lee/renderdoc
1
0
Fork 0
mirror of https://github.com/baldurk/renderdoc.git synced 2026-09-23 14:15:42 +00:00
Code Issues Packages Projects Releases Wiki Activity

Properly handle offsets on memory maps in vulkan

Browse Source
This commit is contained in:
baldurk
2020-08-04 13:59:45 +01:00
parent 27b541c6d5
commit 7b1995a7ba
5 changed files with 94 additions and 37 deletions
Download Patch File Download Diff File Expand all files Collapse all files
renderdoc/driver/vulkan/vk_resources.h
+8 -16
View File
@@ -1080,22 +1080,14 @@ struct PipelineLayoutData
struct MemMapState
{
MemMapState()
: mapOffset(0),
mapSize(0),
needRefData(false),
mapFlushed(false),
mapCoherent(false),
mappedPtr(NULL),
refData(NULL)
{
}
VkDeviceSize mapOffset, mapSize;
bool needRefData;
bool mapFlushed;
bool mapCoherent;
byte *mappedPtr;
byte *refData;
VkDeviceSize mapOffset = 0, mapSize = 0;
bool needRefData = false;
bool mapFlushed = false;
bool mapCoherent = false;
// pointer to base of memory, may not be valid until after mapOffset bytes
byte *mappedPtr = NULL;
// this is map sized, not memory sized, rebased at the map offset.
byte *refData = NULL;
Threading::CriticalSection mrLock;
};
renderdoc/driver/vulkan/wrappers/vk_queue_funcs.cpp
+2 -2
View File
@@ -1079,8 +1079,8 @@ VkResult WrappedVulkan::vkQueueSubmit(VkQueue queue, uint32_t submitCount,
// if we have a previous set of data, compare.
// otherwise just serialise it all
if(state.refData)
found = FindDiffRange((byte *)state.mappedPtr, state.refData, (size_t)state.mapSize,
diffStart, diffEnd);
found = FindDiffRange(((byte *)state.mappedPtr) + state.mapOffset, state.refData,
(size_t)state.mapSize, diffStart, diffEnd);
else
#endif
diffEnd = (size_t)state.mapSize;
renderdoc/driver/vulkan/wrappers/vk_resource_funcs.cpp
+20 -8
View File
@@ -581,9 +581,19 @@ void WrappedVulkan::vkFreeMemory(VkDevice device, VkDeviceMemory memory,
VkResult WrappedVulkan::vkMapMemory(VkDevice device, VkDeviceMemory mem, VkDeviceSize offset,
VkDeviceSize size, VkMemoryMapFlags flags, void **ppData)
{
void *realData = NULL;
VkResult ret =
ObjDisp(device)->MapMemory(Unwrap(device), Unwrap(mem), offset, size, flags, &realData);
// ensure we always map on a 16-byte boundary. This is for our own purposes so we can
// FindDiffRange against the mapped region. We adjust the pointer returned to the user but
// otherwise we act as if the mapped region was 16-byte aligned. Fortunately flushed regions in
// vkFlushMappedMemoryRanges are relative to the memory base, not the mapped region, so this
// offset effectively only modifies the returned pointer and has no other side-effects.
VkDeviceSize misalignedOffset = offset & 0xf;
offset &= ~0xf;
// need to adjust the size so the end-point is still the same!
size += misalignedOffset;
byte *realData = NULL;
VkResult ret = ObjDisp(device)->MapMemory(Unwrap(device), Unwrap(mem), offset, size, flags,
(void **)&realData);
if(ret == VK_SUCCESS && realData)
{
@@ -599,9 +609,11 @@ VkResult WrappedVulkan::vkMapMemory(VkDevice device, VkDeviceMemory mem, VkDevic
MemMapState &state = *memrecord->memMapState;
// ensure size is valid
RDCASSERT(size == VK_WHOLE_SIZE || (size > 0 && size <= memrecord->Length), GetResID(mem),
size, memrecord->Length);
RDCASSERT(size == VK_WHOLE_SIZE || (size > 0 && offset + size <= memrecord->Length),
GetResID(mem), size, memrecord->Length);
// flush range offsets are relative to the start of the memory so keep mappedPtr at that
// basis. We'll only access within the mapped range
state.mappedPtr = (byte *)realData - (size_t)offset;
state.refData = NULL;
@@ -609,7 +621,7 @@ VkResult WrappedVulkan::vkMapMemory(VkDevice device, VkDeviceMemory mem, VkDevic
state.mapSize = size == VK_WHOLE_SIZE ? (memrecord->Length - offset) : size;
state.mapFlushed = false;
*ppData = realData;
*ppData = realData + misalignedOffset;
if(state.mapCoherent)
{
@@ -619,7 +631,7 @@ VkResult WrappedVulkan::vkMapMemory(VkDevice device, VkDeviceMemory mem, VkDevic
}
else
{
*ppData = realData;
*ppData = realData + misalignedOffset;
}
}
else
@@ -816,7 +828,7 @@ bool WrappedVulkan::Serialise_vkFlushMappedMemoryRanges(SerialiserType &ser, VkD
if(!state->refData)
{
// if we're in this case, the range should be for the whole memory region.
RDCASSERT(MemRange.offset == 0 && memRangeSize == state->mapSize);
RDCASSERT(MemRange.offset == state->mapOffset && memRangeSize == state->mapSize);
// allocate ref data so we can compare next time to minimise serialised data
state->refData = AllocAlignedBuffer((size_t)state->mapSize);
util/test/demos/vk/vk_misaligned_dirty.cpp
+29 -11
View File
@@ -58,27 +58,39 @@ RD_TEST(VK_Misaligned_Dirty, VulkanGraphicsTest)
const float val = 2.0f / 3.0f;
const DefaultA2V tri[4] = {
DefaultA2V tri[4] = {
{Vec3f(-val, -val, val), Vec4f(0.0f, 1.0f, 0.0f, 1.0f), Vec2f(0.0f, 0.0f)},
{Vec3f(0.0f, val, val), Vec4f(0.0f, 1.0f, 0.0f, 1.0f), Vec2f(0.0f, 1.0f)},
{Vec3f(val, -val, val), Vec4f(0.0f, 1.0f, 0.0f, 1.0f), Vec2f(1.0f, 0.0f)},
{},
};
AllocatedBuffer vb(this, vkh::BufferCreateInfo(sizeof(tri), VK_BUFFER_USAGE_VERTEX_BUFFER_BIT |
VK_BUFFER_USAGE_TRANSFER_DST_BIT),
VmaAllocationCreateInfo({0, VMA_MEMORY_USAGE_CPU_TO_GPU}));
vb.upload(tri);
AllocatedBuffer copy_src(
this, vkh::BufferCreateInfo(
sizeof(tri), VK_BUFFER_USAGE_TRANSFER_SRC_BIT | VK_BUFFER_USAGE_TRANSFER_DST_BIT),
VmaAllocationCreateInfo({0, VMA_MEMORY_USAGE_CPU_TO_GPU}));
setName(copy_src.buffer, "copy_src");
AllocatedBuffer vb(this, vkh::BufferCreateInfo(sizeof(tri), VK_BUFFER_USAGE_VERTEX_BUFFER_BIT |
VK_BUFFER_USAGE_TRANSFER_DST_BIT),
VmaAllocationCreateInfo({0, VMA_MEMORY_USAGE_CPU_TO_GPU}));
setName(vb.buffer, "vb");
vb.upload(tri);
tri[0].pos = Vec3f(0.0f, 0.0f, 10.0f);
copy_src.upload(tri);
float *mapped = (float *)(copy_src.map() + sizeof(DefaultA2V) * 3);
VmaAllocationInfo alloc_info;
vmaGetAllocationInfo(copy_src.allocator, copy_src.alloc, &alloc_info);
float *mapped = NULL;
vkMapMemory(device, alloc_info.deviceMemory, alloc_info.offset + sizeof(DefaultA2V) * 3,
sizeof(Vec4f), 0, (void **)&mapped);
float counter = 0;
while(Running())
@@ -91,6 +103,7 @@ RD_TEST(VK_Misaligned_Dirty, VulkanGraphicsTest)
// create a dummy submit which uses the memory. This will serialise the whole memory contents
// (we don't create reference data until after this)
vkBeginCommandBuffer(cmd, vkh::CommandBufferBeginInfo());
setMarker(cmd, "First Submit");
vkCmdUpdateBuffer(cmd, copy_src.buffer, sizeof(Vec3f), sizeof(Vec4f), &tri[0].col);
vkEndCommandBuffer(cmd);
Submit(0, 2, {cmd});
@@ -110,10 +123,14 @@ RD_TEST(VK_Misaligned_Dirty, VulkanGraphicsTest)
vkh::ImageSubresourceRange());
VkBufferCopy region = {};
region.srcOffset = 3;
region.dstOffset = 3;
region.srcOffset = 3 + sizeof(DefaultA2V);
region.dstOffset = 3 + sizeof(DefaultA2V);
region.size = 7;
vkCmdCopyBuffer(cmd, copy_src.buffer, vb.buffer, 1, &region);
region.srcOffset = sizeof(DefaultA2V) * 3;
region.dstOffset = sizeof(DefaultA2V) * 3;
region.size = sizeof(DefaultA2V);
vkCmdCopyBuffer(cmd, copy_src.buffer, vb.buffer, 1, &region);
vkCmdBeginRenderPass(
cmd, vkh::RenderPassBeginInfo(mainWindow->rp, mainWindow->GetFB(), mainWindow->scissor),
@@ -123,6 +140,7 @@ RD_TEST(VK_Misaligned_Dirty, VulkanGraphicsTest)
vkCmdSetViewport(cmd, 0, 1, &mainWindow->viewport);
vkCmdSetScissor(cmd, 0, 1, &mainWindow->scissor);
vkh::cmdBindVertexBuffers(cmd, 0, {vb.buffer}, {0});
setMarker(cmd, "Second Submit");
vkCmdDraw(cmd, 3, 1, 0, 0);
vkCmdEndRenderPass(cmd);
@@ -136,7 +154,7 @@ RD_TEST(VK_Misaligned_Dirty, VulkanGraphicsTest)
Present();
}
copy_src.unmap();
vkUnmapMemory(device, alloc_info.deviceMemory);
return 0;
}
util/test/tests/Vulkan/VK_Misaligned_Dirty.py
+35
View File
@@ -1,4 +1,5 @@
import renderdoc as rd
import struct
import rdtest
@@ -74,3 +75,37 @@ class VK_Misaligned_Dirty(rdtest.TestCase):
self.check_pixel_value(tex, coord[0], coord[1], [0.0, 1.0, 0.0, 1.0])
rdtest.log.success("picked values are as expected")
checkpoint1 = self.find_draw("First Submit")
checkpoint2 = self.find_draw("Second Submit")
self.check(checkpoint1 is not None)
self.check(checkpoint2 is not None)
resources = self.controller.GetResources()
copy_src = None
vb = None
for r in resources:
if r.name == 'copy_src':
copy_src = r.resourceId
elif r.name == 'vb':
vb = r.resourceId
self.check(copy_src is not None)
self.check(vb is not None)
self.controller.SetFrameEvent(checkpoint1.eventId, False)
val = struct.unpack('f', self.controller.GetBufferData(copy_src, 116, 4))
self.check(val[0] == 10.0)
self.controller.SetFrameEvent(checkpoint2.eventId, False)
val = struct.unpack('f', self.controller.GetBufferData(copy_src, 116, 4))
self.check(val[0] == 11.0)
val = struct.unpack('f', self.controller.GetBufferData(vb, 116, 4))
self.check(val[0] == 11.0)
rdtest.log.success("buffers have correct values in both submits")