On 21/06/10 19:38, Delio Brignoli wrote:
Hello Nick,
On 21/06/2010, at 11:27, Nick Thompson wrote:
The following restructuring and optimisations increase the SPI read performance from 1.3MiB/s (on da850) to 2.87MiB/s (on da830):
Using this patch I get 2.21MiB/s on my L138 EVM (da850), quite an improvement! I would like to see how much my original patch can be improved using some of your changes without splitting the code to handle the three cases. I will try later this week.
Not testing the txp and rxp pointers in the loop was a significant gain for me and pipe-lining the TX and RX operations is going to be a little trickier, but give it a go by all means.
[...]
- if (!dout)
return davinci_spi_read(slave, len, din, flags);- else if (!din)
return davinci_spi_write(slave, len, dout, flags);+#ifndef CONFIG_SPI_HALF_DUPLEX
- else
return davinci_spi_read_write(slave, len, din, dout, flags);+#endif
I think there should always be an else branch at the end even if CONFIG_SPI_HALF_DUPLEX is not defined. Something like:
#else flags |= SPI_XFER_END; #endif
Hmmm, yes, you are correct. That must be added, else in the unexpected case, the transaction will be opened and left open. I'll fix that.
to terminate the transfer instead of failing silently. In fact it should signal the error condition somehow, but I do not know enough about u-boot to provide an advice on this.
I think maybe a printk(KERN_ERR ...) will do it. I'll add that too.
Thanks
Delio
Thanks for the review.
Nick.