VFPU: problem with vtfm3.t

[davidgf]

Hello I have a problem. I'm implementing some functions in VFPU hardware but this function:

Code: Select all

PURE_INLINE void dMULTIPLY0_331_ (dReal* vout, const dReal* mat, const dReal* v) {
	    __asm__ volatile (
	       "ulv.q    R000, 0+  %1\n"

	       "ulv.q    R001, 16+ %1\n"
	       "ulv.q    R002, 32+ %1\n"
	       "ulv.q    R100, %2\n"

	       "vtfm3.t  R003, M000, R100\n"

	       "usv.q    R003, %0\n"
	       : "+m"(vout[0]): "m"(mat[0]), "m"(v[0]));
}

does not work, but this one (which doesn't use vtfm3.t) works like a charm:

Code: Select all

PURE_INLINE void dMULTIPLY0_331_ (dReal* vout, const dReal* mat, const dReal* v) {
	    __asm__ volatile (
	       "ulv.q    R000, 0+  %1\n"

	       "ulv.q    R001, 16+ %1\n"
	       "ulv.q    R002, 32+ %1\n"
	       "ulv.q    R100, %2\n"

	       "vdot.t   S003, R000, R100\n"
	       "vdot.t   S013, R001, R100\n"
	       "vdot.t   S023, R002, R100\n"

	       "usv.q    R003, %0\n"
	       : "+m"(vout[0]): "m"(mat[0]), "m"(v[0]));
}

[jsharrad]

I read this a while back...
PLEASE READ THIS: There are 2 instructions, ulv.q and usv.q, that perform unaligned ran transfers to/from the vfpu. These have been found to be faulty so it is not recommended to use them. You have been warned.

[davidgf]

I think this has nothing to do with the error.
The error is within the tfm and dot functions. And it's almost impossible to avoid using ulv.q function, because aligning data in the stack is quite complicated!

[hlide]

on PSP1000, using ULV wiill scratch FPU register with the same index so if you are using VFPU register $C000 to load it from memory, you're modifying $fpr0 with the same value as $S000 as well. That's why you MUST never use ULV if you plan to keep compatibily with psp1000. The reason is ULV is mapped to MIPS64 LDC1 (FPU double load) and sony engineers probably forgot to cut off the output link to FPU registers.

[hlide]

can you be more precise about what you get with vtfm3.t ?