octo/citra
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

morton_swizzle: Move out of bounds texture check out of the decode loop

Browse Source 
* Running relative expensive checks like this on a hot path causes small but measurable performance loss. Tested SMD wit this and it doesn't crash
This commit is contained in:
emufan4568
2022-09-07 22:57:11 +03:00
committed by GPUCode
parent e464507b7a
commit 1f450d6d1d
3 changed files with 62 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files

41
src/video_core/rasterizer_cache/morton_swizzle.h
View File

@@ -73,7 +73,7 @@ static void MortonCopy(u32 stride, u32 height, u8* linear_buffer, PAddr base, PA
const PAddr aligned_down_start = base + Common::AlignDown(start - base, tile_size);
const PAddr aligned_start = base + Common::AlignUp(start - base, tile_size);
const PAddr aligned_end = base + Common::AlignDown(end - base, tile_size);
PAddr aligned_end = base + Common::AlignDown(end - base, tile_size);
ASSERT(!morton_to_linear || (aligned_start == start && aligned_end == end));
@@ -81,9 +81,12 @@ static void MortonCopy(u32 stride, u32 height, u8* linear_buffer, PAddr base, PA
u32 x = (begin_pixel_index % (stride * 8)) / 8;
u32 y = (begin_pixel_index / (stride * 8)) * 8;
// In OpenGL the texture origin is in the bottom left corner as opposed to other
// APIs that have it at the top left. To avoid flipping texture coordinates in
// the shader we read/write the linear buffer backwards
linear_buffer += ((height - 8 - y) * stride + x) * aligned_bytes_per_pixel;
auto linearbuf_next_tile = [&] {
auto linear_next_tile = [&] {
x = (x + 8) % stride;
linear_buffer += 8 * aligned_bytes_per_pixel;
if (!x) {
@@ -94,36 +97,38 @@ static void MortonCopy(u32 stride, u32 height, u8* linear_buffer, PAddr base, PA
u8* tile_buffer = VideoCore::g_memory->GetPhysicalPointer(start);
// If during a texture download the start coordinate is inside a tile, swizzle
// the tile to a temporary buffer and copy the part we are interested in
if (start < aligned_start && !morton_to_linear) {
std::array<u8, tile_size> tmp_buf;
MortonCopyTile<morton_to_linear, format>(stride, &tmp_buf[0], linear_buffer);
std::memcpy(tile_buffer, &tmp_buf[start - aligned_down_start],
MortonCopyTile<morton_to_linear, format>(stride, tmp_buf.data(), linear_buffer);
std::memcpy(tile_buffer, tmp_buf.data() + start - aligned_down_start,
std::min(aligned_start, end) - start);
tile_buffer += aligned_start - start;
linearbuf_next_tile();
linear_next_tile();
}
const u8* const buffer_end = tile_buffer + aligned_end - aligned_start;
PAddr current_paddr = aligned_start;
// Pokemon Super Mystery Dungeon will try to use textures that go beyond
// the end address of VRAM. Clamp the address to the end of VRAM if that happens
// TODO: Move this to the rasterizer cache
if (const u32 clamped_end = VideoCore::g_memory->ClampPhysicalAddress(aligned_start, aligned_end);
clamped_end != aligned_end) {
LOG_ERROR(Render_OpenGL, "Out of bound texture read address {:#x}, clamping to {:#x}", aligned_end, clamped_end);
aligned_end = clamped_end;
}
const u8* buffer_end = tile_buffer + aligned_end - aligned_start;
while (tile_buffer < buffer_end) {
// Pokemon Super Mystery Dungeon will try to use textures that go beyond
// the end address of VRAM. Stop reading if reaches invalid address
if (!VideoCore::g_memory->IsValidPhysicalAddress(current_paddr) ||
!VideoCore::g_memory->IsValidPhysicalAddress(current_paddr + tile_size)) {
LOG_ERROR(Render_OpenGL, "Out of bound texture");
break;
}
MortonCopyTile<morton_to_linear, format>(stride, tile_buffer, linear_buffer);
tile_buffer += tile_size;
current_paddr += tile_size;
linearbuf_next_tile();
linear_next_tile();
}
if (end > std::max(aligned_start, aligned_end) && !morton_to_linear) {
std::array<u8, tile_size> tmp_buf;
MortonCopyTile<morton_to_linear, format>(stride, &tmp_buf[0], linear_buffer);
std::memcpy(tile_buffer, &tmp_buf[0], end - aligned_end);
MortonCopyTile<morton_to_linear, format>(stride, tmp_buf.data(), linear_buffer);
std::memcpy(tile_buffer, tmp_buf.data(), end - aligned_end);
}
}