From f6c6c04648f2177f1ddece65e8585cd7661c6a4b Mon Sep 17 00:00:00 2001
From: kd-11 <karokidii@gmail.com>
Date: Sat, 15 Aug 2020 21:37:10 +0300
Subject: [PATCH] vk: Implement transport for D24S8_FLOAT data

---
 rpcs3/Emu/RSX/VK/VKCompute.h   | 37 +++++++++++++++++++++++++++++-----
 rpcs3/Emu/RSX/VK/VKTexture.cpp | 34 ++++++++++++++++++++++++++++++-
 2 files changed, 65 insertions(+), 6 deletions(-)

diff --git a/rpcs3/Emu/RSX/VK/VKCompute.h b/rpcs3/Emu/RSX/VK/VKCompute.h
index 6ab1515aa5..0d8bf03252 100644
--- a/rpcs3/Emu/RSX/VK/VKCompute.h
+++ b/rpcs3/Emu/RSX/VK/VKCompute.h
@@ -280,6 +280,8 @@ namespace vk
 				"// Depth format conversions\n"
 				"#define d24_to_f32(bits)             floatBitsToUint(float(bits) / 16777215.f)\n"
 				"#define f32_to_d24(bits)             uint(uintBitsToFloat(bits) * 16777215.f)\n"
+				"#define d24f_to_f32(bits)            (bits << 7)\n"
+				"#define f32_to_d24f(bits)            (bits >> 7)\n"
 				"#define d24x8_to_f32(bits)           d24_to_f32(bits >> 8)\n"
 				"#define d24x8_to_d24x8_swapped(bits) (bits & 0xFF00) | (bits & 0xFF0000) >> 16 | (bits & 0xFF) << 16\n"
 				"#define f32_to_d24x8_swapped(bits)   d24x8_to_d24x8_swapped(f32_to_d24(bits))\n"
@@ -492,7 +494,7 @@ namespace vk
 		}
 	};
 
-	template<bool _SwapBytes = false>
+	template<bool _SwapBytes = false, bool _DepthFloat = false>
 	struct cs_gather_d32x8 : cs_interleave_task
 	{
 		cs_gather_d32x8()
@@ -500,8 +502,20 @@ namespace vk
 			work_kernel =
 				"		if (index >= block_length)\n"
 				"			return;\n"
-				"\n"
-				"		depth = f32_to_d24(data[index + z_offset]);\n"
+				"\n";
+
+			if constexpr (!_DepthFloat)
+			{
+				work_kernel +=
+				"		depth = f32_to_d24(data[index + z_offset]);\n";
+			}
+			else
+			{
+				work_kernel +=
+				"		depth = f32_to_d24f(data[index + z_offset]);\n";
+			}
+
+			work_kernel +=
 				"		stencil_offset = (index / 4);\n"
 				"		stencil_shift = (index % 4) * 8;\n"
 				"		stencil = data[stencil_offset + s_offset];\n"
@@ -542,6 +556,7 @@ namespace vk
 		}
 	};
 
+	template<bool _DepthFloat = false>
 	struct cs_scatter_d32x8 : cs_interleave_task
 	{
 		cs_scatter_d32x8()
@@ -550,8 +565,20 @@ namespace vk
 				"		if (index >= block_length)\n"
 				"			return;\n"
 				"\n"
-				"		value = data[index];\n"
-				"		data[index + z_offset] = d24_to_f32(value >> 8);\n"
+				"		value = data[index];\n";
+
+			if constexpr (!_DepthFloat)
+			{
+				work_kernel +=
+				"		data[index + z_offset] = d24_to_f32(value >> 8);\n";
+			}
+			else
+			{
+				work_kernel +=
+				"		data[index + z_offset] = d24f_to_f32(value >> 8);\n";
+			}
+
+			work_kernel +=
 				"		stencil_offset = (index / 4);\n"
 				"		stencil_shift = (index % 4) * 8;\n"
 				"		stencil = (value & 0xFF) << stencil_shift;\n"
diff --git a/rpcs3/Emu/RSX/VK/VKTexture.cpp b/rpcs3/Emu/RSX/VK/VKTexture.cpp
index b8f8d4e067..2225c5bc4d 100644
--- a/rpcs3/Emu/RSX/VK/VKTexture.cpp
+++ b/rpcs3/Emu/RSX/VK/VKTexture.cpp
@@ -78,6 +78,11 @@ namespace vk
 			vkCmdCopyImageToBuffer(cmd, src->value, src->current_layout, dst->value, 1, &region);
 			break;
 		}
+		case VK_FORMAT_D32_SFLOAT:
+		{
+			fmt::throw_exception("Unsupported transfer (D16_FLOAT");
+			break;
+		}
 		case VK_FORMAT_D24_UNORM_S8_UINT:
 		case VK_FORMAT_D32_SFLOAT_S8_UINT:
 		{
@@ -113,6 +118,11 @@ namespace vk
 				{
 					job = vk::get_compute_task<vk::cs_gather_d24x8<false>>();
 				}
+				else if (auto dsv = dynamic_cast<const vk::render_target*>(src);
+						 dsv && dsv->get_format_type() == rsx::format_type::depth_float)
+				{
+					job = vk::get_compute_task<vk::cs_gather_d32x8<false, true>>();
+				}
 				else
 				{
 					job = vk::get_compute_task<vk::cs_gather_d32x8<false>>();
@@ -124,6 +134,11 @@ namespace vk
 				{
 					job = vk::get_compute_task<vk::cs_gather_d24x8<true>>();
 				}
+				else if (auto dsv = dynamic_cast<const vk::render_target*>(src);
+						 dsv && dsv->get_format_type() == rsx::format_type::depth_float)
+				{
+					job = vk::get_compute_task<vk::cs_gather_d32x8<true, true>>();
+				}
 				else
 				{
 					job = vk::get_compute_task<vk::cs_gather_d32x8<true>>();
@@ -162,6 +177,11 @@ namespace vk
 			vkCmdCopyBufferToImage(cmd, src->value, dst->value, dst->current_layout, 1, &region);
 			break;
 		}
+		case VK_FORMAT_D32_SFLOAT:
+		{
+			fmt::throw_exception("Unsupported transfer (D16_FLOAT");
+			break;
+		}
 		case VK_FORMAT_D24_UNORM_S8_UINT:
 		case VK_FORMAT_D32_SFLOAT_S8_UINT:
 		{
@@ -191,9 +211,14 @@ namespace vk
 			{
 				job = vk::get_compute_task<vk::cs_scatter_d24x8>();
 			}
+			else if (auto dsv = dynamic_cast<const vk::render_target*>(dst);
+					 dsv && dsv->get_format_type() == rsx::format_type::depth_float)
+			{
+				job = vk::get_compute_task<vk::cs_scatter_d32x8<true>>();
+			}
 			else
 			{
-				job = vk::get_compute_task<vk::cs_scatter_d32x8>();
+				job = vk::get_compute_task<vk::cs_scatter_d32x8<false>>();
 			}
 
 			job->run(cmd, src, data_offset, packed_length, z_offset, s_offset);
@@ -518,6 +543,13 @@ namespace vk
 					stretch_image_typeless_unsafe(src, dst, typeless->value, src_rect, dst_rect, VK_IMAGE_ASPECT_DEPTH_BIT);
 					break;
 				}
+				case VK_FORMAT_D32_SFLOAT:
+				{
+					auto typeless = vk::get_typeless_helper(VK_FORMAT_R32_SFLOAT, typeless_w, typeless_h);
+					change_image_layout(cmd, typeless, VK_IMAGE_LAYOUT_GENERAL);
+					stretch_image_typeless_unsafe(src, dst, typeless->value, src_rect, dst_rect, VK_IMAGE_ASPECT_DEPTH_BIT);
+					break;
+				}
 				case VK_FORMAT_D24_UNORM_S8_UINT:
 				{
 					const VkImageAspectFlags depth_stencil = VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT;