Minor, micro-optimization

1 files changed, 14 insertions, 3 deletions

diff --git a/Whisper/Whisper/ContextImpl.diarize.cpp b/Whisper/Whisper/ContextImpl.diarize.cpp
index 90acb04..9d88fac 100644
--- a/Whisper/Whisper/ContextImpl.diarize.cpp
+++ b/Whisper/Whisper/ContextImpl.diarize.cpp
@@ -16,22 +16,33 @@ namespace
 	// and return left / right numbers in the lower 2 lanes of the SSE vector
 	inline __m128 __vectorcall computeChannelsEnergy( const std::vector<StereoSample>& sourceVector )
 	{
+		// Might be possible to implement way more sophisticated, and precise, version of this function.
+		// For example, compute these 3 metrics with VAD code, and cluster the numbers somehow.
+		// Not doing that currently; instead, replicating the simple version from the whisper.cpp original version.
+
 		const StereoSample* rsi = sourceVector.data();
 		const StereoSample* const rsiEnd = rsi + sourceVector.size();
 		const StereoSample* const rsiEndAligned = rsi + ( sourceVector.size() & ( ~(size_t)1 ) );
 
-		const __m128 absMask = _mm_set1_ps( -0.0f );
+		// Move 0x7FFFFFFF to lowest lane of the int32 vector;
+		// unlike float scalars or all vectors, integer scalar constants are in the instruction stream
+		__m128i absMaskInt = _mm_cvtsi32_si128( (int)0x7FFFFFFFu );
+		// Broadcast over the complete vector
+		absMaskInt = _mm_shuffle_epi32( absMaskInt, 0 );
+		// Bitcast to FP32 vector, for _mm_and_ps instruction
+		const __m128 absMask = _mm_castsi128_ps( absMaskInt );
+
 		__m128 acc = _mm_setzero_ps();
 		for( ; rsi < rsiEndAligned; rsi += 2 )
 		{
 			__m128 v = _mm_loadu_ps( (const float*)rsi );
-			v = _mm_andnot_ps( absMask, v );
+			v = _mm_and_ps( v, absMask );
 			acc = _mm_add_ps( acc, v );
 		}
 		if( rsi != rsiEnd )
 		{
 			__m128 v = _mm_castpd_ps( _mm_load_sd( (const double*)rsi ) );
-			v = _mm_andnot_ps( absMask, v );
+			v = _mm_and_ps( v, absMask );
 			acc = _mm_add_ps( acc, v );
 		}