AMDGPU: Document denormal behavior

[arsenm]

Summary

Not sure this is the right place for it next to the table.

[arsenm]

ping

[artem.tamazov]

Just in case, am I correct that the compiler engages output modifiers only when "denormal-fp-math"="positive-zero"? AFAIU this is necessary because output modifiers force -0 to +0.

[arsenm]

In D139455#3996356, @artem.tamazov wrote:

Just in case, am I correct that the compiler engages output modifiers only when "denormal-fp-math"="positive-zero"? AFAIU this is necessary because output modifiers force -0 to +0.

We have preserve-sign behavior and ignore positive-zero mode entirely. Also the modifiers are silently ignored for ieee mode =0. I don't remember what the wrong behavior depending on the denormal mode was, beyond don't use these without flushing

[artem.tamazov]

In D139455#3996367, @arsenm wrote:

We have preserve-sign behavior and ignore positive-zero mode entirely. Also the modifiers are silently ignored for ieee mode =0. I don't remember what the wrong behavior depending on the denormal mode was, beyond don't use these without flushing

I see. What I would like to emphasize is that the compiler should NOT use output modifiers for optimization when "denormal-fp-math"="preserve-sign" (in spite of modifiers will work in this mode), because output modifiers break sign preservation behavior. The compiler should use output modifiers only in "denormal-fp-math"="positive-zero" mode.

[arsenm]

In D139455#3998100, @artem.tamazov wrote:

In D139455#3996367, @arsenm wrote:

We have preserve-sign behavior and ignore positive-zero mode entirely. Also the modifiers are silently ignored for ieee mode =0. I don't remember what the wrong behavior depending on the denormal mode was, beyond don't use these without flushing

I see. What I would like to emphasize is that the compiler should NOT use output modifiers for optimization when "denormal-fp-math"="preserve-sign" (in spite of modifiers will work in this mode), because output modifiers break sign preservation behavior. The compiler should use output modifiers only in "denormal-fp-math"="positive-zero" mode.

The whole point of this change is to document that the hardware has preserve sign denormal behavior and we do not support positive-zero mode. We only apply output modifiers with nsz and don't consider the denormal mode

[arsenm]

ping

[arsenm]

ping

[arsenm]

ping

[foad]

Seems like an obvious incremental improvement to mention this, irrespective of any nits in the wording.

llvm/docs/AMDGPUUsage.rst

4596

If we can't implement flushing-to-positive-zero then surely we should treat "denormal-fp-math"="positive-zero" the same as "denormal-fp-math"="ieee", since according to LangRef "it is not mandated that flushing to zero occurs" at all. So I don't see where undefined behaviour comes into it.

[arsenm]

llvm/docs/AMDGPUUsage.rst

4596

It's informative of the mode, and we can be a bit lax on the outputs. However, for the input treatment, this acts more like a constraint (e.g. you can't introduce a divide by 0 for a denormal input)

[arsenm]

c0db240e5167885cc8083d72db85c80f5b7eb936

llvm/docs/AMDGPUUsage.rst

.. table:: Floating Point Denorm Mode Enumeration Values
      ====================================== ===== ==============================
      FLOAT_DENORM_MODE_FLUSH_SRC_DST        0     Flush Source and Destination
                                                   Denorms
      FLOAT_DENORM_MODE_FLUSH_DST            1     Flush Output Denorms
      FLOAT_DENORM_MODE_FLUSH_SRC            2     Flush Source Denorms
      FLOAT_DENORM_MODE_FLUSH_NONE           3     No Flush
      ====================================== ===== ==============================
 
+     Denormal flushing is sign respecting. i.e. the behavior expected
+     by ``"denormal-fp-math"="preserve-sign"``. The behavior is
+     undefined with ``"denormal-fp-math"="positive-zero"``
+
 ..
 
   .. table:: System VGPR Work-Item ID Enumeration Values
      :name: amdgpu-amdhsa-system-vgpr-work-item-id-enumeration-values-table
 
      ======================================== ===== ============================
      Enumeration Name                         Value Description
      ======================================== ===== ============================