Khronos Group has published two new extensions for Vulkan: VK_KHR_shader_float16_int8 and VK_KHR_shader_float_controls. In this post, I will talk about VK_KHR_shader_float_controls, which is the extension I have been implementing on Anvil driver, the open-source Intel Vulkan driver, as part of my job at Igalia. For information about VK_KHR_shader_float16_int8 and its implementation in Mesa, you can read Iago’s blogpost.
The Vulkan Working Group has defined a new extension VK_KHR_shader_float_controls, which allows applications to query and override the implementation’s default floating point behavior for rounding modes, denormals, signed zero and infinity. From the Vulkan application developer perspective, VK_shader_float_controls defines a new structure called VkPhysicalDeviceFloatControlsPropertiesKHR where the drivers expose the supported capabilities such as the rounding modes for each floating point data type, how the denormals are expected to be handled by the hardware (either flush to zero or preserve their bits) and if the value is a signed zero, infinity and NaN, whether it will preserve their bits.
typedef struct VkPhysicalDeviceFloatControlsPropertiesKHR {
VkStructureType sType;
void* pNext;
VkBool32 separateDenormSettings;
VkBool32 separateRoundingModeSettings;
VkBool32 shaderSignedZeroInfNanPreserveFloat16;
VkBool32 shaderSignedZeroInfNanPreserveFloat32;
VkBool32 shaderSignedZeroInfNanPreserveFloat64;
VkBool32 shaderDenormPreserveFloat16;
VkBool32 shaderDenormPreserveFloat32;
VkBool32 shaderDenormPreserveFloat64;
VkBool32 shaderDenormFlushToZeroFloat16;
VkBool32 shaderDenormFlushToZeroFloat32;
VkBool32 shaderDenormFlushToZeroFloat64;
VkBool32 shaderRoundingModeRTEFloat16;
VkBool32 shaderRoundingModeRTEFloat32;
VkBool32 shaderRoundingModeRTEFloat64;
VkBool32 shaderRoundingModeRTZFloat16;
VkBool32 shaderRoundingModeRTZFloat32;
VkBool32 shaderRoundingModeRTZFloat64;
} VkPhysicalDeviceFloatControlsPropertiesKHR;
This structure will be filled by the driver when calling vkGetPhysicalDeviceProperties2(), with a pointer to such structure as one of the pNext pointers of VkPhysicalDeviceProperties2 structure. With that, we know if the driver will support the SPIR-V capabilities we want to use in our shaders, if separate*Settings are true, remember to check the value of the property for the floating point bit-size types you are planning to work with.
The required bits to enable such capabilities in a SPIR-V shader are the following:
- Enable the extension:
OpExtension "SPV_KHR_float_controls" - Enable the desired capability. For example:
OpCapability DenormFlushToZero - Specify where to apply it. For example, we would like to flush to zero all fp64 denormalss in the %main function of a shader:
OpExecutionMode %main DenormFlushToZero 64. If we want to apply different modes, we would repeat that line with the needed ones. - Profit!
I implemented the support of this extensions for the Anvil’s supported GPUs (Broadwell, Skylake, Kabylake and newer), although we don’t support all the capabilities. For example on Broadwell, float16 denormals are not supported, and the support for flushing to zero the float16 denormals is not supported for all the instructions in the rest of generations.
If you are interested, the patches are now under review :-) As there are not real world code using this feature yet, please fill any bug you find about this in our bugzilla.