compute_kernel_hw_startup

template<SrcOrder src_order = SrcOrder::Regular>
void ckernel::compute_kernel_hw_startup(uint32_t icb0, uint32_t icb1, uint32_t ocb)

Performs the required hardware initialization for all subsequent operations in the compute kernel. This function should be called exactly once at the very beginning of the kernel, before any operation-specific initialization functions (such as reduce_init, tilize_init, etc.). The circular buffer (CB) IDs provided to this function must match those used in the next operation-specific initialization function. If the operands for the next operation require a different data format than what was configured here, you must call one of the reconfig_data_format functions before proceeding with the next initialization. Similarly, if the next operation requires different properties (such as tile or face dimensions), you must ensure that the same CB IDs are used as in this function.

The src_order template parameter selects how (icb0, icb1) map onto SrcA/SrcB; this is the single piece of operation-specific knowledge startup needs, because the per-source-register state it programs (formats, tile/face dimensions, tile sizes) depends on that mapping. Use SrcOrder::Regular for all operations except matmul, which must use SrcOrder::Reverse (see the SrcOrder documentation). The (icb0, icb1) arguments are always passed in natural operand order (in0, in1) regardless of the tag.

NOTE: This function performs MMIO writes, which are slow and almost exclusively require the idle state of the execution units that should be configured (PACK, MATH, UNPACK, CFG, etc.). This is why it is unsafe to call this function in the middle of a kernel execution. This function should be called only once at the beginning of the kernel, before any other calls to Compute API are made (either init or other). Calling this function after other API calls may lead cause race conditions and undefined behavior which can be hard to debug.

Return value: None

Param Type Name

Template src_order

Function icb0

Function icb1

Function ocb

void ckernel::compute_kernel_hw_startup(uint32_t icb0, uint32_t ocb)

Convenience overload for hardware initialization when only one input circular buffer is used. Both input operands (srcA and srcB) will be programmed using the same circular buffer identifier (icb0). Internally, this calls the three-parameter version with icb0 passed for both input operands.

Param Type Name

Function icb0

Function ocb