Actually make MobileBERT run in the test. (#5264)

Previously added MobileBERT test was only compiling the model, but not
running it due to `iree.symbol.export` not being set on the
function.

Fixing that leads to the issue of too many descriptors being used
(Issue #5152). For now
- drop `noinline` from some constants in the test file
- add a flag to increase the size at which constants are inlined into
  the dispatch region

iree/compiler/Dialect/Flow/IR/FlowOps.cpp

@@ -19,6 +19,7 @@
 #include "iree/compiler/Dialect/Shape/IR/Builders.h"
 #include "llvm/ADT/BitVector.h"
 #include "llvm/ADT/StringExtras.h"
+#include "llvm/Support/CommandLine.h"
 #include "mlir/Dialect/StandardOps/IR/Ops.h"
 #include "mlir/IR/Attributes.h"
 #include "mlir/IR/BlockAndValueMapping.h"
@@ -33,6 +34,12 @@
 #include "mlir/Support/LogicalResult.h"
 #include "mlir/Transforms/RegionUtils.h"
 
+static llvm::cl::opt<int> clInlineConstantByteLength(
+    "iree-flow-inline-constants-max-byte-length",
+    llvm::cl::desc("Maximum byte-length of constant that can be inlined into a "
+                   "dispatch region"),
+    llvm::cl::init(256));
+
 namespace mlir {
 namespace iree_compiler {
 namespace IREE {
@@ -831,7 +838,8 @@ static bool canDispatchRegionContainOp(Operation *op) {
       uint64_t estimatedByteLength =
           (shapedType.getNumElements() * shapedType.getElementTypeBitWidth()) /
           8;
-      return denseAttr.isSplat() || estimatedByteLength <= 256;  // or whatever
+      return denseAttr.isSplat() ||
+             estimatedByteLength <= clInlineConstantByteLength;
     } else if (constantType.isIntOrIndexOrFloat()) {
       return true;
     }

iree/test/e2e/models/BUILD

@@ -68,6 +68,7 @@ iree_check_single_backend_test_suite(
     srcs = CHECK_FRAMEWORK_TESTS,
     compiler_flags = [
         "-iree-flow-dispatch-linalg-on-tensors",
+        "-iree-flow-inline-constants-max-byte-length=2048",
         "-iree-codegen-spirv-experimental-linalg-on-tensors",
         "-iree-spirv-enable-vectorization",
     ],
@@ -80,6 +81,7 @@ iree_check_single_backend_test_suite(
     srcs = CHECK_FRAMEWORK_TESTS,
     compiler_flags = [
         "-iree-flow-dispatch-linalg-on-tensors",
+        "-iree-flow-inline-constants-max-byte-length=2048",
         "-iree-codegen-llvm-experimental-linalg-on-tensors",
     ],
     driver = "dylib",

iree/test/e2e/models/CMakeLists.txt

@@ -53,6 +53,7 @@ iree_check_single_backend_test_suite(
     "vulkan"
   COMPILER_FLAGS
     "-iree-flow-dispatch-linalg-on-tensors"
+    "-iree-flow-inline-constants-max-byte-length=2048"
     "-iree-codegen-spirv-experimental-linalg-on-tensors"
     "-iree-spirv-enable-vectorization"
 )
@@ -69,6 +70,7 @@ iree_check_single_backend_test_suite(
     "dylib"
   COMPILER_FLAGS
     "-iree-flow-dispatch-linalg-on-tensors"
+    "-iree-flow-inline-constants-max-byte-length=2048"
     "-iree-codegen-llvm-experimental-linalg-on-tensors"
 )