From f0e902a7d6df72531fbd01d27756f9875dc3c65d Mon Sep 17 00:00:00 2001
From: FernandoS27 <fsahmkow27@gmail.com>
Date: Fri, 26 Oct 2018 19:18:03 -0400
Subject: [PATCH] Fixed mipmap block autosizing algorithm

---
 .../renderer_opengl/gl_rasterizer_cache.cpp   |  8 +++----
 .../renderer_opengl/gl_rasterizer_cache.h     | 24 ++++++++++++-------
 src/video_core/textures/decoders.h            |  6 +++++
 3 files changed, 25 insertions(+), 13 deletions(-)

diff --git a/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp b/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp
index 4d89d0f671..b2250e6bed 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp
@@ -103,7 +103,7 @@ std::size_t SurfaceParams::InnerMipmapMemorySize(u32 mip_level, bool force_gl, b
                    ? m_height
                    : std::max(1U, (m_height + compression_factor - 1) / compression_factor);
     m_depth = std::max(1U, m_depth >> mip_level);
-    u32 m_block_height = MipBlockHeight(mip_level, m_height);
+    u32 m_block_height = MipBlockHeight(mip_level);
     u32 m_block_depth = MipBlockDepth(mip_level);
     return Tegra::Texture::CalculateSize(force_gl ? false : is_tiled, bytes_per_pixel, m_width,
                                          m_height, m_depth, m_block_height, m_block_depth);
@@ -111,7 +111,7 @@ std::size_t SurfaceParams::InnerMipmapMemorySize(u32 mip_level, bool force_gl, b
 
 std::size_t SurfaceParams::InnerMemorySize(bool force_gl, bool layer_only,
                                            bool uncompressed) const {
-    std::size_t block_size_bytes = 512 * block_height * block_depth; // 512 is GOB size
+    std::size_t block_size_bytes = Tegra::Texture::GetGOBSize() * block_height * block_depth;
     std::size_t size = 0;
     for (u32 i = 0; i < max_mip_level; i++) {
         size += InnerMipmapMemorySize(i, force_gl, layer_only, uncompressed);
@@ -1043,8 +1043,8 @@ void CachedSurface::FlushGLBuffer() {
     glPixelStorei(GL_PACK_ROW_LENGTH, static_cast<GLint>(params.width));
     ASSERT(!tuple.compressed);
     glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
-    glGetTextureImage(texture.handle, 0, tuple.format, tuple.type, static_cast<GLsizei>(gl_buffer[0].size()),
-                      gl_buffer[0].data());
+    glGetTextureImage(texture.handle, 0, tuple.format, tuple.type,
+                      static_cast<GLsizei>(gl_buffer[0].size()), gl_buffer[0].data());
     glPixelStorei(GL_PACK_ROW_LENGTH, 0);
     ConvertFormatAsNeeded_FlushGLBuffer(gl_buffer[0], params.pixel_format, params.width,
                                         params.height);
diff --git a/src/video_core/renderer_opengl/gl_rasterizer_cache.h b/src/video_core/renderer_opengl/gl_rasterizer_cache.h
index 951e03ba60..15ac4a1b16 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer_cache.h
+++ b/src/video_core/renderer_opengl/gl_rasterizer_cache.h
@@ -917,14 +917,14 @@ struct SurfaceParams {
 
     // Auto block resizing algorithm from:
     // https://cgit.freedesktop.org/mesa/mesa/tree/src/gallium/drivers/nouveau/nv50/nv50_miptree.c
-    u32 MipBlockHeight(u32 mip_level, u32 alt_height = 0) const {
+    u32 MipBlockHeight(u32 mip_level) const {
         if (mip_level == 0)
             return block_height;
-        if (alt_height == 0)
-            alt_height = MipHeight(mip_level);
-        u32 blocks_in_y = (alt_height + 7) / 8;
-        u32 bh = 32;
-        while (bh > 1 && blocks_in_y <= bh * 2) {
+        u32 alt_height = MipHeight(mip_level);
+        u32 h = GetDefaultBlockHeight(pixel_format);
+        u32 blocks_in_y = (alt_height + h - 1) / h;
+        u32 bh = 16;
+        while (bh > 1 && blocks_in_y <= bh * 4) {
             bh >>= 1;
         }
         return bh;
@@ -933,11 +933,17 @@ struct SurfaceParams {
     u32 MipBlockDepth(u32 mip_level) const {
         if (mip_level == 0)
             return block_depth;
+        if (is_layered)
+            return 1;
         u32 depth = MipDepth(mip_level);
         u32 bd = 32;
-        // Magical block resizing algorithm, needs more testing.
-        while (bd > 1 && depth / depth <= bd) {
-            bd = bd >> 1;
+        while (bd > 1 && depth * 2 <= bd) {
+            bd >>= 1;
+        }
+        if (bd == 32) {
+            u32 bh = MipBlockHeight(mip_level);
+            if (bh >= 4)
+                return 16;
         }
         return bd;
     }
diff --git a/src/video_core/textures/decoders.h b/src/video_core/textures/decoders.h
index 4726f54a5e..b390219e4b 100644
--- a/src/video_core/textures/decoders.h
+++ b/src/video_core/textures/decoders.h
@@ -10,6 +10,12 @@
 
 namespace Tegra::Texture {
 
+// GOBSize constant. Calculated by 64 bytes in x multiplied by 8 y coords, represents
+// an small rect of (64/bytes_per_pixel)X8.
+inline std::size_t GetGOBSize() {
+    return 512;
+}
+
 /**
  * Unswizzles a swizzled texture without changing its format.
  */