[clang] [mlir] [llvm] [mlir][verifyMemref] Fix bug and support more types for verifyMemref (PR #77682)
Yinying Li via cfe-commits
cfe-commits at lists.llvm.org
Wed Jan 10 12:22:20 PST 2024
https://github.com/yinying-lisa-li updated https://github.com/llvm/llvm-project/pull/77682
>From abb4cfff1b74a34f074e3feccb5c03598ba48e64 Mon Sep 17 00:00:00 2001
From: Yinying Li <yinyingli at google.com>
Date: Wed, 10 Jan 2024 20:00:32 +0000
Subject: [PATCH] [mlir][verifyMemref] Fix bug and support more types for
verifyMemref in RunnerUtils
1. Fix a bug in verifyMemref to pass in data instead of bareptr, which didn't verify data correctly.
2. Add == for f16 and bf16.
3. Add a comprehensive tests of verifyMemref for all supported types.
---
.../mlir/ExecutionEngine/Float16bits.h | 3 +
.../mlir/ExecutionEngine/RunnerUtils.h | 12 +-
mlir/lib/ExecutionEngine/Float16bits.cpp | 8 +
mlir/lib/ExecutionEngine/RunnerUtils.cpp | 30 ++++
.../Dialect/Memref/verify-memref.mlir | 167 ++++++++++++++++++
5 files changed, 219 insertions(+), 1 deletion(-)
create mode 100644 mlir/test/Integration/Dialect/Memref/verify-memref.mlir
diff --git a/mlir/include/mlir/ExecutionEngine/Float16bits.h b/mlir/include/mlir/ExecutionEngine/Float16bits.h
index 5eb1f2ce07639d..ad409841b2a927 100644
--- a/mlir/include/mlir/ExecutionEngine/Float16bits.h
+++ b/mlir/include/mlir/ExecutionEngine/Float16bits.h
@@ -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);
diff --git a/mlir/include/mlir/ExecutionEngine/RunnerUtils.h b/mlir/include/mlir/ExecutionEngine/RunnerUtils.h
index ebf95f90f37450..965dff83816dbf 100644
--- a/mlir/include/mlir/ExecutionEngine/RunnerUtils.h
+++ b/mlir/include/mlir/ExecutionEngine/RunnerUtils.h
@@ -332,7 +332,7 @@ 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,
+ std::cerr, actual.data, expected.data, actual.rank, actual.offset,
actual.sizes, actual.strides, printCounter);
}
@@ -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(
diff --git a/mlir/lib/ExecutionEngine/Float16bits.cpp b/mlir/lib/ExecutionEngine/Float16bits.cpp
index 38a05fe86bbdda..2ac9c18b6821ac 100644
--- a/mlir/lib/ExecutionEngine/Float16bits.cpp
+++ b/mlir/lib/ExecutionEngine/Float16bits.cpp
@@ -150,6 +150,14 @@ 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
diff --git a/mlir/lib/ExecutionEngine/RunnerUtils.cpp b/mlir/lib/ExecutionEngine/RunnerUtils.cpp
index 378aa7ce35ef16..d93ec2b7af0c0d 100644
--- a/mlir/lib/ExecutionEngine/RunnerUtils.cpp
+++ b/mlir/lib/ExecutionEngine/RunnerUtils.cpp
@@ -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) {
diff --git a/mlir/test/Integration/Dialect/Memref/verify-memref.mlir b/mlir/test/Integration/Dialect/Memref/verify-memref.mlir
new file mode 100644
index 00000000000000..b7e2a46688f475
--- /dev/null
+++ b/mlir/test/Integration/Dialect/Memref/verify-memref.mlir
@@ -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
+ }
+}
More information about the cfe-commits
mailing list