Nvidia target via NVRTC and Nim ↦ CUDA DSL #487

Merged

mratsim merged 60 commits intomasterfrom

cudaJitCompiler

May 14, 2025

Commits on May 14, 2025

commit the initial Nim ⇒ CUDA DSL, NVRTC & CUDA execution helpers

Vindaar
authored and
mratsim
committed
[tests/examples] add example for a BigInt modular addition

Vindaar
authored and
mratsim
committed
remove `nimcuda` dependency, wrap everything we need manually

Vindaar
authored and
mratsim
committed
merge `execCuda` logic for LLVM & NVRTC

Vindaar
authored and
mratsim
committed
add `quitOnFailure` for the `check` calls to avoid regression

Vindaar
authored and
mratsim
committed
[tests] turn big int `modadd` example into real test

Vindaar
authored and
mratsim
committed
copy `libpaths.nim` over from nimcuda

Vindaar
authored and
mratsim
committed
[cuda] add partial support for `const` in CUDA generator

Vindaar
authored and
mratsim
committed
[cuda] fix minor issue in `if` statements in CUDA generator

Vindaar
authored and
mratsim
committed
[cuda] add support for named blocks

Vindaar
authored and
mratsim
committed
[cuda] add support for `bool`

Vindaar
authored and
mratsim
committed
[cuda] remove unnecessary semicolon

Vindaar
authored and
mratsim
committed
[cuda] add support for `{.volatile.}` variables

Vindaar
authored and
mratsim
committed
[nvrtc] add `modadd`, `modsub`, `mtymul` implementations using inline PTX

Vindaar
authored and
mratsim
committed
[cuda] support basic type conversions

Vindaar
authored and
mratsim
committed
[cuda] support `var` parameters in procs

Vindaar
authored and
mratsim
committed
[cuda] make sure proc body is a block

Vindaar
authored and
mratsim
committed
[cuda] support boolean / bitwise AND/OR and XOR, NOT

Vindaar
authored and
mratsim
committed
[cuda] support int32 literals

Vindaar
authored and
mratsim
committed
[cuda] handle prefix `not`

Vindaar
authored and
mratsim
committed
[cuda] make sure to pass `array` types by pointer instead of copy

Vindaar
authored and
mratsim
committed
[nvrtc] add more helpers, add TODO to investigate `slct` calls

Vindaar
authored and
mratsim
committed
[nvrtc] add many more field arithmetic / bigint operations

Vindaar
authored and
mratsim
committed
[tests] add test to pass by pointer and `var`

Vindaar
authored and
mratsim
committed
[tests] add test case for modadd/sub/mtymul

Vindaar
authored and
mratsim
committed
[tests] add basic test cases for all new NVRTC operations

Vindaar
authored and
mratsim
committed
[tests] update modadd/sub, mtymul test for new `getFieldModulus`

Vindaar
authored and
mratsim
committed
add BabyBear field

Vindaar
authored and
mratsim
committed
[nvrtc] handle `mtymul` for fields with 1 limb

Vindaar
authored and
mratsim
committed
[cuda] support nested array types, unpack generic instantiatons

Vindaar
authored and
mratsim
committed
[cuda] correctly generate ptr to array & ptr to array return types

Vindaar
authored and
mratsim
committed
[cuda] automatically generate `memcpy` for static array types

Vindaar
authored and
mratsim
committed
[nvrtc] use `const` for field modulus and other CT constants

Vindaar
authored and
mratsim
committed
[cuda] extend error message for non copyable inputs

Vindaar
authored and
mratsim
committed
[cuda] disable passStructByPointer also for CUDA

Vindaar
authored and
mratsim
committed
[cuda] add `{.nimonly.}` pragma one can use in `cuda` block

Vindaar
authored and
mratsim
committed
[cuda] better logic for detection of type names

Vindaar
authored and
mratsim
committed
[cuda] support `const` by mapping it to a `__constant__`

Vindaar
authored and
mratsim
committed
[cuda] allow type determination from array literal

Vindaar
authored and
mratsim
committed
[nvrtc] get rid of complexity with custom uint32 constants

Vindaar
authored and
mratsim
committed
[staticFor] add stepped variant of `staticFor`

Vindaar
authored and
mratsim
committed
[cuda] support func, discard and command nnkCommand

Vindaar
authored and
mratsim
committed
[cuda] better handle required semicolons

Vindaar
authored and
mratsim
committed
[cuda] extract type from `getType` for execution helper

Vindaar
authored and
mratsim
committed
[cuda] special case `CUdeviceptr` as a type that *must not* be copied

Vindaar
authored and
mratsim
committed
[cuda] allow passing in shared memory size for a kernel

Vindaar
authored and
mratsim
committed
[nvidia ABI] wrap cuModuleGetGlobal, cudaMemcpyKind and a couple more

Vindaar
authored and
mratsim
committed
[cuda] support while loops

Vindaar
authored and
mratsim
committed
[cuda] support void pointers and `nil` literals

Vindaar
authored and
mratsim
committed
[cuda] refactor out module loading from execution

Vindaar
authored and
mratsim
committed
[cuda] store PTX before echoeing it

Vindaar
authored and
mratsim
committed
[cuda] add `copyToSymbol` helper to copy to constant symbol in CUDA code

Vindaar
authored and
mratsim
committed
[cuda] generalize `volatile` annotation to support other pragmas

Vindaar
authored and
mratsim
committed
[cuda] `cudaName` pragma for custom name for a proc, eg __syncthreads

Vindaar
authored and
mratsim
committed
[cuda] support float literals

Vindaar
authored and
mratsim
committed
[cuda] map arrays of explicit length 0 to `[]` arrays in CUDA

Vindaar
authored and
mratsim
committed
[cuda] explicitly support constants, mapped to `__constant__`

Vindaar
authored and
mratsim
committed
[cuda] minor cleanup

Vindaar
authored and
mratsim
committed
[cuda] add `gridDim`, `cuExtern` and `share` + device malloc/free

Vindaar
authored and
mratsim
committed
force compilation with `-d:CTT_32` for the moment

Vindaar
authored and
mratsim
committed