VK_EXT_shader_float8

Other Extension Metadata

Last Modified Date

2025-04-16

IP Status

No known IP claims.

Contributors

• Kévin Petit, Arm Ltd.
• Stu Smith, AMD
• Jeff Bolz, NVIDIA
• Craig Graham, Samsung

Description

This extension enables support for 8-bit floating-point data types as defined in SPV_EXT_float8.

New Structures

• Extending VkPhysicalDeviceFeatures2, VkDeviceCreateInfo:
  • VkPhysicalDeviceShaderFloat8FeaturesEXT

New Enum Constants

• VK_EXT_SHADER_FLOAT8_EXTENSION_NAME
• VK_EXT_SHADER_FLOAT8_SPEC_VERSION
• Extending VkStructureType:
  • VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_FLOAT8_FEATURES_EXT

If VK_KHR_cooperative_matrix is supported:

• Extending VkComponentTypeKHR:
  • VK_COMPONENT_TYPE_FLOAT8_E4M3_EXT
  • VK_COMPONENT_TYPE_FLOAT8_E5M2_EXT

New SPIR-V Capabilities

• Float8EXT
• Float8CooperativeMatrixEXT

Issues

1) Resolve interactions with the changes VK_KHR_shader_float16 makes to rules for denorm flushing (always allowed by default for all FP formats). How to describe the requirement to preserve subnormals?

+ RESOLVED: Subnormals are always preserved when converting FP8 values to IEEE 754 binary 16. In all other cases, subnormals may be flushed to zero.

+

Version History

• Revision 1, 2025-04-16 (Kévin Petit)
  • Initial draft