docs: host-batched vendor + TF32 + fusion + latency results on sweep-… #8

	# GPU-less compile smoke: every push/PR compiles the full header surface and
	# all examples with nvcc on a hosted runner (compilation needs no GPU). The
	# actual GPU test suite runs on the lab box — see verify-gpu-proof.yml for the
	# signed-receipt attestation that those tests passed.
	name: CI
	on:
	push:
	branches: [main]
	pull_request:

	jobs:
	compile-smoke:
	runs-on: ubuntu-latest
	strategy:
	matrix:
	arch: [sm_90, sm_120]
	steps:
	- uses: actions/checkout@v4

	- name: Install CUDA toolkit (nvcc only, no GPU needed)
	uses: Jimver/[email protected]
	with:
	cuda: '13.2.0'
	method: network
	sub-packages: '["nvcc"]'

	- name: Compile the full header surface
	run: \|
	cat > smoke.cu <<'EOF'
	#include "glass.cuh"
	#include "glass-cgrps.cuh"
	#include "glass-defaults.cuh"
	__global__ void k(float* a, float* b, float* s) {
	glass::gemm<float, 4, 4, 4>(1.f, a, b, 0.f, s);
	glass::potrf<float, 4>(a);
	glass::posv<float>(4, a, b);
	glass::trsm<float, 4, 2>(a, b);
	glass::warp::trsv<float, 4>(a, b);
	glass::inv<float, 4>(a, s);
	}
	EOF
	nvcc -std=c++17 -arch=${{ matrix.arch }} -I. -c smoke.cu -o smoke.o

	- name: Compile all examples (MathDx-less set)
	run: make -C examples -j2 ARCH=${{ matrix.arch }}

Navigation Menu