rsx/vk: Add support for 8 and 16-bit texel GPU-accelerated deswizzle

rpcs3/Emu/RSX/Common/TextureUtils.cpp

@@ -1096,53 +1096,35 @@ namespace rsx
 			fmt::throw_exception("Wrong format 0x%x", format);
 		}
 
-		if (word_size)
+		if (!word_size)
 		{
-			if (word_size == 1)
+			return result;
-			{
-				if (is_swizzled)
-				{
-					copy_unmodified_block_swizzled::copy_mipmap_level(dst_buffer.as_span<u8>(), src_layout.data.as_span<const u8>(), words_per_block, w, h, depth, src_layout.border, dst_pitch_in_block);
 		}
-				else if (caps.supports_zero_copy)
+
-				{
-					result.require_upload = true;
-					result.deferred_cmds = build_transfer_cmds(src_layout.data.data(), words_per_block, w, h, depth, src_layout.border, dst_pitch_in_block, src_layout.pitch_in_block);
-				}
-				else
-				{
-					copy_unmodified_block::copy_mipmap_level(dst_buffer.as_span<u8>(), src_layout.data.as_span<const u8>(), words_per_block, w, h, depth, src_layout.border, dst_pitch_in_block, src_layout.pitch_in_block);
-				}
-			}
-			else
-			{
 		result.element_size = word_size;
 		result.block_length = words_per_block;
 
 		bool require_cpu_swizzle = !caps.supports_hw_deswizzle && is_swizzled;
-				bool require_cpu_byteswap = !caps.supports_byteswap;
+		bool require_cpu_byteswap = word_size > 1 && !caps.supports_byteswap;
 
 		if (is_swizzled && caps.supports_hw_deswizzle)
-				{
-					if (word_size == 4 || (((word_size * words_per_block) & 3) == 0))
 		{
 			result.require_deswizzle = true;
 		}
-					else
-					{
-						require_cpu_swizzle = true;
-					}
-				}
 
 		if (!require_cpu_byteswap && !require_cpu_swizzle)
 		{
-					result.require_swap = true;
+			result.require_swap = (word_size > 1);
 
 			if (caps.supports_zero_copy)
 			{
 				result.require_upload = true;
 				result.deferred_cmds = build_transfer_cmds(src_layout.data.data(), word_size * words_per_block, w, h, depth, src_layout.border, dst_pitch_in_block, src_layout.pitch_in_block);
 			}
+			else if (word_size == 1)
+			{
+				copy_unmodified_block::copy_mipmap_level(dst_buffer.as_span<u8>(), src_layout.data.as_span<const u8>(), words_per_block, w, h, depth, src_layout.border, dst_pitch_in_block, src_layout.pitch_in_block);
+			}
 			else if (word_size == 2)
 			{
 				copy_unmodified_block::copy_mipmap_level(dst_buffer.as_span<u16>(), src_layout.data.as_span<const u16>(), words_per_block, w, h, depth, src_layout.border, dst_pitch_in_block, src_layout.pitch_in_block);
@@ -1151,10 +1133,16 @@ namespace rsx
 			{
 				copy_unmodified_block::copy_mipmap_level(dst_buffer.as_span<u32>(), src_layout.data.as_span<const u32>(), words_per_block, w, h, depth, src_layout.border, dst_pitch_in_block, src_layout.pitch_in_block);
 			}
+
+			return result;
 		}
-				else
+
+		if (word_size == 1)
 		{
-					if (word_size == 2)
+			ensure(is_swizzled);
+			copy_unmodified_block_swizzled::copy_mipmap_level(dst_buffer.as_span<u8>(), src_layout.data.as_span<const u8>(), words_per_block, w, h, depth, src_layout.border, dst_pitch_in_block);
+		}
+		else if (word_size == 2)
 		{
 			if (is_swizzled)
 				copy_unmodified_block_swizzled::copy_mipmap_level(dst_buffer.as_span<u16>(), src_layout.data.as_span<const be_t<u16>>(), words_per_block, w, h, depth, src_layout.border, dst_pitch_in_block);
@@ -1168,9 +1156,6 @@ namespace rsx
 			else
 				copy_unmodified_block::copy_mipmap_level(dst_buffer.as_span<u32>(), src_layout.data.as_span<const be_t<u32>>(), words_per_block, w, h, depth, src_layout.border, dst_pitch_in_block, src_layout.pitch_in_block);
 		}
-				}
-			}
-		}
 
 		return result;
 	}

rpcs3/Emu/RSX/Common/surface_store.h

@@ -1219,7 +1219,7 @@ namespace rsx
 
 			if (result.size() > 1)
 			{
-				std::sort(result.begin(), result.end(), [](const auto &a, const auto &b)
+				result.sort([](const auto &a, const auto &b)
 				{
 					if (a.surface->last_use_tag == b.surface->last_use_tag)
 					{

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

@@ -3,6 +3,9 @@ R"(
 
 #define SSBO_LOCATION(x) (x + %loc)
 
+#define USE_8BIT_ADDRESSING %_8bit
+#define USE_16BIT_ADDRESSING %_16bit
+
 layout(local_size_x = %ws, local_size_y = 1, local_size_z = 1) in;
 
 layout(%set, binding=SSBO_LOCATION(0), std430) buffer ssbo0{ uint data_in[]; };
@@ -98,12 +101,57 @@ uint get_z_index(const in uint x_, const in uint y_, const in uint z_)
 	return offset;
 }
 
+#if USE_16BIT_ADDRESSING
+
+void write16(inout uint accumulator, const in uint subword, const in uint src_id, const in uint dst_id)
+{
+	const uint masks[] = { 0x0000FFFF, 0xFFFF0000 };
+	accumulator |= data_in[src_id / 2] & masks[subword];
+
+	if (subword == 1)
+	{
+		data_out[dst_id / 2] = accumulator;
+	}
+}
+
+#elif USE_8BIT_ADDRESSING
+
+void write8(inout uint accumulator, const in uint subword, const in uint src_id, const in uint dst_id)
+{
+	const uint masks[] = { 0x000000FF, 0x0000FF00, 0x00FF0000, 0xFF000000 };
+	accumulator |= data_in[src_id / 4] & masks[subword];
+
+	if (subword == 3)
+	{
+		data_out[dst_id / 4] = accumulator;
+	}
+}
+
+#else
+
+void write32(const in uint word_count, in uint src_id, in uint dst_id)
+{
+	for (uint i = 0; i < word_count; ++i)
+	{
+		uint value = data_in[src_id++];
+		data_out[dst_id++] = %f(value);
+	}
+}
+
+#endif
+
 void main()
 {
 	uint invocations_x = (gl_NumWorkGroups.x * gl_WorkGroupSize.x);
 	uint texel_id = (gl_GlobalInvocationID.y * invocations_x) + gl_GlobalInvocationID.x;
 	uint word_count = %_wordcount;
 
+#if USE_8BIT_ADDRESSING
+	texel_id *= 4;        // Each invocation consumes 4 texels
+#elif USE_16BIT_ADDRESSING
+	texel_id *= 2;        // Each invocation consumes 2 texels
+#endif
+
 	if (!init_invocation_properties(texel_id))
 		return;
 
@@ -116,14 +164,25 @@ void main()
 	uint y = (slice_offset / row_length);
 	uint x = (slice_offset % row_length);
 
+#if USE_8BIT_ADDRESSING
+	for (uint subword = 0, accumulator = 0; subword < 4; ++subword, ++x) {
+#elif USE_16BIT_ADDRESSING
+	for (uint subword = 0, temp = 0; subword < 2; ++subword, ++x) {
+#endif
+
 		uint src_texel_id = get_z_index(x, y, z);
 		uint dst_id = (texel_id * word_count);
 		uint src_id = (src_texel_id + invocation.data_offset) * word_count;
 
-	for (uint i = 0; i < word_count; ++i)
+#if USE_8BIT_ADDRESSING
-	{
+		write8(accumulator, subword, src_id, dst_id);
-		uint value = data_in[src_id++];
-		data_out[dst_id++] = %f(value);
 	}
+#elif USE_16BIT_ADDRESSING
+		write16(accumulator, subword, src_id, dst_id);
+	}
+#else
+	write32(word_count, src_id, dst_id);
+#endif
+
 }
 )"

rpcs3/Emu/RSX/VK/VKCompute.h

@@ -403,8 +403,6 @@ namespace vk
 
 		cs_deswizzle_3d()
 		{
-			ensure((sizeof(_BlockType) & 3) == 0); // "Unsupported block type"
-
 			ssbo_count = 2;
 			use_push_constants = true;
 			push_constants_size = 28;
@@ -438,8 +436,10 @@ namespace vk
 				{ "%set", "set = 0" },
 				{ "%push_block", "push_constant" },
 				{ "%ws", std::to_string(optimal_group_size) },
-				{ "%_wordcount", std::to_string(sizeof(_BlockType) / 4) },
+				{ "%_wordcount", std::to_string(std::max<u32>(sizeof(_BlockType) / 4u, 1u)) },
-				{ "%f", transform }
+				{ "%f", transform },
+				{ "%_8bit", sizeof(_BlockType) == 1 ? "1" : "0" },
+				{ "%_16bit", sizeof(_BlockType) == 2 ? "1" : "0" },
 			};
 
 			m_src = fmt::replace_all(m_src, syntax_replace);
@@ -476,7 +476,21 @@ namespace vk
 			params.logd = rsx::ceil_log2(depth);
 
 			const u32 num_bytes_per_invocation = (sizeof(_BlockType) * optimal_group_size);
-			const u32 linear_invocations = utils::aligned_div(data_length, num_bytes_per_invocation);
+			u32 linear_invocations = utils::aligned_div(data_length, num_bytes_per_invocation);
+
+			// Check if we need to do subaddressing and adjust invocation count accordingly
+			switch (sizeof(_BlockType))
+			{
+			case 1:
+				linear_invocations /= 4;
+				break;
+			case 2:
+				linear_invocations /= 2;
+				break;
+			default:
+				break;
+			}
+
 			compute_task::run(cmd, linear_invocations);
 		}
 	};

rpcs3/Emu/RSX/VK/VKTexture.cpp

@@ -759,6 +759,10 @@ namespace vk
 	{
 		switch (block_size)
 		{
+		case 1:
+			return vk::get_compute_task<cs_deswizzle_3d<u8, u8, false>>();
+		case 2:
+			return vk::get_compute_task<cs_deswizzle_3d<u16, WordType, SwapBytes>>();
 		case 4:
 			return vk::get_compute_task<cs_deswizzle_3d<u32, WordType, SwapBytes>>();
 		case 8:
@@ -776,21 +780,27 @@ namespace vk
 		vk::cs_deswizzle_base* job = nullptr;
 		const auto block_size = (word_size * word_count);
 
-		ensure(word_size == 4 || word_size == 2);
-
 		if (!swap_bytes)
 		{
-			if (word_size == 4)
+			switch (word_size)
-			{
-				job = get_deswizzle_transformation<u32, false>(block_size);
-			}
-			else
 			{
+			case 1:
+				job = get_deswizzle_transformation<u8, false>(block_size);
+				break;
+			case 2:
 				job = get_deswizzle_transformation<u16, false>(block_size);
+				break;
+			case 4:
+				job = get_deswizzle_transformation<u32, false>(block_size);
+				break;
+			default:
+				fmt::throw_exception("Unimplemented deswizzle for format.");
 			}
 		}
 		else
 		{
+			ensure(word_size == 2 || word_size == 4);
+
 			if (word_size == 4)
 			{
 				job = get_deswizzle_transformation<u32, true>(block_size);