all,+cl_intel_subgroup_2d_block_io" | FileCheck %s --check-prefix=CHECK-VISAASM-8B-16R-32X2C
3838

39-
// CHECK-VISAASM-8B-16R-32X2C: lsc_load_block2d.ugm (M1, 1) %null:d8.2x32x16nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
39+
// CHECK-VISAASM-8B-16R-32X2C: lsc_load_block2d.ugm (M1, 1) %null:d8.64x16nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
4040

4141
// RUN: ocloc compile -file %s -device pvc -options "-igc_opts 'DumpVISAASMToConsole=1' \
4242
// RUN: -DINPUT_TYPE=ushort -DFUNCTION=intel_sub_group_2d_block_prefetch_8b_32r32x2c" \
4343
// RUN: -internal_options "-cl-ext=-all,+cl_intel_subgroup_2d_block_io" | FileCheck %s --check-prefix=CHECK-VISAASM-8B-32R-32X2C
4444

45-
// CHECK-VISAASM-8B-32R-32X2C: lsc_load_block2d.ugm (M1, 1) %null:d8.2x32x32nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
45+
// CHECK-VISAASM-8B-32R-32X2C: lsc_load_block2d.ugm (M1, 1) %null:d8.64x32nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
4646

4747
// RUN: ocloc compile -file %s -device pvc -options "-igc_opts 'DumpVISAASMToConsole=1' \
4848
// RUN: -DINPUT_TYPE=ushort -DFUNCTION=intel_sub_group_2d_block_prefetch_16b_1r16x1c" \
@@ -84,37 +84,37 @@ SPDX-License-Identifier: MIT
8484
// RUN: -DINPUT_TYPE=ushort -DFUNCTION=intel_sub_group_2d_block_prefetch_16b_1r16x2c" \
8585
// RUN: -internal_options "-cl-ext=-all,+cl_intel_subgroup_2d_block_io" | FileCheck %s --check-prefix=CHECK-VISAASM-16B-1R-16X2C
8686

87-
// CHECK-VISAASM-16B-1R-16X2C: lsc_load_block2d.ugm (M1, 1) %null:d16.2x16x1nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
87+
// CHECK-VISAASM-16B-1R-16X2C: lsc_load_block2d.ugm (M1, 1) %null:d16.32x1nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
8888

8989
// RUN: ocloc compile -file %s -device pvc -options "-igc_opts 'DumpVISAASMToConsole=1' \
9090
// RUN: -DINPUT_TYPE=ushort -DFUNCTION=intel_sub_group_2d_block_prefetch_16b_2r16x2c" \
9191
// RUN: -internal_options "-cl-ext=-all,+cl_intel_subgroup_2d_block_io" | FileCheck %s --check-prefix=CHECK-VISAASM-16B-2R-16X2C
9292

93-
// CHECK-VISAASM-16B-2R-16X2C: lsc_load_block2d.ugm (M1, 1) %null:d16.2x16x2nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
93+
// CHECK-VISAASM-16B-2R-16X2C: lsc_load_block2d.ugm (M1, 1) %null:d16.32x2nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
9494

9595
// RUN: ocloc compile -file %s -device pvc -options "-igc_opts 'DumpVISAASMToConsole=1' \
9696
// RUN: -DINPUT_TYPE=ushort -DFUNCTION=intel_sub_group_2d_block_prefetch_16b_4r16x2c" \
9797
// RUN: -internal_options "-cl-ext=-all,+cl_intel_subgroup_2d_block_io" | FileCheck %s --check-prefix=CHECK-VISAASM-16B-4R-16X2C
9898

99-
// CHECK-VISAASM-16B-4R-16X2C: lsc_load_block2d.ugm (M1, 1) %null:d16.2x16x4nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
99+
// CHECK-VISAASM-16B-4R-16X2C: lsc_load_block2d.ugm (M1, 1) %null:d16.32x4nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
100100

101101
// RUN: ocloc compile -file %s -device pvc -options "-igc_opts 'DumpVISAASMToConsole=1' \
102102
// RUN: -DINPUT_TYPE=ushort -DFUNCTION=intel_sub_group_2d_block_prefetch_16b_8r16x2c" \
103103
// RUN: -internal_options "-cl-ext=-all,+cl_intel_subgroup_2d_block_io" | FileCheck %s --check-prefix=CHECK-VISAASM-16B-8R-16X2C
104104

105-
// CHECK-VISAASM-16B-8R-16X2C: lsc_load_block2d.ugm (M1, 1) %null:d16.2x16x8nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
105+
// CHECK-VISAASM-16B-8R-16X2C: lsc_load_block2d.ugm (M1, 1) %null:d16.32x8nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
106106

107107
// RUN: ocloc compile -file %s -device pvc -options "-igc_opts 'DumpVISAASMToConsole=1' \
108108
// RUN: -DINPUT_TYPE=ushort -DFUNCTION=intel_sub_group_2d_block_prefetch_16b_16r16x2c" \
109109
// RUN: -internal_options "-cl-ext=-all,+cl_intel_subgroup_2d_block_io" | FileCheck %s --check-prefix=CHECK-VISAASM-16B-16R-16X2C
110110

111-
// CHECK-VISAASM-16B-16R-16X2C: lsc_load_block2d.ugm (M1, 1) %null:d16.2x16x16nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
111+
// CHECK-VISAASM-16B-16R-16X2C: lsc_load_block2d.ugm (M1, 1) %null:d16.32x16nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
112112

113113
// RUN: ocloc compile -file %s -device pvc -options "-igc_opts 'DumpVISAASMToConsole=1' \
114114
// RUN: -DINPUT_TYPE=ushort -DFUNCTION=intel_sub_group_2d_block_prefetch_16b_32r16x2c" \
115115
// RUN: -internal_options "-cl-ext=-all,+cl_intel_subgroup_2d_block_io" | FileCheck %s --check-prefix=CHECK-VISAASM-16B-32R-16X2C
116116

117-
// CHECK-VISAASM-16B-32R-16X2C: lsc_load_block2d.ugm (M1, 1) %null:d16.2x16x32nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
117+
// CHECK-VISAASM-16B-32R-16X2C: lsc_load_block2d.ugm (M1, 1) %null:d16.32x32nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
118118

119119
// RUN: ocloc compile -file %s -device pvc -options "-igc_opts 'DumpVISAASMToConsole=1' \
120120
// RUN: -DINPUT_TYPE=ushort -DFUNCTION=intel_sub_group_2d_block_prefetch_8b_32r16x1c" \
@@ -234,49 +234,49 @@ SPDX-License-Identifier: MIT
234234
// RUN: -DINPUT_TYPE=uint -DFUNCTION=intel_sub_group_2d_block_prefetch_32b_1r8x2c" \
235235
// RUN: -internal_options "-cl-ext=-all,+cl_intel_subgroup_2d_block_io" | FileCheck %s --check-prefix=CHECK-VISAASM-32B-1R-8X2C
236236

237-
// CHECK-VISAASM-32B-1R-8X2C: lsc_load_block2d.ugm (M1, 1) %null:d32.2x8x1nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
237+
// CHECK-VISAASM-32B-1R-8X2C: lsc_load_block2d.ugm (M1, 1) %null:d32.16x1nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
238238

239239
// RUN: ocloc compile -file %s -device pvc -options "-igc_opts 'DumpVISAASMToConsole=1' \
240240
// RUN: -DINPUT_TYPE=uint -DFUNCTION=intel_sub_group_2d_block_prefetch_32b_2r8x2c" \
241241
// RUN: -internal_options "-cl-ext=-all,+cl_intel_subgroup_2d_block_io" | FileCheck %s --check-prefix=CHECK-VISAASM-32B-2R-8X2C
242242

243-
// CHECK-VISAASM-32B-2R-8X2C: lsc_load_block2d.ugm (M1, 1) %null:d32.2x8x2nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
243+
// CHECK-VISAASM-32B-2R-8X2C: lsc_load_block2d.ugm (M1, 1) %null:d32.16x2nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
244244

245245
// RUN: ocloc compile -file %s -device pvc -options "-igc_opts 'DumpVISAASMToConsole=1' \
246246
// RUN: -DINPUT_TYPE=uint -DFUNCTION=intel_sub_group_2d_block_prefetch_32b_4r8x2c" \
247247
// RUN: -internal_options "-cl-ext=-all,+cl_intel_subgroup_2d_block_io" | FileCheck %s --check-prefix=CHECK-VISAASM-32B-4R-8X2C
248248

249-
// CHECK-VISAASM-32B-4R-8X2C: lsc_load_block2d.ugm (M1, 1) %null:d32.2x8x4nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
249+
// CHECK-VISAASM-32B-4R-8X2C: lsc_load_block2d.ugm (M1, 1) %null:d32.16x4nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
250250

251251
// RUN: ocloc compile -file %s -device pvc -options "-igc_opts 'DumpVISAASMToConsole=1' \
252252
// RUN: -DINPUT_TYPE=uint -DFUNCTION=intel_sub_group_2d_block_prefetch_32b_8r8x2c" \
253253
// RUN: -internal_options "-cl-ext=-all,+cl_intel_subgroup_2d_block_io" | FileCheck %s --check-prefix=CHECK-VISAASM-32B-8R-8X2C
254254

255-
// CHECK-VISAASM-32B-8R-8X2C: lsc_load_block2d.ugm (M1, 1) %null:d32.2x8x8nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
255+
// CHECK-VISAASM-32B-8R-8X2C: lsc_load_block2d.ugm (M1, 1) %null:d32.16x8nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
256256

257257
// RUN: ocloc compile -file %s -device pvc -options "-igc_opts 'DumpVISAASMToConsole=1' \
258258
// RUN: -DINPUT_TYPE=uint -DFUNCTION=intel_sub_group_2d_block_prefetch_32b_16r8x2c" \
259259
// RUN: -internal_options "-cl-ext=-all,+cl_intel_subgroup_2d_block_io" | FileCheck %s --check-prefix=CHECK-VISAASM-32B-16R-8X2C
260260

261-
// CHECK-VISAASM-32B-16R-8X2C: lsc_load_block2d.ugm (M1, 1) %null:d32.2x8x16nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
261+
// CHECK-VISAASM-32B-16R-8X2C: lsc_load_block2d.ugm (M1, 1) %null:d32.16x16nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
262262

263263
// RUN: ocloc compile -file %s -device pvc -options "-igc_opts 'DumpVISAASMToConsole=1' \
264264
// RUN: -DINPUT_TYPE=uint -DFUNCTION=intel_sub_group_2d_block_prefetch_32b_32r8x2c" \
265265
// RUN: -internal_options "-cl-ext=-all,+cl_intel_subgroup_2d_block_io" | FileCheck %s --check-prefix=CHECK-VISAASM-32B-32R-8X2C
266266

267-
// CHECK-VISAASM-32B-32R-8X2C: lsc_load_block2d.ugm (M1, 1) %null:d32.2x8x32nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
267+
// CHECK-VISAASM-32B-32R-8X2C: lsc_load_block2d.ugm (M1, 1) %null:d32.16x32nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
268268

269269
// RUN: ocloc compile -file %s -device pvc -options "-igc_opts 'DumpVISAASMToConsole=1' \
270270
// RUN: -DINPUT_TYPE=uint -DFUNCTION=intel_sub_group_2d_block_prefetch_8b_32r16x2c" \
271271
// RUN: -internal_options "-cl-ext=-all,+cl_intel_subgroup_2d_block_io" | FileCheck %s --check-prefix=CHECK-VISAASM-8B-32R-16X2C
272272

273-
// CHECK-VISAASM-8B-32R-16X2C: lsc_load_block2d.ugm (M1, 1) %null:d8.2x16x32nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
273+
// CHECK-VISAASM-8B-32R-16X2C: lsc_load_block2d.ugm (M1, 1) %null:d8.32x32nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
274274

275275
// RUN: ocloc compile -file %s -device pvc -options "-igc_opts 'DumpVISAASMToConsole=1' \
276276
// RUN: -DINPUT_TYPE=uint -DFUNCTION=intel_sub_group_2d_block_prefetch_8b_32r16x4c" \
277277
// RUN: -internal_options "-cl-ext=-all,+cl_intel_subgroup_2d_block_io" | FileCheck %s --check-prefix=CHECK-VISAASM-8B-32R-16X4C
278278

279-
// CHECK-VISAASM-8B-32R-16X4C: lsc_load_block2d.ugm (M1, 1) %null:d8.4x16x32nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
279+
// CHECK-VISAASM-8B-32R-16X4C: lsc_load_block2d.ugm (M1, 1) %null:d8.64x32nn flat[{{.+}},0x1FF,0x2D,0x1FF,V{{[0-9]+}},V{{[0-9]+}}
280280

281281
__attribute__((intel_reqd_sub_group_size(16)))
282282
kernel void test_default(global INPUT_TYPE* input, const global int2* coord) {

0 commit comments

Comments
 (0)