Example

Setting Up the Environment

• Configure environment variables by referring to Environment Setup.
• Click here to obtain the sample project for operator check.
  

  • ${git_clone_path} is the installation path of the sample repository.
  • This sample project uses Atlas A2 training products/Atlas A2 inference products as an example.
  • When downloading the code sample, run the following command to specify the branch version:

    git clone https://gitee.com/ascend/samples.git -b master

Procedure

1. Run the install.sh script in the ${git_clone_path}/operator/ascendc/0_introduction/12_matmulleakyrelu_frameworklaunch directory to generate a custom operator project and implement operators on the host and kernel. For details, see the sample project in Setting Up the Environment.

   bash install.sh -v Ascendxxxyy    # xxxyy indicates the type of the chip used by the user.

2. Switch to the custom operator project directory.

   cd CustomOp

3. Edit the operator startup script matmulleakyrelu.py.

   import numpy as np
   import mskpp
   # The input parameters of this function must be the same as those of the kernel function.
   def run_kernel(input_a, input_b, input_bias, output, workspace, tiling_data):
       kernel_binary_file = "MatmulLeakyreluCustom.o"   # The .o file name varies depending on the hardware and operating system. For details about the path, see the parameter -reloc in Table 1.
       kernel = mskpp.get_kernel_from_binary(kernel_binary_file, 'mix')
       return kernel(input_a, input_b, input_bias, output, workspace, tiling_data)
   
   if __name__ == "__main__":
       # input/output tensor
       M = 1024
       N = 640
       K = 256
       input_a = np.random.randint(1, 10, [M, K]).astype(np.float16)
       input_b = np.random.randint(1, 10, [K, N]).astype(np.float16)
       input_bias = np.random.randint(1, 10, [N]).astype(np.float32)
       output = np.zeros([M, N]).astype(np.float32)
       # shape info
       inputs_info = [{"shape": [M, K], "dtype": "float16", "format": "ND"},
                      {"shape": [K, N], "dtype": "float16", "format": "ND"},
                      {"shape": [N], "dtype": "float32", "format": "ND"}]
       outputs_info = [{"shape": [M, N], "dtype": "float32", "format": "ND"}]
       attr = {}
       # Calling the tiling function
       tiling_output = mskpp.tiling_func(
           op_type="MatmulLeakyreluCustom",
           inputs_info=inputs_info, outputs_info=outputs_info, # Optional
           inputs=[input_a, input_b, input_bias], outputs=[output],
           attr=attr # Optional
           lib_path="liboptiling.so",  # Compilation product of the tiling code. For details about the directory structure, see Step 2 in DPP Deployment.
           # soc_version="", # Optional
       )
       blockdim = tiling_output.blockdim
       workspace_size = tiling_output.workspace_size
       tiling_data = tiling_output.tiling_data # numpy array.
       workspace = np.zeros(workspace_size).astype(np.uint8) # The workspace needs to be applied for by the user.
       # Calling the kernel function
       run_kernel(input_a, input_b, input_bias, output, workspace, tiling_data)
   
       # Verification output
       alpha = 0.001
       golden = (np.matmul(input_a.astype(np.float32), input_b.astype(np.float32)) + input_bias).astype(np.float32)
       golden = np.where(golden >= 0, golden, golden * alpha)
       is_equal = np.array_equal(output, golden)
       result = "success" if is_equal else "failed"
       print("compare {}.".format(result))

4. Run the script.

   python3 matmulleakyrelu.py