ShuriZma
/
rpcs3
mirror of https://github.com/RPCS3/rpcs3.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

vk/rsx: Fix some more bugs

This commit is contained in:
kd-11 2020-12-29 23:16:21 +03:00 committed by kd-11
parent 59e46f09b7
commit a1ab6c28c1
2 changed files with 48 additions and 23 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
rpcs3/Emu/RSX/Common/TextureUtils.cpp
View File

@ -735,15 +735,15 @@ namespace rsx
{
if (word_size == 1)
{
if (caps.supports_zero_copy)
if (is_swizzled)
{
copy_unmodified_block_swizzled::copy_mipmap_level(as_span_workaround<u8>(dst_buffer), as_const_span<const u8>(src_layout.data), words_per_block, w, h, depth, src_layout.border, dst_pitch_in_block);
}
else if (caps.supports_zero_copy)
{
result.require_upload = true;
result.deferred_cmds = build_transfer_cmds(src_layout.data.data(), words_per_block, w, h, depth, src_layout.border, dst_pitch_in_block, src_layout.pitch_in_block);
}
else if (is_swizzled)
{
copy_unmodified_block_swizzled::copy_mipmap_level(as_span_workaround<u8>(dst_buffer), as_const_span<const u8>(src_layout.data), words_per_block, w, h, depth, src_layout.border, dst_pitch_in_block);
}
else
{
copy_unmodified_block::copy_mipmap_level(as_span_workaround<u8>(dst_buffer), as_const_span<const u8>(src_layout.data), words_per_block, w, h, depth, src_layout.border, dst_pitch_in_block, src_layout.pitch_in_block);
@ -751,6 +751,9 @@ namespace rsx
}
else
{
result.element_size = word_size;
result.block_length = words_per_block;
bool require_cpu_swizzle = !caps.supports_hw_deswizzle;
bool require_cpu_byteswap = !caps.supports_byteswap;
@ -768,9 +771,7 @@ namespace rsx
if (!require_cpu_byteswap && !require_cpu_swizzle)
{
result.require_deswizzle = is_swizzled;
result.require_swap = true;
result.element_size = word_size;
if (caps.supports_zero_copy)
{

56
rpcs3/Emu/RSX/VK/VKTexture.cpp
View File

@ -809,7 +809,7 @@ namespace vk
u32 block_in_pixel = rsx::get_format_block_size_in_texel(format);
u8 block_size_in_bytes = rsx::get_format_block_size_in_bytes(format);
rsx::texture_uploader_capabilities caps{ true, false, true, true, heap_align };
rsx::texture_uploader_capabilities caps{ .alignment = heap_align };
rsx::texture_memory_info opt{};
bool check_caps = true;
@ -817,13 +817,14 @@ namespace vk
u32 scratch_offset = 0;
u32 row_pitch, image_linear_size;
vk::buffer* upload_buffer = nullptr;
usz offset_in_upload_buffer = 0;
std::vector<VkBufferImageCopy> copy_regions;
std::vector<VkBufferCopy> buffer_copies;
std::vector<std::pair<VkBuffer, u32>> upload_commands;
copy_regions.reserve(subresource_layout.size());
VkBuffer read_buffer = upload_heap.heap->value;
VkDeviceSize offset_in_read_buffer = 0;
if (vk::is_renderpass_open(cmd))
{
vk::end_renderpass(cmd);
@ -854,14 +855,16 @@ namespace vk
image_linear_size = row_pitch * layout.height_in_block * layout.depth;
// Map with extra padding bytes in case of realignment
usz offset_in_buffer = upload_heap.alloc<512>(image_linear_size + 8);
void* mapped_buffer = upload_heap.map(offset_in_buffer, image_linear_size + 8);
offset_in_upload_buffer = upload_heap.alloc<512>(image_linear_size + 8);
void* mapped_buffer = upload_heap.map(offset_in_upload_buffer, image_linear_size + 8);
// Only do GPU-side conversion if occupancy is good
if (check_caps)
{
caps.supports_byteswap = (image_linear_size >= 1024);
caps.supports_hw_deswizzle = caps.supports_byteswap;
caps.supports_zero_copy = caps.supports_byteswap;
caps.supports_vtc_decoding = false;
check_caps = false;
}
@ -871,7 +874,7 @@ namespace vk
copy_regions.push_back({});
auto& copy_info = copy_regions.back();
copy_info.bufferOffset = offset_in_buffer;
copy_info.bufferOffset = offset_in_upload_buffer;
copy_info.imageExtent.height = layout.height_in_texel;
copy_info.imageExtent.width = layout.width_in_texel;
copy_info.imageExtent.depth = layout.depth;
@ -881,6 +884,8 @@ namespace vk
copy_info.imageSubresource.mipLevel = layout.level;
copy_info.bufferRowLength = std::max<u32>(block_in_pixel * row_pitch / block_size_in_bytes, layout.width_in_texel);
upload_buffer = upload_heap.heap.get();
if (opt.require_upload)
{
ensure(!opt.deferred_cmds.empty());
@ -900,12 +905,11 @@ namespace vk
}
auto dma_mapping = vk::map_dma(cmd, static_cast<u32>(src_address), static_cast<u32>(data_length));
vk::load_dma(src_address, data_length);
vk::load_dma(::narrow<u32>(src_address), data_length);
read_buffer = dma_mapping.second->value;
offset_in_read_buffer = dma_mapping.first;
copy_info.bufferOffset = offset_in_read_buffer;
upload_buffer = dma_mapping.second;
offset_in_upload_buffer = dma_mapping.first;
copy_info.bufferOffset = offset_in_upload_buffer;
}
if (opt.require_swap || opt.require_deswizzle || requires_depth_processing)
@ -929,7 +933,7 @@ namespace vk
{
buffer_copies.push_back({});
auto& copy = buffer_copies.back();
copy.srcOffset = uptr(copy_cmd.dst) + offset_in_read_buffer;
copy.srcOffset = uptr(copy_cmd.dst) + offset_in_upload_buffer;
copy.dstOffset = scratch_offset;
copy.size = copy_cmd.length;
}
@ -938,7 +942,7 @@ namespace vk
{
buffer_copies.push_back({});
auto& copy = buffer_copies.back();
copy.srcOffset = offset_in_buffer;
copy.srcOffset = offset_in_upload_buffer;
copy.dstOffset = scratch_offset;
copy.size = image_linear_size;
}
@ -951,15 +955,26 @@ namespace vk
}
else if (opt.require_upload)
{
if (upload_commands.empty() || upload_buffer->value != upload_commands.back().first)
{
upload_commands.emplace_back(upload_buffer->value, 1);
}
else
{
upload_commands.back().second++;
}
copy_info.bufferRowLength = std::max<u32>(block_in_pixel * layout.pitch_in_block, layout.width_in_texel);
}
}
ensure(upload_buffer);
if (opt.require_swap || opt.require_deswizzle || requires_depth_processing)
{
ensure(scratch_buf);
vkCmdCopyBuffer(cmd, read_buffer, scratch_buf->value, static_cast<u32>(buffer_copies.size()), buffer_copies.data());
vkCmdCopyBuffer(cmd, upload_buffer->value, scratch_buf->value, static_cast<u32>(buffer_copies.size()), buffer_copies.data());
insert_buffer_memory_barrier(cmd, scratch_buf->value, 0, scratch_offset, VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
VK_ACCESS_TRANSFER_WRITE_BIT, VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT);
@ -1005,9 +1020,18 @@ namespace vk
vkCmdCopyBufferToImage(cmd, scratch_buf->value, dst_image->value, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, static_cast<u32>(copy_regions.size()), copy_regions.data());
}
else if (opt.require_upload)
{
auto region_ptr = copy_regions.data();
for (const auto& op : upload_commands)
{
vkCmdCopyBufferToImage(cmd, op.first, dst_image->value, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, op.second, region_ptr);
region_ptr += op.second;
}
}
else
{
vkCmdCopyBufferToImage(cmd, read_buffer, dst_image->value, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, static_cast<u32>(copy_regions.size()), copy_regions.data());
vkCmdCopyBufferToImage(cmd, upload_buffer->value, dst_image->value, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, static_cast<u32>(copy_regions.size()), copy_regions.data());
}
}