Please support AVX512_FP16

[Elijah-777]

Chips supporting AVX512_FP16 have been released for more than a year. Why does Intel's open source GPU computing sub-computing library still not support AVX512_FP16? AVX512_FP16 is the instruction I expect to use

[vpirogov]

oneDNN uses instructions from AVX512_FP16 ISA extension on processors with Intel AVX 10.1/512 instruction set support (4th and 5th generation Intel Xeon Scalable Processors and Intel Xeon 6 processors).

Default numerical behavior for oneDNN functions requires fp32 accumulation, which is not supported by FMA instructions in AVX512_FP16 extension. This implementation can be added in relaxed accumulation mode, but it's not a priority for the core engineering team at the moment.

[shu1chen]

Hello @DaiShaoJie77, the configuration of oneDNN functions might prevent the test from utilizing the AVX512_FP16 ISA in some use cases. For more details of the issue, could you please provide the oneDNN verbose log by setting ONEDNN_VERBOSE=dispatch in the test environment?

[Elijah-777]

oneDNN uses instructions from AVX512_FP16 ISA extension on processors with Intel AVX 10.1/512 instruction set support (4th and 5th generation Intel Xeon Scalable Processors and Intel Xeon 6 processors).

Default numerical behavior for oneDNN functions requires fp32 accumulation, which is not supported by FMA instructions in AVX512_FP16 extension. This implementation can be added in relaxed accumulation mode, but it's not a priority for the core engineering team at the moment.

Hi, I don't understand what you mean. Do you mean to add an option somewhere to enable fp16? Which file and where is it?@vpirogov

[Elijah-777]

Hello @DaiShaoJie77, the configuration of oneDNN functions might prevent the test from utilizing the AVX512_FP16 ISA in some use cases. For more details of the issue, could you please provide the oneDNN verbose log by setting ONEDNN_VERBOSE=dispatch in the test environment?

I tried to set this parameter in the environment variable, but it just printed some data types and did not use the instruction set I wanted.@shu1chen

[shu1chen]

I tried to set this parameter in the environment variable, but it just printed some data types and did not use the instruction set I wanted.

Please send us the oneDNN verbose log so we can identify the exact issue you're experiencing.

Since we haven't received the verbose log, we're unable to determine how you're using oneDNN. For example, if your input data to oneDNN is FP32 and you wish to use AVX512_FP16 ISA, you'll need to set the fpmath_mode to f16 and the accumulation mode to relaxed in dnnl::primitive_attr. For example:

    dnnl::primitive_attr attr;
    attr.set_fpmath_mode(dnnl::fpmath_mode::f16, true);
    attr.set_accumulation_mode(dnnl::accumulation_mode::relaxed);

You can find the documentation for Primitive Attributes in the link.