[Mesa-dev] [PATCH] radv: enable denorms for 64-bit and 16-bit floats

[Samuel Pitoiset]

On 12/28/2017 11:08 PM, Matt Arsenault wrote:
> 
> 
>> On Dec 28, 2017, at 16:55, Samuel Pitoiset <samuel.pitoiset at gmail.com> wrote:
>>
>> Similar to RadeonSI.
>>
>> This fixes:
>> dEQP-VK.image.texel_view_compatible.graphic.basic.attachment_read.bc*r16g16b16a16_sfloat
>> dEQP-VK.image.extended_usage_bit.attachment_write.r16_sfloat
>>
>> Signed-off-by: Samuel Pitoiset <samuel.pitoiset at gmail.com>
>> ---
>> src/amd/common/ac_nir_to_llvm.c | 14 ++++++++++++++
>> 1 file changed, 14 insertions(+)
>>
>> diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
>> index d9f2cb408c..9d9a1f911b 100644
>> --- a/src/amd/common/ac_nir_to_llvm.c
>> +++ b/src/amd/common/ac_nir_to_llvm.c
>> @@ -6879,6 +6879,20 @@ static void ac_compile_llvm_module(LLVMTargetMachineRef tm,
>> 	/* +3 for scratch wave offset and VCC */
>> 	config->num_sgprs = MAX2(config->num_sgprs,
>> 	                         shader_info->num_input_sgprs + 3);
>> +
>> +	/* Enable 64-bit and 16-bit denormals, because there is no performance
>> +	 * cost.
>> +	 *
>> +	 * If denormals are enabled, all floating-point output modifiers are
>> +	 * ignored.
>> +	 *
>> +	 * Don't enable denormals for 32-bit floats, because:
>> +	 * - Floating-point output modifiers would be ignored by the hw.
>> +	 * - Some opcodes don't support denormals, such as v_mad_f32. We would
>> +	 *   have to stop using those.
>> +	 * - SI & CI would be very slow.
>> +	 */
>> +	config->float_mode |= V_00B028_FP_64_DENORMS;
>> }
> 
> This is set in the program binary. You should use that directly rather than ignoring it

Not sure to understand where that flag is set actually, and RadeonSI 
does a similar thing.

>