And given the required SIMD to scalar code translation I guess the Peg's 600 MHz G3 could also have problems to do this.
On the other hand, Gekko is a G3 running at 486 MHz, so the Efika 5200B with its 400 MHz G2 surely wouldn't be able to emulate/virtualize a Gamecube at full speed anyway (that's even before taking the heavy RAM constraints into account). Yes, that's why earlier in this thread I mentioned the possibility of translating the SIMD code into scalar code with the drawback of even slower execution on slower CPUs such as G3. > eFika users from using the program, since both lack Altivec.
> of course, it would eliminate the possibility of G3 Peg and But then, I'm not a coder and even less a SIMD expert, so there might indeed be people who'd do that as a warm-up exercise -) And I don't think it was "simple" to map Gekko's/Broadway's SIMD to SSE in Dolphin. > the SIMD instructions into Altivec instructions. > since Dolphin already does that it may be just as simple as adapting If my suspicion is correct we wouldn't gain much from looking at this part of Dolphin's source code. I didn't have a look at it but I suspect that there's a JIT in place that translates scalar PPC code directly into (scalar) x86 code and Gekko SIMD code directly into SSE code. > Looking at Dolphin's source would be a good start. That means the SIMD instructions would have to be emulated by translating them into AltiVec code (or scalar code which would be even slower of course).ītw, if anybody has a reliable source (IBM, Nintendo or game developers) showing Gekko's SIMD as being a subset of AltiVec/VMX I'd be glad to be corrected :-) But I don't think the G4's AltiVec unit is able to directly execute Gekko's SIMD instructions as that would mean Gekko's SIMD is a subset of AltiVec, which I strongly doubt. > maybe the G4's altivec could pick up some of the work. So you had a look at Gekko's SIMD instructions?
> The SIMD instructions would not be but so hard all but the SIMD ones) should be directly executable by a physical G4. > Both run in Big Endian, so much of the code may be ableĪctually, all scalar instructions (i.e. Basic/low level mapping shoule be close, if not the same