code/pointcept_framework/pointcept/models/quantization/backend/binary_gemm.cpp

// PyTorch binding for packed binary GEMM (±1 dot product via uint32 packing).
#include <torch/extension.h>

torch::Tensor binary_gemm_forward_cuda(
    torch::Tensor packed_input,
    torch::Tensor packed_weight_t,
    int64_t original_n);

torch::Tensor binary_gemm_forward(
    torch::Tensor packed_input,
    torch::Tensor packed_weight_t,
    int64_t original_n) {
  TORCH_CHECK(packed_input.is_cuda(), "binary_gemm: packed_input must be CUDA");
  TORCH_CHECK(packed_weight_t.is_cuda(), "binary_gemm: packed_weight_t must be CUDA");
  TORCH_CHECK(packed_input.scalar_type() == torch::kUInt32,
              "binary_gemm: packed_input must be uint32");
  TORCH_CHECK(packed_weight_t.scalar_type() == torch::kUInt32,
              "binary_gemm: packed_weight_t must be uint32");
  packed_input = packed_input.contiguous();
  packed_weight_t = packed_weight_t.contiguous();
  return binary_gemm_forward_cuda(packed_input, packed_weight_t, original_n);
}

PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
  m.def("forward", &binary_gemm_forward, "Binary GEMM forward (packed uint32)");
}