2024UD24 RC2 Content (#110)

Author: Maxim-Doronin

src/vpux_compiler/include/vpux/compiler/NPU37XX/dialect/IE/transforms/passes.hpp

@@ -94,9 +94,6 @@ struct DefaultHWOptions : public IE::DefaultHWOptionsDialectBase, virtual vpux::
     BoolOption mergeUnrolledMatmul{*this, "merge-unrolled-matmul", llvm::cl::desc("Enable merging urolled Matmul ops"),
                                    llvm::cl::init(false)};
 
-    BoolOption enableRuntimeDequant{*this, "enable-runtime-dequant",
-                                    llvm::cl::desc("Enable runtime dequantization of asymmetricly quantized weight"),
-                                    llvm::cl::init(false)};
     Int64Option runtimeDequantizationLimit{
             *this, "runtime-dequantization-limit",
             llvm::cl::desc("Lower limit on weight size for runtime dequantization"

src/vpux_compiler/include/vpux/compiler/NPU40XX/dialect/IE/transforms/passes.hpp

@@ -43,9 +43,6 @@ struct DefaultHWOptions : public IE::DefaultHWOptionsDialectBase, virtual vpux::
     BoolOption mergeUnrolledMatmul{*this, "merge-unrolled-matmul", llvm::cl::desc("Enable merging urolled Matmul ops"),
                                    llvm::cl::init(true)};
 
-    BoolOption enableRuntimeDequant{*this, "enable-runtime-dequant",
-                                    llvm::cl::desc("Enable runtime dequantization of asymmetricly quantized weight"),
-                                    llvm::cl::init(true)};
     BoolOption enableApplyDynamicBoundaryCorrection{*this, "enable-apply-dynamic-boundary-correction",
                                                     llvm::cl::desc("Enable apply-dynamic-boundary-correction pass"),
                                                     llvm::cl::init(false)};

src/vpux_compiler/include/vpux/compiler/compilation_options.hpp

@@ -13,11 +13,8 @@
 namespace vpux {
 
 constexpr bool arePrivateOptionsEnabled() {
-#ifdef PRIVATE_COMPILER_OPTIONS_ENABLED
+    // Note: private options have been intentionally made public for the time being
     return true;
-#else
-    return false;
-#endif
 }
 
 template <typename T>

src/vpux_compiler/include/vpux/compiler/core/pipelines_options.hpp

@@ -149,9 +149,6 @@ struct ReferenceSWOptions : mlir::PassPipelineOptions<T> {
                                                   llvm::cl::desc("Enable weights dequantization for weights as input"),
                                                   llvm::cl::init(false)};
 
-    BoolOption enableRuntimeDequant{*this, "enable-runtime-dequant",
-                                    llvm::cl::desc("Enable runtime dequantization of asymmetricly quantized weight"),
-                                    llvm::cl::init(false)};
     Int64Option runtimeDequantizationLimit{
             *this, "runtime-dequantization-limit",
             llvm::cl::desc("Lower limit on weight size for runtime dequantization"

src/vpux_compiler/include/vpux/compiler/core/public_options.hpp

@@ -141,6 +141,18 @@ struct PublicOptions : mlir::PassPipelineOptions<PublicOptions> {
     BoolOption shaveDryRun{*this, "shave-dry-run", llvm::cl::desc("Enable shave dry run stripping"),
                            llvm::cl::init(false)};
 
+    BoolOption enableRuntimeDequant{*this, "enable-runtime-dequant",
+                                    llvm::cl::desc("Enable runtime dequantization of asymmetricly quantized weight"),
+                                    llvm::cl::init(false)};
+    static bool getDefaultEnableRuntimeDequant(VPU::ArchKind arch) {
+        switch (arch) {
+        case VPU::ArchKind::NPU40XX:
+            return true;
+        default:
+            return false;
+        }
+    }
+
     //
     // Constructors
     //
@@ -152,6 +164,7 @@ struct PublicOptions : mlir::PassPipelineOptions<PublicOptions> {
             workloadManagementMode = getDefaultWorkloadManagementMode(arch);
         }
         enableDMAProfiling = getDefaultEnableDMAProfiling(arch);
+        enableRuntimeDequant = getDefaultEnableRuntimeDequant(arch);
     }
 
     static std::unique_ptr<PublicOptions> createFromString(StringRef options, VPU::ArchKind arch) {