From b220c8989b0a8006360c6d2c7bd2a5cda9064e7c Mon Sep 17 00:00:00 2001
From: Robin Kertels <robin.kertels@gmail.com>
Date: Thu, 27 Jan 2022 17:27:25 +0100
Subject: [PATCH] [d3d9] Only enable FETCH4 for single channel formats

---
 src/d3d9/d3d9_common_texture.cpp | 18 +++++++++++++++---
 src/d3d9/d3d9_common_texture.h   | 11 +++++++++++
 src/d3d9/d3d9_device.cpp         | 24 ++++++++++++++++++++++--
 3 files changed, 48 insertions(+), 5 deletions(-)

diff --git a/src/d3d9/d3d9_common_texture.cpp b/src/d3d9/d3d9_common_texture.cpp
index 63943c62..8e5d237b 100644
--- a/src/d3d9/d3d9_common_texture.cpp
+++ b/src/d3d9/d3d9_common_texture.cpp
@@ -30,8 +30,9 @@ namespace dxvk {
 
     m_mapping = pDevice->LookupFormat(m_desc.Format);
 
-    m_mapMode = DetermineMapMode();
-    m_shadow  = DetermineShadowState();
+    m_mapMode        = DetermineMapMode();
+    m_shadow         = DetermineShadowState();
+    m_supportsFetch4 = DetermineFetch4Compatibility();
 
     if (m_mapMode == D3D9_COMMON_TEXTURE_MAP_MODE_BACKED) {
       bool plainSurface = m_type == D3DRTYPE_SURFACE &&
@@ -316,7 +317,7 @@ namespace dxvk {
 
 
   BOOL D3D9CommonTexture::DetermineShadowState() const {
-    static std::array<D3D9Format, 3> blacklist = {
+    constexpr std::array<D3D9Format, 3> blacklist = {
       D3D9Format::INTZ, D3D9Format::DF16, D3D9Format::DF24
     };
 
@@ -325,6 +326,17 @@ namespace dxvk {
   }
 
 
+  BOOL D3D9CommonTexture::DetermineFetch4Compatibility() const {
+    constexpr std::array<D3D9Format, 8> singleChannelFormats = {
+      D3D9Format::INTZ, D3D9Format::DF16, D3D9Format::DF24,
+      D3D9Format::R16F, D3D9Format::R32F, D3D9Format::A8,
+      D3D9Format::L8, D3D9Format::L16
+    };
+
+    return std::find(singleChannelFormats.begin(), singleChannelFormats.end(), m_desc.Format) != singleChannelFormats.end();
+  }
+
+
   BOOL D3D9CommonTexture::CheckImageSupport(
     const DxvkImageCreateInfo*  pImageInfo,
           VkImageTiling         Tiling) const {
diff --git a/src/d3d9/d3d9_common_texture.h b/src/d3d9/d3d9_common_texture.h
index e3da26fd..f4d14115 100644
--- a/src/d3d9/d3d9_common_texture.h
+++ b/src/d3d9/d3d9_common_texture.h
@@ -190,6 +190,14 @@ namespace dxvk {
       return m_shadow;
     }
 
+    /**
+     * \brief FETCH4 compatibility
+     * \returns Whether the format of the texture supports the FETCH4 hack
+     */
+    bool SupportsFetch4() const {
+      return m_supportsFetch4;
+    }
+
     /**
      * \brief Null
      * \returns Whether the texture is D3DFMT_NULL or not
@@ -444,6 +452,7 @@ namespace dxvk {
     D3D9_VK_FORMAT_MAPPING        m_mapping;
 
     bool                          m_shadow; //< Depth Compare-ness
+    bool                          m_supportsFetch4;
 
     int64_t                       m_size = 0;
 
@@ -483,6 +492,8 @@ namespace dxvk {
 
     BOOL DetermineShadowState() const;
 
+    BOOL DetermineFetch4Compatibility() const;
+
     BOOL CheckImageSupport(
       const DxvkImageCreateInfo*  pImageInfo,
             VkImageTiling         Tiling) const;
diff --git a/src/d3d9/d3d9_device.cpp b/src/d3d9/d3d9_device.cpp
index 0699d8f3..ffba18d1 100644
--- a/src/d3d9/d3d9_device.cpp
+++ b/src/d3d9/d3d9_device.cpp
@@ -3682,10 +3682,14 @@ namespace dxvk {
     constexpr DWORD Fetch4Disabled = MAKEFOURCC('G', 'E', 'T', '1');
 
     if (unlikely(Type == D3DSAMP_MIPMAPLODBIAS)) {
+      auto texture = GetCommonTexture(m_state.textures[StateSampler]);
+      bool textureSupportsFetch4 = texture != nullptr && texture->SupportsFetch4();
+
       if (unlikely(Value == Fetch4Enabled)) {
         m_fetch4Enabled |= 1u << StateSampler;
-        if (state[StateSampler][D3DSAMP_MAGFILTER] == D3DTEXF_POINT)
+        if (textureSupportsFetch4 && state[StateSampler][D3DSAMP_MAGFILTER] == D3DTEXF_POINT) {
           m_fetch4 |= 1u << StateSampler;
+        }
       }
       else if (unlikely(Value == Fetch4Disabled)) {
         m_fetch4Enabled &= ~(1u << StateSampler);
@@ -3694,7 +3698,10 @@ namespace dxvk {
     }
 
     if (unlikely(Type == D3DSAMP_MAGFILTER && (m_fetch4Enabled & (1u << StateSampler)))) {
-      if (Value == D3DTEXF_POINT)
+      auto texture = GetCommonTexture(m_state.textures[StateSampler]);
+      bool textureSupportsFetch4 = texture != nullptr && texture->SupportsFetch4();
+
+      if (Value == D3DTEXF_POINT && textureSupportsFetch4)
         m_fetch4 |=   1u << StateSampler;
       else
         m_fetch4 &= ~(1u << StateSampler);
@@ -3748,6 +3755,19 @@ namespace dxvk {
 
         m_dirtySamplerStates |= 1u << StateSampler;
       }
+
+      if (unlikely(m_fetch4Enabled & (1u << StateSampler) && !(m_fetch4 & (1u << StateSampler)))) {
+        bool textureSupportsFetch4 = newTexture->SupportsFetch4();
+        if (textureSupportsFetch4
+          && m_state.samplerStates[StateSampler][D3DSAMP_MAGFILTER] == D3DTEXF_POINT
+          && m_state.samplerStates[StateSampler][D3DSAMP_MINFILTER] == D3DTEXF_POINT) {
+          m_fetch4 |= 1u << StateSampler;
+          m_dirtySamplerStates |= 1u << StateSampler;
+        }
+      }
+    } else if (unlikely(m_fetch4 & (1u << StateSampler))) {
+      m_fetch4 &= ~(1u << StateSampler);
+      m_dirtySamplerStates |= 1u << StateSampler;
     }
 
     DWORD combinedUsage = oldUsage | newUsage;