b6c453c13f
This commit fixes memory leaks in sparse tensor integration tests by adding `bufferization.dealloc_tensor` ops. Note: Buffer deallocation will be automated in the future with the ownership-based buffer deallocation pass, making `dealloc_tensor` obsolete (only codegen path, not when using the runtime library).
275 lines
10 KiB
MLIR
275 lines
10 KiB
MLIR
//--------------------------------------------------------------------------------------------------
|
|
// WHEN CREATING A NEW TEST, PLEASE JUST COPY & PASTE WITHOUT EDITS.
|
|
//
|
|
// Set-up that's shared across all tests in this directory. In principle, this
|
|
// config could be moved to lit.local.cfg. However, there are downstream users that
|
|
// do not use these LIT config files. Hence why this is kept inline.
|
|
//
|
|
// DEFINE: %{sparsifier_opts} = enable-runtime-library=true
|
|
// DEFINE: %{sparsifier_opts_sve} = enable-arm-sve=true %{sparsifier_opts}
|
|
// DEFINE: %{compile} = mlir-opt %s --sparsifier="%{sparsifier_opts}"
|
|
// DEFINE: %{compile_sve} = mlir-opt %s --sparsifier="%{sparsifier_opts_sve}"
|
|
// DEFINE: %{run_libs} = -shared-libs=%mlir_c_runner_utils,%mlir_runner_utils
|
|
// DEFINE: %{run_opts} = -e main -entry-point-result=void
|
|
// DEFINE: %{run} = mlir-cpu-runner %{run_opts} %{run_libs}
|
|
// DEFINE: %{run_sve} = %mcr_aarch64_cmd --march=aarch64 --mattr="+sve" %{run_opts} %{run_libs}
|
|
//
|
|
// DEFINE: %{env} =
|
|
//--------------------------------------------------------------------------------------------------
|
|
|
|
// REDEFINE: %{env} = TENSOR0=%mlir_src_dir/test/Integration/data/mttkrp_b.tns
|
|
// RUN: %{compile} | env %{env} %{run} | FileCheck %s
|
|
//
|
|
// Do the same run, but now with direct IR generation.
|
|
// REDEFINE: %{sparsifier_opts} = enable-runtime-library=false
|
|
// RUN: %{compile} | env %{env} %{run} | FileCheck %s
|
|
|
|
!Filename = !llvm.ptr
|
|
|
|
#S1 = #sparse_tensor.encoding<{
|
|
map = (d0, d1, d2) -> (d0 : compressed, d1 : compressed, d2 : compressed)
|
|
}>
|
|
|
|
#S2 = #sparse_tensor.encoding<{
|
|
map = (d0, d1, d2) -> (d0 : compressed, d2 : compressed, d1 : compressed)
|
|
}>
|
|
|
|
#S3 = #sparse_tensor.encoding<{
|
|
map = (d0, d1, d2) -> (d1 : compressed, d0 : compressed, d2 : compressed)
|
|
}>
|
|
|
|
#S4 = #sparse_tensor.encoding<{
|
|
map = (d0, d1, d2) -> (d1 : compressed, d2 : compressed, d0 : compressed)
|
|
}>
|
|
|
|
#S5 = #sparse_tensor.encoding<{
|
|
map = (d0, d1, d2) -> (d2 : compressed, d0 : compressed, d1 : compressed)
|
|
}>
|
|
|
|
#S6 = #sparse_tensor.encoding<{
|
|
map = (d0, d1, d2) -> (d2 : compressed, d1 : compressed, d0 : compressed)
|
|
}>
|
|
|
|
#trait_3d = {
|
|
indexing_maps = [
|
|
affine_map<(i,j,k) -> (i,j,k)>, // B
|
|
affine_map<(i,j,k) -> (i,j,k)> // A (out)
|
|
],
|
|
iterator_types = ["parallel", "parallel", "parallel"],
|
|
doc = "A(i,j,k) = B(i,j,k)"
|
|
}
|
|
|
|
//
|
|
// Integration test that lowers a kernel annotated as sparse to
|
|
// actual sparse code, initializes a matching sparse storage scheme
|
|
// from file, and runs the resulting code with the JIT compiler.
|
|
//
|
|
module {
|
|
func.func private @getTensorFilename(index) -> (!Filename)
|
|
|
|
func.func @dump(%a: tensor<2x3x4xf64>) {
|
|
%c0 = arith.constant 0 : index
|
|
%f0 = arith.constant 0.0 : f64
|
|
%v = vector.transfer_read %a[%c0, %c0, %c0], %f0 : tensor<2x3x4xf64>, vector<2x3x4xf64>
|
|
vector.print %v : vector<2x3x4xf64>
|
|
return
|
|
}
|
|
|
|
//// S1
|
|
|
|
func.func @linalg1(%b: tensor<2x3x4xf64, #S1>)-> tensor<2x3x4xf64> {
|
|
%0 = arith.constant dense<0.000000e+00> : tensor<2x3x4xf64>
|
|
%a = linalg.generic #trait_3d
|
|
ins(%b: tensor<2x3x4xf64, #S1>)
|
|
outs(%0: tensor<2x3x4xf64>) {
|
|
^bb(%x: f64, %y: f64):
|
|
linalg.yield %x : f64
|
|
} -> tensor<2x3x4xf64>
|
|
return %a : tensor<2x3x4xf64>
|
|
}
|
|
|
|
func.func @convert1(%b: tensor<2x3x4xf64, #S1>) -> tensor<2x3x4xf64> {
|
|
%a = sparse_tensor.convert %b : tensor<2x3x4xf64, #S1> to tensor<2x3x4xf64>
|
|
return %a : tensor<2x3x4xf64>
|
|
}
|
|
|
|
func.func @foo1(%fileName : !Filename) {
|
|
%b = sparse_tensor.new %fileName : !Filename to tensor<2x3x4xf64, #S1>
|
|
%0 = call @linalg1(%b) : (tensor<2x3x4xf64, #S1>) -> tensor<2x3x4xf64>
|
|
call @dump(%0) : (tensor<2x3x4xf64>) -> ()
|
|
%1 = call @convert1(%b) : (tensor<2x3x4xf64, #S1>) -> tensor<2x3x4xf64>
|
|
call @dump(%1) : (tensor<2x3x4xf64>) -> ()
|
|
bufferization.dealloc_tensor %0 : tensor<2x3x4xf64>
|
|
bufferization.dealloc_tensor %b : tensor<2x3x4xf64, #S1>
|
|
bufferization.dealloc_tensor %1 : tensor<2x3x4xf64>
|
|
return
|
|
}
|
|
|
|
//// S2
|
|
|
|
func.func @linalg2(%b: tensor<2x3x4xf64, #S2>)-> tensor<2x3x4xf64> {
|
|
%0 = arith.constant dense<0.000000e+00> : tensor<2x3x4xf64>
|
|
%a = linalg.generic #trait_3d
|
|
ins(%b: tensor<2x3x4xf64, #S2>)
|
|
outs(%0: tensor<2x3x4xf64>) {
|
|
^bb(%x: f64, %y: f64):
|
|
linalg.yield %x : f64
|
|
} -> tensor<2x3x4xf64>
|
|
return %a : tensor<2x3x4xf64>
|
|
}
|
|
|
|
func.func @convert2(%b: tensor<2x3x4xf64, #S2>) -> tensor<2x3x4xf64> {
|
|
%a = sparse_tensor.convert %b : tensor<2x3x4xf64, #S2> to tensor<2x3x4xf64>
|
|
return %a : tensor<2x3x4xf64>
|
|
}
|
|
|
|
func.func @foo2(%fileName : !Filename) {
|
|
%b = sparse_tensor.new %fileName : !Filename to tensor<2x3x4xf64, #S2>
|
|
%0 = call @linalg2(%b) : (tensor<2x3x4xf64, #S2>) -> tensor<2x3x4xf64>
|
|
call @dump(%0) : (tensor<2x3x4xf64>) -> ()
|
|
%2 = call @convert2(%b) : (tensor<2x3x4xf64, #S2>) -> tensor<2x3x4xf64>
|
|
call @dump(%2) : (tensor<2x3x4xf64>) -> ()
|
|
bufferization.dealloc_tensor %0 : tensor<2x3x4xf64>
|
|
bufferization.dealloc_tensor %b : tensor<2x3x4xf64, #S2>
|
|
bufferization.dealloc_tensor %2 : tensor<2x3x4xf64>
|
|
return
|
|
}
|
|
|
|
//// S3
|
|
|
|
func.func @linalg3(%b: tensor<2x3x4xf64, #S3>)-> tensor<2x3x4xf64> {
|
|
%0 = arith.constant dense<0.000000e+00> : tensor<2x3x4xf64>
|
|
%a = linalg.generic #trait_3d
|
|
ins(%b: tensor<2x3x4xf64, #S3>)
|
|
outs(%0: tensor<2x3x4xf64>) {
|
|
^bb(%x: f64, %y: f64):
|
|
linalg.yield %x : f64
|
|
} -> tensor<2x3x4xf64>
|
|
return %a : tensor<2x3x4xf64>
|
|
}
|
|
|
|
func.func @convert3(%b: tensor<2x3x4xf64, #S3>) -> tensor<2x3x4xf64> {
|
|
%a = sparse_tensor.convert %b : tensor<2x3x4xf64, #S3> to tensor<2x3x4xf64>
|
|
return %a : tensor<2x3x4xf64>
|
|
}
|
|
|
|
func.func @foo3(%fileName : !Filename) {
|
|
%b = sparse_tensor.new %fileName : !Filename to tensor<2x3x4xf64, #S3>
|
|
%0 = call @linalg3(%b) : (tensor<2x3x4xf64, #S3>) -> tensor<2x3x4xf64>
|
|
call @dump(%0) : (tensor<2x3x4xf64>) -> ()
|
|
%3 = call @convert3(%b) : (tensor<2x3x4xf64, #S3>) -> tensor<2x3x4xf64>
|
|
call @dump(%3) : (tensor<2x3x4xf64>) -> ()
|
|
bufferization.dealloc_tensor %0 : tensor<2x3x4xf64>
|
|
bufferization.dealloc_tensor %b : tensor<2x3x4xf64, #S3>
|
|
bufferization.dealloc_tensor %3 : tensor<2x3x4xf64>
|
|
return
|
|
}
|
|
|
|
//// S4
|
|
|
|
func.func @linalg4(%b: tensor<2x3x4xf64, #S4>)-> tensor<2x3x4xf64> {
|
|
%0 = arith.constant dense<0.000000e+00> : tensor<2x3x4xf64>
|
|
%a = linalg.generic #trait_3d
|
|
ins(%b: tensor<2x3x4xf64, #S4>)
|
|
outs(%0: tensor<2x3x4xf64>) {
|
|
^bb(%x: f64, %y: f64):
|
|
linalg.yield %x : f64
|
|
} -> tensor<2x3x4xf64>
|
|
return %a : tensor<2x3x4xf64>
|
|
}
|
|
|
|
func.func @convert4(%b: tensor<2x3x4xf64, #S4>) -> tensor<2x3x4xf64> {
|
|
%a = sparse_tensor.convert %b : tensor<2x3x4xf64, #S4> to tensor<2x3x4xf64>
|
|
return %a : tensor<2x3x4xf64>
|
|
}
|
|
|
|
func.func @foo4(%fileName : !Filename) {
|
|
%b = sparse_tensor.new %fileName : !Filename to tensor<2x3x4xf64, #S4>
|
|
%0 = call @linalg4(%b) : (tensor<2x3x4xf64, #S4>) -> tensor<2x3x4xf64>
|
|
call @dump(%0) : (tensor<2x3x4xf64>) -> ()
|
|
%4 = call @convert4(%b) : (tensor<2x3x4xf64, #S4>) -> tensor<2x3x4xf64>
|
|
call @dump(%4) : (tensor<2x3x4xf64>) -> ()
|
|
bufferization.dealloc_tensor %0 : tensor<2x3x4xf64>
|
|
bufferization.dealloc_tensor %b : tensor<2x3x4xf64, #S4>
|
|
bufferization.dealloc_tensor %4 : tensor<2x3x4xf64>
|
|
return
|
|
}
|
|
|
|
//// S5
|
|
|
|
func.func @linalg5(%b: tensor<2x3x4xf64, #S5>)-> tensor<2x3x4xf64> {
|
|
%0 = arith.constant dense<0.000000e+00> : tensor<2x3x4xf64>
|
|
%a = linalg.generic #trait_3d
|
|
ins(%b: tensor<2x3x4xf64, #S5>)
|
|
outs(%0: tensor<2x3x4xf64>) {
|
|
^bb(%x: f64, %y: f64):
|
|
linalg.yield %x : f64
|
|
} -> tensor<2x3x4xf64>
|
|
return %a : tensor<2x3x4xf64>
|
|
}
|
|
|
|
func.func @convert5(%b: tensor<2x3x4xf64, #S5>) -> tensor<2x3x4xf64> {
|
|
%a = sparse_tensor.convert %b : tensor<2x3x4xf64, #S5> to tensor<2x3x4xf64>
|
|
return %a : tensor<2x3x4xf64>
|
|
}
|
|
|
|
func.func @foo5(%fileName : !Filename) {
|
|
%b = sparse_tensor.new %fileName : !Filename to tensor<2x3x4xf64, #S5>
|
|
%0 = call @linalg5(%b) : (tensor<2x3x4xf64, #S5>) -> tensor<2x3x4xf64>
|
|
call @dump(%0) : (tensor<2x3x4xf64>) -> ()
|
|
%5 = call @convert5(%b) : (tensor<2x3x4xf64, #S5>) -> tensor<2x3x4xf64>
|
|
call @dump(%5) : (tensor<2x3x4xf64>) -> ()
|
|
bufferization.dealloc_tensor %0 : tensor<2x3x4xf64>
|
|
bufferization.dealloc_tensor %b : tensor<2x3x4xf64, #S5>
|
|
bufferization.dealloc_tensor %5 : tensor<2x3x4xf64>
|
|
return
|
|
}
|
|
|
|
//// S6
|
|
|
|
func.func @linalg6(%b: tensor<2x3x4xf64, #S6>)-> tensor<2x3x4xf64> {
|
|
%0 = arith.constant dense<0.000000e+00> : tensor<2x3x4xf64>
|
|
%a = linalg.generic #trait_3d
|
|
ins(%b: tensor<2x3x4xf64, #S6>)
|
|
outs(%0: tensor<2x3x4xf64>) {
|
|
^bb(%x: f64, %y: f64):
|
|
linalg.yield %x : f64
|
|
} -> tensor<2x3x4xf64>
|
|
return %a : tensor<2x3x4xf64>
|
|
}
|
|
|
|
func.func @convert6(%b: tensor<2x3x4xf64, #S6>) -> tensor<2x3x4xf64> {
|
|
%a = sparse_tensor.convert %b : tensor<2x3x4xf64, #S6> to tensor<2x3x4xf64>
|
|
return %a : tensor<2x3x4xf64>
|
|
}
|
|
|
|
func.func @foo6(%fileName : !Filename) {
|
|
%b = sparse_tensor.new %fileName : !Filename to tensor<2x3x4xf64, #S6>
|
|
%0 = call @linalg6(%b) : (tensor<2x3x4xf64, #S6>) -> tensor<2x3x4xf64>
|
|
call @dump(%0) : (tensor<2x3x4xf64>) -> ()
|
|
%6 = call @convert6(%b) : (tensor<2x3x4xf64, #S6>) -> tensor<2x3x4xf64>
|
|
call @dump(%6) : (tensor<2x3x4xf64>) -> ()
|
|
bufferization.dealloc_tensor %0 : tensor<2x3x4xf64>
|
|
bufferization.dealloc_tensor %b : tensor<2x3x4xf64, #S6>
|
|
bufferization.dealloc_tensor %6 : tensor<2x3x4xf64>
|
|
return
|
|
}
|
|
|
|
//
|
|
// Main driver.
|
|
//
|
|
// CHECK-COUNT-12: ( ( ( 0, 0, 3, 63 ), ( 0, 11, 100, 0 ), ( 66, 61, 13, 43 ) ), ( ( 77, 0, 10, 46 ), ( 61, 53, 3, 75 ), ( 0, 22, 18, 0 ) ) )
|
|
//
|
|
func.func @main() {
|
|
%c0 = arith.constant 0 : index
|
|
%fileName = call @getTensorFilename(%c0) : (index) -> (!Filename)
|
|
call @foo1(%fileName) : (!Filename) -> ()
|
|
call @foo2(%fileName) : (!Filename) -> ()
|
|
call @foo3(%fileName) : (!Filename) -> ()
|
|
call @foo4(%fileName) : (!Filename) -> ()
|
|
call @foo5(%fileName) : (!Filename) -> ()
|
|
call @foo6(%fileName) : (!Filename) -> ()
|
|
return
|
|
}
|
|
}
|