Support double-precision floating point

[kpet]

clspv generally doesn't have support for double-precision floating point. The following at least are missing:

• [x] support double literals (see #157)
• [x] support double constants
• [ ] add support for double to the built-in functions
• [ ] support double with operators and add tests
• [ ] update the documentation
• [ ] anything else?

[Trass3r]

add support for double to the built-in functions

Do they need to be emulated? Currently some are mapped to ExtInsts (like here) which are not specified for fp64: https://www.khronos.org/registry/spir-v/specs/unified1/GLSL.std.450.html#_extended_instructions

[Cazadorro]

add support for double to the built-in functions

Do they need to be emulated? Currently some are mapped to ExtInsts (like here) which are not specified for fp64: https://www.khronos.org/registry/spir-v/specs/unified1/GLSL.std.450.html#_extended_instructions

I know this is years late, but my understanding is that none of the current desktop GPU makers have hardware accelerated double precision trigonometry (ie they don't have special function units for). Nvidia appears to only have single precision SFU, even on their server GPUs.

• https://forums.developer.nvidia.com/t/sfu-performance-in-a100/197699/12

• https://docs.nvidia.com/cuda/pdf/CUDA_C_Programming_Guide.pdf see "Maximize Instruction Throughput"

This means "emulated" double precision functions are literally how special functions are handled on Nvidia cards (aside from basic arithmetic operations). Note not all double precision functions need to be "emulated" https://registry.khronos.org/SPIR-V/specs/unified1/GLSL.std.450.html#_introduction

For GLSL functions, this means Trig, Log, Pow, and Exp functions all required to be "emulated", for OpenCL, additional functions would need to be emulated (like gamma etc...).