Skip to content

Workaround for memory types and linear images on Nvidia #28

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
K0bin wants to merge 10 commits into from
Closed

Workaround for memory types and linear images on Nvidia #28

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
b7e36c2
mantle: Ensure buffer barrier region is within buffer bounds
K0bin Oct 22, 2021
c8565e6
mantle: Enforce minimum size of images and viewport
K0bin Oct 22, 2021
1d4a7b0
mantle: Add texem and tile size of additional formats
K0bin Oct 22, 2021
a02a371
mantle: Refactor format check in grCreateImage
K0bin Oct 22, 2021
2a8900b
Revert "mantle: don't defer buffer creation"
K0bin Oct 22, 2021
5d13cde
mantle: don't defer buffer creation
K0bin Oct 22, 2021
bd5471b
mantle: Virtualize memory allocation with vma on Nvidia GPUs
K0bin Oct 13, 2021
b6f838f
mantle: Track linear images that need to be synced
K0bin Oct 22, 2021
519aa33
mantle: Keep linear images that have associated buffers in LAYOUT_GEN…
K0bin Oct 22, 2021
bfe5ffb
mantle: Sync linear image workaround buffer
K0bin Oct 22, 2021
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
19,857 changes: 19,857 additions & 0 deletions include/vk_mem_alloc.h
View file
Open in desktop

Large diffs are not rendered by default.

2 changes: 1 addition & 1 deletion meson.build
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,4 +1,4 @@
project('grvk', 'c', version : '0.5.0', meson_version : '>= 0.46')
project('grvk', ['c', 'cpp'], version : '0.5.0', meson_version : '>= 0.46')

cpu_family = target_machine.cpu_family()

Expand Down
87 changes: 80 additions & 7 deletions src/mantle/mantle_cmd_buf.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -314,6 +314,19 @@ static void grCmdBufferUpdateResources(
bindPoint->dirtyFlags = 0;
}

void grCmdTrackImage(GrCmdBuffer* grCmdBuffer, GrImage* image) {
if (image->image == VK_NULL_HANDLE || image->buffer == VK_NULL_HANDLE) {
return;
}

if (grCmdBuffer->linearImageCapacity == grCmdBuffer->linearImageCount) {
grCmdBuffer->linearImageCapacity += 8;
grCmdBuffer->linearImages = realloc(grCmdBuffer->linearImages, sizeof(GrImage*) * grCmdBuffer->linearImageCapacity);
}
grCmdBuffer->linearImages[grCmdBuffer->linearImageCount] = image;
grCmdBuffer->linearImageCount++;
}

// Command Buffer Building Functions

GR_VOID GR_STDCALL grCmdBindPipeline(
Expand Down Expand Up @@ -491,6 +504,15 @@ GR_VOID GR_STDCALL grCmdBindDescriptorSet(
bindPoint->grDescriptorSet = grDescriptorSet;
bindPoint->slotOffset = slotOffset;
bindPoint->dirtyFlags |= FLAG_DIRTY_DESCRIPTOR_SETS;

if (grDescriptorSet->linearImageCount != 0) {
if (grCmdBuffer->linearImageCapacity < grCmdBuffer->linearImageCount + grDescriptorSet->linearImageCount) {
grCmdBuffer->linearImageCapacity = max(grCmdBuffer->linearImageCount + grDescriptorSet->linearImageCount, grCmdBuffer->linearImageCapacity + 8);
grCmdBuffer->linearImages = realloc(grCmdBuffer->linearImages, grCmdBuffer->linearImageCapacity * sizeof(GrImage*));
}
memcpy(grCmdBuffer->linearImages + grCmdBuffer->linearImageCount, grDescriptorSet->linearImages, grDescriptorSet->linearImageCount * sizeof(GrImage*));
grCmdBuffer->linearImageCount += grDescriptorSet->linearImageCount;
}
}

GR_VOID GR_STDCALL grCmdBindDynamicMemoryView(
Expand Down Expand Up @@ -543,6 +565,8 @@ GR_VOID GR_STDCALL grCmdBindIndexData(
const GrDevice* grDevice = GET_OBJ_DEVICE(grCmdBuffer);
GrGpuMemory* grGpuMemory = (GrGpuMemory*)mem;

grGpuMemoryBindBuffer(grGpuMemory);

VKD.vkCmdBindIndexBuffer(grCmdBuffer->commandBuffer, grGpuMemory->buffer, offset,
getVkIndexType(indexType));
}
Expand All @@ -566,6 +590,8 @@ GR_VOID GR_STDCALL grCmdPrepareMemoryRegions(
const GR_MEMORY_STATE_TRANSITION* stateTransition = &pStateTransitions[i];
GrGpuMemory* grGpuMemory = (GrGpuMemory*)stateTransition->mem;

grGpuMemoryBindBuffer(grGpuMemory);

barriers[i] = (VkBufferMemoryBarrier) {
.sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER,
.pNext = NULL,
Expand All @@ -575,7 +601,7 @@ GR_VOID GR_STDCALL grCmdPrepareMemoryRegions(
.dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
.buffer = grGpuMemory->buffer,
.offset = stateTransition->offset,
.size = stateTransition->regionSize > 0 ? stateTransition->regionSize : VK_WHOLE_SIZE,
.size = stateTransition->regionSize > 0 ? min(grGpuMemory->deviceSize - stateTransition->offset, stateTransition->regionSize) : VK_WHOLE_SIZE,
};

srcStageMask |= getVkPipelineStageFlagsMemory(stateTransition->oldState);
Expand Down Expand Up @@ -667,6 +693,15 @@ GR_VOID GR_STDCALL grCmdBindTargets(
minExtent.depth = MIN(minExtent.depth, grDepthStencilView->extent.depth);
}

for (unsigned i = 0; i < colorTargetCount; i++) {
const GrColorTargetView* grColorTargetView = (GrColorTargetView*)pColorTargets[i].view;

if (grColorTargetView != NULL) {
GrImage* grImage = (GrImage*)grColorTargetView->image;
grCmdTrackImage(grCmdBuffer, grImage);
}
}

if (colorAttachmentCount != grCmdBuffer->colorAttachmentCount ||
memcmp(colorAttachments, grCmdBuffer->colorAttachments, colorAttachmentCount * sizeof(colorAttachments[0])) ||
hasDepthStencil != grCmdBuffer->hasDepthStencil ||
Expand Down Expand Up @@ -709,7 +744,11 @@ GR_VOID GR_STDCALL grCmdPrepareImages(

grCmdBufferEndRenderPass(grCmdBuffer);

unsigned bufferBarrierCount = 0;
unsigned imageBarrierCount = 0;

STACK_ARRAY(VkImageMemoryBarrier, barriers, 128, transitionCount);
STACK_ARRAY(VkBufferMemoryBarrier, bufferBarriers, 128, transitionCount);
VkPipelineStageFlags srcStageMask = 0;
VkPipelineStageFlags dstStageMask = 0;

Expand All @@ -718,27 +757,47 @@ GR_VOID GR_STDCALL grCmdPrepareImages(
GrImage* grImage = (GrImage*)stateTransition->image;
bool isDepthStencil = isVkFormatDepthStencil(grImage->format);

barriers[i] = (VkImageMemoryBarrier) {
if (grImage->buffer != VK_NULL_HANDLE) {
bufferBarriers[bufferBarrierCount] = (VkBufferMemoryBarrier) {
.sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER,
.pNext = NULL,
.srcAccessMask = getVkAccessFlagsImage(stateTransition->oldState, false),
.dstAccessMask = getVkAccessFlagsImage(stateTransition->newState, false),
.srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
.dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
.buffer = grImage->buffer,
.offset = 0,
.size = VK_WHOLE_SIZE,
};
bufferBarrierCount++;

if (grImage->image == VK_NULL_HANDLE) {
continue;
}
}

barriers[imageBarrierCount] = (VkImageMemoryBarrier) {
.sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
.pNext = NULL,
.srcAccessMask = getVkAccessFlagsImage(stateTransition->oldState, isDepthStencil),
.dstAccessMask = getVkAccessFlagsImage(stateTransition->newState, isDepthStencil),
.oldLayout = getVkImageLayout(stateTransition->oldState, isDepthStencil),
.newLayout = getVkImageLayout(stateTransition->newState, isDepthStencil),
.oldLayout = grImage->buffer == VK_NULL_HANDLE ? getVkImageLayout(stateTransition->oldState, isDepthStencil) : VK_IMAGE_LAYOUT_GENERAL,
.newLayout = grImage->buffer == VK_NULL_HANDLE ? getVkImageLayout(stateTransition->newState, isDepthStencil) : VK_IMAGE_LAYOUT_GENERAL,
.srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
.dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
.image = grImage->image,
.subresourceRange = getVkImageSubresourceRange(stateTransition->subresourceRange,
grImage->multiplyCubeLayers),
grImage->multiplyCubeLayers),
};

srcStageMask |= getVkPipelineStageFlagsImage(stateTransition->oldState);
dstStageMask |= getVkPipelineStageFlagsImage(stateTransition->newState);
imageBarrierCount++;
}

VKD.vkCmdPipelineBarrier(grCmdBuffer->commandBuffer, srcStageMask, dstStageMask,
0, 0, NULL, 0, NULL, transitionCount, barriers);

0, 0, NULL, bufferBarrierCount, bufferBarriers, imageBarrierCount, barriers);
STACK_ARRAY_FINISH(bufferBarriers);
STACK_ARRAY_FINISH(barriers);
}

Expand Down Expand Up @@ -840,6 +899,7 @@ GR_VOID GR_STDCALL grCmdDispatchIndirect(

grCmdBufferUpdateResources(grCmdBuffer, VK_PIPELINE_BIND_POINT_COMPUTE);
grCmdBufferEndRenderPass(grCmdBuffer);
grGpuMemoryBindBuffer(grGpuMemory);

VKD.vkCmdDispatchIndirect(grCmdBuffer->commandBuffer, grGpuMemory->buffer, offset);
}
Expand All @@ -858,6 +918,8 @@ GR_VOID GR_STDCALL grCmdCopyMemory(
GrGpuMemory* grDstGpuMemory = (GrGpuMemory*)destMem;

grCmdBufferEndRenderPass(grCmdBuffer);
grGpuMemoryBindBuffer(grSrcGpuMemory);
grGpuMemoryBindBuffer(grDstGpuMemory);

STACK_ARRAY(VkBufferCopy, vkRegions, 128, regionCount);

Expand Down Expand Up @@ -889,6 +951,8 @@ GR_VOID GR_STDCALL grCmdCopyImage(
const GrDevice* grDevice = GET_OBJ_DEVICE(grCmdBuffer);
GrImage* grSrcImage = (GrImage*)srcImage;
GrImage* grDstImage = (GrImage*)destImage;
grCmdTrackImage(grCmdBuffer, grSrcImage);
grCmdTrackImage(grCmdBuffer, grDstImage);
unsigned srcTileSize = getVkFormatTileSize(grSrcImage->format);
unsigned dstTileSize = getVkFormatTileSize(grDstImage->format);
unsigned extentTileSize = srcTileSize > dstTileSize ? dstTileSize : srcTileSize;
Expand Down Expand Up @@ -996,13 +1060,15 @@ GR_VOID GR_STDCALL grCmdCopyMemoryToImage(
const GrDevice* grDevice = GET_OBJ_DEVICE(grCmdBuffer);
GrGpuMemory* grSrcGpuMemory = (GrGpuMemory*)srcMem;
GrImage* grDstImage = (GrImage*)destImage;
grCmdTrackImage(grCmdBuffer, grDstImage);
unsigned dstTileSize = getVkFormatTileSize(grDstImage->format);

if (quirkHas(QUIRK_COMPRESSED_IMAGE_COPY_IN_TEXELS)) {
dstTileSize = 1;
}

grCmdBufferEndRenderPass(grCmdBuffer);
grGpuMemoryBindBuffer(grSrcGpuMemory);

STACK_ARRAY(VkBufferImageCopy, vkRegions, 128, regionCount);

Expand Down Expand Up @@ -1049,13 +1115,15 @@ GR_VOID GR_STDCALL grCmdCopyImageToMemory(
const GrDevice* grDevice = GET_OBJ_DEVICE(grCmdBuffer);
GrImage* grSrcImage = (GrImage*)srcImage;
GrGpuMemory* grDstGpuMemory = (GrGpuMemory*)destMem;
grCmdTrackImage(grCmdBuffer, grSrcImage);
unsigned srcTileSize = getVkFormatTileSize(grSrcImage->format);

if (quirkHas(QUIRK_COMPRESSED_IMAGE_COPY_IN_TEXELS)) {
srcTileSize = 1;
}

grCmdBufferEndRenderPass(grCmdBuffer);
grGpuMemoryBindBuffer(grDstGpuMemory);

STACK_ARRAY(VkBufferImageCopy, vkRegions, 128, regionCount);

Expand Down Expand Up @@ -1102,6 +1170,7 @@ GR_VOID GR_STDCALL grCmdUpdateMemory(
GrGpuMemory* grDstGpuMemory = (GrGpuMemory*)destMem;

grCmdBufferEndRenderPass(grCmdBuffer);
grGpuMemoryBindBuffer(grDstGpuMemory);

VKD.vkCmdUpdateBuffer(grCmdBuffer->commandBuffer, grDstGpuMemory->buffer, destOffset,
dataSize, pData);
Expand All @@ -1120,6 +1189,7 @@ GR_VOID GR_STDCALL grCmdFillMemory(
GrGpuMemory* grDstGpuMemory = (GrGpuMemory*)destMem;

grCmdBufferEndRenderPass(grCmdBuffer);
grGpuMemoryBindBuffer(grDstGpuMemory);

VKD.vkCmdFillBuffer(grCmdBuffer->commandBuffer, grDstGpuMemory->buffer, destOffset,
fillSize, data);
Expand All @@ -1137,6 +1207,7 @@ GR_VOID GR_STDCALL grCmdClearColorImage(
GrCmdBuffer* grCmdBuffer = (GrCmdBuffer*)cmdBuffer;
const GrDevice* grDevice = GET_OBJ_DEVICE(grCmdBuffer);
GrImage* grImage = (GrImage*)image;
grCmdTrackImage(grCmdBuffer, grImage);

grCmdBufferEndRenderPass(grCmdBuffer);

Expand Down Expand Up @@ -1169,6 +1240,7 @@ GR_VOID GR_STDCALL grCmdClearColorImageRaw(
GrCmdBuffer* grCmdBuffer = (GrCmdBuffer*)cmdBuffer;
const GrDevice* grDevice = GET_OBJ_DEVICE(grCmdBuffer);
GrImage* grImage = (GrImage*)image;
grCmdTrackImage(grCmdBuffer, grImage);

grCmdBufferEndRenderPass(grCmdBuffer);

Expand Down Expand Up @@ -1201,6 +1273,7 @@ GR_VOID GR_STDCALL grCmdClearDepthStencil(
GrCmdBuffer* grCmdBuffer = (GrCmdBuffer*)cmdBuffer;
const GrDevice* grDevice = GET_OBJ_DEVICE(grCmdBuffer);
GrImage* grImage = (GrImage*)image;
grCmdTrackImage(grCmdBuffer, grImage);

grCmdBufferEndRenderPass(grCmdBuffer);

Expand Down
68 changes: 68 additions & 0 deletions src/mantle/mantle_cmd_buf_man.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,6 +12,12 @@ void grCmdBufferResetState(
VKD.vkResetDescriptorPool(grDevice->device, grCmdBuffer->descriptorPools[i], 0);
}

// Reset tracked linear images
grCmdBuffer->linearImageCount = 0;
free(grCmdBuffer->linearImages);
grCmdBuffer->linearImageCapacity = 8;
grCmdBuffer->linearImages = calloc(sizeof(GrImage*), grCmdBuffer->linearImageCapacity);

// Clear state
unsigned stateOffset = OFFSET_OF(GrCmdBuffer, isBuilding);
memset(&((uint8_t*)grCmdBuffer)[stateOffset], 0, sizeof(GrCmdBuffer) - stateOffset);
Expand Down Expand Up @@ -135,6 +141,9 @@ GR_RESULT GR_STDCALL grCreateCommandBuffer(
.stencilAttachment = { 0 },
.depthStencilFormat = 0,
.minExtent = { 0, 0, 0 },
.linearImages = calloc(sizeof(GrImage*), 8),
.linearImageCount = 0,
.linearImageCapacity = 8,
};

*pCmdBuffer = (GR_CMD_BUFFER)grCmdBuffer;
Expand Down Expand Up @@ -203,6 +212,65 @@ GR_RESULT GR_STDCALL grEndCommandBuffer(

grCmdBufferEndRenderPass(grCmdBuffer);

if (grCmdBuffer->linearImageCount != 0) {
VkBufferImageCopy* regions = NULL;
int regionsCapacity = 0;
for (int i = 0; i < grCmdBuffer->linearImageCount; i++) {
GrImage* image = grCmdBuffer->linearImages[i];
int imageRegionsCount = image->arrayLayers * image->mipLevels;
if (regions == NULL) {
regions = malloc(sizeof(VkBufferImageCopy) * imageRegionsCount);
regionsCapacity = imageRegionsCount;
} else if (regionsCapacity < imageRegionsCount) {
regions = realloc(regions, sizeof(VkBufferImageCopy) * imageRegionsCount);
regionsCapacity = imageRegionsCount;
}
for (int a = 0; a < image->arrayLayers; a++) {
for (int m = 0; m < image->mipLevels; m++) {
if (getVkFormatTileSize(image->format) != 1) {
LOGE("Block compressed linear images aren't supported");
continue;
}

VkImageAspectFlags aspectMask = 0;
if (isVkFormatDepth(image->format)) {
aspectMask |= VK_IMAGE_ASPECT_DEPTH_BIT;
}
if (isVkFormatStencil(image->format)) {
aspectMask |= VK_IMAGE_ASPECT_STENCIL_BIT;
}
if (aspectMask == 0) {
aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
}

int layerIndex = a * image->mipLevels + m;
regions[layerIndex] = (VkBufferImageCopy) {
.bufferOffset = grImageGetBufferOffset(image->extent, image->format, a, image->arrayLayers, m),
.bufferRowLength = 0,
.imageSubresource.aspectMask = aspectMask,
.imageSubresource.mipLevel = m,
.imageSubresource.baseArrayLayer = a,
.imageSubresource.layerCount = 1,
.imageOffset = (VkOffset3D) {
.x = 0,
.y = 0,
.z = 0,
},
.imageExtent = (VkExtent3D) {
.width = max(1, image->extent.width >> m),
.height = max(1, image->extent.height >> m),
.depth = max(1, image->extent.depth >> m),
},
};
}
}
VKD.vkCmdCopyImageToBuffer(grCmdBuffer->commandBuffer, image->image, VK_IMAGE_LAYOUT_GENERAL, image->buffer, imageRegionsCount, regions);
}
if (regions != NULL) {
free(regions);
}
}

VkResult res = VKD.vkEndCommandBuffer(grCmdBuffer->commandBuffer);
if (res != VK_SUCCESS) {
LOGE("vkEndCommandBuffer failed (%d)\n", res);
Expand Down
Loading