From 3242352e509bf6525b5d1f6ccc29c9f44db65324 Mon Sep 17 00:00:00 2001
From: psucien <bad_cast@protonmail.com>
Date: Sun, 17 Nov 2024 14:40:15 +0100
Subject: [PATCH] MS micro tile extents LUTs

---
 src/video_core/texture_cache/image_info.cpp | 107 +++++++++++++++++++-
 1 file changed, 102 insertions(+), 5 deletions(-)

diff --git a/src/video_core/texture_cache/image_info.cpp b/src/video_core/texture_cache/image_info.cpp
index 036154d634..78f96b9183 100644
--- a/src/video_core/texture_cache/image_info.cpp
+++ b/src/video_core/texture_cache/image_info.cpp
@@ -47,7 +47,7 @@ static vk::ImageType ConvertImageType(AmdGpu::ImageType type) noexcept {
 
 // clang-format off
 // The table of macro tiles parameters for given tiling index (row) and bpp (column)
-static constexpr std::array macro_tile_extents{
+static constexpr std::array macro_tile_extents_x1{
     std::pair{256u, 128u}, std::pair{256u, 128u}, std::pair{256u, 128u}, std::pair{256u, 128u}, std::pair{256u, 128u}, // 00
     std::pair{256u, 128u}, std::pair{128u, 128u}, std::pair{128u, 128u}, std::pair{128u, 128u}, std::pair{128u, 128u}, // 01
     std::pair{256u, 128u}, std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{128u, 64u},  // 02
@@ -66,26 +66,123 @@ static constexpr std::array macro_tile_extents{
     std::pair{256u, 128u}, std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{64u, 64u},   // 0F
     std::pair{256u, 256u}, std::pair{256u, 128u}, std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{64u, 64u},   // 10
     std::pair{256u, 256u}, std::pair{256u, 128u}, std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{64u, 64u},   // 11
-    std::pair{256u, 256u}, std::pair{256u, 128u}, std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{128u, 64u},  // 12
+    std::pair{256u, 256u}, std::pair{256u, 128u}, std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{64u, 64u},   // 12
+    std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     // 13
+    std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   // 14
+    std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   // 15
+    std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   // 16
+    std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   // 17
+    std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   // 18
+    std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   // 19
+    std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   // 1A
+};
+
+static constexpr std::array macro_tile_extents_x2{
+    std::pair{256u, 128u}, std::pair{256u, 128u}, std::pair{256u, 128u}, std::pair{256u, 128u}, std::pair{256u, 128u}, // 00
+    std::pair{128u, 128u}, std::pair{128u, 128u}, std::pair{128u, 128u}, std::pair{128u, 128u}, std::pair{128u, 128u}, // 01
+    std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{128u, 64u},  // 02
+    std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{128u, 64u},  // 03
+    std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   // 04
+    std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     // 05
+    std::pair{256u, 128u}, std::pair{128u, 128u}, std::pair{128u, 128u}, std::pair{128u, 128u}, std::pair{128u, 128u}, // 06
+    std::pair{256u, 128u}, std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   // 07
+    std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     // 08
+    std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     // 09
+    std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   // 0A
+    std::pair{256u, 128u}, std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   // 0B
+    std::pair{256u, 128u}, std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   // 0C
+    std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     // 0D
+    std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   // 0E
+    std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   // 0F
+    std::pair{256u, 128u}, std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   // 10
+    std::pair{256u, 128u}, std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   // 11
+    std::pair{256u, 128u}, std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   // 12
+    std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     // 13
+    std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   // 14
+    std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   // 15
+    std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   // 16
+    std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   // 17
+    std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   // 18
+    std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   // 19
+    std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   // 1A
+};
+
+static constexpr std::array macro_tile_extents_x4{
+    std::pair{256u, 128u}, std::pair{256u, 128u}, std::pair{256u, 128u}, std::pair{256u, 128u}, std::pair{256u, 128u}, // 00
+    std::pair{128u, 128u}, std::pair{128u, 128u}, std::pair{128u, 128u}, std::pair{128u, 128u}, std::pair{128u, 128u}, // 01
+    std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{128u, 64u},  // 02
+    std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{128u, 64u},  // 03
+    std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   // 04
+    std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     // 05
+    std::pair{128u, 128u}, std::pair{128u, 128u}, std::pair{128u, 128u}, std::pair{128u, 128u}, std::pair{128u, 128u}, // 06
+    std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   // 07
+    std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     // 08
+    std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     // 09
+    std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   // 0A
+    std::pair{128u, 128u}, std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   // 0B
+    std::pair{128u, 128u}, std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   // 0C
+    std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     // 0D
+    std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   // 0E
+    std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   // 0F
+    std::pair{128u, 128u}, std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   // 10
+    std::pair{128u, 128u}, std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   // 11
+    std::pair{128u, 128u}, std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   // 12
     std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     // 13
     std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   // 14
     std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   // 15
     std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   // 16
     std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   // 17
-    std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{128u, 64u},  // 18
+    std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   // 18
     std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   // 19
     std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   // 1A
 };
+
+static constexpr std::array macro_tile_extents_x8{
+    std::pair{256u, 128u}, std::pair{256u, 128u}, std::pair{256u, 128u}, std::pair{256u, 128u}, std::pair{256u, 128u}, // 00
+    std::pair{128u, 128u}, std::pair{128u, 128u}, std::pair{128u, 128u}, std::pair{128u, 128u}, std::pair{128u, 128u}, // 01
+    std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{128u, 64u},  // 02
+    std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{128u, 64u},  // 03
+    std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   // 04
+    std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     // 05
+    std::pair{128u, 128u}, std::pair{128u, 128u}, std::pair{128u, 128u}, std::pair{128u, 128u}, std::pair{128u, 128u}, // 06
+    std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   // 07
+    std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     // 08
+    std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     // 09
+    std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   // 0A
+    std::pair{128u, 128u}, std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   // 0B
+    std::pair{128u, 128u}, std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   // 0C
+    std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     // 0D
+    std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   // 0E
+    std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   // 0F
+    std::pair{128u, 128u}, std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   // 10
+    std::pair{128u, 128u}, std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   // 11
+    std::pair{128u, 128u}, std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   // 12
+    std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     std::pair{0u, 0u},     // 13
+    std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   // 14
+    std::pair{128u, 64u},  std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   // 15
+    std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   // 16
+    std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   // 17
+    std::pair{128u, 128u}, std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   // 18
+    std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   // 19
+    std::pair{128u, 64u},  std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   std::pair{64u, 64u},   // 1A
+};
+
+static constexpr std::array macro_tile_extents{
+    macro_tile_extents_x1,
+    macro_tile_extents_x2,
+    macro_tile_extents_x4,
+    macro_tile_extents_x8,
+};
 // clang-format on
 
 static constexpr std::pair micro_tile_extent{8u, 8u};
 static constexpr auto hw_pipe_interleave = 256u;
 
 static constexpr std::pair<u32, u32> GetMacroTileExtents(u32 tiling_idx, u32 bpp, u32 num_samples) {
-    ASSERT(num_samples == 1);
+    ASSERT(num_samples <= 8);
     const auto row = tiling_idx * 5;
     const auto column = std::bit_width(bpp) - 4; // bpps are 8, 16, 32, 64, 128
-    return macro_tile_extents[row + column];
+    return (macro_tile_extents[std::log2(num_samples)])[row + column];
 }
 
 static constexpr std::pair<u32, size_t> ImageSizeLinearAligned(u32 pitch, u32 height, u32 bpp,