Skip to content

Commit

Permalink
further GPU3D optimisations
Browse files Browse the repository at this point in the history
  • Loading branch information
@RSDuck
RSDuck committed Jul 10, 2020
1 parent 71d89b6 commit 01e8e36
Showing 1 changed file with 79 additions and 23 deletions.
102 changes: 79 additions & 23 deletions src/GPU3D.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -140,7 +140,7 @@ const u8 CmdNumParams[256] =
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
};

typedef union
union CmdFIFOEntry
{
u64 _contents;
struct
Expand All @@ -149,7 +149,7 @@ typedef union
u8 Command;
};

} CmdFIFOEntry;
};

FIFO<CmdFIFOEntry, 256> CmdFIFO;
FIFO<CmdFIFOEntry, 4> CmdPIPE;
Expand Down Expand Up @@ -229,7 +229,7 @@ void UpdateClipMatrix();


u32 PolygonMode;
s16 CurVertex[3];
s16 CurVertex[4]; // last one is padding
u8 VertexColor[3];
s16 TexCoords[2];
s16 RawTexCoords[2];
Expand Down Expand Up @@ -742,7 +742,7 @@ void MatrixMult4x3(s32* m, s32* s)
"ld1 {v4.4s, v5.4s, v6.4s}, [%[s]]\n"
// we could also use shifting for this one
// but the combined latency of shift + add, is comparable to smlal
"dup v7.4s, %w[c]\n"
"movi v7.4s, #0x10, lsl #8\n"

"smull v16.2d, v0.2s, v4.4s[0]\n"
"smull2 v17.2d, v0.4s, v4.4s[0]\n"
Expand Down Expand Up @@ -785,7 +785,7 @@ void MatrixMult4x3(s32* m, s32* s)

"st1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%[m]]\n"
: "+m" (m)
: [m] "r" (m), [s] "r" (s), [c] "r" (0x1000)
: [m] "r" (m), [s] "r" (s)
: "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q16", "q17", "q18", "q19", "q20", "q21", "q22", "q23"
);
PROFILER_END_SECTION
Expand Down Expand Up @@ -1110,14 +1110,17 @@ int ClipAgainstPlane(Vertex* vertices, int nverts, int clipstart)
vertices[c++] = vtx;
}

// checkme
for (int i = 0; i < c; i++)
if (attribs)
{
Vertex* vtx = &vertices[i];
// checkme
for (int i = 0; i < c; i++)
{
Vertex* vtx = &vertices[i];

vtx->Color[0] &= ~0xFFF; vtx->Color[0] += 0xFFF;
vtx->Color[1] &= ~0xFFF; vtx->Color[1] += 0xFFF;
vtx->Color[2] &= ~0xFFF; vtx->Color[2] += 0xFFF;
vtx->Color[0] &= ~0xFFF; vtx->Color[0] += 0xFFF;
vtx->Color[1] &= ~0xFFF; vtx->Color[1] += 0xFFF;
vtx->Color[2] &= ~0xFFF; vtx->Color[2] += 0xFFF;
}
}

return c;
Expand All @@ -1136,14 +1139,39 @@ int ClipPolygon(Vertex* vertices, int nverts, int clipstart)
// some vertices that should get Y=-0x1000 get Y=0x1000 for some reason on hardware. it doesn't make sense.
// clipping seems to process the Y plane before the X plane.

uint32x4_t atleastOneOutside = vdupq_n_u32(0);
uint32x4_t allClippedPositively = vdupq_n_u32(0xFFFFFFFF);
uint32x4_t allClippedNegatively = vdupq_n_u32(0xFFFFFFFF);
for (int i = 0; i < nverts; i++)
{
int32x4_t vert = vld1q_s32(vertices[i].Position);

int32x4_t wValue = vdupq_n_s32(vert[3]);

uint32x4_t positiveClipPlane = vcgtq_s32(vert, wValue);
uint32x4_t negativeClipPlane = vcltq_s32(vert, vnegq_s32(wValue));

atleastOneOutside = vorrq_u32(atleastOneOutside, vorrq_u32(positiveClipPlane, negativeClipPlane));
allClippedPositively = vandq_u32(positiveClipPlane, allClippedPositively);
allClippedNegatively = vandq_u32(negativeClipPlane, allClippedNegatively);
}

if (vmaxvq_u32(allClippedPositively) || vmaxvq_u32(allClippedNegatively))
return 0;

u64 atleastOneOutsideScalar = vreinterpret_u64_u16(vmovn_u32(atleastOneOutside))[0];

// Z clipping
nverts = ClipAgainstPlane<2, attribs>(vertices, nverts, clipstart);
if (atleastOneOutsideScalar & (1ULL << 32))
nverts = ClipAgainstPlane<2, attribs>(vertices, nverts, clipstart);

// Y clipping
nverts = ClipAgainstPlane<1, attribs>(vertices, nverts, clipstart);
if (atleastOneOutsideScalar & (1ULL << 16))
nverts = ClipAgainstPlane<1, attribs>(vertices, nverts, clipstart);

// X clipping
nverts = ClipAgainstPlane<0, attribs>(vertices, nverts, clipstart);
if (atleastOneOutsideScalar & (1ULL << 0))
nverts = ClipAgainstPlane<0, attribs>(vertices, nverts, clipstart);

return nverts;
}
Expand Down Expand Up @@ -1573,14 +1601,40 @@ void SubmitVertex()
{
PROFILER_SECTION(submitVertex)

s64 vertex[4] = {(s64)CurVertex[0], (s64)CurVertex[1], (s64)CurVertex[2], 0x1000};
Vertex* vertextrans = &TempVertexBuffer[VertexNumInPoly];

UpdateClipMatrix();
vertextrans->Position[0] = (vertex[0]*ClipMatrix[0] + vertex[1]*ClipMatrix[4] + vertex[2]*ClipMatrix[8] + vertex[3]*ClipMatrix[12]) >> 12;
vertextrans->Position[1] = (vertex[0]*ClipMatrix[1] + vertex[1]*ClipMatrix[5] + vertex[2]*ClipMatrix[9] + vertex[3]*ClipMatrix[13]) >> 12;
vertextrans->Position[2] = (vertex[0]*ClipMatrix[2] + vertex[1]*ClipMatrix[6] + vertex[2]*ClipMatrix[10] + vertex[3]*ClipMatrix[14]) >> 12;
vertextrans->Position[3] = (vertex[0]*ClipMatrix[3] + vertex[1]*ClipMatrix[7] + vertex[2]*ClipMatrix[11] + vertex[3]*ClipMatrix[15]) >> 12;
__asm__ volatile
(
"movi v5.4s, #0x10, lsl #8\n" // 1 << 12

"ld1 {v0.16b, v1.16b, v2.16b, v3.16b}, [%[ClipMatrix]]\n"
"ld1 {v4.8b}, [%[curvertex]]\n"

"sxtl v4.4s, v4.4h\n"
"ins v4.4s[3], v5.4s[0]\n"

"smull v5.2d, v0.2s, v4.4s[0]\n"
"smlal v5.2d, v1.2s, v4.4s[1]\n"
"smlal v5.2d, v2.2s, v4.4s[2]\n"
"smlal v5.2d, v3.2s, v4.4s[3]\n"
"smull2 v6.2d, v0.4s, v4.4s[0]\n"
"smlal2 v6.2d, v1.4s, v4.4s[1]\n"
"smlal2 v6.2d, v2.4s, v4.4s[2]\n"
"smlal2 v6.2d, v3.4s, v4.4s[3]\n"

"shrn v5.2s, v5.2d, #12\n"
"shrn2 v5.4s, v6.2d, #12\n"

"st1 {v5.16b}, [%[vertexTrans]]\n"
:
"=m" (vertextrans->Position)
:
[curvertex] "r" (CurVertex), [vertexTrans] "r" (vertextrans->Position),
[ClipMatrix] "r" (ClipMatrix)
:
"q0", "q1", "q2", "q3", "q5", "q6"
);

// this probably shouldn't be.
// the way color is handled during clipping needs investigation. TODO
Expand All @@ -1590,8 +1644,8 @@ void SubmitVertex()

if ((TexParam >> 30) == 3)
{
vertextrans->TexCoords[0] = ((vertex[0]*TexMatrix[0] + vertex[1]*TexMatrix[4] + vertex[2]*TexMatrix[8]) >> 24) + RawTexCoords[0];
vertextrans->TexCoords[1] = ((vertex[0]*TexMatrix[1] + vertex[1]*TexMatrix[5] + vertex[2]*TexMatrix[9]) >> 24) + RawTexCoords[1];
vertextrans->TexCoords[0] = (((s64)CurVertex[0]*TexMatrix[0] + (s64)CurVertex[1]*TexMatrix[4] + (s64)CurVertex[2]*TexMatrix[8]) >> 24) + RawTexCoords[0];
vertextrans->TexCoords[1] = (((s64)CurVertex[0]*TexMatrix[1] + (s64)CurVertex[1]*TexMatrix[5] + (s64)CurVertex[2]*TexMatrix[9]) >> 24) + RawTexCoords[1];
}
else
{
Expand Down Expand Up @@ -2009,7 +2063,9 @@ void ExecuteCommand()

// commands that stall the polygon pipeline: 0x32, 0x70, 0x40

if (CmdNumParams[entry.Command] <= 1)
u8 numParams = CmdNumParams[entry.Command];

if (numParams <= 1)
{
switch (entry.Command)
{
Expand Down Expand Up @@ -2350,7 +2406,7 @@ void ExecuteCommand()
ExecParams[ExecParamCount] = entry.Param;
ExecParamCount++;

if (ExecParamCount >= (u32)CmdNumParams[entry.Command])
if (ExecParamCount >= numParams)
{
AddCycles(1);
/*printf("[GXS:%08X] 0x%02X, ", GXStat, entry.Command);
Expand Down

0 comments on commit 01e8e36

Please sign in to comment.