Dear @Mowlwurf,
Thank you for your patience. I have tested with 4.14.117 kernel with 3.0B2 image, there are quite some improvements in SPI driver and it is really using SPI DMA.
512 bytes with 32 bits per word mode has been transferred by 4 packets of the transaction, each 128 bytes packet. There is a 6.9us pause between the packet transactions. This is better than 40us pause between transaction with 2.8B6 image kernel.
Could you test the same with the 3.0B2 image and let us know is that meeting your requirement.
If you are not happy with 7us pause then try this suggestion, it does the continuous transfer without any pause. However, you need to live with the ERR009165 silicon issue.
Please refer this forum question : https://www.toradex.com/community/questions/2034/can-colibri-imx6-spi-ports-operate-at-up-to-52-mbp.html for SPI max frequency query. This is applicable to i.MX7 also, confirmed from NXP forum question.
Refer i.MX7D datasheet section 4.10.1.1 ECSPI Master mode timing for timing details.
Please let us know if you have any other queries.