OKKernel 用户开发文档

disclaimer

• Disclaimer

Summary

• Introduction
• History

About TPU

• TPU Architecture
• Memory Types
• Heterogeneous Programming
  • Device
  • Host

Address, Strides and Layouts

• Address of Tensor in Local Memory
• Scatter Features to Different NPUs
• Strides in Local Memory
• Strides in System Memory
• Layouts
  • Continuous Layout
  • 128-Byte Aligned Layout
  • Compact Layout
  • Matrix Layout
• Storage Modes
  • 4N-mode
  • 2N-mode
  • 2IC-mode

Atomic Functions

• About Function Names
• Some Definitions
• Common Functions
  • okk_initialize
  • okk_poll
  • okk_parallel_start
  • okk_parallel_end
  • okk_is_parallel_state
  • okk_local_mem_size_per_npu
  • okk_l2_sram_size
  • okk_dtcm_size
  • okk_npu_num
• Utils Functions
  • okk_start_npu_index
  • okk_channle_num_per_npu
  • okk_128_byte_aligned_stride_for_32bit
  • okk_128_byte_aligned_stride_for_16bit
  • okk_128_byte_aligned_stride_for_8bit
  • okk_compact_stride
  • okk_continuous_stride
• GDMA Functions
  • okk_gdma_32bit_cpy_S2L
  • okk_gdma_32bit_cpy_L2S
  • okk_gdma_32bit_cpy_L2L
  • okk_gdma_32bit_cpy_S2S
  • okk_gdma_32bit_matrix_S2L
  • okk_gdma_32bit_matrix_L2S
  • okk_gdma_32bit_set_C_system
  • okk_gdma_32bit_set_C_local
  • okk_gdma_16bit_cpy_S2L
  • okk_gdma_16bit_cpy_L2S
  • okk_gdma_16bit_cpy_L2L
  • okk_gdma_16bit_cpy_S2S
  • okk_gdma_16bit_set_C_system
  • okk_gdma_16bit_set_C_local
  • okk_gdma_8bit_cpy_S2L
  • okk_gdma_8bit_cpy_L2S
  • okk_gdma_8bit_cpy_L2L
  • okk_gdma_8bit_cpy_S2S
  • okk_gdma_8bit_set_C_system
  • okk_gdma_8bit_set_C_local
• Memory Functions
  • okk_bdc_32bit_cpy
  • okk_bdc_32bit_set_C
• Data Type Converting Functions
  • okk_bdc_fp32_to_int32
  • okk_bdc_lookup_int32_to_fp32
  • okk_bdc_4N_int8_to_fp32
  • okk_bdc_int8_to_fp32
• FP32 Binary Functions
  • okk_bdc_add
  • okk_bdc_add_C
  • okk_bdc_sub
  • okk_bdc_sub_C
  • okk_bdc_C_sub
  • okk_bdc_mul
  • okk_bdc_mul_C
  • okk_bdc_div
  • okk_bdc_div_C
  • okk_bdc_C_div
  • okk_bdc_mac
  • okk_bdc_mac_C
  • okk_bdc_max
  • okk_bdc_max_C
  • okk_bdc_min
  • okk_bdc_min_C
  • okk_bdc_greater_select_value
  • okk_bdc_greater_C_select_value
  • okk_bdc_C_greater_select_value
  • okk_bdc_equal_select_value
  • okk_bdc_equal_C_select_value
• 32-Bit Binary Functions
  • okk_bdc_32bit_and
  • okk_bdc_32bit_and_C
  • okk_bdc_32bit_or
  • okk_bdc_32bit_or_C
  • okk_bdc_32bit_xor
  • okk_bdc_32bit_xor_C
  • okk_bdc_32bit_arithmetic_shift
  • okk_bdc_32bit_logical_shift
  • okk_bdc_32bit_arithmetic_shift_C
  • okk_bdc_32bit_logical_shift_C
  • okk_bdc_32bit_C_arithmetic_shift
  • okk_bdc_32bit_C_logical_shift
• FP32 Unary Functions
  • okk_bdc_rsqrt
  • okk_bdc_sqrt
  • okk_bdc_taylor_exp
  • okk_bdc_lookup_exp
  • okk_bdc_exp
  • okk_bdc_exp_tunable
  • okk_bdc_sigmoid
  • okk_bdc_sigmoid_tunable
  • okk_bdc_tanh
  • okk_bdc_tanh_tunable
  • okk_bdc_reciprocal
  • okk_bdc_neg
• FP32 Neural Network Functions
  • okk_bdc_relu
  • okk_bdc_bias
  • okk_bdc_scale
  • okk_bdc_scale_bias
  • okk_bdc_conv2d
  • okk_bdc_depthwise2d
  • okk_bdc_avg_pool2d
  • okk_bdc_avg_pool2d_v2
  • okk_bdc_max_pool2d
  • okk_bdc_matmul
• Fixed Point Binary Functions
  • okk_bdc_fixed_point_packed_add
  • okk_bdc_fixed_point_packed_add_C
  • okk_bdc_fixed_point_packed_sub
  • okk_bdc_fixed_point_packed_sub_C
  • okk_bdc_fixed_point_packed_C_sub
  • okk_bdc_fixed_point_packed_mul
  • okk_bdc_fixed_point_packed_mul_C
  • okk_bdc_fixed_point_packed_mac
  • okk_bdc_fixed_point_packed_mac_C
  • okk_bdc_fixed_point_packed_max
  • okk_bdc_fixed_point_packed_max_C
  • okk_bdc_fixed_point_packed_min
  • okk_bdc_fixed_point_packed_min_C
  • okk_bdc_fixed_point_packed_16bit_arithmetic_shift
  • okk_bdc_fixed_point_packed_16bit_logical_shift
  • okk_bdc_fixed_point_packed_16bit_arithmetic_shift_C
  • okk_bdc_fixed_point_packed_16bit_logical_shift_C
  • okk_bdc_fixed_point_packed_16bit_C_arithmetic_shift
  • okk_bdc_fixed_point_packed_16bit_C_logical_shift
  • okk_bdc_fixed_point_packed_16bit_mul_8bit
• Fixed Point Unary Functions
  • okk_bdc_fixed_point_packed_16bit_split_high_8bit
  • okk_bdc_fixed_point_packed_16bit_split_low_8bit

Programming

• Programming on Device
  • Kernel Function
  • Register Kernel Function
  • Hello World on Device
  • Example of Using GDMA and BDC
  • GDMA and BDC Run Parallel
  • Divide Tensor Into Parts
  • Reshape
• Programming on Host
  • Launch Kernel Function
  • Hello World on Host

Demo

• Introduction
• Compilation Tools
• Header Files and Libraries
• Develop
• Compile and Update Firmware
• Execute