package xla.gpu

Get desktop application:
View/edit binary Protocol Buffers messages

• repeated AlgorithmDenylistEntry entries = 1

Used in: AlgorithmDenylist

• string hlo = 1

• optional tensorflow.ComputeCapability cc = 2

• optional tensorflow.CudnnVersion cudnn_version = 3

• string blas_version = 5

• repeated DenylistedAlgorithm algos = 4

Backend config for bitcast operation generated from MLIR MHLO dialect.

• optional LayoutProto source_layout = 1

• optional LayoutProto result_layout = 2

• optional HloInstructionProto instruction = 1

• repeated ShapeProto operand_shapes = 2

• uint64 result_address = 3

• repeated uint64 operand_addresses = 4

Backend config for a convolution that runs through cudnn.

• optional stream_executor.dnn.AlgorithmProto algorithm = 6

  Opaque algorithm number and tuning knobs chosen for this conv.

• double conv_result_scale = 4

  The scaling factor multiplied with the convolution result.

• int64 activation_mode = 3

  The requested activation (e.g. relu) after the convolution. It is with type stream_executor::dnn::ActivationMode.

• double side_input_scale = 5

  The scaling factor multiplied with the side input. If no side input buffer is provided, this field must be 0.

Used in: AlgorithmDenylistEntry

• int64 id = 1

• bool tensor_ops = 2

Backend config for the GEMM operation running through cuBLAS.

• oneof algorithm

  Opaque optional algorithm number. No chosen number indicates that a different cuBLAS API will be used, which does not allow for choosing an algorithm.

  • int64 selected_algorithm = 1

• double alpha_real = 2

• double alpha_imag = 9

• double beta = 3

• optional DotDimensionNumbers dot_dimension_numbers = 7

• optional PrecisionConfig precision_config = 12

• GemmBackendConfig.Epilogue epilogue = 13

cublasLt matmul epilogue.

Used in: GemmBackendConfig

• DEFAULT = 0

• BIAS = 1

• optional XlaRuntimeExecutableProto xla_runtime_executable = 1

• string gpu_asm_text = 2

  PTX for the compiled GPU kernels.

• bytes gpu_binary = 3

  Corresponding CUBIN for the above PTX.