rooty/rpi-vk-driver
1
0
Fork 0
mirror of https://github.com/Yours3lf/rpi-vk-driver.git synced 2025-01-30 22:52:14 +01:00
Code Issues Releases Wiki Activity

fixed clmarkers (memory could be moved around)

Browse Source
This commit is contained in:
yours3lf 2020-05-17 10:47:23 +01:00
parent 269a1b412c
commit cc894e3342
6 changed files with 42 additions and 36 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
driver/ControlListUtil.c
View File

@ -56,9 +56,9 @@ void clInsertNewCLMarker(ControlList* cl,
CLMarker marker = {};
marker.memGuard = 0xDDDDDDDD;
marker.handlesBufOffset = handlesCL->offset;
marker.shaderRecBufOffset = shaderRecCL->offset;
marker.uniformsBufOffset = uniformsCL->offset;
marker.handlesBufOffset = 0;
marker.shaderRecBufOffset = 0;
marker.uniformsBufOffset = 0;
marker.nextMarkerOffset = -1;
//close current marker
@ -95,9 +95,9 @@ void clCloseCurrentMarker(ControlList* cl, ControlList* handlesCL, ControlList*
assert(uniformsCL);
CLMarker* currMarker = getCPAptrFromOffset(cl->CPA, cl->currMarkerOffset);
currMarker->size = cl->nextFreeByteOffset - (cl->currMarkerOffset + sizeof(CLMarker));
currMarker->handlesSize = handlesCL->nextFreeByteOffset - currMarker->handlesBufOffset;
currMarker->shaderRecSize = shaderRecCL->nextFreeByteOffset - currMarker->shaderRecBufOffset;
currMarker->uniformsSize = uniformsCL->nextFreeByteOffset - currMarker->uniformsBufOffset;
currMarker->handlesSize = handlesCL->nextFreeByteOffset - (currMarker->handlesBufOffset + handlesCL->offset);
currMarker->shaderRecSize = shaderRecCL->nextFreeByteOffset - (currMarker->shaderRecBufOffset + shaderRecCL->offset);
currMarker->uniformsSize = uniformsCL->nextFreeByteOffset - (currMarker->uniformsBufOffset + uniformsCL->offset);
currMarker->shaderRecCount = shaderRecCount - currMarker->shaderRecCount; //update shader rec count to reflect added shader recs
}
@ -789,6 +789,8 @@ uint32_t clGetHandleIndex(ControlList* handlesCl, uint32_t handlesOffset, uint32
*(uint32_t*)getCPAptrFromOffset(handlesCl->CPA, handlesCl->nextFreeByteOffset) = handle;
handlesCl->nextFreeByteOffset += 4;
assert(handlesCl->nextFreeByteOffset < handlesCl->offset + handlesCl->blockSize * handlesCl->numBlocks);
return c;
}

2
driver/ControlListUtil.h
View File

@ -45,7 +45,7 @@ typedef struct CLMarker
//pointers that point to where all the other CL data is
//plus sizes
//uint8_t* handlesBuf; //
uint32_t handlesBufOffset;
uint32_t handlesBufOffset; //relative offset, because underlying buffer could be moved
uint32_t handlesSize;
//uint8_t* shaderRecBuf; //
uint32_t shaderRecBufOffset;

29
driver/command.c
View File

@ -49,8 +49,8 @@ VKAPI_ATTR VkResult VKAPI_CALL rpi_vkCreateCommandPool(
//initial number of command buffers to hold
int numCommandBufs = 128;
//TODO uniforms might need to realloc, which should be handled properly
int consecutiveBlockSize = ARM_PAGE_SIZE;// * 20;
int consecutiveBlockNumber = 128;
int consecutiveBlockSize = ARM_PAGE_SIZE * 20;
int consecutiveBlockNumber = 256;
//int numCommandBufs = 30;
//int consecutiveBlockSize = getCPABlockSize(256);
//int consecutiveBlockNumber = 30;
@ -329,12 +329,13 @@ VKAPI_ATTR VkResult VKAPI_CALL rpi_vkQueueSubmit(
uint32_t readMSAAdepthStencilImage = marker->readMSAAdepthStencilImage;
//This should not result in an insertion!
uint32_t writeImageIdx = writeImage ? clGetHandleIndex(&cmdbuf->handlesCl, marker->handlesBufOffset, marker->handlesSize, writeImage->boundMem->bo) : 0;
uint32_t readImageIdx = readImage ? clGetHandleIndex(&cmdbuf->handlesCl, marker->handlesBufOffset, marker->handlesSize, readImage->boundMem->bo) : 0;
uint32_t writeDepthStencilImageIdx = writeDepthStencilImage ? clGetHandleIndex(&cmdbuf->handlesCl, marker->handlesBufOffset, marker->handlesSize, writeDepthStencilImage->boundMem->bo) : 0;
uint32_t readDepthStencilImageIdx = readDepthStencilImage ? clGetHandleIndex(&cmdbuf->handlesCl, marker->handlesBufOffset, marker->handlesSize, readDepthStencilImage->boundMem->bo) : 0;
uint32_t writeMSAAimageIdx = writeMSAAimage ? clGetHandleIndex(&cmdbuf->handlesCl, marker->handlesSize, marker->handlesBufOffset, writeMSAAimage->boundMem->bo) : 0;
uint32_t writeMSAAdepthStencilImageIdx = writeMSAAdepthStencilImage ? clGetHandleIndex(&cmdbuf->handlesCl, marker->handlesBufOffset, marker->handlesSize, writeMSAAdepthStencilImage->boundMem->bo) : 0;
clFit(cmdbuf, &cmdbuf->handlesCl, 4 * 6); //just to be safe
uint32_t writeImageIdx = writeImage ? clGetHandleIndex(&cmdbuf->handlesCl, marker->handlesBufOffset + cmdbuf->handlesCl.offset, marker->handlesSize, writeImage->boundMem->bo) : 0;
uint32_t readImageIdx = readImage ? clGetHandleIndex(&cmdbuf->handlesCl, marker->handlesBufOffset + cmdbuf->handlesCl.offset, marker->handlesSize, readImage->boundMem->bo) : 0;
uint32_t writeDepthStencilImageIdx = writeDepthStencilImage ? clGetHandleIndex(&cmdbuf->handlesCl, marker->handlesBufOffset + cmdbuf->handlesCl.offset, marker->handlesSize, writeDepthStencilImage->boundMem->bo) : 0;
uint32_t readDepthStencilImageIdx = readDepthStencilImage ? clGetHandleIndex(&cmdbuf->handlesCl, marker->handlesBufOffset + cmdbuf->handlesCl.offset, marker->handlesSize, readDepthStencilImage->boundMem->bo) : 0;
uint32_t writeMSAAimageIdx = writeMSAAimage ? clGetHandleIndex(&cmdbuf->handlesCl, marker->handlesSize, marker->handlesBufOffset + cmdbuf->handlesCl.offset, writeMSAAimage->boundMem->bo) : 0;
uint32_t writeMSAAdepthStencilImageIdx = writeMSAAdepthStencilImage ? clGetHandleIndex(&cmdbuf->handlesCl, marker->handlesBufOffset + cmdbuf->handlesCl.offset, marker->handlesSize, writeMSAAdepthStencilImage->boundMem->bo) : 0;
// fprintf(stderr, "writeImage: %u\n", writeImage);
// fprintf(stderr, "readImage: %u\n", readImage);
@ -504,10 +505,10 @@ VKAPI_ATTR VkResult VKAPI_CALL rpi_vkQueueSubmit(
submitCl.height = height;
submitCl.flags |= marker->flags;
submitCl.bo_handles = getCPAptrFromOffset(cmdbuf->handlesCl.CPA, marker->handlesBufOffset);
submitCl.bo_handles = getCPAptrFromOffset(cmdbuf->handlesCl.CPA, marker->handlesBufOffset + cmdbuf->handlesCl.offset);
submitCl.bin_cl = ((uint8_t*)marker) + sizeof(CLMarker);
submitCl.shader_rec = getCPAptrFromOffset(cmdbuf->shaderRecCl.CPA, marker->shaderRecBufOffset);
submitCl.uniforms = getCPAptrFromOffset(cmdbuf->uniformsCl.CPA, marker->uniformsBufOffset);
submitCl.shader_rec = getCPAptrFromOffset(cmdbuf->shaderRecCl.CPA, marker->shaderRecBufOffset + cmdbuf->shaderRecCl.offset);
submitCl.uniforms = getCPAptrFromOffset(cmdbuf->uniformsCl.CPA, marker->uniformsBufOffset + cmdbuf->uniformsCl.offset);
if(marker->perfmonID)
{
@ -542,15 +543,15 @@ VKAPI_ATTR VkResult VKAPI_CALL rpi_vkQueueSubmit(
printf("BO handles: ");
for(int d = 0; d < marker->handlesSize / 4; ++d)
{
printf("%u ", *(((uint32_t*)getCPAptrFromOffset(cmdbuf->handlesCl.CPA, marker->handlesBufOffset))+d));
printf("%u ", *(((uint32_t*)getCPAptrFromOffset(cmdbuf->handlesCl.CPA, marker->handlesBufOffset + cmdbuf->handlesCl.offset))+d));
}
printf("\nUniforms: ");
for(int d = 0; d < marker->uniformsSize / 4; ++d)
{
printf("%i ", *(((uint32_t*)getCPAptrFromOffset(cmdbuf->uniformsCl.CPA, marker->uniformsBufOffset))+d));
printf("%i ", *(((uint32_t*)getCPAptrFromOffset(cmdbuf->uniformsCl.CPA, marker->uniformsBufOffset + cmdbuf->uniformsCl.offset))+d));
}
printf("\nShader recs: ");
uint8_t* ptr = getCPAptrFromOffset(cmdbuf->shaderRecCl.CPA, marker->shaderRecBufOffset + (3 + 3) * 4);
uint8_t* ptr = getCPAptrFromOffset(cmdbuf->shaderRecCl.CPA, marker->shaderRecBufOffset + cmdbuf->shaderRecCl.offset + (3 + 3) * 4);
for(int d = 0; d < marker->shaderRecCount; ++d)
{
uint8_t flags = *ptr;

19
driver/draw.c
View File

@ -184,7 +184,7 @@ static uint32_t drawCommon(VkCommandBuffer commandBuffer, int32_t vertexOffset)
//TODO
commandBuffer->shaderRecCount++;
clFit(commandBuffer, &commandBuffer->shaderRecCl, V3D21_SHADER_RECORD_length);
clFit(commandBuffer, &commandBuffer->shaderRecCl, 12 * sizeof(uint32_t) + 104 + 8 * 32);
ControlList relocCl = commandBuffer->shaderRecCl;
uint32_t attribCount = 0;
@ -229,10 +229,14 @@ static uint32_t drawCommon(VkCommandBuffer commandBuffer, int32_t vertexOffset)
{
clInsertNop(&commandBuffer->shaderRecCl);
}
assert(((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->memGuard == 0xDDDDDDDD);
clFit(commandBuffer, &commandBuffer->handlesCl, (3 + 8)*4);
clInsertShaderRecord(&commandBuffer->shaderRecCl,
&relocCl,
&commandBuffer->handlesCl,
((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesBufOffset,
((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesBufOffset + cb->handlesCl.offset,
((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesSize,
!fragModule->hasThreadSwitch,
0, //TODO point size included in shaded vertex data?
@ -318,11 +322,10 @@ static uint32_t drawCommon(VkCommandBuffer commandBuffer, int32_t vertexOffset)
+ cb->vertexBuffers[cb->graphicsPipeline->vertexAttributeDescriptions[c].binding]->boundOffset,
};
clFit(commandBuffer, &commandBuffer->shaderRecCl, V3D21_ATTRIBUTE_RECORD_length);
clInsertAttributeRecord(&commandBuffer->shaderRecCl,
&relocCl,
&commandBuffer->handlesCl,
((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesBufOffset,
((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesBufOffset + cb->handlesCl.offset,
((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesSize,
vertexBuffer, //reloc address
formatByteSize,
@ -363,7 +366,7 @@ static uint32_t drawCommon(VkCommandBuffer commandBuffer, int32_t vertexOffset)
//emit reloc for texture BO
clFit(commandBuffer, &commandBuffer->handlesCl, 4);
uint32_t idx = clGetHandleIndex(&commandBuffer->handlesCl, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesBufOffset, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesSize, di->imageView->image->boundMem->bo);
uint32_t idx = clGetHandleIndex(&commandBuffer->handlesCl, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesBufOffset + cb->handlesCl.offset, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesSize, di->imageView->image->boundMem->bo);
//emit tex bo reloc index
clFit(commandBuffer, &commandBuffer->uniformsCl, 4);
@ -382,7 +385,7 @@ static uint32_t drawCommon(VkCommandBuffer commandBuffer, int32_t vertexOffset)
//emit reloc for BO
clFit(commandBuffer, &commandBuffer->handlesCl, 4);
uint32_t idx = clGetHandleIndex(&commandBuffer->handlesCl, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesBufOffset, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesSize, db->buffer->boundMem->bo);
uint32_t idx = clGetHandleIndex(&commandBuffer->handlesCl, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesBufOffset + cb->handlesCl.offset, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesSize, db->buffer->boundMem->bo);
//emit bo reloc index
clFit(commandBuffer, &commandBuffer->uniformsCl, 4);
@ -399,7 +402,7 @@ static uint32_t drawCommon(VkCommandBuffer commandBuffer, int32_t vertexOffset)
//emit reloc for BO
clFit(commandBuffer, &commandBuffer->handlesCl, 4);
uint32_t idx = clGetHandleIndex(&commandBuffer->handlesCl, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesBufOffset, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesSize, dtb->bufferView->buffer->boundMem->bo);
uint32_t idx = clGetHandleIndex(&commandBuffer->handlesCl, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesBufOffset + cb->handlesCl.offset, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesSize, dtb->bufferView->buffer->boundMem->bo);
//emit bo reloc index
clFit(commandBuffer, &commandBuffer->uniformsCl, 4);
@ -614,7 +617,7 @@ VKAPI_ATTR void VKAPI_CALL rpi_vkCmdDrawIndexed(
assert(((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->memGuard == 0xDDDDDDDD);
clFit(commandBuffer, &commandBuffer->handlesCl, 4);
uint32_t idx = clGetHandleIndex(&commandBuffer->handlesCl, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesBufOffset, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesSize, cb->indexBuffer->boundMem->bo);
uint32_t idx = clGetHandleIndex(&commandBuffer->handlesCl, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesBufOffset + cb->handlesCl.offset, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesSize, cb->indexBuffer->boundMem->bo);
clInsertGEMRelocations(&commandBuffer->binCl, idx, 0);

12
driver/renderpass.c
View File

@ -194,37 +194,37 @@ void rpi_vkCmdBeginRenderPass(VkCommandBuffer commandBuffer, const VkRenderPassB
if(writeImage)
{
clFit(commandBuffer, &commandBuffer->handlesCl, 4);
clGetHandleIndex(&commandBuffer->handlesCl, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesBufOffset, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesSize, writeImage->boundMem->bo);
clGetHandleIndex(&commandBuffer->handlesCl, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesBufOffset + cb->handlesCl.offset, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesSize, writeImage->boundMem->bo);
}
if(readImage)
{
clFit(commandBuffer, &commandBuffer->handlesCl, 4);
clGetHandleIndex(&commandBuffer->handlesCl, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesBufOffset, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesSize, readImage->boundMem->bo);
clGetHandleIndex(&commandBuffer->handlesCl, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesBufOffset + cb->handlesCl.offset, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesSize, readImage->boundMem->bo);
}
if(writeDepthStencilImage)
{
clFit(commandBuffer, &commandBuffer->handlesCl, 4);
clGetHandleIndex(&commandBuffer->handlesCl, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesBufOffset, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesSize, writeDepthStencilImage->boundMem->bo);
clGetHandleIndex(&commandBuffer->handlesCl, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesBufOffset + cb->handlesCl.offset, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesSize, writeDepthStencilImage->boundMem->bo);
}
if(readDepthStencilImage)
{
clFit(commandBuffer, &commandBuffer->handlesCl, 4);
clGetHandleIndex(&commandBuffer->handlesCl, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesBufOffset, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesSize, readDepthStencilImage->boundMem->bo);
clGetHandleIndex(&commandBuffer->handlesCl, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesBufOffset + cb->handlesCl.offset, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesSize, readDepthStencilImage->boundMem->bo);
}
if(writeMSAAimage)
{
clFit(commandBuffer, &commandBuffer->handlesCl, 4);
clGetHandleIndex(&commandBuffer->handlesCl, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesBufOffset, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesSize, writeMSAAimage->boundMem->bo);
clGetHandleIndex(&commandBuffer->handlesCl, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesBufOffset + cb->handlesCl.offset, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesSize, writeMSAAimage->boundMem->bo);
}
if(writeMSAAdepthStencilImage)
{
clFit(commandBuffer, &commandBuffer->handlesCl, 4);
clGetHandleIndex(&commandBuffer->handlesCl, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesBufOffset, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesSize, writeMSAAdepthStencilImage->boundMem->bo);
clGetHandleIndex(&commandBuffer->handlesCl, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesBufOffset + cb->handlesCl.offset, ((CLMarker*)getCPAptrFromOffset(cb->binCl.CPA, cb->binCl.currMarkerOffset))->handlesSize, writeMSAAdepthStencilImage->boundMem->bo);
}
uint32_t bpp = 0;

2
driver/stateChange.c
View File

@ -517,7 +517,7 @@ VKAPI_ATTR void VKAPI_CALL rpi_vkCmdClearColorImage(
//insert reloc for render target
clFit(commandBuffer, &commandBuffer->handlesCl, 4);
clGetHandleIndex(&commandBuffer->handlesCl, ((CLMarker*)getCPAptrFromOffset(commandBuffer->binCl.CPA, commandBuffer->binCl.currMarkerOffset))->handlesBufOffset, ((CLMarker*)getCPAptrFromOffset(commandBuffer->binCl.CPA, commandBuffer->binCl.currMarkerOffset))->handlesSize, i->boundMem->bo);
clGetHandleIndex(&commandBuffer->handlesCl, ((CLMarker*)getCPAptrFromOffset(commandBuffer->binCl.CPA, commandBuffer->binCl.currMarkerOffset))->handlesBufOffset + commandBuffer->handlesCl.offset, ((CLMarker*)getCPAptrFromOffset(commandBuffer->binCl.CPA, commandBuffer->binCl.currMarkerOffset))->handlesSize, i->boundMem->bo);
clFit(commandBuffer, &commandBuffer->binCl, V3D21_TILE_BINNING_MODE_CONFIGURATION_length);
clInsertTileBinningModeConfiguration(&commandBuffer->binCl,