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 );
47 uint16 bbs_sqrt32( uint32 valA )  in bbs_sqrt32()
49 	uint32 rootL = 0;  in bbs_sqrt32()
50 	uint32 expL = 0;  in bbs_sqrt32()
97 const uint32 bbs_fastSqrt32_tableG[] = 
125 uint16 bbs_fastSqrt32( uint32 valA )  in bbs_fastSqrt32()
127 	uint32 expL = 0;  in bbs_fastSqrt32()
128 	uint32 valL;  in bbs_fastSqrt32()
129 	uint32 offsL;  in bbs_fastSqrt32()
130 	uint32 indexL = 0;  in bbs_fastSqrt32()
151 const uint32 bbs_invSqrt32_tableG[] = 
179 uint32 bbs_invSqrt32( uint32 valA )  in bbs_invSqrt32()
182 	uint32 expL = 0;  in bbs_invSqrt32()
183 	uint32 valL;  in bbs_invSqrt32()
184 	uint32 offsL;  in bbs_invSqrt32()
185 	uint32 indexL = 0;  in bbs_invSqrt32()
221 	uint32 expL = 0;  in bbs_inv32()
225 	uint32 indexL = 0;  in bbs_inv32()
245 uint32 bbs_intLog2( uint32 valA )  in bbs_intLog2()
247 	uint32 expL = 0;  in bbs_intLog2()
258 const uint32 bbs_pow2M1_tableG[] = 
278 uint32 bbs_pow2M1( uint32 valA )  in bbs_pow2M1()
280 	uint32 offsL = ( valA & 0x03FFFFFF ) >> 10;  in bbs_pow2M1()
287 uint32 bbs_pow2( int32 valA )  in bbs_pow2()
290 	uint32 offsL  = ( uint32 )( valA << 5 );  in bbs_pow2()
297 uint32 bbs_exp( int32 valA )  in bbs_exp()
314 	adjustedL = ( valA >> 16 ) * 94548 + ( ( ( ( ( uint32 )valA ) & 0x0FFFF ) * 47274 ) >> 15 );  in bbs_exp()
318 	offsL  = ( uint32 )( adjustedL << 5 );  in bbs_exp()
344 int32 bbs_dotProduct_intelMMX16( const int16* vec1A, const int16* vec2A, uint32 sizeA )  in bbs_dotProduct_intelMMX16()
418 int32 bbs_dotProduct_intelMMX16( const int16* vec1A, const int16* vec2A, uint32 sizeA )  in bbs_dotProduct_intelMMX16()
497 int32 bbs_dotProduct_dsp( const int16* vec1A, const int16* vec2A, uint32 sizeA )  in bbs_dotProduct_dsp()
519 int32 bbs_dotProduct_EE( const int16* vec1A, const int16* vec2A, uint32 sizeA )  in bbs_dotProduct_EE()
574 int32 bbs_dotProduct_arm9e( const int16* vec1A, const int16* vec2A, uint32 sizeA )  in bbs_dotProduct_arm9e()
611 int32 bbs_dotProduct_stdc( const int16* vec1A, const int16* vec2A, uint32 sizeA )  in bbs_dotProduct_stdc()
638 int32 bbs_dotProductInt16( const int16* vec1A, const int16* vec2A, uint32 sizeA )  in bbs_dotProductInt16()
644 		uint32 size16L = sizeA & 0xfffffff0;  in bbs_dotProductInt16()
647 			if( ( (uint32)vec1A & 0xF ) == 0 && ( (uint32)vec2A & 0xF ) == 0 )   in bbs_dotProductInt16()
658 		uint32 size16L = sizeA & 0xfffffff0;  in bbs_dotProductInt16()
675 	if( ( (uint32)vec1A & 0xF ) == 0 && ( (uint32)vec2A & 0xF ) == 0 )   in bbs_dotProductInt16()
693 	uint32 size16L = sizeA & 0xfffffff0;  in bbs_dotProductInt16()
715 const uint32 bbs_fermi_tableG[] = 
731 	uint32 indexL = ( ( valA >> 15 ) + 20 ) << 1;  in bbs_fermi()
732 	uint32 offsL  = ( ( valA & 0x00007FFF ) + 4 ) >> 3;  in bbs_fermi()
740 void bbs_uint32ReduceToNBits( uint32* argPtrA, int32* bbpPtrA, uint32 nBitsA )  in bbs_uint32ReduceToNBits()
753 void bbs_int32ReduceToNBits( int32* argPtrA, int32* bbpPtrA, uint32 nBitsA )  in bbs_int32ReduceToNBits()
766 uint32 bbs_convertU32( uint32 srcA, int32 srcBbpA, int32 dstBbpA )  in bbs_convertU32()
770 		uint32 shiftL = dstBbpA - srcBbpA;  in bbs_convertU32()
771 		if( srcA > ( ( uint32 )0xFFFFFFFF >> shiftL ) )  in bbs_convertU32()
774 			return ( uint32 )0xFFFFFFFF;  in bbs_convertU32()
783 		uint32 shiftL = srcBbpA - dstBbpA;  in bbs_convertU32()
784 		uint32 addL = 1L << ( shiftL - 1 );  in bbs_convertU32()
803 		uint32 shiftL = ( uint32 )( dstBbpA - srcBbpA );  in bbs_convertS32()
807 			return ( uint32 )0x7FFFFFFF;  in bbs_convertS32()
821 		uint32 shiftL = ( uint32 )( srcBbpA - dstBbpA );  in bbs_convertS32()
857 void bbs_mulU32( uint32 v1A, uint32 v2A, uint32* manPtrA, int32* expPtrA )  in bbs_mulU32()
859 	uint32 log1L = bbs_intLog2( v1A );  in bbs_mulU32()
860 	uint32 log2L = bbs_intLog2( v2A );  in bbs_mulU32()
869 		uint32 v1L = v1A;  in bbs_mulU32()
870 		uint32 v2L = v2A;  in bbs_mulU32()
871 		uint32 exp1L = 0;  in bbs_mulU32()
872 		uint32 exp2L = 0;  in bbs_mulU32()
900 	uint32 log1L = bbs_intLog2( v1A > 0 ? v1A : -v1A );  in bbs_mulS32()
901 	uint32 log2L = bbs_intLog2( v2A > 0 ? v2A : -v2A );  in bbs_mulS32()
939 void bbs_vecSqrNorm32( const int32* vecA, uint32 sizeA, uint32* manPtrA, uint32* expPtrA )  in bbs_vecSqrNorm32()
941 	uint32 sumL = 0;  in bbs_vecSqrNorm32()
944 	uint32 iL;  in bbs_vecSqrNorm32()
950 		uint32 prdL;  in bbs_vecSqrNorm32()
966 			uint32 shrL = sumExpL - expL;  in bbs_vecSqrNorm32()
971 			uint32 shrL = expL - sumExpL;  in bbs_vecSqrNorm32()
998 void bbs_vecSqrNorm16( const int16* vecA, uint32 sizeA, uint32* manPtrA, uint32* expPtrA )  in bbs_vecSqrNorm16()
1000 	uint32 sumL = 0;  in bbs_vecSqrNorm16()
1003 	uint32 iL;  in bbs_vecSqrNorm16()
1007 		uint32 prdL = vL * vL;  in bbs_vecSqrNorm16()
1011 			uint32 shrL = sumExpL;  in bbs_vecSqrNorm16()
1037 uint32 bbs_vecNorm16( const int16* vecA, uint32 sizeA )  in bbs_vecNorm16()
1039 	uint32 manL;  in bbs_vecNorm16()
1040 	uint32 expL;  in bbs_vecNorm16()