Texture blob moved to IPU

2026-03-21 15:09:45 +00:00 · 2023-10-25 11:23:36 +00:00
parent cd8e7c14cc
commit 4746082583
8 changed files with 91 additions and 68 deletions
--- a/src/ipu/ipu_interface.h
+++ b/src/ipu/ipu_interface.h
@@ -31,7 +31,7 @@ extern "C" {
 #define IPUNUMTEXTURETILES (IPUTEXTURETILESPERRENDERTILE * IPUNUMRENDERTILES)
 #define IPUNUMTEXTURECACHELINES (1)
 #define IPUTEXTURECACHELINESIZE (128 / sizeof(int))
-
+#define IPUMAXNUMTEXTURES (130)
 #define IPUCOMMSBUFSIZE (IPUNUMRENDERTILES)


--- a/src/ipu/ipu_texturetiles.cpp
+++ b/src/ipu/ipu_texturetiles.cpp
@@ -14,6 +14,8 @@
 unsigned* tileLocalProgBuf;
 unsigned* tileLocalCommsBuf;
 unsigned* tileLocalTextureBuf;
+const int* tileLocalTextureRange;
+const int* tileLocalTextureOffsets;


 //  -------- Components for the tiles that serve textures ------------ //
--- a/src/ipu/ipu_texturetiles.h
+++ b/src/ipu/ipu_texturetiles.h
@@ -19,6 +19,8 @@ typedef struct {
 extern unsigned* tileLocalProgBuf;
 extern unsigned* tileLocalCommsBuf;
 extern unsigned* tileLocalTextureBuf;
+extern const int* tileLocalTextureRange;
+extern const int* tileLocalTextureOffsets;


 __SUPER__ void IPU_R_InitTextureTile(unsigned* progBuf, int progBufSize);
--- a/src/ipu/r_codelets.cpp
+++ b/src/ipu/r_codelets.cpp
@@ -19,6 +19,19 @@ extern "C" {
 };


+class R_ExecuteSetViewSize_Vertex : public poplar::SupervisorVertex {
+ public:
+  poplar::Input<poplar::Vector<unsigned char>> miscValues;
+  __SUPER__ 
+  void compute() {
+    IPU_R_ExecuteSetViewSize_UnpackMiscValues(
+      (R_ExecuteSetViewSize_MiscValues_t*) &miscValues[0]
+    );
+    R_ExecuteSetViewSize();
+  }
+};
+
+
 struct R_Init_Vertex: public poplar::SupervisorVertex {
  
  poplar::Output<poplar::Vector<unsigned>> progBuf;
@@ -46,29 +59,6 @@ struct R_Init_Vertex: public poplar::SupervisorVertex {
  }
 };

-struct R_Init_TT_Vertex: public poplar::SupervisorVertex {
-  
-  poplar::Input<poplar::Vector<unsigned char>> miscValues;
-  poplar::Input<poplar::Vector<unsigned char>> lumpBuf;
-  poplar::Output<int> lumpNum;
-
-  __SUPER__ 
-  void compute() {
-
-static int step = 0; switch (step++) { case 0:
-
-      *lumpNum = ((R_Init_MiscValues_t*)&miscValues[0])->TEXTURE1_lumpnum;
-    
-break; case 1:
-
-      R_InitTextures_TT((int*)&lumpBuf[0]);
-
-
-*lumpNum = 0; step = 0; }
-
-  }
-};
-

 struct 
 [[
@@ -102,20 +92,14 @@ R_RenderPlayerView_Vertex : public poplar::SupervisorVertex {
  }
 };

+struct R_InitTexture_Vertex : public poplar::SupervisorVertex {
+  poplar::Output<poplar::Vector<unsigned>> progBuf;

-class R_ExecuteSetViewSize_Vertex : public poplar::SupervisorVertex {
- public:
-  poplar::Input<poplar::Vector<unsigned char>> miscValues;
-  __SUPER__ 
-  void compute() {
-    IPU_R_ExecuteSetViewSize_UnpackMiscValues(
-      (R_ExecuteSetViewSize_MiscValues_t*) &miscValues[0]
-    );
-    R_ExecuteSetViewSize();
+  __SUPER__ void compute() {
+    IPU_R_InitTextureTile(&progBuf[0], progBuf.size()); 
  }
 };

-
 struct 
 [[
 poplar::constraint("region(*dummy)  != region(*progBuf)"),
@@ -129,12 +113,28 @@ R_FulfilColumnRequests_Vertex : public poplar::SupervisorVertex {
  poplar::InOut<poplar::Vector<unsigned>> progBuf;
  poplar::InOut<poplar::Vector<unsigned>> commsBuf;
  poplar::Output<poplar::Vector<unsigned char>> textureBuf;
+  // poplar::Input<poplar::Vector<int>> textureOffsets;
+  // poplar::Input<poplar::Vector<int>> textureRange;

  __SUPER__ void compute() {
+    // tileLocalTextureRange = &textureRange[0];
+    // tileLocalTextureOffsets = &textureOffsets[0];
+
    IPU_R_FulfilColumnRequest(&progBuf[0], &textureBuf[0], &commsBuf[0]);
  }
 };

+
+struct R_InitSans_Vertex : public poplar::SupervisorVertex {
+  poplar::Output<poplar::Vector<unsigned>> progBuf;
+
+  __SUPER__ void compute() {
+    // Reuse IPU_R_InitTextureTile because the 'done' flag receiving program
+    // it compiles is perfectly valid for use by a sans tile
+    IPU_R_InitTextureTile(&progBuf[0], progBuf.size());   
+  }
+};
+
 struct 
 [[
 poplar::constraint("region(*dummy) != region(*progBuf)"),
@@ -150,11 +150,3 @@ R_Sans_Vertex : public poplar::SupervisorVertex {
    IPU_R_Sans(&progBuf[0], &commsBuf[0]);
  }
 };
-
-struct R_InitTextureOrSans_Vertex : public poplar::SupervisorVertex {
-  poplar::Output<poplar::Vector<unsigned>> progBuf;
-
-  __SUPER__ void compute() {
-    IPU_R_InitTextureTile(&progBuf[0], progBuf.size());    
-  }
-};
--- a/src/ipu_host.cpp
+++ b/src/ipu_host.cpp
@@ -159,6 +159,15 @@ void IpuDoom::buildIpuGraph() {

  // -------- R_Init ------ //

+  poplar::Tensor textureBuf = m_ipuGraph.addVariable(
+    poplar::UNSIGNED_CHAR, 
+    {IPUNUMTEXTURETILES, IPUTEXTURETILEBUFSIZE}, 
+    "textureBuf");
+  auto textureBufStream = m_ipuGraph.addHostToDeviceFIFO(
+    "textureBuf-stream", 
+    poplar::UNSIGNED_CHAR, 
+    IPUTEXTURETILESPERRENDERTILE * IPUTEXTURETILEBUFSIZE);
+
  poplar::ComputeSet R_Init_CS = m_ipuGraph.addComputeSet("R_Init_CS");
  for (int renderTile = 0; renderTile < IPUNUMRENDERTILES; ++renderTile) {
    int logicalTile = IPUFIRSTRENDERTILE + renderTile;
@@ -171,9 +180,17 @@ void IpuDoom::buildIpuGraph() {
    m_ipuGraph.setTileMapping(vtx, logicalTile);
    m_ipuGraph.setPerfEstimate(vtx, 100);
  }
-  poplar::ComputeSet R_InitTextureOrSans_CS = m_ipuGraph.addComputeSet("R_InitTextureOrSans_CS");
-  for (unsigned tile = IPUFIRSTTEXTURETILE; tile < totalTiles; ++tile) {
-    vtx =  m_ipuGraph.addVertex(R_InitTextureOrSans_CS, "R_InitTextureOrSans_Vertex", {
+  poplar::ComputeSet R_InitTextureAndSans_CS = m_ipuGraph.addComputeSet("R_InitTextureAndSans_CS");
+  for (unsigned textureTile = 0; textureTile < IPUNUMTEXTURETILES; ++textureTile) {
+    int logicalTile = IPUFIRSTTEXTURETILE + textureTile;
+    vtx =  m_ipuGraph.addVertex(R_InitTextureAndSans_CS, "R_InitTexture_Vertex", {
+      {"progBuf", progBuf[logicalTile]}, 
+    });
+    m_ipuGraph.setTileMapping(vtx, logicalTile);
+    m_ipuGraph.setPerfEstimate(vtx, 2000);
+  }
+  for (unsigned tile = IPUFIRSTTEXTURETILE + IPUNUMTEXTURETILES; tile < totalTiles; ++tile) {
+    vtx =  m_ipuGraph.addVertex(R_InitTextureAndSans_CS, "R_InitSans_Vertex", {
      {"progBuf", progBuf[tile]}, 
    });
    m_ipuGraph.setTileMapping(vtx, tile);
@@ -187,8 +204,9 @@ void IpuDoom::buildIpuGraph() {
  );

  poplar::program::Sequence R_Init_prog({
-    poplar::program::Execute(R_InitTextureOrSans_CS),
    poplar::program::Copy(miscValuesStream, m_miscValuesBuf),
+    poplar::program::Copy(textureBufStream, textureBuf.slice(0, IPUTEXTURETILESPERRENDERTILE)),
+    poplar::program::Execute(R_InitTextureAndSans_CS),
    poplar::program::Repeat(2, poplar::program::Sequence({ // <- number of R_Init_CS steps
      poplar::program::Execute(R_Init_CS),
      poplar::program::Call(requestLumpFromHost, {m_lumpNum[0]}, {lumpBuf}),
@@ -273,21 +291,24 @@ void IpuDoom::buildIpuGraph() {

  // -------- R_RenderPlayerView_CS ------ //

-  poplar::Tensor textureBuf = m_ipuGraph.addVariable(
-    poplar::UNSIGNED_CHAR, 
-    {IPUNUMTEXTURETILES, IPUTEXTURETILEBUFSIZE}, 
-    "textureBuf");
  poplar::Tensor textureCache = m_ipuGraph.addVariable(
    poplar::UNSIGNED_INT, 
    { IPUNUMRENDERTILES, 
      IPUNUMTEXTURECACHELINES * IPUTEXTURECACHELINESIZE}, 
    "textureCache");

+  poplar::Tensor textureOffsets = m_ipuGraph.addVariable(poplar::INT, {IPUMAXNUMTEXTURES}, "textureOffsets");
+  poplar::Tensor textureRanges = m_ipuGraph.addVariable(poplar::INT, {IPUTEXTURETILESPERRENDERTILE + 1}, "textureRanges");
+  auto textureOffsetStream = m_ipuGraph.addHostToDeviceFIFO("textureOffsets-stream", poplar::INT, IPUMAXNUMTEXTURES);
+  auto textureRangeStream = m_ipuGraph.addHostToDeviceFIFO("textureRanges-stream", poplar::INT, IPUTEXTURETILESPERRENDERTILE + 1);
+  m_ipuGraph.setTileMapping(textureOffsets, IPUFIRSTTEXTURETILE);
+  m_ipuGraph.setTileMapping(textureRanges, IPUFIRSTTEXTURETILE);
+
  poplar::ComputeSet R_RenderPlayerView_CS = m_ipuGraph.addComputeSet("R_RenderPlayerView_CS");
  for (int renderTile = 0; renderTile < IPUNUMRENDERTILES; ++renderTile) {
    int logicalTile = IPUFIRSTRENDERTILE + renderTile;
    vtx = m_ipuGraph.addVertex(R_RenderPlayerView_CS, "R_RenderPlayerView_Vertex", {
-      {"frame", ipuFrameSlices[renderTile]}, 
+      {"frame", ipuFrameSlices[renderTile]},
      {"textureCache", textureCache[renderTile]},
      {"progBuf", progBuf[logicalTile]}, 
      {"commsBuf", commsBuf[logicalTile]},
@@ -300,11 +321,14 @@ void IpuDoom::buildIpuGraph() {
  }
  for (int textureTile = 0; textureTile < IPUNUMTEXTURETILES; ++textureTile) {
    int logicalTile = IPUFIRSTTEXTURETILE + textureTile;
+    // int textureStripeIdx = textureTile % IPUTEXTURETILESPERRENDERTILE;
    vtx =  m_ipuGraph.addVertex(R_RenderPlayerView_CS, "R_FulfilColumnRequests_Vertex", {
      {"dummy", nonExecutableDummy[logicalTile]}, 
-      {"textureBuf", textureBuf[textureTile]},
      {"progBuf", progBuf[logicalTile]},
      {"commsBuf", commsBuf[logicalTile]},
+      {"textureBuf", textureBuf[textureTile]},
+      // {"textureOffsets", textureOffsets},
+      // {"textureRange", textureRanges.slice(textureStripeIdx, textureStripeIdx + 2)},
    });
    m_ipuGraph.setTileMapping(vtx, logicalTile);
    m_ipuGraph.setPerfEstimate(vtx, 1000);
@@ -361,6 +385,8 @@ void IpuDoom::buildIpuGraph() {

  m_ipuEngine->connectStream("miscValues-stream", m_miscValuesBuf_h);
  m_ipuEngine->connectStream("lumpNum-stream", &m_lumpNum_h);
+  m_ipuEngine->connectStream("textureOffsets-stream", ipuTextureBlobOffsets);
+  m_ipuEngine->connectStream("textureRanges-stream", ipuTextureBlobRanges);
  // Connect frame-instream/outstream later in run_IPU_MiscSetup because 
  // I_VideoBuffer is initialised quite late

@@ -411,6 +437,7 @@ void IpuDoom::run_R_ExecuteSetViewSize() {
  m_ipuEngine->run(6);
 }
 void IpuDoom::run_R_Init() {
+  m_ipuEngine->connectStream("textureBuf-stream", ipuTextureBlob);
  IPU_R_Init_PackMiscValues(m_miscValuesBuf_h);
  m_ipuEngine->run(7);
 }
--- a/src/ipu_transfer.h
+++ b/src/ipu_transfer.h
@@ -8,6 +8,11 @@ extern "C" {
 #include "doomtype.h"
 #include "r_defs.h"

+#include "ipu/ipu_interface.h"
+
+extern byte* ipuTextureBlob;
+extern int ipuTextureBlobOffsets[IPUMAXNUMTEXTURES];
+extern int ipuTextureBlobRanges[IPUTEXTURETILESPERRENDERTILE + 1];

 void IPU_G_LoadLevel_PackMiscValues(void* buf);
 void IPU_G_Ticker_PackMiscValues(void* buf);
--- a/src/r_data.c
+++ b/src/r_data.c
@@ -354,18 +354,22 @@ byte *R_GetColumn_Original(int tex, int col) { // JOSEF: Renamed to `_Original`
 // JOSEF: We dedicate several tiles on the IPU just to storing textures,
 //        so we can afford to preassemble all textures into a big blob
 byte* ipuTextureBlob;
-int* ipuTextureBlobOffsets;
+int ipuTextureBlobOffsets[IPUMAXNUMTEXTURES];
 int ipuTextureBlobRanges[IPUTEXTURETILESPERRENDERTILE + 1];
+
 void GenerateIPUTextureBlob(void) {
+
+  if (numtextures >= IPUMAXNUMTEXTURES) {
+    I_Error("GenerateIPUTextureBlob: numtextures >= IPUMAXNUMTEXTURES");
+  }
  ipuTextureBlob = malloc(IPUTEXTURETILEBUFSIZE * IPUTEXTURETILESPERRENDERTILE);
-  ipuTextureBlobOffsets = malloc(sizeof(*ipuTextureBlobOffsets) * numtextures);
  ipuTextureBlobRanges[0] = 0;
  int tile = 0, pos = 0;
+
  for (int t = 0; t < numtextures; ++t) {
    int tex_width = textures[t]->width;
    int tex_height = textures[t]->height;
    int tex_size = tex_width * tex_height;
-    
    if (pos + tex_size >= IPUTEXTURETILEBUFSIZE) {
      pos = 0;
      tile += 1;
@@ -384,13 +388,13 @@ void GenerateIPUTextureBlob(void) {
    pos += tex_size;
  }
  ipuTextureBlobRanges[tile + 1] = numtextures;
-
 }

 //
 // R_GetColumn : JOSEF: The version that uses the IPU texture dump
 //
 byte *R_GetColumn(int tex, int col) {
+  // return R_GetColumn_Original(tex, col); // For checking old behaviour
  int tile;
  col &= texturewidthmask[tex];
  for (tile = 0; tile < IPUTEXTURETILESPERRENDERTILE; ++tile) {
@@ -551,8 +555,6 @@ void R_InitTextures(void) {
      printf("\b");
  }

-  int totaltexturesize = 0; // JOSEF TMP
-
  for (i = 0; i < numtextures; i++, directory++) {
    if (!(i & 63))
      printf(".");
@@ -580,8 +582,6 @@ void R_InitTextures(void) {
    texture->height = SHORT(mtexture->height);
    texture->patchcount = SHORT(mtexture->patchcount);

-    totaltexturesize += texture->width * texture->height; // JOSEF TMP
-
    memcpy(texture->name, mtexture->name, sizeof(texture->name));
    mpatch = &mtexture->patches[0];
    patch = &texture->patches[0];
@@ -609,12 +609,6 @@ void R_InitTextures(void) {
    totalwidth += texture->width;
  }

-  printf("\n JOSEF: Total Texture Size = %d bytes (%dKb), numtextures = %d\n",
-    totaltexturesize,
-    totaltexturesize / 1000,
-    numtextures
-  );
-
  Z_Free(patchlookup);

  W_ReleaseLumpName(("TEXTURE1"));
--- a/src/r_main.c
+++ b/src/r_main.c
@@ -646,7 +646,6 @@ void R_ExecuteSetViewSize(void) {
 //

 void R_Init(void) {
-  IPU_R_Init();

  R_InitData();
  printf(".");
@@ -666,6 +665,8 @@ void R_Init(void) {
  printf(".");

  framecount = 0;
+  
+  IPU_R_Init();
 }

 //
@@ -741,7 +742,7 @@ void R_RenderPlayerView(player_t *player) {

  // The head node is the last node output.
  R_RenderBSPNode(numnodes - 1);
-  // IPU_R_RenderPlayerView();
+  IPU_R_RenderPlayerView();

  // Check for new console commands.
  NetUpdate();