Optimizing IRQ latency on the STM32H743 @ 480 MHz, perhaps for NES ROM emulation...
-
Optimizing IRQ latency on the STM32H743 @ 480 MHz, perhaps for NES ROM emulation... Best result so far: 100 nanoseconds input-to-output latency when the vector table and the IRQ handler are relocated to Tightly-Coupled Memory without making HAL calls. Not bad, but the GPIO controller (several buses away) looks like the real performance killer here. WARNING: buggy code, see correction https://mk.absturztau.be/notes/ajvb448y305b01i4. #electronics #STM32
-
Optimizing IRQ latency on the STM32H743 @ 480 MHz, perhaps for NES ROM emulation... Best result so far: 100 nanoseconds input-to-output latency when the vector table and the IRQ handler are relocated to Tightly-Coupled Memory without making HAL calls. Not bad, but the GPIO controller (several buses away) looks like the real performance killer here. WARNING: buggy code, see correction https://mk.absturztau.be/notes/ajvb448y305b01i4. #electronics #STM32
Keep optimizing IRQ latency on the STM32H743 @ 480 MHz. Just enabled i-cache and d-cache, and the IRQ latency dropped from 100 ns to 70 ns.
But cache shouldn't work like this. So my code is still touching slow memory somewhere. The stack perhaps, which is still in "normal" RAM. The slow Flash perhaps also makes it slower to abort main()if an instruction is stuck in a wait state. Need to check everything carefully... #electronics #STM32 -
Keep optimizing IRQ latency on the STM32H743 @ 480 MHz. Just enabled i-cache and d-cache, and the IRQ latency dropped from 100 ns to 70 ns.
But cache shouldn't work like this. So my code is still touching slow memory somewhere. The stack perhaps, which is still in "normal" RAM. The slow Flash perhaps also makes it slower to abort main()if an instruction is stuck in a wait state. Need to check everything carefully... #electronics #STM32Keep optimizing IRQ latency on the STM32H743 @ 480 MHz. The 70 ns vs. 100 ns overhead mystery solved. I did not correctly relocate the vector table to Tightly-Coupled Memory properly, it was still in Flash. The STM32 HAL macro
USER_VECT_TAB_ADDRESSis a flag, not a memory address! In fact, only several hardcoded addresses are available, a real user override is not provided (the name "user" is a lie). Solution: just change VTOR manually, don't trust the startup code. I'm now getting 70-ns IRQ without CPU cache. #electronics #STM32 -
Keep optimizing IRQ latency on the STM32H743 @ 480 MHz. The 70 ns vs. 100 ns overhead mystery solved. I did not correctly relocate the vector table to Tightly-Coupled Memory properly, it was still in Flash. The STM32 HAL macro
USER_VECT_TAB_ADDRESSis a flag, not a memory address! In fact, only several hardcoded addresses are available, a real user override is not provided (the name "user" is a lie). Solution: just change VTOR manually, don't trust the startup code. I'm now getting 70-ns IRQ without CPU cache. #electronics #STM32 -
I do not understand how the NES system bus works, even after reading multiple tutorials. Only one way to find out... #electronics #NES #NESdev
Keep optimizing IRQ latency on the STM32H743 @ 480 MHz. I decided to try an event loop using the
WFEinstruction instead of IRQs, and I managed to get 60 ns input-to-output latency. I suspect this is the best possible latency. Latency did not improve by abusing QSPI controller to generate a write request (in fact it slightly degraded), even if the QSPI controller is physically close to the CPU. Clearly, passively monitoring signals is not the way to go for bus emulation. Perhaps the solution is predicting the clock before it even arrives, by internally generating a phase-shifted version of it. #electronics #STM32 -
Keep optimizing IRQ latency on the STM32H743 @ 480 MHz. I decided to try an event loop using the
WFEinstruction instead of IRQs, and I managed to get 60 ns input-to-output latency. I suspect this is the best possible latency. Latency did not improve by abusing QSPI controller to generate a write request (in fact it slightly degraded), even if the QSPI controller is physically close to the CPU. Clearly, passively monitoring signals is not the way to go for bus emulation. Perhaps the solution is predicting the clock before it even arrives, by internally generating a phase-shifted version of it. #electronics #STM32Keep optimizing IRQ latency on the STM32H743 @ 480 MHz. My "zero-latency IRQ" idea is a success, now I'm getting a 17.30 ns "effective" latency! Upon receiving every rising edge of the clock, the hardware immediately starts a timer that fires after a programmed delay, calculated to be slightly before the next clock rising edge. This way, the firmware is triggered from recovered, phase-shifted version of the clock, a little bit like how analog NTSC TVs got their H/VSYNC. Interrupt latency is completely eliminated for all but the first clock cycle (which is also predictable with pre-enabled outputs, since it's always the reset vector) Perfect bus emulation starts looking feasible. #electronics #STM32
-
Keep optimizing IRQ latency on the STM32H743 @ 480 MHz. My "zero-latency IRQ" idea is a success, now I'm getting a 17.30 ns "effective" latency! Upon receiving every rising edge of the clock, the hardware immediately starts a timer that fires after a programmed delay, calculated to be slightly before the next clock rising edge. This way, the firmware is triggered from recovered, phase-shifted version of the clock, a little bit like how analog NTSC TVs got their H/VSYNC. Interrupt latency is completely eliminated for all but the first clock cycle (which is also predictable with pre-enabled outputs, since it's always the reset vector) Perfect bus emulation starts looking feasible. #electronics #STM32
-
M monkee@other.li shared this topic