RPCSX/rpcsx
1
0
Fork 0
mirror of https://github.com/RPCSX/rpcsx.git synced 2026-04-20 22:05:12 +00:00
Code Issues Projects Releases Packages Wiki Activity

rsx/vk: Support ingesting pre-processed GPU data as texture input

Browse source
This commit is contained in:
kd-11 2023-09-14 05:16:36 +03:00 • committed by kd-11
parent bef7d793d3
commit ca054644ef
5 changed files with 104 additions and 26 deletions
Download patch file Download diff file Expand all files Collapse all files

83
rpcs3/Emu/RSX/VK/VKRenderTargets.cpp
View file

@ -1,3 +1,5 @@
#include "VKCompute.h"
#include "VKDMA.h"
#include "VKRenderTargets.h"
#include "VKResourceManager.h"
#include "Emu/RSX/rsx_methods.h"
@ -681,32 +683,75 @@ namespace vk
subres.depth = 1;
subres.data = { vm::get_super_ptr<const std::byte>(base_addr), static_cast<std::span<const std::byte>::size_type>(rsx_pitch * surface_height * samples_y) };
// FIXME: Move to GPU queue
std::vector<std::byte> ext_data;
const auto range = get_memory_range();
rsx::flags32_t upload_flags = upload_contents_inline;
u32 heap_align = rsx_pitch;
if (auto region = rsx::get_current_renderer()->get_tiled_memory_region(range))
if (auto tiled_region = rsx::get_current_renderer()->get_tiled_memory_region(range))
{
auto real_data = vm::get_super_ptr<u8>(range.start);
ext_data.resize(region.tile->size);
rsx::tile_texel_data<u32, true>(
ext_data.data(),
real_data,
region.base_address,
range.start - region.base_address,
region.tile->size,
region.tile->bank,
region.tile->pitch,
subres.width_in_block,
subres.height_in_block
);
subres.data = std::span(ext_data);
const auto available_tile_size = tiled_region.tile->size - (range.start - tiled_region.base_address);
const auto max_content_size = tiled_region.tile->pitch * utils::align<u32>(subres.height_in_block, 64);
const auto section_length = std::min(max_content_size, available_tile_size);
const auto dma_mapping = vk::map_dma(range.start, section_length);
const auto scratch_buf = vk::get_scratch_buffer(cmd, section_length * 3); // 0 = linear data, 1 = padding (deswz), 2 = tiled data
const auto tiled_data_scratch_offset = section_length * 2;
const auto linear_data_scratch_offset = 0;
// Schedule the job
const RSX_detiler_config config =
{
.tile_base_address = tiled_region.base_address,
.tile_base_offset = range.start - tiled_region.base_address,
.tile_size = tiled_region.tile->size,
.tile_pitch = tiled_region.tile->pitch,
.bank = tiled_region.tile->bank,
.dst = scratch_buf,
.dst_offset = linear_data_scratch_offset,
.src = scratch_buf,
.src_offset = section_length * 2,
.image_width = subres.width_in_block,
.image_height = subres.height_in_block,
.image_pitch = subres.width_in_block * static_cast<u32>(get_bpp()),
.image_bpp = get_bpp()
};
// Transfer
VkBufferCopy copy_rgn
{
.srcOffset = dma_mapping.first,
.dstOffset = tiled_data_scratch_offset,
.size = section_length
};
vkCmdCopyBuffer(cmd, dma_mapping.second->value, scratch_buf->value, 1, &copy_rgn);
// Barrier
vk::insert_buffer_memory_barrier(
cmd, scratch_buf->value, linear_data_scratch_offset, section_length,
VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
VK_ACCESS_TRANSFER_WRITE_BIT, VK_ACCESS_SHADER_READ_BIT);
// Detile
vk::get_compute_task<vk::cs_tile_memcpy<RSX_detiler_op::decode>>()->run(cmd, config);
// Barrier
vk::insert_buffer_memory_barrier(
cmd, scratch_buf->value, linear_data_scratch_offset, subres.width_in_block * get_bpp() * subres.height_in_block,
VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT | VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
VK_ACCESS_SHADER_WRITE_BIT, VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_TRANSFER_READ_BIT);
// FIXME: !!EVIL!!
subres.data = { scratch_buf, linear_data_scratch_offset };
upload_flags |= source_is_gpu_resident;
heap_align = subres.width_in_block * get_bpp();
}
if (g_cfg.video.resolution_scale_percent == 100 && spp == 1) [[likely]]
{
push_layout(cmd, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL);
vk::upload_image(cmd, this, { subres }, get_gcm_format(), is_swizzled, 1, aspect(), upload_heap, rsx_pitch, upload_contents_inline);
vk::upload_image(cmd, this, { subres }, get_gcm_format(), is_swizzled, 1, aspect(), upload_heap, heap_align, upload_flags);
pop_layout(cmd);
}
else
@ -735,7 +780,7 @@ namespace vk
}
// Load Cell data into temp buffer
vk::upload_image(cmd, content, { subres }, get_gcm_format(), is_swizzled, 1, aspect(), upload_heap, rsx_pitch, upload_contents_inline);
vk::upload_image(cmd, content, { subres }, get_gcm_format(), is_swizzled, 1, aspect(), upload_heap, heap_align, upload_flags);
// Write into final image
if (content != final_dst)