bsnes/higan/sfc/smp/timing.cpp

69 lines
2.2 KiB
C++
Raw Normal View History

Update to v103r06 release. byuu says: Changelog: - processor/spc700: restored fetch/load/store/pull/push shorthand functions - processor/spc700: split functions that tested the algorithm used (`op != &SPC700:...`) to separate instructions - mostly for code clarity over code size: it was awkward having cycle counts change based on a function parameter - processor/spc700: implemented Overload's new findings on which cycles are truly internal (no bus reads) - sfc/smp: TEST register emulation has been vastly improved¹ ¹: it turns out that TEST.d4,d5 is the external clock divider (used when accessing RAM through the DSP), and TEST.d6,d7 is the internal clock divider (used when accessing IPLROM, IO registers, or during idle cycles.) The DSP (24576khz) feeds its clock / 12 through to the SMP (2048khz). The clock divider setting further divides the clock by 2, 4, 8, or 16. Since 8 and 16 are not cleanly divislbe by 12, the SMP cycle count glitches out and seems to take 10 and 2 clocks instead of 8 or 16. This can on real hardware either cause the SMP to run very slowly, or more likely, crash the SMP completely until reset. What's even stranger is the timers aren't affected by this. They still clock by 2, 4, 8, or 16. Note that technically I could divide my own clock counters by 24 and reduce these to {1,2,5,10} and {1,2,4,8}, I instead chose to divide by 12 to better illustrate this hardware issue and better model that the SMP clock runs at 2048khz and not 1024khz. Further, note that things aren't 100% perfect yet. This seems to throw off some tests, such as blargg's `test_timer_speed`. I can't tell how far off I am because blargg's test tragically doesn't print out fail values. But you can see the improvements in that higan is now passing all of Revenant's tests that were obviously completely wrong before.
2017-07-03 07:24:47 +00:00
//DSP clock (~24576khz) / 12 (~2048khz) is fed into the SMP
//from here, the wait states value is really a clock divider of {2, 4, 8, 16}
//because dividers of 8 and 16 are not evenly divislbe into 12, the SMP glitches
//in these two cases, the SMP ends up consuming 10 and 20 cycles instead
//this causes unpredictable behavior on real hardware
//sometimes the SMP will run far slower than expected
//other times (and more likely), the SMP will deadlock until the system is reset
//the timers are not affected by this and advance by their expected values
auto SMP::wait(maybe<uint16> addr) -> void {
static const uint cycleWaitStates[4] = {2, 4, 10, 20};
static const uint timerWaitStates[4] = {2, 4, 8, 16};
uint waitStates = io.externalWaitStates;
if(!addr) waitStates = io.internalWaitStates; //idle cycles
else if((*addr & 0xfff0) == 0x00f0) waitStates = io.internalWaitStates; //IO registers
else if(*addr >= 0xffc0 && io.iplromEnable) waitStates = io.internalWaitStates; //IPLROM
step(cycleWaitStates[waitStates]);
stepTimers(timerWaitStates[waitStates]);
}
Update to v099r14 release. byuu says: Changelog: - (u)int(max,ptr) abbreviations removed; use _t suffix now [didn't feel like they were contributing enough to be worth it] - cleaned up nall::integer,natural,real functionality - toInteger, toNatural, toReal for parsing strings to numbers - fromInteger, fromNatural, fromReal for creating strings from numbers - (string,Markup::Node,SQL-based-classes)::(integer,natural,real) left unchanged - template<typename T> numeral(T value, long padding, char padchar) -> string for print() formatting - deduces integer,natural,real based on T ... cast the value if you want to override - there still exists binary,octal,hex,pointer for explicit print() formatting - lstring -> string_vector [but using lstring = string_vector; is declared] - would be nice to remove the using lstring eventually ... but that'd probably require 10,000 lines of changes >_> - format -> string_format [no using here; format was too ambiguous] - using integer = Integer<sizeof(int)*8>; and using natural = Natural<sizeof(uint)*8>; declared - for consistency with boolean. These three are meant for creating zero-initialized values implicitly (various uses) - R65816::io() -> idle() and SPC700::io() -> idle() [more clear; frees up struct IO {} io; naming] - SFC CPU, PPU, SMP use struct IO {} io; over struct (Status,Registers) {} (status,registers); now - still some CPU::Status status values ... they didn't really fit into IO functionality ... will have to think about this more - SFC CPU, PPU, SMP now use step() exclusively instead of addClocks() calling into step() - SFC CPU joypad1_bits, joypad2_bits were unused; killed them - SFC PPU CGRAM moved into PPU::Screen; since nothing else uses it - SFC PPU OAM moved into PPU::Object; since nothing else uses it - the raw uint8[544] array is gone. OAM::read() constructs values from the OAM::Object[512] table now - this avoids having to determine how we want to sub-divide the two OAM memory sections - this also eliminates the OAM::synchronize() functionality - probably more I'm forgetting The FPS fluctuations are driving me insane. This WIP went from 128fps to 137fps. Settled on 133.5fps for the final build. But nothing I changed should have affected performance at all. This level of fluctuation makes it damn near impossible to know whether I'm speeding things up or slowing things down with changes.
2016-07-01 11:50:32 +00:00
auto SMP::step(uint clocks) -> void {
Update to v100r14 release. byuu says: (Windows: compile with -fpermissive to silence an annoying error. I'll fix it in the next WIP.) I completely replaced the time management system in higan and overhauled the scheduler. Before, processor threads would have "int64 clock"; and there would be a 1:1 relationship between two threads. When thread A ran for X cycles, it'd subtract X * B.Frequency from clock; and when thread B ran for Y cycles, it'd add Y * A.Frequency from clock. This worked well and allowed perfect precision; but it doesn't work when you have more complicated relationships: eg the 68K can sync to the Z80 and PSG; the Z80 to the 68K and PSG; so the PSG needs two counters. The new system instead uses a "uint64 clock" variable that represents time in attoseconds. Every time the scheduler exits, it subtracts the smallest clock count from all threads, to prevent an overflow scenario. The only real downside is that rounding errors mean that roughly every 20 minutes, we have a rounding error of one clock cycle (one 20,000,000th of a second.) However, this only applies to systems with multiple oscillators, like the SNES. And when you're in that situation ... there's no such thing as a perfect oscillator anyway. A real SNES will be thousands of times less out of spec than 1hz per 20 minutes. The advantages are pretty immense. First, we obviously can now support more complex relationships between threads. Second, we can build a much more abstracted scheduler. All of libco is now abstracted away completely, which may permit a state-machine / coroutine version of Thread in the future. We've basically gone from this: auto SMP::step(uint clocks) -> void { clock += clocks * (uint64)cpu.frequency; dsp.clock -= clocks; if(dsp.clock < 0 && !scheduler.synchronizing()) co_switch(dsp.thread); if(clock >= 0 && !scheduler.synchronizing()) co_switch(cpu.thread); } To this: auto SMP::step(uint clocks) -> void { Thread::step(clocks); synchronize(dsp); synchronize(cpu); } As you can see, we don't have to do multiple clock adjustments anymore. This is a huge win for the SNES CPU that had to update the SMP, DSP, all peripherals and all coprocessors. Likewise, we don't have to synchronize all coprocessors when one runs, now we can just synchronize the active one to the CPU. Third, when changing the frequencies of threads (think SGB speed setting modes, GBC double-speed mode, etc), it no longer causes the "int64 clock" value to be erroneous. Fourth, this results in a fairly decent speedup, mostly across the board. Aside from the GBA being mostly a wash (for unknown reasons), it's about an 8% - 12% speedup in every other emulation core. Now, all of this said ... this was an unbelievably massive change, so ... you know what that means >_> If anyone can help test all types of SNES coprocessors, and some other system games, it'd be appreciated. ---- Lastly, we have a bitchin' new about screen. It unfortunately adds ~200KiB onto the binary size, because the PNG->C++ header file transformation doesn't compress very well, and I want to keep the original resource files in with the higan archive. I might try some things to work around this file size increase in the future, but for now ... yeah, slightly larger archive sizes, sorry. The logo's a bit busted on Windows (the Label control's background transparency and alignment settings aren't working), but works well on GTK. I'll have to fix Windows before the next official release. For now, look on my Twitter feed if you want to see what it's supposed to look like. ---- EDIT: forgot about ICD2::Enter. It's doing some weird inverse run-to-save thing that I need to implement support for somehow. So, save states on the SGB core probably won't work with this WIP.
2016-07-30 03:56:12 +00:00
Thread::step(clocks);
synchronize(dsp);
#if defined(DEBUGGER)
Update to v100r14 release. byuu says: (Windows: compile with -fpermissive to silence an annoying error. I'll fix it in the next WIP.) I completely replaced the time management system in higan and overhauled the scheduler. Before, processor threads would have "int64 clock"; and there would be a 1:1 relationship between two threads. When thread A ran for X cycles, it'd subtract X * B.Frequency from clock; and when thread B ran for Y cycles, it'd add Y * A.Frequency from clock. This worked well and allowed perfect precision; but it doesn't work when you have more complicated relationships: eg the 68K can sync to the Z80 and PSG; the Z80 to the 68K and PSG; so the PSG needs two counters. The new system instead uses a "uint64 clock" variable that represents time in attoseconds. Every time the scheduler exits, it subtracts the smallest clock count from all threads, to prevent an overflow scenario. The only real downside is that rounding errors mean that roughly every 20 minutes, we have a rounding error of one clock cycle (one 20,000,000th of a second.) However, this only applies to systems with multiple oscillators, like the SNES. And when you're in that situation ... there's no such thing as a perfect oscillator anyway. A real SNES will be thousands of times less out of spec than 1hz per 20 minutes. The advantages are pretty immense. First, we obviously can now support more complex relationships between threads. Second, we can build a much more abstracted scheduler. All of libco is now abstracted away completely, which may permit a state-machine / coroutine version of Thread in the future. We've basically gone from this: auto SMP::step(uint clocks) -> void { clock += clocks * (uint64)cpu.frequency; dsp.clock -= clocks; if(dsp.clock < 0 && !scheduler.synchronizing()) co_switch(dsp.thread); if(clock >= 0 && !scheduler.synchronizing()) co_switch(cpu.thread); } To this: auto SMP::step(uint clocks) -> void { Thread::step(clocks); synchronize(dsp); synchronize(cpu); } As you can see, we don't have to do multiple clock adjustments anymore. This is a huge win for the SNES CPU that had to update the SMP, DSP, all peripherals and all coprocessors. Likewise, we don't have to synchronize all coprocessors when one runs, now we can just synchronize the active one to the CPU. Third, when changing the frequencies of threads (think SGB speed setting modes, GBC double-speed mode, etc), it no longer causes the "int64 clock" value to be erroneous. Fourth, this results in a fairly decent speedup, mostly across the board. Aside from the GBA being mostly a wash (for unknown reasons), it's about an 8% - 12% speedup in every other emulation core. Now, all of this said ... this was an unbelievably massive change, so ... you know what that means >_> If anyone can help test all types of SNES coprocessors, and some other system games, it'd be appreciated. ---- Lastly, we have a bitchin' new about screen. It unfortunately adds ~200KiB onto the binary size, because the PNG->C++ header file transformation doesn't compress very well, and I want to keep the original resource files in with the higan archive. I might try some things to work around this file size increase in the future, but for now ... yeah, slightly larger archive sizes, sorry. The logo's a bit busted on Windows (the Label control's background transparency and alignment settings aren't working), but works well on GTK. I'll have to fix Windows before the next official release. For now, look on my Twitter feed if you want to see what it's supposed to look like. ---- EDIT: forgot about ICD2::Enter. It's doing some weird inverse run-to-save thing that I need to implement support for somehow. So, save states on the SGB core probably won't work with this WIP.
2016-07-30 03:56:12 +00:00
synchronize(cpu);
#else
//forcefully sync S-SMP to S-CPU in case chips are not communicating
//sync if S-SMP is more than 1ms ahead of S-CPU
if(clock() - cpu.clock() > Thread::Second / 1'000) synchronize(cpu);
#endif
}
Update to v103r06 release. byuu says: Changelog: - processor/spc700: restored fetch/load/store/pull/push shorthand functions - processor/spc700: split functions that tested the algorithm used (`op != &SPC700:...`) to separate instructions - mostly for code clarity over code size: it was awkward having cycle counts change based on a function parameter - processor/spc700: implemented Overload's new findings on which cycles are truly internal (no bus reads) - sfc/smp: TEST register emulation has been vastly improved¹ ¹: it turns out that TEST.d4,d5 is the external clock divider (used when accessing RAM through the DSP), and TEST.d6,d7 is the internal clock divider (used when accessing IPLROM, IO registers, or during idle cycles.) The DSP (24576khz) feeds its clock / 12 through to the SMP (2048khz). The clock divider setting further divides the clock by 2, 4, 8, or 16. Since 8 and 16 are not cleanly divislbe by 12, the SMP cycle count glitches out and seems to take 10 and 2 clocks instead of 8 or 16. This can on real hardware either cause the SMP to run very slowly, or more likely, crash the SMP completely until reset. What's even stranger is the timers aren't affected by this. They still clock by 2, 4, 8, or 16. Note that technically I could divide my own clock counters by 24 and reduce these to {1,2,5,10} and {1,2,4,8}, I instead chose to divide by 12 to better illustrate this hardware issue and better model that the SMP clock runs at 2048khz and not 1024khz. Further, note that things aren't 100% perfect yet. This seems to throw off some tests, such as blargg's `test_timer_speed`. I can't tell how far off I am because blargg's test tragically doesn't print out fail values. But you can see the improvements in that higan is now passing all of Revenant's tests that were obviously completely wrong before.
2017-07-03 07:24:47 +00:00
auto SMP::stepTimers(uint clocks) -> void {
timer0.step(clocks);
timer1.step(clocks);
timer2.step(clocks);
}
Update to v103r06 release. byuu says: Changelog: - processor/spc700: restored fetch/load/store/pull/push shorthand functions - processor/spc700: split functions that tested the algorithm used (`op != &SPC700:...`) to separate instructions - mostly for code clarity over code size: it was awkward having cycle counts change based on a function parameter - processor/spc700: implemented Overload's new findings on which cycles are truly internal (no bus reads) - sfc/smp: TEST register emulation has been vastly improved¹ ¹: it turns out that TEST.d4,d5 is the external clock divider (used when accessing RAM through the DSP), and TEST.d6,d7 is the internal clock divider (used when accessing IPLROM, IO registers, or during idle cycles.) The DSP (24576khz) feeds its clock / 12 through to the SMP (2048khz). The clock divider setting further divides the clock by 2, 4, 8, or 16. Since 8 and 16 are not cleanly divislbe by 12, the SMP cycle count glitches out and seems to take 10 and 2 clocks instead of 8 or 16. This can on real hardware either cause the SMP to run very slowly, or more likely, crash the SMP completely until reset. What's even stranger is the timers aren't affected by this. They still clock by 2, 4, 8, or 16. Note that technically I could divide my own clock counters by 24 and reduce these to {1,2,5,10} and {1,2,4,8}, I instead chose to divide by 12 to better illustrate this hardware issue and better model that the SMP clock runs at 2048khz and not 1024khz. Further, note that things aren't 100% perfect yet. This seems to throw off some tests, such as blargg's `test_timer_speed`. I can't tell how far off I am because blargg's test tragically doesn't print out fail values. But you can see the improvements in that higan is now passing all of Revenant's tests that were obviously completely wrong before.
2017-07-03 07:24:47 +00:00
template<uint Frequency> auto SMP::Timer<Frequency>::step(uint clocks) -> void {
//stage 0 increment
Update to v103r06 release. byuu says: Changelog: - processor/spc700: restored fetch/load/store/pull/push shorthand functions - processor/spc700: split functions that tested the algorithm used (`op != &SPC700:...`) to separate instructions - mostly for code clarity over code size: it was awkward having cycle counts change based on a function parameter - processor/spc700: implemented Overload's new findings on which cycles are truly internal (no bus reads) - sfc/smp: TEST register emulation has been vastly improved¹ ¹: it turns out that TEST.d4,d5 is the external clock divider (used when accessing RAM through the DSP), and TEST.d6,d7 is the internal clock divider (used when accessing IPLROM, IO registers, or during idle cycles.) The DSP (24576khz) feeds its clock / 12 through to the SMP (2048khz). The clock divider setting further divides the clock by 2, 4, 8, or 16. Since 8 and 16 are not cleanly divislbe by 12, the SMP cycle count glitches out and seems to take 10 and 2 clocks instead of 8 or 16. This can on real hardware either cause the SMP to run very slowly, or more likely, crash the SMP completely until reset. What's even stranger is the timers aren't affected by this. They still clock by 2, 4, 8, or 16. Note that technically I could divide my own clock counters by 24 and reduce these to {1,2,5,10} and {1,2,4,8}, I instead chose to divide by 12 to better illustrate this hardware issue and better model that the SMP clock runs at 2048khz and not 1024khz. Further, note that things aren't 100% perfect yet. This seems to throw off some tests, such as blargg's `test_timer_speed`. I can't tell how far off I am because blargg's test tragically doesn't print out fail values. But you can see the improvements in that higan is now passing all of Revenant's tests that were obviously completely wrong before.
2017-07-03 07:24:47 +00:00
stage0 += clocks;
if(stage0 < Frequency) return;
stage0 -= Frequency;
//stage 1 increment
stage1 ^= 1;
synchronizeStage1();
}
Update to v099r14 release. byuu says: Changelog: - (u)int(max,ptr) abbreviations removed; use _t suffix now [didn't feel like they were contributing enough to be worth it] - cleaned up nall::integer,natural,real functionality - toInteger, toNatural, toReal for parsing strings to numbers - fromInteger, fromNatural, fromReal for creating strings from numbers - (string,Markup::Node,SQL-based-classes)::(integer,natural,real) left unchanged - template<typename T> numeral(T value, long padding, char padchar) -> string for print() formatting - deduces integer,natural,real based on T ... cast the value if you want to override - there still exists binary,octal,hex,pointer for explicit print() formatting - lstring -> string_vector [but using lstring = string_vector; is declared] - would be nice to remove the using lstring eventually ... but that'd probably require 10,000 lines of changes >_> - format -> string_format [no using here; format was too ambiguous] - using integer = Integer<sizeof(int)*8>; and using natural = Natural<sizeof(uint)*8>; declared - for consistency with boolean. These three are meant for creating zero-initialized values implicitly (various uses) - R65816::io() -> idle() and SPC700::io() -> idle() [more clear; frees up struct IO {} io; naming] - SFC CPU, PPU, SMP use struct IO {} io; over struct (Status,Registers) {} (status,registers); now - still some CPU::Status status values ... they didn't really fit into IO functionality ... will have to think about this more - SFC CPU, PPU, SMP now use step() exclusively instead of addClocks() calling into step() - SFC CPU joypad1_bits, joypad2_bits were unused; killed them - SFC PPU CGRAM moved into PPU::Screen; since nothing else uses it - SFC PPU OAM moved into PPU::Object; since nothing else uses it - the raw uint8[544] array is gone. OAM::read() constructs values from the OAM::Object[512] table now - this avoids having to determine how we want to sub-divide the two OAM memory sections - this also eliminates the OAM::synchronize() functionality - probably more I'm forgetting The FPS fluctuations are driving me insane. This WIP went from 128fps to 137fps. Settled on 133.5fps for the final build. But nothing I changed should have affected performance at all. This level of fluctuation makes it damn near impossible to know whether I'm speeding things up or slowing things down with changes.
2016-07-01 11:50:32 +00:00
template<uint Frequency> auto SMP::Timer<Frequency>::synchronizeStage1() -> void {
bool newLine = stage1;
Update to v099r14 release. byuu says: Changelog: - (u)int(max,ptr) abbreviations removed; use _t suffix now [didn't feel like they were contributing enough to be worth it] - cleaned up nall::integer,natural,real functionality - toInteger, toNatural, toReal for parsing strings to numbers - fromInteger, fromNatural, fromReal for creating strings from numbers - (string,Markup::Node,SQL-based-classes)::(integer,natural,real) left unchanged - template<typename T> numeral(T value, long padding, char padchar) -> string for print() formatting - deduces integer,natural,real based on T ... cast the value if you want to override - there still exists binary,octal,hex,pointer for explicit print() formatting - lstring -> string_vector [but using lstring = string_vector; is declared] - would be nice to remove the using lstring eventually ... but that'd probably require 10,000 lines of changes >_> - format -> string_format [no using here; format was too ambiguous] - using integer = Integer<sizeof(int)*8>; and using natural = Natural<sizeof(uint)*8>; declared - for consistency with boolean. These three are meant for creating zero-initialized values implicitly (various uses) - R65816::io() -> idle() and SPC700::io() -> idle() [more clear; frees up struct IO {} io; naming] - SFC CPU, PPU, SMP use struct IO {} io; over struct (Status,Registers) {} (status,registers); now - still some CPU::Status status values ... they didn't really fit into IO functionality ... will have to think about this more - SFC CPU, PPU, SMP now use step() exclusively instead of addClocks() calling into step() - SFC CPU joypad1_bits, joypad2_bits were unused; killed them - SFC PPU CGRAM moved into PPU::Screen; since nothing else uses it - SFC PPU OAM moved into PPU::Object; since nothing else uses it - the raw uint8[544] array is gone. OAM::read() constructs values from the OAM::Object[512] table now - this avoids having to determine how we want to sub-divide the two OAM memory sections - this also eliminates the OAM::synchronize() functionality - probably more I'm forgetting The FPS fluctuations are driving me insane. This WIP went from 128fps to 137fps. Settled on 133.5fps for the final build. But nothing I changed should have affected performance at all. This level of fluctuation makes it damn near impossible to know whether I'm speeding things up or slowing things down with changes.
2016-07-01 11:50:32 +00:00
if(!smp.io.timersEnable) newLine = false;
if(smp.io.timersDisable) newLine = false;
bool oldLine = line;
line = newLine;
if(oldLine != 1 || newLine != 0) return; //only pulse on 1->0 transition
//stage 2 increment
Update to v099r14 release. byuu says: Changelog: - (u)int(max,ptr) abbreviations removed; use _t suffix now [didn't feel like they were contributing enough to be worth it] - cleaned up nall::integer,natural,real functionality - toInteger, toNatural, toReal for parsing strings to numbers - fromInteger, fromNatural, fromReal for creating strings from numbers - (string,Markup::Node,SQL-based-classes)::(integer,natural,real) left unchanged - template<typename T> numeral(T value, long padding, char padchar) -> string for print() formatting - deduces integer,natural,real based on T ... cast the value if you want to override - there still exists binary,octal,hex,pointer for explicit print() formatting - lstring -> string_vector [but using lstring = string_vector; is declared] - would be nice to remove the using lstring eventually ... but that'd probably require 10,000 lines of changes >_> - format -> string_format [no using here; format was too ambiguous] - using integer = Integer<sizeof(int)*8>; and using natural = Natural<sizeof(uint)*8>; declared - for consistency with boolean. These three are meant for creating zero-initialized values implicitly (various uses) - R65816::io() -> idle() and SPC700::io() -> idle() [more clear; frees up struct IO {} io; naming] - SFC CPU, PPU, SMP use struct IO {} io; over struct (Status,Registers) {} (status,registers); now - still some CPU::Status status values ... they didn't really fit into IO functionality ... will have to think about this more - SFC CPU, PPU, SMP now use step() exclusively instead of addClocks() calling into step() - SFC CPU joypad1_bits, joypad2_bits were unused; killed them - SFC PPU CGRAM moved into PPU::Screen; since nothing else uses it - SFC PPU OAM moved into PPU::Object; since nothing else uses it - the raw uint8[544] array is gone. OAM::read() constructs values from the OAM::Object[512] table now - this avoids having to determine how we want to sub-divide the two OAM memory sections - this also eliminates the OAM::synchronize() functionality - probably more I'm forgetting The FPS fluctuations are driving me insane. This WIP went from 128fps to 137fps. Settled on 133.5fps for the final build. But nothing I changed should have affected performance at all. This level of fluctuation makes it damn near impossible to know whether I'm speeding things up or slowing things down with changes.
2016-07-01 11:50:32 +00:00
if(!enable) return;
if(++stage2 != target) return;
//stage 3 increment
stage2 = 0;
stage3++;
}