rsx/vk: Support mismatched pitch (padded) and fix the overscan problem

rpcs3/Emu/RSX/Common/tiled_dma_copy.hpp

@@ -19,6 +19,7 @@ namespace rsx
 		uint32_t tile_bank;
 		uint32_t image_width;
 		uint32_t image_height;
+		uint32_t image_pitch;
 		uint32_t image_bpp;
 	};
 
@@ -98,7 +99,7 @@ namespace rsx
 		tile_address ^= ((tile_address >> 11) & 1) << 10;
 
 		// Calculate relative addresses and sample
-		uint32_t linear_image_offset = (row * conf.tile_pitch) + (col * conf.image_bpp);
+		uint32_t linear_image_offset = (row * conf.image_pitch) + (col * conf.image_bpp);
 		uint32_t tile_data_offset = tile_address - (conf.tile_base_address + conf.tile_offset);
 
 		if (tile_data_offset >= conf.tile_size)
@@ -160,6 +161,7 @@ namespace rsx
 			.tile_bank = bank_sense,
 			.image_width = image_width,
 			.image_height = image_height,
+			.image_pitch = row_pitch_in_bytes,
 			.image_bpp = sizeof(T)
 		};
 

rpcs3/Emu/RSX/Program/GLSLSnippets/RSXMemoryTiling.glsl

@@ -39,6 +39,7 @@ layout(%push_block) uniform Configuration
 	uint tile_bank;
 	uint image_width;
 	uint image_height;
+	uint image_pitch;
 	uint image_bpp;
 };
 #else
@@ -52,6 +53,7 @@ layout(%push_block) uniform Configuration
 	uniform uint tile_bank;
 	uniform uint image_width;
 	uniform uint image_height;
+	uniform uint image_pitch;
 	uniform uint image_bpp;
 #endif
 
@@ -306,7 +308,7 @@ void do_memory_op(const in uint row, const in uint col)
 	tile_address ^= ((tile_address >> 11) & 1) << 10;
 
 	// Calculate relative addresses and sample
-	uint linear_image_offset = (row * tile_pitch) + (col * image_bpp);
+	uint linear_image_offset = (row * image_pitch) + (col * image_bpp);
 	uint tile_data_offset = tile_address - (tile_base_address + tile_offset);
 
 	if (tile_data_offset >= tile_size)
@@ -335,7 +337,7 @@ void main()
 	const uint row = gl_GlobalInvocationID.y;
 	const uint col0 = gl_GlobalInvocationID.x;
 	
-	// for (uint col = col0; col < (col0 + num_iterations); ++col)
+	for (uint col = col0; col < (col0 + num_iterations); ++col)
 	{
 		if (row >= image_height || col0 >= image_width)
 		{

rpcs3/Emu/RSX/VK/VKCompute.h

@@ -524,6 +524,7 @@ namespace vk
 		u16 image_width;
 		u16 image_height;
 		u32 image_pitch;
+		u8  image_bpp;
 	};
 
 	template <RSX_detiler_op Op>
@@ -542,6 +543,7 @@ namespace vk
 			u32 tile_bank;
 			u32 image_width;
 			u32 image_height;
+			u32 image_pitch;
 			u32 image_bpp;
 		} params;
 #pragma pack (pop)
@@ -557,7 +559,7 @@ namespace vk
 		{
 			ssbo_count = 2;
 			use_push_constants = true;
-			push_constants_size = 44;
+			push_constants_size = 48;
 
 			create();
 
@@ -565,7 +567,6 @@ namespace vk
 			#include "../Program/GLSLSnippets/RSXMemoryTiling.glsl"
 				;
 
-			optimal_group_size = 1;
 			const std::pair<std::string_view, std::string> syntax_replace[] =
 			{
 				{ "%loc", "0" },
@@ -646,11 +647,14 @@ namespace vk
 			params.tile_pitch = config.tile_pitch;
 			params.tile_bank = config.bank;
 			params.image_width = config.image_width;
-			params.image_height = config.image_height;
+			params.image_height = tiled_height;
-			params.image_bpp = config.image_pitch / config.image_width;
+			params.image_pitch = config.image_pitch;
+			params.image_bpp = config.image_bpp;
 			set_parameters(cmd);
 
-			const u32 invocations_x = utils::aligned_div(config.image_width, optimal_group_size);
+			const u32 subtexels_per_invocation = (config.image_bpp < 4) ? (4 / config.image_bpp) : 1;
+			const u32 virtual_width = config.image_width / subtexels_per_invocation;
+			const u32 invocations_x = utils::aligned_div(virtual_width, optimal_group_size);
 			compute_task::run(cmd, invocations_x, config.image_height, 1);
 		}
 	};

rpcs3/Emu/RSX/VK/VKTextureCache.cpp

@@ -99,13 +99,24 @@ namespace vk
 
 		if (require_gpu_transform)
 		{
-			const auto section_length = valid_range.length();
+			auto section_length = valid_range.length();
 			const auto transfer_pitch = real_pitch;
 			const auto task_length = transfer_pitch * src_area.height();
 			auto working_buffer_length = calculate_working_buffer_size(task_length, src->aspect());
 
-			if (require_tiling) {
+			if (require_tiling)
+			{
 				working_buffer_length += tiled_region.tile->size;
+
+				// Calculate actual section length
+				const auto available_tile_size = tiled_region.tile->size - (valid_range.start - tiled_region.base_address);
+				const auto max_content_size = tiled_region.tile->pitch * utils::align(height, 64);
+				section_length = std::min(max_content_size, available_tile_size);
+
+				if (section_length > valid_range.length()) [[ likely ]]
+				{
+					dma_mapping = vk::map_dma(valid_range.start, section_length);
+				}
 			}
 
 			auto working_buffer = vk::get_scratch_buffer(cmd, working_buffer_length);
@@ -185,14 +196,17 @@ namespace vk
 
 					.image_width = width,
 					.image_height = height,
-					.image_pitch = real_pitch
+					.image_pitch = real_pitch,
+					.image_bpp = rsx::get_format_block_size_in_bytes(gcm_format)
 				};
 
 				// Execute
 				const auto job = vk::get_compute_task<vk::cs_tile_memcpy<RSX_detiler_op::encode>>();
 				job->run(cmd, config);
 
+				// Update internal variables
 				result_offset = task_length;
+				real_pitch = tiled_region.tile->pitch;
 				require_rw_barrier = true;
 #endif
 			}

rpcs3/Emu/RSX/VK/VKTextureCache.h

@@ -291,7 +291,17 @@ namespace vk
 
 			// Calculate smallest range to flush - for framebuffers, the raster region is enough
 			const auto range = (context == rsx::texture_upload_context::framebuffer_storage) ? get_section_range() : get_confirmed_range();
-			vk::flush_dma(range.start, range.length());
+			auto flush_length = range.length();
+
+			const auto tiled_region = rsx::get_current_renderer()->get_tiled_memory_region(range);
+			if (tiled_region)
+			{
+				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(height, 64);
+				flush_length = std::min(max_content_size, available_tile_size);
+			}
+
+			vk::flush_dma(range.start, flush_length);
 
 #if DEBUG_DMA_TILING
 			// Are we a tiled region?
@@ -310,10 +320,7 @@ namespace vk
 					width,
 					height
 				);
-				const auto available_tile_size = tiled_region.tile->size - (range.start - tiled_region.base_address);
+				std::memcpy(real_data, out_data.data(), flush_length);
-				const auto max_content_size = tiled_region.tile->pitch * utils::align(height, 64);
-				const auto write_length = std::min(max_content_size, available_tile_size);
-				std::memcpy(real_data, out_data.data(), write_length);
 			}
 #endif