Optional startup flags to override performance-related defaults for the compute shaders

3 files changed, 17 insertions, 2 deletions

diff --git a/Whisper/API/eGpuModelFlags.h b/Whisper/API/eGpuModelFlags.h
new file mode 100644
index 0000000..96f5a76
--- /dev/null
+++ b/Whisper/API/eGpuModelFlags.h
@@ -0,0 +1,13 @@
+#pragma once
+#include <stdint.h>
+
+namespace Whisper
+{
+	enum struct eGpuModelFlags : uint32_t
+	{
+		Wave32 = 1,
+		Wave64 = 2,
+		NoReshapedMatMul = 4,
+		UseReshapedMatMul = 8,
+	};
+}
\ No newline at end of file
diff --git a/Whisper/API/iContext.cl.h b/Whisper/API/iContext.cl.h
index 97d34c7..fdb15ce 100644
--- a/Whisper/API/iContext.cl.h
+++ b/Whisper/API/iContext.cl.h
@@ -5,6 +5,7 @@
 #include "loggerApi.h"
 #include "sLanguageList.h"
 #include "sLoadModelCallbacks.h"
+#include "eGpuModelFlags.h"
 
 namespace Whisper
 {
@@ -55,7 +56,7 @@ namespace Whisper
 	};
 
 	HRESULT COMLIGHTCALL setupLogger( const sLoggerSetup& setup );
-	HRESULT COMLIGHTCALL loadModel( const wchar_t* path, eModelImplementation impl, const sLoadModelCallbacks* callbacks, iModel** pp );
+	HRESULT COMLIGHTCALL loadModel( const wchar_t* path, eModelImplementation impl, uint32_t flags, const sLoadModelCallbacks* callbacks, iModel** pp );
 
 	uint32_t COMLIGHTCALL findLanguageKeyW( const wchar_t* lang );
 	uint32_t COMLIGHTCALL findLanguageKeyA( const char* lang );
diff --git a/Whisper/API/iContext.h b/Whisper/API/iContext.h
index 9661093..d6ca29c 100644
--- a/Whisper/API/iContext.h
+++ b/Whisper/API/iContext.h
@@ -4,6 +4,7 @@
 #include "loggerApi.h"
 #include "sLanguageList.h"
 #include "sLoadModelCallbacks.h"
+#include "eGpuModelFlags.h"
 
 namespace Whisper
 {
@@ -50,7 +51,7 @@ namespace Whisper
 	};
 
 	HRESULT __stdcall setupLogger( const sLoggerSetup& setup );
-	HRESULT __stdcall loadModel( const wchar_t* path, eModelImplementation impl, const sLoadModelCallbacks* callbacks, iModel** pp );
+	HRESULT __stdcall loadModel( const wchar_t* path, eModelImplementation impl, uint32_t flags, const sLoadModelCallbacks* callbacks, iModel** pp );
 
 	uint32_t __stdcall findLanguageKeyW( const wchar_t* lang );
 	uint32_t __stdcall findLanguageKeyA( const char* lang );