Skip to content

Commit

Permalink
Vulkan: Fix bugs about linear images
Browse files Browse the repository at this point in the history 
1. When piping the image data to UI for the texture view, if the image
is linear and has row pitch, depth pitch larger than a tightly packed
blob, we need to cut off the extra data.

2. For MEC, when pulling image data at the tracing side, if the image is
linear, we may need to change the image level pool size, and when
mapping the data to the pool, consider the row pitch and depth pitch.
  • Loading branch information
@Qining
Qining committed Aug 3, 2018
1 parent 7ff6745 commit 0cdc8ef
Show file tree
Hide file tree
Showing 4 changed files with 183 additions and 65 deletions.
4 changes: 4 additions & 0 deletions core/image/format.go
View file
Original file line number Diff line number Diff line change
Expand Up @@ -64,6 +64,10 @@ var _ = []format{
&FmtETC2_RG_U11_NORM{},
&FmtETC2_R_S11_NORM{},
&FmtETC2_RG_S11_NORM{},
&FmtRGTC1_BC4_R_U8_NORM{},
&FmtRGTC1_BC4_R_S8_NORM{},
&FmtRGTC2_BC5_RG_U8_NORM{},
&FmtRGTC2_BC5_RG_S8_NORM{},
&FmtS3_DXT1_RGB{},
&FmtS3_DXT1_RGBA{},
&FmtS3_DXT3_RGBA{},
Expand Down
175 changes: 126 additions & 49 deletions gapii/cc/vulkan_mid_execution.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -452,38 +452,76 @@ void VulkanSpy::serializeGPUBuffers(StateSerializer *serializer) {
};

struct pitch {
uint32_t height_pitch;
uint32_t depth_pitch;
size_t row_pitch;
size_t depth_pitch;
size_t linear_layout_row_pitch;
size_t linear_layout_depth_pitch;
uint32_t texel_width;
uint32_t texel_height;
uint32_t element_size;
};

// block pitch is calculated with the in-image element size.
auto block_pitch = [this, &get_element_size](
const VkExtent3D &extent, uint32_t format,
uint32_t mip_level, uint32_t aspect_bit) -> pitch {
auto level_pitch = [this, &get_element_size](
gapil::Ref<gapii::ImageObject> img,
uint32_t aspect_bit, uint32_t layer,
uint32_t level) -> pitch {
auto &info = img->mInfo;
auto &lev = img->mAspects[aspect_bit]->mLayers[layer]->mLevels[level];
const bool has_linear_layout =
(lev->mLinearLayout != nullptr) && (lev->mLinearLayout->msize != 0);
auto elementAndTexelBlockSize =
subGetElementAndTexelBlockSize(nullptr, nullptr, format);
subGetElementAndTexelBlockSize(nullptr, nullptr, info.mFormat);
const uint32_t texel_width =
elementAndTexelBlockSize.mTexelBlockSize.mWidth;
const uint32_t texel_height =
elementAndTexelBlockSize.mTexelBlockSize.mHeight;

const uint32_t width =
subGetMipSize(nullptr, nullptr, extent.mWidth, mip_level);
subGetMipSize(nullptr, nullptr, info.mExtent.mWidth, level);
const uint32_t height =
subGetMipSize(nullptr, nullptr, extent.mHeight, mip_level);
subGetMipSize(nullptr, nullptr, info.mExtent.mHeight, level);
const uint32_t width_in_blocks =
subRoundUpTo(nullptr, nullptr, width, texel_width);
const uint32_t height_in_blocks =
subRoundUpTo(nullptr, nullptr, height, texel_height);
const uint32_t element_size =
get_element_size(info.mFormat, aspect_bit, false);
const size_t row_pitch = width_in_blocks * element_size;
const size_t depth_pitch =
width_in_blocks * height_in_blocks * element_size;
pitch p{row_pitch, depth_pitch, 0, 0,
texel_width, texel_height, element_size};
if (has_linear_layout) {
if (lev->mLinearLayout->mdepthPitch != 0) {
p.linear_layout_depth_pitch = lev->mLinearLayout->mdepthPitch;
}
if (lev->mLinearLayout->mrowPitch != 0) {
p.linear_layout_row_pitch = lev->mLinearLayout->mrowPitch;
}
}
return p;
};

// extent pitch is calculated with the in-image element size.
auto extent_pitch = [this, &get_element_size](
const VkExtent3D &extent, uint32_t format,
uint32_t aspect_bit) -> pitch {
auto elementAndTexelBlockSize =
subGetElementAndTexelBlockSize(nullptr, nullptr, format);
const uint32_t texel_width =
elementAndTexelBlockSize.mTexelBlockSize.mWidth;
const uint32_t texel_height =
elementAndTexelBlockSize.mTexelBlockSize.mHeight;

const uint32_t width_in_blocks =
subRoundUpTo(nullptr, nullptr, extent.mWidth, texel_width);
const uint32_t height_in_blocks =
subRoundUpTo(nullptr, nullptr, extent.mHeight, texel_height);
const uint32_t element_size = get_element_size(format, aspect_bit, false);
const size_t size = width_in_blocks * height_in_blocks * element_size;

return pitch{
uint32_t(width_in_blocks * element_size),
uint32_t(size),
uint32_t(width_in_blocks * height_in_blocks * element_size),
uint32_t(elementAndTexelBlockSize.mTexelBlockSize.mWidth),
uint32_t(elementAndTexelBlockSize.mTexelBlockSize.mHeight),
uint32_t(element_size),
Expand Down Expand Up @@ -552,8 +590,12 @@ void VulkanSpy::serializeGPUBuffers(StateSerializer *serializer) {
img->mAspects[aspect]->mLayers[layer]->mLevels[level];
level_sizes[img_level.get()] =
level_size(img->mInfo.mExtent, img->mInfo.mFormat, level, aspect);
serializer->encodeBuffer(level_sizes[img_level.get()].level_size,
&img_level->mData, nullptr);
uint64_t pool_size = level_sizes[img_level.get()].level_size;
if (img_level->mLinearLayout != nullptr &&
img_level->mLinearLayout->msize > pool_size) {
pool_size = img_level->mLinearLayout->msize;
}
serializer->encodeBuffer(pool_size, &img_level->mData, nullptr);
});

if (img->mIsSwapchainImage) {
Expand Down Expand Up @@ -777,29 +819,33 @@ void VulkanSpy::serializeGPUBuffers(StateSerializer *serializer) {
this](
const std::vector<VkBufferImageCopy> &copies,
gapil::Ref<QueueObject> queue) {
const uint32_t queue_family = queue->mFamily;
StagingCommandBuffer commandBuffer(device_functions, img->mDevice,
queue->mFamily);
queue_family);
std::vector<VkImageMemoryBarrier> img_barriers;
std::vector<uint32_t> old_layouts;
walkImageSubRng(
img, img_whole_rng,
[&img, &img_barriers, &old_layouts](
[&img, &img_barriers, &old_layouts, queue_family](
uint32_t aspect_bit, uint32_t layer, uint32_t level) {
auto &img_level =
img->mAspects[aspect_bit]->mLayers[layer]->mLevels[level];
img_barriers.push_back(VkImageMemoryBarrier{
VkStructureType::VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
nullptr,
(VkAccessFlagBits::VK_ACCESS_MEMORY_WRITE_BIT << 1) - 1,
VkAccessFlagBits::VK_ACCESS_TRANSFER_READ_BIT,
img_level->mLayout,
VkImageLayout::VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
kQueueFamilyIgnore,
kQueueFamilyIgnore,
img->mVulkanHandle,
{VkImageAspectFlags(aspect_bit), level, 1, layer, 1},
});
old_layouts.push_back(img_level->mLayout);
if (img_level->mLastBoundQueue != nullptr &&
img_level->mLastBoundQueue->mFamily == queue_family) {
img_barriers.push_back(VkImageMemoryBarrier{
VkStructureType::VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
nullptr,
(VkAccessFlagBits::VK_ACCESS_MEMORY_WRITE_BIT << 1) - 1,
VkAccessFlagBits::VK_ACCESS_TRANSFER_READ_BIT,
img_level->mLayout,
VkImageLayout::VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
kQueueFamilyIgnore,
kQueueFamilyIgnore,
img->mVulkanHandle,
{VkImageAspectFlags(aspect_bit), level, 1, layer, 1},
});
old_layouts.push_back(img_level->mLayout);
}
});
device_functions.vkCmdPipelineBarrier(
commandBuffer.GetBuffer(),
Expand Down Expand Up @@ -838,19 +884,10 @@ void VulkanSpy::serializeGPUBuffers(StateSerializer *serializer) {
: copies_in_order[i + 1].mbufferOffset;
const uint32_t aspect_bit =
(uint32_t)copy.mimageSubresource.maspectMask;
const uint32_t mip_level = copy.mimageSubresource.mmipLevel;
const uint32_t array_layer = copy.mimageSubresource.mbaseArrayLayer;
byte_size_and_extent e =
level_size(copy.mimageExtent, image_info.mFormat, 0, aspect_bit);
auto bp = block_pitch(copy.mimageExtent, image_info.mFormat,
copy.mimageSubresource.mmipLevel, aspect_bit);

if ((copy.mimageOffset.mx % bp.texel_width != 0) ||
(copy.mimageOffset.my % bp.texel_height != 0)) {
// We cannot place partial blocks
return;
}
uint32_t x = (copy.mimageOffset.mx / bp.texel_width) * bp.element_size;
uint32_t y = (copy.mimageOffset.my / bp.texel_height) * bp.height_pitch;
uint32_t z = copy.mimageOffset.mz * bp.depth_pitch;

if ((image_info.mFormat == VkFormat::VK_FORMAT_X8_D24_UNORM_PACK32 ||
image_info.mFormat == VkFormat::VK_FORMAT_D24_UNORM_S8_UINT) &&
Expand Down Expand Up @@ -881,16 +918,56 @@ void VulkanSpy::serializeGPUBuffers(StateSerializer *serializer) {
}
}

memory::Observation observation;
const uint32_t mip_level = copy.mimageSubresource.mmipLevel;
const uint32_t array_layer = copy.mimageSubresource.mbaseArrayLayer;
observation.set_base(x + y + z);
observation.set_pool(img->mAspects[aspect_bit]
->mLayers[array_layer]
->mLevels[mip_level]
->mData.pool_id());
serializer->sendData(&observation, true, pData + new_offset,
e.level_size);
auto bp = level_pitch(img, aspect_bit, array_layer, mip_level);
if ((copy.mimageOffset.mx % bp.texel_width != 0) ||
(copy.mimageOffset.my % bp.texel_height != 0)) {
// We cannot place partial blocks
return;
}
auto &img_level =
img->mAspects[aspect_bit]->mLayers[array_layer]->mLevels[mip_level];
// If the image has linear layout and its row pitch and depth pitch is
// larger than the piches for tightly packed image, we need to set the
// observation row by row. Otherwise, we can use just one observation
// for the extent of this copy.
if (bp.linear_layout_depth_pitch <= bp.depth_pitch &&
bp.linear_layout_row_pitch <= bp.row_pitch) {
uint32_t x =
(copy.mimageOffset.mx / bp.texel_width) * bp.element_size;
uint32_t y = (copy.mimageOffset.my / bp.texel_height) * bp.row_pitch;
uint32_t z = copy.mimageOffset.mz * bp.depth_pitch;
memory::Observation observation;
observation.set_base(x + y + z);
observation.set_pool(img_level->mData.pool_id());
serializer->sendData(&observation, true, pData + new_offset,
e.level_size);

} else {
// Need to set base row by row for linear layout images which have
// larger row pitch and depth pitch
pitch ep =
extent_pitch(copy.mimageExtent, img->mInfo.mFormat, aspect_bit);
for (uint32_t zd = 0; zd < copy.mimageExtent.mDepth; zd++) {
for (uint32_t yd = 0;
yd < subRoundUpTo(nullptr, nullptr, copy.mimageExtent.mHeight,
bp.texel_height);
yd++) {
uint32_t x =
(copy.mimageOffset.mx / bp.texel_width) * bp.element_size;
uint32_t y = ((copy.mimageOffset.my / bp.texel_height) + yd) *
bp.linear_layout_row_pitch;
uint32_t z =
(copy.mimageOffset.mz + zd) * bp.linear_layout_depth_pitch;
uint32_t mem_row_offset =
zd * ep.depth_pitch + yd * ep.row_pitch + new_offset;
memory::Observation observation;
observation.set_base(x + y + z);
observation.set_pool(img_level->mData.pool_id());
serializer->sendData(&observation, true, pData + mem_row_offset,
ep.row_pitch);
}
}
}
new_offset = next_offset;
}
}
Expand Down
14 changes: 8 additions & 6 deletions gapis/api/vulkan/api/image.api
View file
Original file line number Diff line number Diff line change
Expand Up @@ -352,23 +352,25 @@ cmd VkResult vkBindImageMemory(
// stencil element is always 1 byte wide
as!u32(1)
}
size := widthInBlocks * heightInBlocks * level.Depth * elementSize
tightlyPackedSize := widthInBlocks * heightInBlocks * level.Depth * elementSize

// If the image has LINEAR tiling and the image level has layout
// PREINITIALIZED, link the data back to the bound device memory.
// Otherwise creates its own shadow memory pool.
// PREINITIALIZED and size larger than our calculated tightly packed
// size, link the data back to the bound device memory. Otherwise
// creates its own shadow memory pool.
// TODO: If the image as a whole requires more memory than we
// calculated, we should link the data back to the bound device memory
// no matter whether the tiling is LINEAR or OPTIMAL. But we need to
// come up with a 'linear layout' used in GAPID.
// come up with a 'linear layout' in GAPID.
if (imageObject.Info.Tiling == VK_IMAGE_TILING_LINEAR) &&
(level.Layout == VK_IMAGE_LAYOUT_PREINITIALIZED) &&
(level.LinearLayout != null) {
(level.LinearLayout != null) &&
(as!u64(level.LinearLayout.size) > as!u64(tightlyPackedSize)) {
loffset := as!u64(memoryOffset + level.LinearLayout.offset)
lsize := as!u64(level.LinearLayout.size)
level.Data = imageObject.BoundMemory.Data[loffset:loffset + lsize]
} else {
level.Data = make!u8(size)
level.Data = make!u8(tightlyPackedSize)
}
}
}
Expand Down
Loading

0 comments on commit 0cdc8ef

Please sign in to comment.