shylie/llvm-project
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[mlir][verifyMemref] Fix bug and support more types for verifyMemref (#77682)

Browse Source 
1. Fix a bug in verifyMemref to pass in `data` instead of `baseptr`,
which didn't verify data correctly.
2. Add `==` for f16 and bf16.
3. Add a comprehensive test of verifyMemref for all supported types.
This commit is contained in:
Yinying Li 2024-01-10 20:04:43 -05:00 committed by GitHub
parent 03be448cce
commit 753dc0a01c
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
5 changed files with 217 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
mlir/include/mlir/ExecutionEngine/Float16bits.h
View File

@ -48,6 +48,9 @@ MLIR_FLOAT16_EXPORT std::ostream &operator<<(std::ostream &os, const f16 &f);
// Outputs a bfloat value.
MLIR_FLOAT16_EXPORT std::ostream &operator<<(std::ostream &os, const bf16 &d);
MLIR_FLOAT16_EXPORT bool operator==(const f16 &f1, const f16 &f2);
MLIR_FLOAT16_EXPORT bool operator==(const bf16 &bf1, const bf16 &bf2);
extern "C" MLIR_FLOAT16_EXPORT void printF16(uint16_t bits);
extern "C" MLIR_FLOAT16_EXPORT void printBF16(uint16_t bits);

16
mlir/include/mlir/ExecutionEngine/RunnerUtils.h
View File

@ -331,9 +331,9 @@ int64_t verifyMemRef(const DynamicMemRefType<T> &actual,
}
// Return the number of errors.
int64_t printCounter = 0;
return MemRefDataVerifier<T>::verify(
std::cerr, actual.basePtr, expected.basePtr, actual.rank, actual.offset,
actual.sizes, actual.strides, printCounter);
return MemRefDataVerifier<T>::verify(std::cerr, actual.data, expected.data,
actual.rank, actual.offset, actual.sizes,
actual.strides, printCounter);
}
/// Verify the equivalence of two unranked memrefs and return the number of
@ -429,8 +429,18 @@ _mlir_ciface_printMemref1dC64(StridedMemRefType<impl::complex64, 1> *m);
extern "C" MLIR_RUNNERUTILS_EXPORT void _mlir_ciface_printMemrefVector4x4xf32(
StridedMemRefType<Vector2D<4, 4, float>, 2> *m);
extern "C" MLIR_RUNNERUTILS_EXPORT int64_t _mlir_ciface_verifyMemRefI8(
UnrankedMemRefType<int8_t> *actual, UnrankedMemRefType<int8_t> *expected);
extern "C" MLIR_RUNNERUTILS_EXPORT int64_t _mlir_ciface_verifyMemRefI16(
UnrankedMemRefType<int16_t> *actual, UnrankedMemRefType<int16_t> *expected);
extern "C" MLIR_RUNNERUTILS_EXPORT int64_t _mlir_ciface_verifyMemRefI32(
UnrankedMemRefType<int32_t> *actual, UnrankedMemRefType<int32_t> *expected);
extern "C" MLIR_RUNNERUTILS_EXPORT int64_t _mlir_ciface_verifyMemRefI64(
UnrankedMemRefType<int64_t> *actual, UnrankedMemRefType<int64_t> *expected);
extern "C" MLIR_RUNNERUTILS_EXPORT int64_t _mlir_ciface_verifyMemRefBF16(
UnrankedMemRefType<bf16> *actual, UnrankedMemRefType<bf16> *expected);
extern "C" MLIR_RUNNERUTILS_EXPORT int64_t _mlir_ciface_verifyMemRefF16(
UnrankedMemRefType<f16> *actual, UnrankedMemRefType<f16> *expected);
extern "C" MLIR_RUNNERUTILS_EXPORT int64_t _mlir_ciface_verifyMemRefF32(
UnrankedMemRefType<float> *actual, UnrankedMemRefType<float> *expected);
extern "C" MLIR_RUNNERUTILS_EXPORT int64_t _mlir_ciface_verifyMemRefF64(

4
mlir/lib/ExecutionEngine/Float16bits.cpp
View File

@ -150,6 +150,10 @@ std::ostream &operator<<(std::ostream &os, const bf16 &d) {
return os;
}
bool operator==(const f16 &f1, const f16 &f2) { return f1.bits == f2.bits; }
bool operator==(const bf16 &f1, const bf16 &f2) { return f1.bits == f2.bits; }
// Mark these symbols as weak so they don't conflict when compiler-rt also
// defines them.
#define ATTR_WEAK

30
mlir/lib/ExecutionEngine/RunnerUtils.cpp
View File

@ -219,12 +219,42 @@ _mlir_ciface_printMemref1dC64(StridedMemRefType<impl::complex64, 1> *M) {
impl::printMemRef(*M);
}
extern "C" int64_t
_mlir_ciface_verifyMemRefI8(UnrankedMemRefType<int8_t> *actual,
UnrankedMemRefType<int8_t> *expected) {
return impl::verifyMemRef(*actual, *expected);
}
extern "C" int64_t
_mlir_ciface_verifyMemRefI16(UnrankedMemRefType<int16_t> *actual,
UnrankedMemRefType<int16_t> *expected) {
return impl::verifyMemRef(*actual, *expected);
}
extern "C" int64_t
_mlir_ciface_verifyMemRefI32(UnrankedMemRefType<int32_t> *actual,
UnrankedMemRefType<int32_t> *expected) {
return impl::verifyMemRef(*actual, *expected);
}
extern "C" int64_t
_mlir_ciface_verifyMemRefI64(UnrankedMemRefType<int64_t> *actual,
UnrankedMemRefType<int64_t> *expected) {
return impl::verifyMemRef(*actual, *expected);
}
extern "C" int64_t
_mlir_ciface_verifyMemRefF16(UnrankedMemRefType<f16> *actual,
UnrankedMemRefType<f16> *expected) {
return impl::verifyMemRef(*actual, *expected);
}
extern "C" int64_t
_mlir_ciface_verifyMemRefBF16(UnrankedMemRefType<bf16> *actual,
UnrankedMemRefType<bf16> *expected) {
return impl::verifyMemRef(*actual, *expected);
}
extern "C" int64_t
_mlir_ciface_verifyMemRefF32(UnrankedMemRefType<float> *actual,
UnrankedMemRefType<float> *expected) {

167
mlir/test/Integration/Dialect/Memref/verify-memref.mlir Normal file
View File

@ -0,0 +1,167 @@
// RUN: mlir-opt %s \
// RUN: -func-bufferize -arith-bufferize --canonicalize \
// RUN: -convert-vector-to-scf -convert-scf-to-cf -convert-vector-to-llvm -finalize-memref-to-llvm\
// RUN: -convert-func-to-llvm -reconcile-unrealized-casts |\
// RUN: mlir-cpu-runner \
// RUN: -e entry -entry-point-result=void \
// RUN: -shared-libs=%mlir_c_runner_utils,%mlir_runner_utils |\
// RUN: FileCheck %s
module {
func.func private @verifyMemRefI8(%a : tensor<*xi8>, %b : tensor<*xi8>) -> i64 attributes { llvm.emit_c_interface }
func.func private @verifyMemRefI16(%a : tensor<*xi16>, %b : tensor<*xi16>) -> i64 attributes { llvm.emit_c_interface }
func.func private @verifyMemRefI32(%a : tensor<*xi32>, %b : tensor<*xi32>) -> i64 attributes { llvm.emit_c_interface }
func.func private @verifyMemRefI64(%a : tensor<*xi64>, %b : tensor<*xi64>) -> i64 attributes { llvm.emit_c_interface }
func.func private @verifyMemRefBF16(%a : tensor<*xbf16>, %b : tensor<*xbf16>) -> i64 attributes { llvm.emit_c_interface }
func.func private @verifyMemRefF16(%a : tensor<*xf16>, %b : tensor<*xf16>) -> i64 attributes { llvm.emit_c_interface }
func.func private @verifyMemRefF32(%a : tensor<*xf32>, %b : tensor<*xf32>) -> i64 attributes { llvm.emit_c_interface }
func.func private @verifyMemRefF64(%a : tensor<*xf64>, %b : tensor<*xf64>) -> i64 attributes { llvm.emit_c_interface }
func.func private @verifyMemRefC32(%a : tensor<*xcomplex<f32>>, %b : tensor<*xcomplex<f32>>) -> i64 attributes { llvm.emit_c_interface }
func.func private @verifyMemRefC64(%a : tensor<*xcomplex<f64>>, %b : tensor<*xcomplex<f64>>) -> i64 attributes { llvm.emit_c_interface }
func.func private @verifyMemRefInd(%a : tensor<*xindex>, %b : tensor<*xindex>) -> i64 attributes { llvm.emit_c_interface }
func.func @entry() {
%i8 = arith.constant dense<90> : tensor<3x3xi8>
%i16 = arith.constant dense<1> : tensor<3x3xi16>
%i32 = arith.constant dense<2> : tensor<3x3xi32>
%i64 = arith.constant dense<3> : tensor<3x3xi64>
%f16 = arith.constant dense<1.5> : tensor<3x3xf16>
%bf16 = arith.constant dense<2.5> : tensor<3x3xbf16>
%f32 = arith.constant dense<3.5> : tensor<3x3xf32>
%f64 = arith.constant dense<4.5> : tensor<3x3xf64>
%c32 = arith.constant dense<(1.000000e+01,5.000000e+00)> : tensor<3x3xcomplex<f32>>
%c64 = arith.constant dense<(2.000000e+01,5.000000e+00)> : tensor<3x3xcomplex<f64>>
%ind = arith.constant dense<4> : tensor<3x3xindex>
%1 = tensor.cast %i8 : tensor<3x3xi8> to tensor<*xi8>
%2 = tensor.cast %i16 : tensor<3x3xi16> to tensor<*xi16>
%3 = tensor.cast %i32 : tensor<3x3xi32> to tensor<*xi32>
%4 = tensor.cast %i64 : tensor<3x3xi64> to tensor<*xi64>
%5 = tensor.cast %f16 : tensor<3x3xf16> to tensor<*xf16>
%6 = tensor.cast %bf16 : tensor<3x3xbf16> to tensor<*xbf16>
%7 = tensor.cast %f32 : tensor<3x3xf32> to tensor<*xf32>
%8 = tensor.cast %f64 : tensor<3x3xf64> to tensor<*xf64>
%9 = tensor.cast %c32 : tensor<3x3xcomplex<f32>> to tensor<*xcomplex<f32>>
%10 = tensor.cast %c64 : tensor<3x3xcomplex<f64>> to tensor<*xcomplex<f64>>
%11 = tensor.cast %ind : tensor<3x3xindex> to tensor<*xindex>
//
// Ensure that verifyMemRef could detect equal memrefs.
//
// CHECK: 0
%res0 = call @verifyMemRefI8(%1, %1) : (tensor<*xi8>, tensor<*xi8>) -> (i64)
vector.print %res0 : i64
// CHECK-NEXT: 0
%res1 = call @verifyMemRefI16(%2, %2) : (tensor<*xi16>, tensor<*xi16>) -> (i64)
vector.print %res1 : i64
// CHECK-NEXT: 0
%res2 = call @verifyMemRefI32(%3, %3) : (tensor<*xi32>, tensor<*xi32>) -> (i64)
vector.print %res2 : i64
// CHECK-NEXT: 0
%res3 = call @verifyMemRefI64(%4, %4) : (tensor<*xi64>, tensor<*xi64>) -> (i64)
vector.print %res3 : i64
// CHECK-NEXT: 0
%res4 = call @verifyMemRefF16(%5, %5) : (tensor<*xf16>, tensor<*xf16>) -> (i64)
vector.print %res4 : i64
// CHECK-NEXT: 0
%res5 = call @verifyMemRefBF16(%6, %6) : (tensor<*xbf16>, tensor<*xbf16>) -> (i64)
vector.print %res5 : i64
// CHECK-NEXT: 0
%res6 = call @verifyMemRefF32(%7, %7) : (tensor<*xf32>, tensor<*xf32>) -> (i64)
vector.print %res6 : i64
// CHECK-NEXT: 0
%res7 = call @verifyMemRefF64(%8, %8) : (tensor<*xf64>, tensor<*xf64>) -> (i64)
vector.print %res7 : i64
// CHECK-NEXT: 0
%res8 = call @verifyMemRefC32(%9, %9) : (tensor<*xcomplex<f32>>, tensor<*xcomplex<f32>>) -> (i64)
vector.print %res8 : i64
// CHECK-NEXT: 0
%res9 = call @verifyMemRefC64(%10, %10) : (tensor<*xcomplex<f64>>, tensor<*xcomplex<f64>>) -> (i64)
vector.print %res9 : i64
// CHECK-NEXT: 0
%res10 = call @verifyMemRefInd(%11, %11) : (tensor<*xindex>, tensor<*xindex>) -> (i64)
vector.print %res10 : i64
//
// Ensure that verifyMemRef could detect the correct number of errors
// for unequal memrefs.
//
%m1 = arith.constant dense<100> : tensor<3x3xi8>
%f1 = tensor.cast %m1 : tensor<3x3xi8> to tensor<*xi8>
%fail_res1 = call @verifyMemRefI8(%1, %f1) : (tensor<*xi8>, tensor<*xi8>) -> (i64)
// CHECK-NEXT: 9
vector.print %fail_res1 : i64
%m2 = arith.constant dense<100> : tensor<3x3xi16>
%f2 = tensor.cast %m2 : tensor<3x3xi16> to tensor<*xi16>
%fail_res2 = call @verifyMemRefI16(%2, %f2) : (tensor<*xi16>, tensor<*xi16>) -> (i64)
// CHECK-NEXT: 9
vector.print %fail_res2 : i64
%m3 = arith.constant dense<100> : tensor<3x3xi32>
%f3 = tensor.cast %m3 : tensor<3x3xi32> to tensor<*xi32>
%fail_res3 = call @verifyMemRefI32(%3, %f3) : (tensor<*xi32>, tensor<*xi32>) -> (i64)
// CHECK-NEXT: 9
vector.print %fail_res3 : i64
%m4 = arith.constant dense<100> : tensor<3x3xi64>
%f4 = tensor.cast %m4 : tensor<3x3xi64> to tensor<*xi64>
%fail_res4 = call @verifyMemRefI64(%4, %f4) : (tensor<*xi64>, tensor<*xi64>) -> (i64)
// CHECK-NEXT: 9
vector.print %fail_res4 : i64
%m5 = arith.constant dense<100.0> : tensor<3x3xf16>
%f5 = tensor.cast %m5 : tensor<3x3xf16> to tensor<*xf16>
%fail_res5 = call @verifyMemRefF16(%5, %f5) : (tensor<*xf16>, tensor<*xf16>) -> (i64)
// CHECK-NEXT: 9
vector.print %fail_res5 : i64
%m6 = arith.constant dense<100.0> : tensor<3x3xbf16>
%f6 = tensor.cast %m6 : tensor<3x3xbf16> to tensor<*xbf16>
%fail_res6 = call @verifyMemRefBF16(%6, %f6) : (tensor<*xbf16>, tensor<*xbf16>) -> (i64)
// CHECK-NEXT: 9
vector.print %fail_res6 : i64
%m7 = arith.constant dense<100.0> : tensor<3x3xf32>
%f7 = tensor.cast %m7 : tensor<3x3xf32> to tensor<*xf32>
%fail_res7 = call @verifyMemRefF32(%7, %f7) : (tensor<*xf32>, tensor<*xf32>) -> (i64)
// CHECK-NEXT: 9
vector.print %fail_res7 : i64
%m8 = arith.constant dense<100.0> : tensor<3x3xf64>
%f8 = tensor.cast %m8 : tensor<3x3xf64> to tensor<*xf64>
%fail_res8 = call @verifyMemRefF64(%8, %f8) : (tensor<*xf64>, tensor<*xf64>) -> (i64)
// CHECK-NEXT: 9
vector.print %fail_res8 : i64
%m9 = arith.constant dense<(5.000000e+01,1.000000e+00)> : tensor<3x3xcomplex<f32>>
%f9 = tensor.cast %m9 : tensor<3x3xcomplex<f32>> to tensor<*xcomplex<f32>>
%fail_res9 = call @verifyMemRefC32(%9, %f9) : (tensor<*xcomplex<f32>>, tensor<*xcomplex<f32>>) -> (i64)
// CHECK-NEXT: 9
vector.print %fail_res9 : i64
%m10 = arith.constant dense<(5.000000e+01,1.000000e+00)> : tensor<3x3xcomplex<f64>>
%f10 = tensor.cast %m10 : tensor<3x3xcomplex<f64>> to tensor<*xcomplex<f64>>
%fail_res10 = call @verifyMemRefC64(%10, %f10) : (tensor<*xcomplex<f64>>, tensor<*xcomplex<f64>>) -> (i64)
// CHECK-NEXT: 9
vector.print %fail_res10 : i64
%m11 = arith.constant dense<100> : tensor<3x3xindex>
%f11 = tensor.cast %m11 : tensor<3x3xindex> to tensor<*xindex>
%fail_res11 = call @verifyMemRefInd(%11, %f11) : (tensor<*xindex>, tensor<*xindex>) -> (i64)
// CHECK-NEXT: 9
vector.print %fail_res11 : i64
return
}
}