Skip to content

Commit b301a98

Browse files
angelz913angelz913
authored
[mlir][spirv] Add integration tests for vector.interleave and vector.shuffle (#93858)
This PR tries to reland #93595 which was reverted in #93732 due to some issues. The original PR: - Add integration test for `vector.shuffle` and `vector.interleave` - Add `VectorToSPIRV` patterns to `GPUToSPIRVPass` Description of the issue: - #93595 (comment) - Using either `vector.load` or `vector.store` in the kernel function will cause the validation layer to report an error - Trying to bypass the issue by using `memref.load` and `memref.store` to load/store individual elements from/to the vectors, and populate the vectors using `vector.insertelement` and `vector.extractelement` instead.
1 parent 8aa33f1 commit b301a98

File tree

4 files changed

+161
-0
lines changed

4 files changed

+161
-0
lines changed

‎mlir/lib/Conversion/GPUToSPIRV/CMakeLists.txt

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -13,6 +13,7 @@ add_mlir_conversion_library(MLIRGPUToSPIRV
1313
MLIRIR
1414
MLIRPass
1515
MLIRSCFToSPIRV
16+
MLIRVectorToSPIRV
1617
MLIRSPIRVDialect
1718
MLIRSPIRVConversion
1819
MLIRSupport

‎mlir/lib/Conversion/GPUToSPIRV/GPUToSPIRVPass.cpp

Lines changed: 2 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -18,6 +18,7 @@
1818
#include "mlir/Conversion/GPUToSPIRV/GPUToSPIRV.h"
1919
#include "mlir/Conversion/MemRefToSPIRV/MemRefToSPIRV.h"
2020
#include "mlir/Conversion/SCFToSPIRV/SCFToSPIRV.h"
21+
#include "mlir/Conversion/VectorToSPIRV/VectorToSPIRV.h"
2122
#include "mlir/Dialect/Func/IR/FuncOps.h"
2223
#include "mlir/Dialect/GPU/IR/GPUDialect.h"
2324
#include "mlir/Dialect/SPIRV/IR/SPIRVDialect.h"
@@ -132,6 +133,7 @@ void GPUToSPIRVPass::runOnOperation() {
132133
mlir::arith::populateArithToSPIRVPatterns(typeConverter, patterns);
133134
populateMemRefToSPIRVPatterns(typeConverter, patterns);
134135
populateFuncToSPIRVPatterns(typeConverter, patterns);
136+
populateVectorToSPIRVPatterns(typeConverter, patterns);
135137

136138
if (failed(applyFullConversion(gpuModule, *target, std::move(patterns))))
137139
return signalPassFailure();

‎mlir/test/mlir-vulkan-runner/vector-interleave.mlir

Lines changed: 79 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,79 @@
1+
// RUN: mlir-vulkan-runner %s \
2+
// RUN: --shared-libs=%vulkan-runtime-wrappers,%mlir_runner_utils \
3+
// RUN: --entry-point-result=void | FileCheck %s
4+
5+
// CHECK: [0, 2, 1, 3]
6+
module attributes {
7+
gpu.container_module,
8+
spirv.target_env = #spirv.target_env<
9+
#spirv.vce<v1.0, [Shader], [SPV_KHR_storage_buffer_storage_class]>, #spirv.resource_limits<>>
10+
} {
11+
gpu.module @kernels {
12+
gpu.func @kernel_vector_interleave(%arg0 : memref<2xi32>, %arg1 : memref<2xi32>, %arg2 : memref<4xi32>)
13+
kernel attributes { spirv.entry_point_abi = #spirv.entry_point_abi<workgroup_size = [1, 1, 1]>} {
14+
%idx0 = arith.constant 0 : index
15+
%idx1 = arith.constant 1 : index
16+
%idx2 = arith.constant 2 : index
17+
%idx3 = arith.constant 3 : index
18+
%idx4 = arith.constant 4 : index
19+
20+
%lhs = arith.constant dense<[0, 0]> : vector<2xi32>
21+
%rhs = arith.constant dense<[0, 0]> : vector<2xi32>
22+
23+
%val0 = memref.load %arg0[%idx0] : memref<2xi32>
24+
%val1 = memref.load %arg0[%idx1] : memref<2xi32>
25+
%val2 = memref.load %arg1[%idx0] : memref<2xi32>
26+
%val3 = memref.load %arg1[%idx1] : memref<2xi32>
27+
28+
%lhs0 = vector.insertelement %val0, %lhs[%idx0 : index] : vector<2xi32>
29+
%lhs1 = vector.insertelement %val1, %lhs0[%idx1 : index] : vector<2xi32>
30+
%rhs0 = vector.insertelement %val2, %rhs[%idx0 : index] : vector<2xi32>
31+
%rhs1 = vector.insertelement %val3, %rhs0[%idx1 : index] : vector<2xi32>
32+
33+
%interleave = vector.interleave %lhs1, %rhs1 : vector<2xi32> -> vector<4xi32>
34+
35+
%res0 = vector.extractelement %interleave[%idx0 : index] : vector<4xi32>
36+
%res1 = vector.extractelement %interleave[%idx1 : index] : vector<4xi32>
37+
%res2 = vector.extractelement %interleave[%idx2 : index] : vector<4xi32>
38+
%res3 = vector.extractelement %interleave[%idx3 : index] : vector<4xi32>
39+
40+
memref.store %res0, %arg2[%idx0]: memref<4xi32>
41+
memref.store %res1, %arg2[%idx1]: memref<4xi32>
42+
memref.store %res2, %arg2[%idx2]: memref<4xi32>
43+
memref.store %res3, %arg2[%idx3]: memref<4xi32>
44+
45+
gpu.return
46+
}
47+
}
48+
49+
func.func @main() {
50+
// Allocate 3 buffers.
51+
%buf0 = memref.alloc() : memref<2xi32>
52+
%buf1 = memref.alloc() : memref<2xi32>
53+
%buf2 = memref.alloc() : memref<4xi32>
54+
55+
%idx0 = arith.constant 0 : index
56+
%idx1 = arith.constant 1 : index
57+
%idx4 = arith.constant 4 : index
58+
59+
// Initialize input buffer.
60+
%buf0_vals = arith.constant dense<[0, 1]> : vector<2xi32>
61+
%buf1_vals = arith.constant dense<[2, 3]> : vector<2xi32>
62+
vector.store %buf0_vals, %buf0[%idx0] : memref<2xi32>, vector<2xi32>
63+
vector.store %buf1_vals, %buf1[%idx0] : memref<2xi32>, vector<2xi32>
64+
65+
// Initialize output buffer.
66+
%value0 = arith.constant 0 : i32
67+
%buf3 = memref.cast %buf2 : memref<4xi32> to memref<?xi32>
68+
call @fillResource1DInt(%buf3, %value0) : (memref<?xi32>, i32) -> ()
69+
70+
gpu.launch_func @kernels::@kernel_vector_interleave
71+
blocks in (%idx4, %idx1, %idx1) threads in (%idx1, %idx1, %idx1)
72+
args(%buf0 : memref<2xi32>, %buf1 : memref<2xi32>, %buf2 : memref<4xi32>)
73+
%buf4 = memref.cast %buf3 : memref<?xi32> to memref<*xi32>
74+
call @printMemrefI32(%buf4) : (memref<*xi32>) -> ()
75+
return
76+
}
77+
func.func private @fillResource1DInt(%0 : memref<?xi32>, %1 : i32)
78+
func.func private @printMemrefI32(%ptr : memref<*xi32>)
79+
}

‎mlir/test/mlir-vulkan-runner/vector-shuffle.mlir

Lines changed: 79 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,79 @@
1+
// RUN: mlir-vulkan-runner %s \
2+
// RUN: --shared-libs=%vulkan-runtime-wrappers,%mlir_runner_utils \
3+
// RUN: --entry-point-result=void | FileCheck %s
4+
5+
// CHECK: [2, 1, 3, 3]
6+
module attributes {
7+
gpu.container_module,
8+
spirv.target_env = #spirv.target_env<
9+
#spirv.vce<v1.0, [Shader], [SPV_KHR_storage_buffer_storage_class]>, #spirv.resource_limits<>>
10+
} {
11+
gpu.module @kernels {
12+
gpu.func @kernel_vector_shuffle(%arg0 : memref<2xi32>, %arg1 : memref<2xi32>, %arg2 : memref<4xi32>)
13+
kernel attributes { spirv.entry_point_abi = #spirv.entry_point_abi<workgroup_size = [1, 1, 1]>} {
14+
%idx0 = arith.constant 0 : index
15+
%idx1 = arith.constant 1 : index
16+
%idx2 = arith.constant 2 : index
17+
%idx3 = arith.constant 3 : index
18+
%idx4 = arith.constant 4 : index
19+
20+
%lhs = arith.constant dense<[0, 0]> : vector<2xi32>
21+
%rhs = arith.constant dense<[0, 0]> : vector<2xi32>
22+
23+
%val0 = memref.load %arg0[%idx0] : memref<2xi32>
24+
%val1 = memref.load %arg0[%idx1] : memref<2xi32>
25+
%val2 = memref.load %arg1[%idx0] : memref<2xi32>
26+
%val3 = memref.load %arg1[%idx1] : memref<2xi32>
27+
28+
%lhs0 = vector.insertelement %val0, %lhs[%idx0 : index] : vector<2xi32>
29+
%lhs1 = vector.insertelement %val1, %lhs0[%idx1 : index] : vector<2xi32>
30+
%rhs0 = vector.insertelement %val2, %rhs[%idx0 : index] : vector<2xi32>
31+
%rhs1 = vector.insertelement %val3, %rhs0[%idx1 : index] : vector<2xi32>
32+
33+
%shuffle = vector.shuffle %lhs1, %rhs1[2, 1, 3, 3] : vector<2xi32>, vector<2xi32>
34+
35+
%res0 = vector.extractelement %shuffle[%idx0 : index] : vector<4xi32>
36+
%res1 = vector.extractelement %shuffle[%idx1 : index] : vector<4xi32>
37+
%res2 = vector.extractelement %shuffle[%idx2 : index] : vector<4xi32>
38+
%res3 = vector.extractelement %shuffle[%idx3 : index] : vector<4xi32>
39+
40+
memref.store %res0, %arg2[%idx0]: memref<4xi32>
41+
memref.store %res1, %arg2[%idx1]: memref<4xi32>
42+
memref.store %res2, %arg2[%idx2]: memref<4xi32>
43+
memref.store %res3, %arg2[%idx3]: memref<4xi32>
44+
45+
gpu.return
46+
}
47+
}
48+
49+
func.func @main() {
50+
// Allocate 3 buffers.
51+
%buf0 = memref.alloc() : memref<2xi32>
52+
%buf1 = memref.alloc() : memref<2xi32>
53+
%buf2 = memref.alloc() : memref<4xi32>
54+
55+
%idx0 = arith.constant 0 : index
56+
%idx1 = arith.constant 1 : index
57+
%idx4 = arith.constant 4 : index
58+
59+
// Initialize input buffer.
60+
%buf0_vals = arith.constant dense<[0, 1]> : vector<2xi32>
61+
%buf1_vals = arith.constant dense<[2, 3]> : vector<2xi32>
62+
vector.store %buf0_vals, %buf0[%idx0] : memref<2xi32>, vector<2xi32>
63+
vector.store %buf1_vals, %buf1[%idx0] : memref<2xi32>, vector<2xi32>
64+
65+
// Initialize output buffer.
66+
%value0 = arith.constant 0 : i32
67+
%buf3 = memref.cast %buf2 : memref<4xi32> to memref<?xi32>
68+
call @fillResource1DInt(%buf3, %value0) : (memref<?xi32>, i32) -> ()
69+
70+
gpu.launch_func @kernels::@kernel_vector_shuffle
71+
blocks in (%idx4, %idx1, %idx1) threads in (%idx1, %idx1, %idx1)
72+
args(%buf0 : memref<2xi32>, %buf1 : memref<2xi32>, %buf2 : memref<4xi32>)
73+
%buf4 = memref.cast %buf3 : memref<?xi32> to memref<*xi32>
74+
call @printMemrefI32(%buf4) : (memref<*xi32>) -> ()
75+
return
76+
}
77+
func.func private @fillResource1DInt(%0 : memref<?xi32>, %1 : i32)
78+
func.func private @printMemrefI32(%ptr : memref<*xi32>)
79+
}

0 commit comments

Comments
 (0)