RPCS3/rpcs3
1
0
Fork 0
mirror of https://github.com/RPCS3/rpcs3.git synced 2024-11-22 18:53:28 +01:00
Code Issues Releases Wiki Activity

rsx/blit: Remove workarounds/hacks added for master. Start implementation/stubs for blit engine rotations in GPU

Browse Source
This commit is contained in:
kd-11 2019-03-11 20:34:34 +03:00 committed by kd-11
parent 745f8f9627
commit 17c49d21a5
2 changed files with 59 additions and 33 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

80
rpcs3/Emu/RSX/Common/texture_cache.h
View File

@ -2119,8 +2119,7 @@ namespace rsx
texaddr, tex.format(), tex_width, tex_height, depth, tex_pitch, slice_h,
extended_dimension, tex.remap(), tex.decoded_remap(), _pool);
if (!result.external_subresource_desc.sections_to_copy.empty() &&
(_pool == 0 || result.atlas_covers_target_area()))
if (!result.external_subresource_desc.sections_to_copy.empty() && result.atlas_covers_target_area())
{
// TODO: Investigate why a full re-upload can cause problems in some games (yellow flicker in SCV)
// Unimplemented readback formats?
@ -2176,8 +2175,18 @@ namespace rsx
const u32 src_address = (u32)((u64)src.pixels - (u64)vm::base(0));
const u32 dst_address = (u32)((u64)dst.pixels - (u64)vm::base(0));
f32 scale_x = dst.scale_x;
f32 scale_y = dst.scale_y;
const f32 scale_x = fabsf(dst.scale_x);
const f32 scale_y = fabsf(dst.scale_y);
if (dst.scale_y < 0.f)
{
// TODO
}
if (dst.scale_x < 0.f)
{
// TODO
}
// Offset in x and y for src is 0 (it is already accounted for when getting pixels_src)
// Reproject final clip onto source...
@ -2210,6 +2219,32 @@ namespace rsx
if (!g_cfg.video.use_gpu_texture_scaling && !(src_is_render_target || dst_is_render_target))
return false;
// Check if trivial memcpy can perform the same task
// Used to copy programs and arbitrary data to the GPU in some cases
if (!src_is_render_target && !dst_is_render_target && dst_is_argb8 == src_is_argb8 && !dst.swizzled)
{
if ((src.slice_h == 1 && dst.clip_height == 1) ||
(dst.clip_width == src.width && dst.clip_height == src.slice_h && src.pitch == dst.pitch))
{
if (dst.scale_x > 0.f && dst.scale_y > 0.f)
{
const u8 bpp = dst_is_argb8 ? 4 : 2;
const u32 memcpy_bytes_length = dst.clip_width * bpp * dst.clip_height;
std::lock_guard lock(m_cache_mutex);
invalidate_range_impl_base(cmd, address_range::start_length(src_address, memcpy_bytes_length), invalidation_cause::read, std::forward<Args>(extras)...);
invalidate_range_impl_base(cmd, address_range::start_length(dst_address, memcpy_bytes_length), invalidation_cause::write, std::forward<Args>(extras)...);
memcpy(dst.pixels, src.pixels, memcpy_bytes_length);
return true;
}
else
{
// Rotation transform applied, use fallback
return false;
}
}
}
if (src_is_render_target)
{
src_subres.surface->read_barrier(cmd);
@ -2260,42 +2295,33 @@ namespace rsx
}
}
//Check if trivial memcpy can perform the same task
//Used to copy programs to the GPU in some cases
if (!src_is_render_target && !dst_is_render_target && dst_is_argb8 == src_is_argb8 && !dst.swizzled)
{
if ((src.slice_h == 1 && dst.clip_height == 1) ||
(dst.clip_width == src.width && dst.clip_height == src.slice_h && src.pitch == dst.pitch))
{
const u8 bpp = dst_is_argb8 ? 4 : 2;
const u32 memcpy_bytes_length = dst.clip_width * bpp * dst.clip_height;
std::lock_guard lock(m_cache_mutex);
invalidate_range_impl_base(cmd, address_range::start_length(src_address, memcpy_bytes_length), invalidation_cause::read, std::forward<Args>(extras)...);
invalidate_range_impl_base(cmd, address_range::start_length(dst_address, memcpy_bytes_length), invalidation_cause::write, std::forward<Args>(extras)...);
memcpy(dst.pixels, src.pixels, memcpy_bytes_length);
return true;
}
}
section_storage_type* cached_dest = nullptr;
u16 max_dst_width = dst.width;
u16 max_dst_height = dst.height;
areai src_area = { 0, 0, src_w, src_h };
areai dst_area = { 0, 0, dst_w, dst_h };
// 1024 height is a hack (for ~720p buffers)
// It is possible to have a large buffer that goes up to around 4kx4k but anything above 1280x720 is rare
// RSX only handles 512x512 tiles so texture 'stitching' will eventually be needed to be completely accurate
// Sections will be submitted as (512x512 + 512x512 + 256x512 + 512x208 + 512x208 + 256x208) to blit a 720p surface to the backbuffer for example
size2i dst_dimensions = { dst.pitch / (dst_is_argb8 ? 4 : 2), dst.height };
if (src_is_render_target)
{
if (dst_dimensions.width == src_subres.surface->get_surface_width())
{
dst_dimensions.height = std::max(src_subres.surface->get_surface_height(), dst.height);
}
else if (dst.max_tile_h > dst.height)
{
// Optimizations table based on common width/height pairings. If we guess wrong, the upload resolver will fix it anyway
// TODO: Add more entries based on empirical data
if (LIKELY(dst.width == 1280))
{
dst_dimensions.height = std::max<s32>(dst.height, 720);
}
else
{
dst_dimensions.height = std::min((s32)dst.max_tile_h, 1024);
}
}
}
reader_lock lock(m_cache_mutex);
@ -2435,7 +2461,7 @@ namespace rsx
{
// TODO: Rejecting unlocked blit_engine dst causes stutter in SCV
// Surfaces marked as dirty have already been removed, leaving only flushed blit_dst data
// continue;
continue;
}
const auto this_address = surface->get_section_base();

2
rpcs3/Emu/RSX/rsx_methods.cpp
View File

@ -988,7 +988,7 @@ namespace rsx
dst_info.max_tile_h = static_cast<u16>((dst_region.tile->size - dst_region.base) / out_pitch);
}
if (!g_cfg.video.force_cpu_blit_processing && (dst_dma == CELL_GCM_CONTEXT_DMA_MEMORY_FRAME_BUFFER || src_dma == CELL_GCM_CONTEXT_DMA_MEMORY_FRAME_BUFFER) && scale_x > 0 && scale_y > 0)
if (!g_cfg.video.force_cpu_blit_processing && (dst_dma == CELL_GCM_CONTEXT_DMA_MEMORY_FRAME_BUFFER || src_dma == CELL_GCM_CONTEXT_DMA_MEMORY_FRAME_BUFFER))
{
//For now, only use this for actual scaled images, there are use cases that should not go through 3d engine, e.g program ucode transfer
//TODO: Figure out more instances where we can use this without problems