Simulation Event Trace - 2022.2 English

AI Engine Tools and Flows User Guide (UG1076)

Document ID
UG1076
Release Date
2022-10-19
Version
2022.2 English

An event in the AI Engine simulator gives information about a system at a specific instant in time. An event that is associated with a time stamp, type and set of data values, is referred to as a payload. The interpretation of the payload depends on the type of the event. A time stamp allows ordering of events, computation of causality relationships, and implementation of verifiers on a sequence of events.

For event modeling purposes, the key modules of the AI Engine array are processor cores, DMA, lock modules, memory and I/O streams. Each module can be viewed as an event generator/responder. Each module receives an event and responds to the event. As a response, new events could be generated. Events are classified based on the generator of the events. The time stamp is not mentioned, explicitly, in the event definitions. Each event is described with a payload, which is a couple of values associated with the event. Each AI Engine, memory, DMA, or lock is addressable by a 2-dimensional index <col, row>, which is the column and row index in the AI Engine array. Some of the AI Engine events are shown in the following table. All events have a time stamp, which is omitted from the table.

Table 1. AI Engine Events
Event Name Values Description
CORE_RESET <col, row> Core <col, row> is out of reset and starts executing code from PC0.
CORE_WAIT <col, row> Core <col, row> is waiting for lock, stream input, or memory response.
CORE_READY <col, row> Core <col, row> is ready after a stall is serviced.

ACQ_READ_LOCK_REQ

ACQ_WRITE_LOCK_REQ

<col, row, lockid, dir, pc, info> Core <col, row> requests for a lock to one of its neighboring memory blocks. The <pc> is the current program counter. The <info> is the symbolic information.
REL_LOCK_REQ <col, row, lockid, dir> Core <col, row> releases a lock in a neighboring lock module.
CORE_READ_REQ <col, row, address, dir, port, bank, data, pcinfo, addrinfo> Core <col, row> requests a read at a memory address. The address is a 15-bit address local to the memory. Together, the <dir, address> forms a 17-bit address. The <dir> field specifies [E, W, N, S] direction. The <port> specifies the load port (A,B), and <bank> specifies access to an odd or even bank. Symbolic information about program variables and the <pc> value is available in the event.
CORE_WRITE_REQ <col, row, address, dir, bank, data, size.pcinfo, addrinfo> Core <col, row> writes data at a particular 15-bit address. The <dir> specifies the direction of the memory being accessed. Symbolic information about the <pc> value and data symbol are available in the event.
PC_CHANGE <col, row, pc, info> Core <col, row> shows the change in the <pc> value. It is useful for identifying loops.
DMA_S2MM_ACQ_LOCK <col, row, ch> DMA requests a lock. The <ch> field represents the DMA channel number. The supported channels are 0 and 1.
DMA_S2MM_IDLE <col, row, bd, ch> DMA in an idle state. <bd> is the buffer descriptor.
DMA_S2MM_START <col, row, bd, ch, start> Start DMA transfer at the start address
DMA_S2MM_DONE <col, row, bd, ch> End DMA transfer at the end address
DMA_S2MM_LOCKSTALL <col, row, bd, ch> DMA is stalled to get a lock
DMA_S2MM_LOCKSTALL_RELEASE <col, row, bd, ch> DMA gets a lock and stall is released
DMA_MM2S_ACQ_LOCK <col, row, ch> DMA <col, row> requests a lock
DMA_MM2S_IDLE <col, row, ch> DMA <col, row> in an idle state
DMA_MM2S_START <col, row, bd, ch, start> Start DMA transfer at <col, row> at the start address, the <bd> is the buffer descriptor.
DMA_MM2S_DONE <col, row, bd, ch> End DMA transfer at <col, row> at the end address, the <bd> is the buffer descriptor.
DMA_MM2S_LOCKSTALL <col, row, bd, ch> DMA is stalled to get a lock
DMA_MM2S_LOCKSTALL_RELEASE <col, row, bd, ch> DMA gets a lock and stall is released
IO <dir, id, data > An I/O event represents new data on an input stream. The <dir> [E, W, N, S] and <id> [0–15] together represent a physical stream, the data field represents 32-bit data appearing on the stream.
DATA_HEAD <col, row, name, netid, pktid, idx> Core <col, row> transfers packet switched data header with routing <name>, on net<netid>. This packet has packet ID <pktid>. <idx> is reserved.
DATA_START <col, row, name,netid, pktid, idx> Data payload (either packet switched or circuit switched) start. All fields follow the meaning of DATA_HEAD fields.
DATA_RESUME <col, row, name,netid, pktid, idx> Data payload resumed after being stalled. All fields follow the meaning of DATA_HEAD fields.
DATA_STALL <col, row, name, netid, idx> Data payload stalled within packet payload. All fields follow the meaning of DATA_HEAD fields.
DATA_END <col, row, name, netid, idx> Data payload end (either packet switched or circuit switched). All fields follow the meaning of DATA_HEAD fields.

PL_TO_SHIM, SHIM_TO_PL

<name, col, channelId, data0, data1, tlast> Data movement between the PL and AI Engine array interface (SHIM). Name shows the PL block name.
PL2PL, PL2PL_E <name, port, data0, data1, tlast, tkeep> PL port activity start and end. Name and port show PL block name/port.
DM_READ_REQ <col, row, portname> A read request event on a port of a memory module. A read port can be a core, DMA, or memory-mapped AXI4 interface read request port.
DM_WRITE_REQ <col, row, portname> A write request event on a port of a memory module. A write port can be a core, DMA, or memory-mapped AXI4 interface write request port.
DM_BANK_CONFLICT <col, row, bankid, bank> A bank conflict event shows a bank conflict because multiple requests are sent to the same bank of a memory module.