New LLE RSP Plugin (accurate, fast but untested?)

[the_randomizer]

Originally Posted by FatCat

To be perfectly honest, I had no idea about that.

I had only just noticed a few days ago when I got desktop access again so that I would try to post a GitHub messaging comment, that it mysteriously voided and never appeared on my activity or anywhere.

Later, I realized that if I'm not actually signed into my GitHub, my entire subdirectory on the site seems to be 404.

Really I have no idea what's going on.
Maybe I should migrate to some other Git?

Either way, I still have the source code posted with the plugin release in this thread, and I can still see my own commits. I have no idea why nobody else can (including me if I'm not logged onto my own account).

Out of curiosity, did you ever get around to renaming the internal file names so they wouldn't show up the same in PJ64?

[HatCat]

#if defined(EXTERN_COMMAND_LIST_GBI) && defined(EXTERN_COMMAND_LIST_ABI)
#define L_NAME "Iconoclast's SP Interpreter (HLE)"
#elif defined(EXTERN_COMMAND_LIST_GBI)
#define L_NAME "Iconoclast's SP Interpreter (MLE)"
#elif defined(EXTERN_COMMAND_LIST_ABI)
#define L_NAME "Iconoclast's SP Interpreter (LLE)"
#elif defined(SEMAPHORE_LOCK_CORRECTIONS) || defined(WAIT_FOR_CPU_HOST)
#define L_NAME "Iconoclast's SP Interpreter (PJ642)"
#elif defined(SP_EXECUTE_LOG) || defined(VU_EMULATE_SCALAR_ACCUMULATOR_READ)
#define L_NAME "Iconoclast's SP Interpreter (debug)"
#else
#define L_NAME "Iconoclast's SP Interpreter"
#endif

@ECHO OFF
TITLE MinGW Compiler Suite Invocation
CD ..\..\MINGW\BIN\

ECHO Building RSP interpreter... (LLE gfx, LLE audio)
gcc --shared -s -O3 -m3dnow -mmmx -msse -msse2 -o ../rsp/rsp.dll ../../rsp/rsp.c

ECHO Building RSP interpreter... (LLE gfx, HLE audio)
gcc --shared -s -O3 -m3dnow -mmmx -msse -msse2 -DEXTERN_COMMAND_LIST_ABI -o ../rsp/rsp_lle.dll ../../rsp/rsp.c

ECHO Building RSP interpreter... (HLE gfx, LLE audio)
gcc --shared -s -O3 -m3dnow -mmmx -msse -msse2 -DEXTERN_COMMAND_LIST_GBI -o ../rsp/rsp_mle.dll ../../rsp/rsp.c

ECHO Building RSP interpreter... (HLE gfx, HLE audio)
gcc --shared -s -O3 -m3dnow -mmmx -msse -msse2 -DEXTERN_COMMAND_LIST_GBI -DEXTERN_COMMAND_LIST_ABI -o ../rsp/rsp_hle.dll ../../rsp/rsp.c

ECHO Building RSP interpreter... (need Project64 2.x)
gcc --shared -s -O3 -m3dnow -mmmx -msse -msse2 -DSEMAPHORE_LOCK_CORRECTIONS -DWAIT_FOR_CPU_HOST -o ../rsp/rsp_pj64.dll ../../rsp/rsp.c

ECHO Building RSP interpreter... (debugging module)
gcc --shared -s -O3 -m3dnow -mmmx -msse -msse2 -DSP_EXECUTE_LOG -DVU_EMULATE_SCALAR_ACCUMULATOR_READ -o ../rsp/rsp_dbg.dll ../../rsp/rsp.c

PAUSE

The other two files (`dummyvid.dll` and `rsp_free.dll`) are unrelated to the GNU compiler toolchain for the RSP plugin. `rsp_free.dll` is the Microsoft Visual Studio 2010 build (slower than the GCC build ), but I might throw in the VS2010 project files for that since the development interface by Microsoft is a bit more comfortable and attractive. I did spent a few hours organizing all the opcode headers for the Microsoft source tree, after all, so why not release that.

Plus, encouraging the activity around inferior compilers is worth it if it means I can get extra feedback.

[the_randomizer]

Originally Posted by FatCat

#if defined(EXTERN_COMMAND_LIST_GBI) && defined(EXTERN_COMMAND_LIST_ABI)
#define L_NAME "Iconoclast's SP Interpreter (HLE)"
#elif defined(EXTERN_COMMAND_LIST_GBI)
#define L_NAME "Iconoclast's SP Interpreter (MLE)"
#elif defined(EXTERN_COMMAND_LIST_ABI)
#define L_NAME "Iconoclast's SP Interpreter (LLE)"
#elif defined(SEMAPHORE_LOCK_CORRECTIONS) || defined(WAIT_FOR_CPU_HOST)
#define L_NAME "Iconoclast's SP Interpreter (PJ642)"
#elif defined(SP_EXECUTE_LOG) || defined(VU_EMULATE_SCALAR_ACCUMULATOR_READ)
#define L_NAME "Iconoclast's SP Interpreter (debug)"
#else
#define L_NAME "Iconoclast's SP Interpreter"
#endif

@ECHO OFF
TITLE MinGW Compiler Suite Invocation
CD ..\..\MINGW\BIN\

ECHO Building RSP interpreter... (LLE gfx, LLE audio)
gcc --shared -s -O3 -m3dnow -mmmx -msse -msse2 -o ../rsp/rsp.dll ../../rsp/rsp.c

ECHO Building RSP interpreter... (LLE gfx, HLE audio)
gcc --shared -s -O3 -m3dnow -mmmx -msse -msse2 -DEXTERN_COMMAND_LIST_ABI -o ../rsp/rsp_lle.dll ../../rsp/rsp.c

ECHO Building RSP interpreter... (HLE gfx, LLE audio)
gcc --shared -s -O3 -m3dnow -mmmx -msse -msse2 -DEXTERN_COMMAND_LIST_GBI -o ../rsp/rsp_mle.dll ../../rsp/rsp.c

ECHO Building RSP interpreter... (HLE gfx, HLE audio)
gcc --shared -s -O3 -m3dnow -mmmx -msse -msse2 -DEXTERN_COMMAND_LIST_GBI -DEXTERN_COMMAND_LIST_ABI -o ../rsp/rsp_hle.dll ../../rsp/rsp.c

ECHO Building RSP interpreter... (need Project64 2.x)
gcc --shared -s -O3 -m3dnow -mmmx -msse -msse2 -DSEMAPHORE_LOCK_CORRECTIONS -DWAIT_FOR_CPU_HOST -o ../rsp/rsp_pj64.dll ../../rsp/rsp.c

ECHO Building RSP interpreter... (debugging module)
gcc --shared -s -O3 -m3dnow -mmmx -msse -msse2 -DSP_EXECUTE_LOG -DVU_EMULATE_SCALAR_ACCUMULATOR_READ -o ../rsp/rsp_dbg.dll ../../rsp/rsp.c

PAUSE

The other two files (`dummyvid.dll` and `rsp_free.dll`) are unrelated to the GNU compiler toolchain for the RSP plugin. `rsp_free.dll` is the Microsoft Visual Studio 2010 build (slower than the GCC build ), but I might throw in the VS2010 project files for that since the development interface by Microsoft is a bit more comfortable and attractive. I did spent a few hours organizing all the opcode headers for the Microsoft source tree, after all, so why not release that.

Plus, encouraging the activity around inferior compilers is worth it if it means I can get extra feedback.

Never mind, I didn't know how busy you were, so you probably didn't see my post. Oh well

[HatCat]

Actually I was replying to your post.

The internal names of the plugins mov'd to the GetDllName buffer when you go to change plugins are listed in the C macro preprocessing template in that first code block I replied to your post with.

So those will be the various names of the different DLL versions.
Any complaints about those TELL ME NAOOOOOOOOOOOOOOOOOOOO *throws a donkey at a mule*

[the_randomizer]

Originally Posted by FatCat

Actually I was replying to your post.

The internal names of the plugins mov'd to the GetDllName buffer when you go to change plugins are listed in the C macro preprocessing template in that first code block I replied to your post with.

So those will be the various names of the different DLL versions.
Any complaints about those TELL ME NAOOOOOOOOOOOOOOOOOOOO *throws a donkey at a mule*

I wasn't sure, but now I know

No complaints from me.

[clobber]

Originally Posted by FatCat

To be perfectly honest, I had no idea about that.

I've heard of this happening to other people on github before where their public-facing account/repo vanished. Could you open a support ticket with them?

[HatCat]

Originally Posted by squall_leonhart

maybe you should raise an issue on the github site support......

Originally Posted by clobber

I've heard of this happening to other people on github before where their public-facing account/repo vanished. Could you open a support ticket with them?

Meh.....maybe

But, you gotta admit.
It's kind of l33t to have an invisible GitHub account.
I'm basically getting private GitHub repositories for free right now.

The plugin is still, after all, open-source, like everything else I have done so far.


It happened while I had no Internet access...
Around April 30th, Sven at Mupen64Plus updated his branch of my RSP plugin.
I suppose that might have been around the time that my GitHub account broke to users who aren't me.

Say I want the GitHub issue to go away.
How should I report it to the GitHub staff / raise a ticket / whatever?

[HatCat]

I don't have much else blog space so am going to keep to this thread.

I gained a 3-6 VI/s speed-up last night by multiplying the DMA transfer into MIPS double-words, instead of 8-bit safe DMA transfers. All the crazy DMA-obsessed games seem to remain stable; this speed-up is safe.

Still, I am not quite satisfied, as somewhere there is bound to be another chance at speeding up the vector instructions generalized emulation, namely the vector decode and management.

method A -- currently used method
Use two-dimensional scalar index look-up table, then split writeback to the accumulator in a separate phase.

#include "vu.h"

static void VMADH(int vd, int vs, int vt, int e)
{
    register signed long long product;
    register int i;

    for (i = 0; i < 8; i++)
    {
        product = VR[vs][i]*VR[vt][ei[e][i]];
        VACC[i].DW += product << 16;
    }
    SIGNED_CLAMP(vd, 0);
    return;
}

method B -- no significant speed change (less than like 0.5 VI/s delta)
Iterate the entire vector-scalar coefficient multiplier using a switch jump table.

#include "vu.h"

static void VMADH(int vd, int vs, int vt, int e)
{
    register signed long long product;
    register int i;

    switch (e)
	{
        case 0x0:
        case 0x1:
            for (i = 0; i < 8; i++)
                VACC[i].DW += (long long)(VR[vs][i]*VR[vt][i]) << 16;
            break;
        case 0x2:
            VACC[00].DW += (long long)(VR[vs][00]*VR[vt][00]) << 16;
            VACC[01].DW += (long long)(VR[vs][01]*VR[vt][00]) << 16;
            VACC[02].DW += (long long)(VR[vs][02]*VR[vt][02]) << 16;
            VACC[03].DW += (long long)(VR[vs][03]*VR[vt][02]) << 16;
            VACC[04].DW += (long long)(VR[vs][04]*VR[vt][04]) << 16;
            VACC[05].DW += (long long)(VR[vs][05]*VR[vt][04]) << 16;
            VACC[06].DW += (long long)(VR[vs][06]*VR[vt][06]) << 16;
            VACC[07].DW += (long long)(VR[vs][07]*VR[vt][06]) << 16;
            break;
        case 0x3:
            VACC[00].DW += (long long)(VR[vs][00]*VR[vt][01]) << 16;
            VACC[01].DW += (long long)(VR[vs][01]*VR[vt][01]) << 16;
            VACC[02].DW += (long long)(VR[vs][02]*VR[vt][03]) << 16;
            VACC[03].DW += (long long)(VR[vs][03]*VR[vt][03]) << 16;
            VACC[04].DW += (long long)(VR[vs][04]*VR[vt][05]) << 16;
            VACC[05].DW += (long long)(VR[vs][05]*VR[vt][05]) << 16;
            VACC[06].DW += (long long)(VR[vs][06]*VR[vt][07]) << 16;
            VACC[07].DW += (long long)(VR[vs][07]*VR[vt][07]) << 16;
            break;
        case 0x4:
            VACC[00].DW += (long long)(VR[vs][00]*VR[vt][00]) << 16;
            VACC[01].DW += (long long)(VR[vs][01]*VR[vt][00]) << 16;
            VACC[02].DW += (long long)(VR[vs][02]*VR[vt][00]) << 16;
            VACC[03].DW += (long long)(VR[vs][03]*VR[vt][00]) << 16;
            VACC[04].DW += (long long)(VR[vs][04]*VR[vt][04]) << 16;
            VACC[05].DW += (long long)(VR[vs][05]*VR[vt][04]) << 16;
            VACC[06].DW += (long long)(VR[vs][06]*VR[vt][04]) << 16;
            VACC[07].DW += (long long)(VR[vs][07]*VR[vt][04]) << 16;
            break;
        case 0x5:
            VACC[00].DW += (long long)(VR[vs][00]*VR[vt][01]) << 16;
            VACC[01].DW += (long long)(VR[vs][01]*VR[vt][01]) << 16;
            VACC[02].DW += (long long)(VR[vs][02]*VR[vt][01]) << 16;
            VACC[03].DW += (long long)(VR[vs][03]*VR[vt][01]) << 16;
            VACC[04].DW += (long long)(VR[vs][04]*VR[vt][05]) << 16;
            VACC[05].DW += (long long)(VR[vs][05]*VR[vt][05]) << 16;
            VACC[06].DW += (long long)(VR[vs][06]*VR[vt][05]) << 16;
            VACC[07].DW += (long long)(VR[vs][07]*VR[vt][05]) << 16;
            break;
        case 0x6:
            VACC[00].DW += (long long)(VR[vs][00]*VR[vt][02]) << 16;
            VACC[01].DW += (long long)(VR[vs][01]*VR[vt][02]) << 16;
            VACC[02].DW += (long long)(VR[vs][02]*VR[vt][02]) << 16;
            VACC[03].DW += (long long)(VR[vs][03]*VR[vt][02]) << 16;
            VACC[04].DW += (long long)(VR[vs][04]*VR[vt][06]) << 16;
            VACC[05].DW += (long long)(VR[vs][05]*VR[vt][06]) << 16;
            VACC[06].DW += (long long)(VR[vs][06]*VR[vt][06]) << 16;
            VACC[07].DW += (long long)(VR[vs][07]*VR[vt][06]) << 16;
            break;
        case 0x7:
            VACC[00].DW += (long long)(VR[vs][00]*VR[vt][03]) << 16;
            VACC[01].DW += (long long)(VR[vs][01]*VR[vt][03]) << 16;
            VACC[02].DW += (long long)(VR[vs][02]*VR[vt][03]) << 16;
            VACC[03].DW += (long long)(VR[vs][03]*VR[vt][03]) << 16;
            VACC[04].DW += (long long)(VR[vs][04]*VR[vt][07]) << 16;
            VACC[05].DW += (long long)(VR[vs][05]*VR[vt][07]) << 16;
            VACC[06].DW += (long long)(VR[vs][06]*VR[vt][07]) << 16;
            VACC[07].DW += (long long)(VR[vs][07]*VR[vt][07]) << 16;
            break;
        case 0x8:
        case 0x9:
        case 0xA:
        case 0xB:
        case 0xC:
        case 0xD:
        case 0xE:
        case 0xF:
            e &= 07;
            VACC[00].DW += (long long)(VR[vs][00]*VR[vt][e]) << 16;
            VACC[01].DW += (long long)(VR[vs][01]*VR[vt][e]) << 16;
            VACC[02].DW += (long long)(VR[vs][02]*VR[vt][e]) << 16;
            VACC[03].DW += (long long)(VR[vs][03]*VR[vt][e]) << 16;
            VACC[04].DW += (long long)(VR[vs][04]*VR[vt][e]) << 16;
            VACC[05].DW += (long long)(VR[vs][05]*VR[vt][e]) << 16;
            VACC[06].DW += (long long)(VR[vs][06]*VR[vt][e]) << 16;
            VACC[07].DW += (long long)(VR[vs][07]*VR[vt][e]) << 16;
            break;
	}
    SIGNED_CLAMP(vd, 0);
    return;
}

method C -- the most desirable method to parallel vector transactions
Do a vector shuffle instruction in pure software (no SSE) to allocate the vector coefficient register, vt, such that loop iterations count 0..7 on all three accesses to the vector register file.

static short VR[32][8];
static short VC[8]; /* vector/scalar coefficient */

int sub_mask[16] = {
    0x0,
    0x0,
    0x1, 0x1,
    0x3, 0x3, 0x3, 0x3,
    0x7, 0x7, 0x7, 0x7, 0x7, 0x7, 0x7, 0x7
};

inline void SHUFFLE_VECTOR(int vt, int e)
{
    register int i, j;

    j = sub_mask[e];
    e = j ^ 07;
    for (i = 0; i < 8; i++)
        VC[i] = VR[vt][(i & e) | j];
/*
    if (e & 0x8)
        for (i = 0; i < 8; i++)
            VC[i] = VR[vt][(i & 00) | (e &= 0x7)];
    else if (e & 0x4)
        for (i = 0; i < 8; i++)
            VC[i] = VR[vt][(i & 04) | (e & 0x3)];
    else if (e & 0x2)
        for (i = 0; i < 8; i++)
            VC[i] = VR[vt][(i & 06) | (e & 0x1)];
    else // e == 0b0000 || e == 0b0001
        for (i = 0; i < 8; i++)
            VC[i] = VR[vt][(i & 07) | (e & 0x0)];
*/
    return;
}

[... example vector instruction]
void VAND(int vd, int vs, int vt, int e)
{
    register int i;

    for (i = 0; i < 8; i++)
        VR[vd][i] = VR[vs][i] & VC[i];
    for (i = 0; i < 8; i++)
        VACC[i].s[LO] = VR[vd][i];
    return;
}

This last method is by far the most preferable for several reasons.
Chiefly, it enables singularly accessing between the three instances of the vector register file.
`VR[vd] = VR[vs] * VR[vt];`
The previous two methods would conflict a premature source overwrite due to special scalar element decoders, in which case the accumulator needed writing to first.

I'll see about blending this in to the RSP core methodology before I do the next release with the fixes you guys helped with.

[shunyuan]

Did anyone else except Ziggy and Jabo have written a LLE graphics plugin?

[Lithium]

What is the difference between LLE Plugin and Software Plugin?