37 	extern int32 bbs_dotProduct_128SSE2( const int16* vec1A, const int16* vec2A, uint32 sizeA );
38 	extern int32 bbs_dotProduct_u128SSE2( const int16* vec1A, const int16* vec2A, uint32 sizeA );
42 	int32 bbs_dotProduct_fr71( const int16* vec1A, const int16* vec2A, uint32 sizeA );
344 int32 bbs_dotProduct_intelMMX16( const int16* vec1A, const int16* vec2A, uint32 sizeA )  in bbs_dotProduct_intelMMX16()  argument
351 			mov     eax, vec1A  in bbs_dotProduct_intelMMX16()
418 int32 bbs_dotProduct_intelMMX16( const int16* vec1A, const int16* vec2A, uint32 sizeA )  in bbs_dotProduct_intelMMX16()  argument
479 		: "g" ( vec1A ), "g" ( vec2A ), "g" ( sizeA )  in bbs_dotProduct_intelMMX16()
497 int32 bbs_dotProduct_dsp( const int16* vec1A, const int16* vec2A, uint32 sizeA )  in bbs_dotProduct_dsp()  argument
502 		resultL = dotProduct_C62x( vec1A, vec2A, sizeA - 1 );  in bbs_dotProduct_dsp()
503 		return resultL + ( int32 ) *( vec1A + sizeA - 1 ) * *( vec2A + sizeA - 1 );  in bbs_dotProduct_dsp()
507 		return dotProduct_C62x( vec1A, vec2A, sizeA );  in bbs_dotProduct_dsp()
519 int32 bbs_dotProduct_EE( const int16* vec1A, const int16* vec2A, uint32 sizeA )  in bbs_dotProduct_EE()  argument
551 		: "=r" ( vec1A ), "=r" ( vec2A ), "=r" ( t1L ), "=r" ( t2L ), "=r" ( t3L )  in bbs_dotProduct_EE()
552 		: "0" ( vec1A ), "1" ( vec2A ), "r" ( iL ), "r" ( accL )  in bbs_dotProduct_EE()
560 	for( ; jL--; ) resultL += ( int32 ) *vec1A++ * *vec2A++;  in bbs_dotProduct_EE()
574 int32 bbs_dotProduct_arm9e( const int16* vec1A, const int16* vec2A, uint32 sizeA )  in bbs_dotProduct_arm9e()  argument
578 	int32* v1PtrL = ( int32* )vec1A;  in bbs_dotProduct_arm9e()
595 	vec1A = ( int16* )v1PtrL;  in bbs_dotProduct_arm9e()
599 	for( ; sizeA > 0; sizeA-- ) accuL += ( int32 )*vec1A++ * *vec2A++;  in bbs_dotProduct_arm9e()
611 int32 bbs_dotProduct_stdc( const int16* vec1A, const int16* vec2A, uint32 sizeA )  in bbs_dotProduct_stdc()  argument
617 		accuL += ( int32 ) *vec1A * *vec2A;  in bbs_dotProduct_stdc()
618 		accuL += ( int32 ) *( vec1A + 1 ) * *( vec2A + 1 );  in bbs_dotProduct_stdc()
619 		accuL += ( int32 ) *( vec1A + 2 ) * *( vec2A + 2 );  in bbs_dotProduct_stdc()
620 		accuL += ( int32 ) *( vec1A + 3 ) * *( vec2A + 3 );  in bbs_dotProduct_stdc()
622 		accuL += ( int32 ) *( vec1A + 4 ) * *( vec2A + 4 );  in bbs_dotProduct_stdc()
623 		accuL += ( int32 ) *( vec1A + 5 ) * *( vec2A + 5 );  in bbs_dotProduct_stdc()
624 		accuL += ( int32 ) *( vec1A + 6 ) * *( vec2A + 6 );  in bbs_dotProduct_stdc()
625 		accuL += ( int32 ) *( vec1A + 7 ) * *( vec2A + 7 );  in bbs_dotProduct_stdc()
627 		vec1A += 8;  in bbs_dotProduct_stdc()
631 	for( ; sizeA; sizeA-- ) accuL += ( int32 ) *vec1A++ * *vec2A++;  in bbs_dotProduct_stdc()
638 int32 bbs_dotProductInt16( const int16* vec1A, const int16* vec2A, uint32 sizeA )  in bbs_dotProductInt16()  argument
647 			if( ( (uint32)vec1A & 0xF ) == 0 && ( (uint32)vec2A & 0xF ) == 0 )   in bbs_dotProductInt16()
649 				return bbs_dotProduct_128SSE2( vec1A, vec2A, sizeA );  in bbs_dotProductInt16()
653 				return bbs_dotProduct_u128SSE2( vec1A, vec2A, sizeA );  in bbs_dotProductInt16()
663 				return bbs_dotProduct_intelMMX16( vec1A, vec2A, size16L );  in bbs_dotProductInt16()
665 			return bbs_dotProduct_intelMMX16( vec1A, vec2A, size16L )  in bbs_dotProductInt16()
666 					+ bbs_dotProduct_stdc( vec1A + size16L, vec2A + size16L, sizeA - size16L );  in bbs_dotProductInt16()
670 	return bbs_dotProduct_stdc( vec1A, vec2A, sizeA );  in bbs_dotProductInt16()
675 	if( ( (uint32)vec1A & 0xF ) == 0 && ( (uint32)vec2A & 0xF ) == 0 )   in bbs_dotProductInt16()
677 		return bbs_dotProduct_EE( vec1A, vec2A, sizeA );  in bbs_dotProductInt16()
679 	return bbs_dotProduct_stdc( vec1A, vec2A, sizeA );  in bbs_dotProductInt16()
684 	return bbs_dotProduct_arm9e( vec1A, vec2A, sizeA );  in bbs_dotProductInt16()
689 	return bbs_dotProduct_dsp( vec1A, vec2A, sizeA );  in bbs_dotProductInt16()
698 			return bbs_dotProduct_fr71( vec1A, vec2A, size16L );  in bbs_dotProductInt16()
700 		return bbs_dotProduct_fr71( vec1A, vec2A, size16L )  in bbs_dotProductInt16()
701 				+ bbs_dotProduct_stdc( vec1A + size16L, vec2A + size16L, sizeA - size16L );  in bbs_dotProductInt16()
704 	return bbs_dotProduct_stdc( vec1A, vec2A, sizeA );  in bbs_dotProductInt16()
708 	return bbs_dotProduct_stdc( vec1A, vec2A, sizeA );  in bbs_dotProductInt16()