Opus Kernel Best Practice

OPUS Kernel Compile-Time Best Practices

Techniques for reducing HIP/C++ kernel compile time when using opus.hpp. These patterns were developed while optimizing a GQA flash attention kernel from 4.8s to 1.5s (70% reduction) in device-only compilation.

Required headers and include paths

For kernel development with OPUS, use these headers from csrc/include/:

• opus/opus.hpp — the OPUS template library + device intrinsic wrappers. This is the only include needed for device code. Provides opus::thread_id_x(), opus::block_id_x(), opus::sync_threads(), opus::warp_all(), etc.
• opus/hip_minimal.hpp — minimal HIP host-side only declarations (dim3, hipMalloc, hipLaunchKernelGGL, etc.). Use on the host pass instead of <hip/hip_runtime.h>.

hipcc my_kernel.cu -I<aiter_root>/csrc/include -D__HIPCC_RTC__ -std=c++20 -O3 --offload-arch=gfx950

If anything is missing, contact the maintainer (carlus.huang@amd.com) for adding support.

0. Always Separate Device and Host Code (Most Important)

This is the single most impactful technique. hipcc always performs two compilation passes on every .hip/.cu file — one for the host (x86_64) and one for the device (AMDGPU). The heavy opus.hpp template library is only needed on the device side, but without a guard, hipcc parses it on BOTH passes, doubling the frontend cost.

Always structure your kernel files like this:

// my_kernel.cu
#ifdef __HIP_DEVICE_COMPILE__
// ── Device pass: include opus.hpp and define kernels ──
#include "opus/opus.hpp"

__global__ __launch_bounds__(256, 2)
void my_kernel(const float* src, float* dst, int n) {
    // ... opus layout, load, store, MMA, etc.
}

#else
// ── Host pass: minimal declarations + launcher only ──
#include "opus/hip_minimal.hpp"

__global__ void my_kernel(const float* src, float* dst, int n);  // declaration only

extern "C" void run_my_kernel(const void* d_src, void* d_dst, int n) {
    dim3 grid((n + 255) / 256), block(256);
    hipLaunchKernelGGL(my_kernel, grid, block, 0, 0,
                       (const float*)d_src, (float*)d_dst, n);
    hipDeviceSynchronize();
}
#endif

Why this works:

• The device pass sees opus.hpp + kernel definitions — full template expansion
• The host pass sees only opus/hip_minimal.hpp (~70 lines) + kernel declaration + launch wrapper
• Saves ~50% of total compile time by eliminating opus.hpp parsing on the host pass
• The extern "C" launcher can be called from Python via ctypes.CDLL — no pybind11/torch extension needed

Compile flags:

hipcc my_kernel.cu \
  -I<aiter_root>/csrc/include \
  -D__HIPCC_RTC__ \
  -std=c++20 -O3 -ffast-math \
  --offload-arch=gfx950 \
  -fPIC -shared -o my_kernel.so

1. Minimize Header Overhead

Replace <hip/hip_runtime.h> with opus/hip_minimal.hpp

Standard <hip/hip_runtime.h> expands to 190K preprocessed lines. The aiter-provided opus/hip_minimal.hpp (80 lines) declares only what's needed — dim3, hipLaunchKernelGGL, hipMalloc/hipFree, __launch_bounds__, __shared__/__device__/__global__, and __all(). Use AMDGCN compiler builtins for device intrinsics:

int tid = __builtin_amdgcn_workitem_id_x();     // threadIdx.x
int bid = __builtin_amdgcn_workgroup_id_x();     // blockIdx.x
int bsz = __builtin_amdgcn_workgroup_size_x();   // blockDim.x
__builtin_amdgcn_s_barrier();                     // __syncthreads()

Use -D__HIPCC_RTC__ to suppress implicit includes

Even with minimal headers, hipcc's implicit __clang_hip_runtime_wrapper.h pulls in <cmath>, <cstdlib>, etc. The -D__HIPCC_RTC__ flag skips these. Provide #define INFINITY __builtin_huge_valf() if needed.

Use ctypes instead of pybind11/torch extension for Python bindings

The C++ binding layer is often the biggest compile cost. The extern "C" + ctypes.CDLL pattern from Section 0 eliminates it entirely:

2. Reduce Template Instantiation Count

Use runtime loops instead of static_for where compile-time indices aren't needed

Each iteration of static_for<N>([&](auto I){...}) creates a unique lambda instantiation. For large N, this dominates compile time. Replace with plain for loops when the loop body doesn't need compile-time I:

// SLOW: N unique lambda instantiations
static_for<N>([&](auto I) {
    r[I.value] = load<vec>(offsets[I.value]);
});

// FAST: 1 instantiation, compiler unrolls identically
for (index_t i = 0; i < N; i++) {
    r[i] = load<vec>(offsets[i]);
}

When you still need static_for: If the body uses I as a template argument (e.g., number<I.value>{} for set_slice, slice, or immediate-offset _tr_load<vec, off>), you must keep static_for.

Use runtime flat_to_coords instead of compile-time multi-index decomposition

layout_to_offsets converts a layout into a precomputed offset array using a runtime loop with flat_to_coords, which produces tuple<index_t, ...> (one type for all iterations) instead of tuple<number<a>, number<b>, ...> (unique type per iteration):

// SLOW: N unique coord_to_linear instantiations (one per multi-index combination)
static_ford(issue_space_vec, [&](auto... ids) {
    offsets[u_linear(ids...)] = u(ids...);
});

// FAST: 1 coord_to_linear instantiation (all iterations use tuple<index_t, ...>)
for (index_t i = 0; i < num_issues; i++) {
    offsets[i] = u(flat_to_coords(i, make_index_seq<ndim>{}, issue_space_vec));
}

Cache constexpr computations in struct members

Repeated constexpr evaluations in multiple methods trigger re-evaluation in each:

// SLOW: y_shape_a() + reduce_tuple_mul evaluated in every operator()/step_k() overload
constexpr auto a_len = get<0>(reduce_tuple_mul(MMA::y_shape_a()));

// FAST: cached once as class member
static constexpr index_t mma_a_len = get<0>(reduce_tuple_mul(MMA::y_shape_a())).value;

3. Use LLVM Builtins for Vector Operations

__builtin_convertvector for type conversion

Replaces N-element element-by-element cast_impl pack expansion with a single LLVM intrinsic:

// SLOW: 64-element pack expansion
return vector_return_type<D, decltype(cast<D>(get<Is>(s)))...>{cast<D>(get<Is>(s))...};

// FAST: single builtin call
return __builtin_convertvector(s, vector_t<D, size<S>()>);

__builtin_shufflevector for vector slice/concat

Replaces element-by-element make_vector(get<Is>(c)...) with a single shuffle:

// SLOW: N-element braced init
return make_vector(get<Is>(c)...);

// FAST: single shuffle (returns GCC-style vector, bit_cast to ext_vector_type)
using R = vector_t<scalar_type, sizeof...(Is)>;
return __builtin_bit_cast(R, __builtin_shufflevector(c, c, Is...));

4. Avoid Intermediate Type Creation

Bypass concat_tuple with direct indexing

concat_tuple creates intermediate tuple types when concatenating >4 tuples. Replace with direct per-element computation:

// unfold_x_stride: instead of concat_tuple(per_group_results...)
// compute each element's stride directly via unfold_x_stride_at<J>()

// pickup_shape: instead of concat_tuple(conditional<match, tuple<T>, tuple<>>{}...)
// build a filtered index sequence, then make_tuple(get<filtered_indices>(Shape{})...)

// flatten_tuple: instead of concat_tuple(explode_tuple(get<Is>(t))...)
// directly index as get<local>(get<group>(t)) via flatten_at<T, J, GS>()

Specify return type explicitly to avoid std::common_type

// SLOW: triggers recursive std::common_type<D, D, D, ..., D> with 64 types
return vector_return_type<void, decltype(cast<D>(get<Is>(s)))...>{...};

// FAST: D is already known, skip common_type entirely
return vector_return_type<D, decltype(cast<D>(get<Is>(s)))...>{...};

Add fold-expression fast paths for common patterns

// reduce_tuple_mul for tuple<number<>...>: fold expression instead of recursive reduction
template<typename... Ns, std::enable_if_t<(is_constant_v<Ns> && ...), bool> = true>
constexpr auto reduce_tuple_mul(const tuple<Ns...>&) { return tuple<number<(Ns::value * ...)>>{}; }

5. Parallel Compilation

Split device test files by template-instantiation cost

One file with 14 MFMA template instantiations (~3.9s) bottlenecks parallel builds. Split into per-type files (f16/f32/f8) to balance workload:

test_mfma.cu (3.9s) -> test_mfma_f16.cu (0.9s) + test_mfma_f32.cu (0.5s) + test_mfma_f8.cu (0.9s)

Use hipcc --genco for device-only compilation when launching from Python

Eliminates the host pass entirely. Python loads the .hsaco via hipModuleLoad and launches with hipModuleLaunchKernel (HIP driver API).

Compile-Time Measurement

Use -ftime-trace for profiling

hipcc kernel.cc --cuda-device-only -c -o /dev/null \
  -Xclang -ftime-trace=trace.json

Analyze with chrome://tracing or a script:

import json
with open('trace.json') as f: data = json.load(f)
events = data.get('traceEvents', data)
inst = [(e['dur'], e['args']['detail']) for e in events
        if e.get('name') == 'InstantiateFunction' and 'dur' in e]
inst.sort(key=lambda x: -x[0])
for dur, name in inst[:20]:
    print(f"{dur/1000:8.1f}ms  {name[:100]}")

Key metrics to track

• Function instantiations: total count and per-function time
• Frontend vs Backend: frontend = template instantiation, backend = LLVM optimizer + codegen
• Critical path: the single slowest template chain determines wall-clock time

Summary Table