#include <torch/library.h>

#include "core/registration.h"
#include "torch_binding.h"

TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
  // CUTLASS w8a8 GEMM, supporting symmetric per-tensor or per-row/column                                                                                                                                            
  // quantization, as well as bias                                                                                                                                                                                   
  ops.def(                                                                                                                                                                                                           
      "cutlass_scaled_mm(Tensor! out, Tensor a,"                                                                                                                                                                     
      "                  Tensor b, Tensor a_scales,"                                                                                                                                                                 
      "                  Tensor b_scales, Tensor? bias) -> ()");                                                                                                                                                     
  ops.impl("cutlass_scaled_mm", torch::kCUDA, &cutlass_scaled_mm);                                                                                                                                                   
                                                                                                                                                                                                                     
  // CUTLASS w8a8 GEMM, supporting asymmetric per-tensor or per-row/column                                                                                                                                           
  // quantization.                                                                                                                                                                                                   
  ops.def(                                                                                                                                                                                                           
      "cutlass_scaled_mm_azp(Tensor! out, Tensor a,"                                                                                                                                                                 
      "                  Tensor b, Tensor a_scales,"                                                                                                                                                                 
      "                  Tensor b_scales, Tensor azp_adj,"                                                                                                                                                           
      "                  Tensor? azp, Tensor? bias) -> ()");                                                                                                                                                         
  ops.impl("cutlass_scaled_mm_azp", torch::kCUDA, &cutlass_scaled_mm_azp);                                                                                                                                           
                                                                                                                                                                                                                     
  // Check if cutlass scaled_mm is supported for CUDA devices of the given                                                                                                                                           
  // capability                                                                                                                                                                                                      
  ops.def("cutlass_scaled_mm_supports_fp8(int cuda_device_capability) -> bool");                                                                                                                                     
  ops.impl("cutlass_scaled_mm_supports_fp8", &cutlass_scaled_mm_supports_fp8);                                            

  // Compute FP8 quantized tensor for given scaling factor.
  ops.def(
      "static_scaled_fp8_quant(Tensor! result, Tensor input, Tensor scale) -> "
      "()");
  ops.impl("static_scaled_fp8_quant", torch::kCUDA, &static_scaled_fp8_quant);

  // Compute dynamic-per-tensor FP8 quantized tensor and scaling factor.
  ops.def(
      "dynamic_scaled_fp8_quant(Tensor! result, Tensor input, Tensor! scale) "
      "-> "
      "()");
  ops.impl("dynamic_scaled_fp8_quant", torch::kCUDA, &dynamic_scaled_fp8_quant);

  // Compute dynamic-per-token FP8 quantized tensor and scaling factor.
  ops.def(
      "dynamic_per_token_scaled_fp8_quant(Tensor! result, Tensor input, "
      "Tensor! scale, Tensor? scale_ub) -> "
      "()");
  ops.impl("dynamic_per_token_scaled_fp8_quant", torch::kCUDA,
           &dynamic_per_token_scaled_fp8_quant);

  // Compute int8 quantized tensor for given scaling factor.
  ops.def(
      "static_scaled_int8_quant(Tensor! result, Tensor input, Tensor scale,"
      "Tensor? azp) -> ()");
  ops.impl("static_scaled_int8_quant", torch::kCUDA, &static_scaled_int8_quant);

  // Compute int8 quantized tensor and scaling factor
  ops.def(
      "dynamic_scaled_int8_quant(Tensor! result, Tensor input, Tensor! scale, "
      "Tensor!? azp) -> ()");
  ops.impl("dynamic_scaled_int8_quant", torch::kCUDA,
           &dynamic_scaled_int8_quant);

  // fp8_marlin Optimized Quantized GEMM for FP8 weight-only.
  ops.def(
      "fp8_marlin_gemm(Tensor a, Tensor b_q_weight, Tensor b_scales, "
      "Tensor! workspace, int num_bits, SymInt size_m, SymInt size_n, "
      "SymInt size_k) -> Tensor");
  ops.impl("fp8_marlin_gemm", &fp8_marlin_gemm);
}

REGISTER_EXTENSION(TORCH_EXTENSION_NAME)