fabi/Xenia-Canary
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge branch 'master' of https://github.com/xenia-project/xenia into canary_experimental

Browse Source
This commit is contained in:
Gliniak
2022-07-23 13:26:31 +02:00
parent 1d480f74fa 3c12814276
commit 1fcac00924
66 changed files with 44142 additions and 43718 deletions
Download Patch File Download Diff File Expand all files Collapse all files

98
src/xenia/gpu/vulkan/vulkan_render_target_cache.cc
View File

@@ -2657,17 +2657,24 @@ VkShaderModule VulkanRenderTargetCache::GetTransferShader(
-int32_t(source_32bpp_tile_half_pixels)))));
}
// Transform the destination 32bpp tile index into the source.
spv::Id source_tile_index = builder.createUnaryOp(
spv::OpBitcast, type_uint,
builder.createBinOp(
spv::OpIAdd, type_int,
builder.createUnaryOp(spv::OpBitcast, type_int, dest_tile_index),
builder.createTriOp(
spv::OpBitFieldSExtract, type_int,
builder.createUnaryOp(spv::OpBitcast, type_int, address_constant),
builder.makeUintConstant(xenos::kEdramPitchTilesBits * 2),
builder.makeUintConstant(xenos::kEdramBaseTilesBits))));
// Transform the destination 32bpp tile index into the source. After the
// addition, it may be negative - in which case, the transfer is done across
// EDRAM addressing wrapping, and xenos::kEdramTileCount must be added to it,
// but `& (xenos::kEdramTileCount - 1)` handles that regardless of the sign.
spv::Id source_tile_index = builder.createBinOp(
spv::OpBitwiseAnd, type_uint,
builder.createUnaryOp(
spv::OpBitcast, type_uint,
builder.createBinOp(
spv::OpIAdd, type_int,
builder.createUnaryOp(spv::OpBitcast, type_int, dest_tile_index),
builder.createTriOp(
spv::OpBitFieldSExtract, type_int,
builder.createUnaryOp(spv::OpBitcast, type_int,
address_constant),
builder.makeUintConstant(xenos::kEdramPitchTilesBits * 2),
builder.makeUintConstant(xenos::kEdramBaseTilesBits + 1)))),
builder.makeUintConstant(xenos::kEdramTileCount - 1));
// Split the source 32bpp tile index into X and Y tile index within the source
// image.
spv::Id source_pitch_tiles = builder.createTriOp(
@@ -3724,19 +3731,28 @@ VkShaderModule VulkanRenderTargetCache::GetTransferShader(
push_constants, id_vector_temp),
spv::NoPrecision);
// Transform the destination tile index into the host depth source.
spv::Id host_depth_source_tile_index = builder.createUnaryOp(
spv::OpBitcast, type_uint,
builder.createBinOp(
spv::OpIAdd, type_int,
builder.createUnaryOp(spv::OpBitcast, type_int,
dest_tile_index),
builder.createTriOp(
spv::OpBitFieldSExtract, type_int,
// After the addition, it may be negative - in which case, the
// transfer is done across EDRAM addressing wrapping, and
// xenos::kEdramTileCount must be added to it, but
// `& (xenos::kEdramTileCount - 1)` handles that regardless of the
// sign.
spv::Id host_depth_source_tile_index = builder.createBinOp(
spv::OpBitwiseAnd, type_uint,
builder.createUnaryOp(
spv::OpBitcast, type_uint,
builder.createBinOp(
spv::OpIAdd, type_int,
builder.createUnaryOp(spv::OpBitcast, type_int,
host_depth_address_constant),
builder.makeUintConstant(xenos::kEdramPitchTilesBits *
2),
builder.makeUintConstant(xenos::kEdramBaseTilesBits))));
dest_tile_index),
builder.createTriOp(
spv::OpBitFieldSExtract, type_int,
builder.createUnaryOp(spv::OpBitcast, type_int,
host_depth_address_constant),
builder.makeUintConstant(
xenos::kEdramPitchTilesBits * 2),
builder.makeUintConstant(
xenos::kEdramBaseTilesBits + 1)))),
builder.makeUintConstant(xenos::kEdramTileCount - 1));
// Split the host depth source tile index into X and Y tile index
// within the source image.
spv::Id host_depth_source_pitch_tiles = builder.createTriOp(
@@ -3837,6 +3853,8 @@ VkShaderModule VulkanRenderTargetCache::GetTransferShader(
type_uint, id_vector_temp);
}
// Combine the tile sample index and the tile index.
// The tile index doesn't need to be wrapped, as the host depth is
// written to the beginning of the buffer, without the base offset.
spv::Id host_depth_offset = builder.createBinOp(
spv::OpIAdd, type_uint,
builder.createBinOp(
@@ -5427,28 +5445,33 @@ VkPipeline VulkanRenderTargetCache::GetDumpPipeline(DumpPipelineKey key) {
const_edram_pitch_tiles_bits),
rectangle_tile_index_y),
rectangle_tile_index_x);
// Add the base tile in the dispatch to the dispatch-local tile index.
// Add the base tile in the dispatch to the dispatch-local tile index, not
// wrapping yet so in case of a wraparound, the address relative to the base
// in the image after subtraction of the base won't be negative.
id_vector_temp.clear();
id_vector_temp.push_back(builder.makeIntConstant(kDumpPushConstantOffsets));
spv::Id offsets_constant = builder.createLoad(
builder.createAccessChain(spv::StorageClassPushConstant, push_constants,
id_vector_temp),
spv::NoPrecision);
spv::Id const_edram_base_tiles_bits =
builder.makeUintConstant(xenos::kEdramBaseTilesBits);
spv::Id edram_tile_index = builder.createBinOp(
spv::Id const_edram_base_tiles_bits_plus_1 =
builder.makeUintConstant(xenos::kEdramBaseTilesBits + 1);
spv::Id edram_tile_index_non_wrapped = builder.createBinOp(
spv::OpIAdd, type_uint,
builder.createTriOp(spv::OpBitFieldUExtract, type_uint, offsets_constant,
const_uint_0, const_edram_base_tiles_bits),
const_uint_0, const_edram_base_tiles_bits_plus_1),
rectangle_tile_index);
// Combine the tile sample index and the tile index into the EDRAM sample
// index.
// Combine the tile sample index and the tile index, wrapping the tile
// addressing, into the EDRAM sample index.
spv::Id edram_sample_address = builder.createBinOp(
spv::OpIAdd, type_uint,
builder.createBinOp(spv::OpIMul, type_uint,
builder.makeUintConstant(tile_width * tile_height),
edram_tile_index),
builder.createBinOp(
spv::OpIMul, type_uint,
builder.makeUintConstant(tile_width * tile_height),
builder.createBinOp(
spv::OpBitwiseAnd, type_uint, edram_tile_index_non_wrapped,
builder.makeUintConstant(xenos::kEdramTileCount - 1))),
builder.createBinOp(spv::OpIAdd, type_uint,
builder.createBinOp(spv::OpIMul, type_uint,
const_tile_width, tile_sample_y),
@@ -5474,10 +5497,11 @@ VkPipeline VulkanRenderTargetCache::GetDumpPipeline(DumpPipelineKey key) {
// Get the linear tile index within the source texture.
spv::Id source_tile_index = builder.createBinOp(
spv::OpISub, type_uint, edram_tile_index,
builder.createTriOp(spv::OpBitFieldUExtract, type_uint, offsets_constant,
const_edram_base_tiles_bits,
const_edram_base_tiles_bits));
spv::OpISub, type_uint, edram_tile_index_non_wrapped,
builder.createTriOp(
spv::OpBitFieldUExtract, type_uint, offsets_constant,
const_edram_base_tiles_bits_plus_1,
builder.makeUintConstant(xenos::kEdramBaseTilesBits)));
// Split the linear tile index in the source texture into X and Y in tiles.
spv::Id source_pitch_tiles = builder.createTriOp(
spv::OpBitFieldUExtract, type_uint, pitches_constant,

21
src/xenia/gpu/vulkan/vulkan_render_target_cache.h
View File

@@ -412,13 +412,13 @@ class VulkanRenderTargetCache final : public RenderTargetCache {
// Same as RenderTargetKey::pitch_tiles_at_32bpp.
uint32_t pitch_tiles_at_32bpp : 8; // 8
// [0, 2047].
uint32_t depth_base_tiles : xenos::kEdramBaseTilesBits - 1; // 19
uint32_t color_0_base_tiles : xenos::kEdramBaseTilesBits - 1; // 30
uint32_t depth_base_tiles : xenos::kEdramBaseTilesBits; // 19
uint32_t color_0_base_tiles : xenos::kEdramBaseTilesBits; // 30
uint32_t color_1_base_tiles : xenos::kEdramBaseTilesBits - 1; // 43
uint32_t color_2_base_tiles : xenos::kEdramBaseTilesBits - 1; // 54
uint32_t color_1_base_tiles : xenos::kEdramBaseTilesBits; // 43
uint32_t color_2_base_tiles : xenos::kEdramBaseTilesBits; // 54
uint32_t color_3_base_tiles : xenos::kEdramBaseTilesBits - 1; // 75
uint32_t color_3_base_tiles : xenos::kEdramBaseTilesBits; // 75
// Including all the padding, for a stable hash.
FramebufferKey() { Reset(); }
@@ -619,16 +619,13 @@ class VulkanRenderTargetCache final : public RenderTargetCache {
// All in tiles.
uint32_t dest_pitch : xenos::kEdramPitchTilesBits;
uint32_t source_pitch : xenos::kEdramPitchTilesBits;
// Safe to use 12 bits for signed difference - no ownership transfer can
// ever occur between render targets with EDRAM base >= 2048 as this would
// result in 0-length spans. 10 + 10 + 12 is exactly 32, any more bits,
// and more root 32-bit constants will be used.
// Destination base in tiles minus source base in tiles (not vice versa
// because this is a transform of the coordinate system, not addresses
// themselves).
// + 1 bit because this is a signed difference between two EDRAM bases.
// 0 for host_depth_source_is_copy (ignored in this case anyway as
// destination == source anyway).
int32_t source_to_dest : xenos::kEdramBaseTilesBits;
int32_t source_to_dest : xenos::kEdramBaseTilesBits + 1;
};
TransferAddressConstant() : constant(0) {
static_assert_size(*this, sizeof(constant));
@@ -746,7 +743,9 @@ class VulkanRenderTargetCache final : public RenderTargetCache {
union DumpOffsets {
uint32_t offsets;
struct {
uint32_t dispatch_first_tile : xenos::kEdramBaseTilesBits;
// May be beyond the EDRAM tile count in case of EDRAM addressing
// wrapping, thus + 1 bit.
uint32_t dispatch_first_tile : xenos::kEdramBaseTilesBits + 1;
uint32_t source_base_tiles : xenos::kEdramBaseTilesBits;
};
DumpOffsets() : offsets(0) { static_assert_size(*this, sizeof(offsets)); }