From 8f19912c641435cb23c83ecb33a066f8be59294d Mon Sep 17 00:00:00 2001
From: Stenzek <stenzek@gmail.com>
Date: Thu, 12 Dec 2024 20:19:27 +1000
Subject: [PATCH] GPU/ShaderGen: Use sample instead of load at 1x as well

Consistency. Mali ends up ever-so-slightly faster with sample versus
texel loads, apparently.

Also fixes compile errors when using texture filtering on GLSL ES.
---
 src/core/gpu_hw_shadergen.cpp   | 4 ++--
 src/core/shader_cache_version.h | 2 +-
 src/util/shadergen.cpp          | 2 ++
 3 files changed, 5 insertions(+), 3 deletions(-)

diff --git a/src/core/gpu_hw_shadergen.cpp b/src/core/gpu_hw_shadergen.cpp
index e13cf4dbd..b3fd5fdd7 100644
--- a/src/core/gpu_hw_shadergen.cpp
+++ b/src/core/gpu_hw_shadergen.cpp
@@ -883,7 +883,7 @@ float4 SampleFromVRAM(TEXPAGE_VALUE texpage, float2 coords)
     #if !UPSCALED
       uint2 icoord = ApplyTextureWindow(FloatToIntegerCoords(coords));
       uint2 vicoord = (texpage.xy + icoord) & uint2(1023, 511);
-      return LOAD_TEXTURE(samp0, int2(vicoord), 0);
+      return SAMPLE_TEXTURE_LEVEL(samp0, float2(vicoord) * RCP_VRAM_SIZE, 0.0);
     #else
       // Coordinates are already upscaled, we need to downscale them to apply the texture
       // window, then re-upscale/offset. We can't round here, because it could result in
@@ -979,7 +979,7 @@ float4 SampleFromVRAM(TEXPAGE_VALUE texpage, float2 coords)
       ialpha = 1.0;
     #elif TEXTURE_FILTERING
       #if PAGE_TEXTURE
-        FilteredSampleFromVRAM(int2(0, 0), v_tex0, v_uv_limits, texcol, ialpha);
+        FilteredSampleFromVRAM(VECTOR_BROADCAST(TEXPAGE_VALUE, 0u), v_tex0, v_uv_limits, texcol, ialpha);
       #else
         FilteredSampleFromVRAM(v_texpage, v_tex0, v_uv_limits, texcol, ialpha);
       #endif
diff --git a/src/core/shader_cache_version.h b/src/core/shader_cache_version.h
index f48ee7f87..dd969ce79 100644
--- a/src/core/shader_cache_version.h
+++ b/src/core/shader_cache_version.h
@@ -5,4 +5,4 @@
 
 #include "common/types.h"
 
-static constexpr u32 SHADER_CACHE_VERSION = 23;
+static constexpr u32 SHADER_CACHE_VERSION = 24;
diff --git a/src/util/shadergen.cpp b/src/util/shadergen.cpp
index 98907cbb1..c77826c0d 100644
--- a/src/util/shadergen.cpp
+++ b/src/util/shadergen.cpp
@@ -298,6 +298,7 @@ void ShaderGen::WriteHeader(std::stringstream& ss, bool enable_rov /* = false */
     ss << "#define LOAD_TEXTURE_BUFFER(name, index) texelFetch(name, index)\n";
     ss << "#define BEGIN_ARRAY(type, size) type[size](\n";
     ss << "#define END_ARRAY )\n";
+    ss << "#define VECTOR_BROADCAST(type, value) (type(value))\n";
 
     ss << "float saturate(float value) { return clamp(value, 0.0, 1.0); }\n";
     ss << "float2 saturate(float2 value) { return clamp(value, float2(0.0, 0.0), float2(1.0, 1.0)); }\n";
@@ -344,6 +345,7 @@ void ShaderGen::WriteHeader(std::stringstream& ss, bool enable_rov /* = false */
     ss << "#define LOAD_TEXTURE_BUFFER(name, index) name.Load(index)\n";
     ss << "#define BEGIN_ARRAY(type, size) {\n";
     ss << "#define END_ARRAY }\n";
+    ss << "#define VECTOR_BROADCAST(type, value) ((type)(value))\n";
   }
 
   ss << "\n";