Hi,

could you please specify what performance degradation I can expect executing my matrix-vector operations where my numbers are float i.e. 32-bit versus 64-bit doubles. I know that SHARC family support 64-bit floats only via floating point emulation library hence execution will be much slower, the question I have how much slower? What sustained performance I can expect i.e. how far from 3.2GFlops pick for 32-bit floats?

Can anyone comment on performance using 40-bit extended precision? Are there any libraries that will support arthementic for such extended precision or the hand crafted assembler is only the option?

Are you aware of BLAS library for Tiger SHARC?

Regards

Adam

> could you please specify what performance degradation I can expect

> executing my matrix-vector operations where my numbers are float i.e.

> 32-bit versus 64-bit doubles. I know that SHARC family support 64-bit

> floats only via floating point emulation library hence execution will

> be much slower, the question I have how much slower? What sustained performance I can expect i.e. how far from 3.2GFlops pick for 32-bit floats?

On TigerSHARC, 64-bit doubles are operated on using relatively slow floating-point emulation software and so this mode is recommended only for code that requires that doubles have more than 32 bits of precision.

The performance degradation between using 32-bit doubles and 64-bit doubles when executing your matrix-vector operations would depend on both the size of the matrix-vectors and the data contained in them so we are not able to provide an exact figure, but it can be considerably slower to use 64-bit doubles. In an example we tried, multiplying two 10x10 matrices took 3283 cycles when using 32-bit doubles, but took 107205 cycles when using 64-bit doubles.

> Can anyone comment on performance using 40-bit extended precision? Are

> there any libraries that will support arthementic for such extended precision or the hand crafted assembler is only the option?

As far as we are aware, no such libraries exist. The VisualDSP++ TigerSHARC compiler and libraries only contain support for 32-bit and 64-bit floating-point arithmetic.

> Are you aware of BLAS library for Tiger SHARC?

No.