mirrors/IPUDOOM
1
0
Fork 0
mirror of https://github.com/jndean/IPUDOOM.git synced 2026-08-18 16:28:53 +00:00
Code Issues Packages Projects Releases Wiki Activity

Intermediate: Add flag aggregation exchange

Browse Source
This commit is contained in:
jndean
2023-10-20 18:11:49 +00:00
parent 65e35ef19b
commit b5e35b40a6
3 changed files with 102 additions and 34 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/ipu/ipu_interface.h
+1 -1
View File
@@ -18,7 +18,7 @@ extern "C" {
#define IPUMAXEVENTSPERTIC (5)
#define IPUAMMARKBUFSIZE (544)
#define IPUMAPPEDLINEUPDATES (2)
#define IPUPROGBUFSIZE (32)
#define IPUPROGBUFSIZE (128)
#define IPUFIRSTRENDERTILE (0)
#define IPUNUMRENDERTILES (32)
src/ipu/ipu_texturetiles.cpp
+99 -33
View File
@@ -67,6 +67,12 @@ void IPU_R_FulfilColumnRequest(unsigned* progBuf, unsigned* textureBuf) {
textureBuf[i] = 0x20202020;
}
XCOM_Execute(sendProgram, textureBuf, NULL);
// TMP
asm volatile(R"(
sans 1
sync 0x1
)");
}
}
@@ -103,61 +109,115 @@ void IPU_R_InitColumnRequester(unsigned* progBuf, int progBufSize) {
textureTileLUT[i] = XCOM_logical2physical(firstTextureTile + i);
}
// Prog Buf starts with directory of programs
// ProgBuf starts with directory of programs
unsigned* progHead = progBuf + 3;
unsigned* progEnd = &progBuf[progBufSize];
{
progBuf[0] = progHead - progBuf;
XCOMAssembler assembler;
int messageSize = sizeof(IPUColRequest_t) / sizeof(int);
assembler.addSend(0, messageSize, XCOM_BROADCAST, XCOM_WORSTSENDDELAY);
progHead = assembler.assemble(progHead, progEnd - progHead);
}
{
XCOMAssembler assembler;
int messageSize = IPUTEXTURECACHELINESIZE;
assembler.addRecv(0, messageSize, 0, XCOM_WORSTRECVDELAY);
unsigned* newProgHead = assembler.assemble(progHead, progEnd - progHead);
for (unsigned* inst = progHead; inst < newProgHead; ++inst) {
if ((*inst & 0xfc003fffu) == 0x64000000u) {
muxInstructionOffset = (inst - &progBuf[progBuf[0]]);
break;
}
XCOMAssembler assembler;
// First Program: performs the request and receives the response
progBuf[0] = progHead - progBuf;
int messageSize = sizeof(IPUColRequest_t) / sizeof(int);
assembler.addSend(0, messageSize, XCOM_BROADCAST, XCOM_WORSTSENDDELAY);
progHead = assembler.assemble(progHead, progEnd - progHead);
assembler.reset();
assembler.addRecv(0, IPUTEXTURECACHELINESIZE, 0, XCOM_WORSTRECVDELAY);
unsigned* newProgHead = assembler.assemble(progHead, progEnd - progHead);
// Record the mux location for later live patching
for (unsigned* inst = progHead; inst < newProgHead; ++inst) {
if ((*inst & 0xfc003fffu) == 0x64000000u) {
muxInstructionOffset = (inst - &progBuf[progBuf[0]]);
break;
}
progHead = newProgHead;
}
// Third program sends finished flag to flag reducer
// {
// progBuf[2] = progHead - progBuf;
// XCOMAssembler assembler;
// int messageSize = 1;
// int recvCycle = XCOM_WORSTRECVDELAY;
// int sendCycle = recvCycle - XCOM_TimeToMux(__builtin_ipu_get_tile_id(), renderTile);
// int direction = XCOM_EastOrWest(__builtin_ipu_get_tile_id(), renderTile);
// assembler.addSend(0, messageSize, direction, sendCycle);
// progHead = assembler.assemble(progHead, progEnd - progHead);
// progHead++;
// }
}
progHead = newProgHead + 1; // +1 => don't overwrite the return statement
assembler.reset();
// Second Program: aggregates the 'done' flag across render tiles
progBuf[1] = progHead - progBuf;
int aggrTile = XCOM_logical2physical(IPUFIRSTRENDERTILE);
if (__builtin_ipu_get_tile_id() == aggrTile) {
for (int slot = 0; slot < IPUNUMRENDERTILES - 1; ++slot) {
int senderID = XCOM_logical2physical(IPUFIRSTRENDERTILE + 1 + slot);
assembler.addRecv((unsigned*)(4*(slot + 1)), 1, senderID, XCOM_WORSTRECVDELAY + slot,
slot == IPUNUMRENDERTILES - 2);
}
progHead = assembler.assemble(progHead, progEnd - progHead);
progHead++;
// Third Program: aggregation tile distributes the answer
progBuf[2] = progHead - progBuf;
assembler.reset();
assembler.addSend(0, 1, XCOM_BROADCAST, XCOM_WORSTSENDDELAY);
progHead = assembler.assemble(progHead, progEnd - progHead);
progHead++;
} else {
int myTimeSlot = tileID - (IPUFIRSTRENDERTILE + 1);
int direction = XCOM_EastOrWest(__builtin_ipu_get_tile_id(), aggrTile);
int muxCycle = XCOM_WORSTRECVDELAY + myTimeSlot;
int sendCycle = muxCycle - XCOM_TimeToMux(__builtin_ipu_get_tile_id(), aggrTile);
assembler.addSend(0, 1, direction, sendCycle);
progHead = assembler.assemble(progHead, progEnd - progHead);
assembler.reset();
sendCycle = XCOM_WORSTSENDDELAY;
muxCycle = sendCycle + XCOM_TimeToMux(aggrTile, __builtin_ipu_get_tile_id());
assembler.addRecv(0, 1, aggrTile, muxCycle);
progHead = assembler.assemble(progHead, progEnd - progHead);
progHead++;
}
// assembler.reset();
}
__SUPER__
static bool checkRenderingFinished() {
auto aggregateProgA = &tileLocalProgBuf[tileLocalProgBuf[1]];
auto aggregateProgB = &tileLocalProgBuf[tileLocalProgBuf[2]];
tileLocalTextureBuf[0] = tileID + 100;
XCOM_Execute(
aggregateProgA,
tileLocalTextureBuf,
tileLocalTextureBuf
);
if (tileID == IPUFIRSTRENDERTILE) {
for (int i = 1; i < IPUNUMRENDERTILES; i++) {
tileLocalTextureBuf[0] += tileLocalTextureBuf[i];
}
XCOM_Execute(
aggregateProgB,
tileLocalTextureBuf,
NULL
);
}
printf("Aggregation result: %d\n", tileLocalTextureBuf[0]);
return tileLocalTextureBuf[0];
}
extern "C"
__SUPER__
byte* IPU_R_RequestColumn(int texture, int column) {
// progBuff starts with a program directory
auto requestProg = &tileLocalProgBuf[tileLocalProgBuf[0]];
auto aggregateProgA = &tileLocalProgBuf[tileLocalProgBuf[1]];
auto aggregateProgB = &tileLocalProgBuf[tileLocalProgBuf[2]];
if (textureFetchCount++ < 6) {
int sourceTile = 0;
XCOM_PatchMuxAndExecute(
requestProg, // Prog
requestProg, // Prog
tileLocalTextureBuf, // Read offset
tileLocalTextureBuf, // Write offset
muxInstructionOffset, // Patch offset
textureTileLUT[sourceTile] // Mux value
);
checkRenderingFinished();
return (byte*) tileLocalTextureBuf;
}
return columnCache[texture % NUMCACHECOLS];
@@ -192,6 +252,12 @@ void IPU_R_Sans(unsigned* progBuf, int progBufSize) {
sans 1
sync 0x1
)");
// TMP
asm volatile(R"(
sans 1
sync 0x1
)");
}
}
src/ipu_host.cpp
+2
View File
@@ -310,6 +310,8 @@ void IpuDoom::buildIpuGraph() {
for (unsigned tile = IPUFIRSTTEXTURETILE + IPUNUMTEXTURETILES; tile < totalTiles; ++tile) {
m_ipuGraph.setTileMapping(m_ipuGraph.addVertex(R_RenderPlayerView_CS, "R_Sans_Vertex"), tile);
}
// Cache line is used as the aggregation buffer, make sure it's big enough
assert(IPUTEXTURECACHELINESIZE >= IPUNUMRENDERTILES);
poplar::program::Sequence R_RenderPlayerView_prog({
poplar::program::Copy(miscValuesStream, m_miscValuesBuf),