Minor, micro-optimization

3 files changed, 26 insertions, 15 deletions

diff --git a/Whisper/ML/Context.ops.cpp b/Whisper/ML/Context.ops.cpp
index a94497e..f6309f0 100644
--- a/Whisper/ML/Context.ops.cpp
+++ b/Whisper/ML/Context.ops.cpp
@@ -84,9 +84,20 @@ Tensor __declspec( noinline ) MlContext::view2d( const Tensor& a, uint32_t ne0,
 
 Tensor MlContext::transpose( const Tensor& a )
 {
-	Tensor result = a;
-	std::swap( result.ne[ 0 ], result.ne[ 1 ] );
-	std::swap( result.nb[ 0 ], result.nb[ 1 ] );
+	Tensor result;
+
+	// A magic number for _mm_shuffle_epi32 SSE2 instruction to swap two lower int32 lanes in a vector
+	constexpr int swapXy = _MM_SHUFFLE( 3, 2, 0, 1 );
+
+	__m128i v = a.sizeVec();
+	v = _mm_shuffle_epi32( v, swapXy );
+	store( result.ne, v );
+
+	v = a.stridesVec();
+	v = _mm_shuffle_epi32( v, swapXy );
+	store( result.nb, v );
+
+	result.setGpuViews( a, a );
 	return result;
 }
 
diff --git a/Whisper/ML/TensorShape.h b/Whisper/ML/TensorShape.h
index 473b0c9..5749764 100644
--- a/Whisper/ML/TensorShape.h
+++ b/Whisper/ML/TensorShape.h
@@ -28,11 +28,11 @@ namespace DirectCompute
 		HRESULT create( const ggml_tensor& ggml );
 		TensorShape( const ggml_tensor& ggml );
 
-		__m128i sizeVec() const
+		__m128i __vectorcall sizeVec() const
 		{
 			return load( ne );
 		}
-		__m128i stridesVec() const
+		__m128i __vectorcall stridesVec() const
 		{
 			return load( nb );
 		}
diff --git a/Whisper/Utils/miscUtils.h b/Whisper/Utils/miscUtils.h
index ff52e21..77f04e9 100644
--- a/Whisper/Utils/miscUtils.h
+++ b/Whisper/Utils/miscUtils.h
@@ -10,48 +10,48 @@ inline void check( HRESULT hr )
 	throw hr;
 }
 
-inline __m128i load16( const int* rsi )
+inline __m128i __vectorcall load16( const int* rsi )
 {
 	return _mm_loadu_si128( ( const __m128i* )rsi );
 }
-inline __m128i load16( const uint32_t* rsi )
+inline __m128i __vectorcall load16( const uint32_t* rsi )
 {
 	return _mm_loadu_si128( ( const __m128i* )rsi );
 }
-inline __m128i load( const std::array<uint32_t, 4>& arr )
+inline __m128i __vectorcall load( const std::array<uint32_t, 4>& arr )
 {
 	return load16( arr.data() );
 }
-inline void store16( void* rdi, __m128i v )
+inline void __vectorcall store16( void* rdi, __m128i v )
 {
 	_mm_storeu_si128( ( __m128i* )rdi, v );
 }
-inline void store12( void* rdi, __m128i v )
+inline void __vectorcall store12( void* rdi, __m128i v )
 {
 	_mm_storel_epi64( ( __m128i* )rdi, v );
 	( (int*)rdi )[ 2 ] = _mm_extract_epi32( v, 2 );
 }
-inline void store( std::array<uint32_t, 4>& arr, __m128i v )
+inline void __vectorcall store( std::array<uint32_t, 4>& arr, __m128i v )
 {
 	store16( arr.data(), v );
 }
-inline bool vectorEqual( __m128i a, __m128i b )
+inline bool __vectorcall vectorEqual( __m128i a, __m128i b )
 {
 	__m128i xx = _mm_xor_si128( a, b );
 	return (bool)_mm_testz_si128( xx, xx );
 }
 
-inline __m128i setLow_size( size_t low )
+inline __m128i __vectorcall setLow_size( size_t low )
 {
 	return _mm_cvtsi64_si128( (int64_t)low );
 }
-inline __m128i setr_size( size_t low, size_t high )
+inline __m128i __vectorcall setr_size( size_t low, size_t high )
 {
 	__m128i v = setLow_size( low );
 	v = _mm_insert_epi64( v, (int64_t)high, 1 );
 	return v;
 }
-inline __m128i setHigh_size( size_t high )
+inline __m128i __vectorcall setHigh_size( size_t high )
 {
 	__m128i v = _mm_setzero_si128();
 	v = _mm_insert_epi64( v, (int64_t)high, 1 );