fix blended textures for dx, opengl, and metal

src/graphic/Fast3D/gfx_direct3d_common.cpp

@@ -350,6 +350,19 @@ void gfx_direct3d_common_build_shader(char buf[8192], size_t& len, size_t& num_f
                 }
                 len += sprintf(buf + len, "    } else {\r\n");
                 len += sprintf(buf + len, "        texVal%d = g_texture%d.Sample(g_sampler%d, tc%d);\r\n", i, i, i, i);
+                if (cc_features.used_masks[i]) {
+                    if (cc_features.used_blend[i]) {
+                        len += sprintf(buf + len,
+                                       "        float4 blendVal%d = g_textureBlend%d.Sample(g_sampler%d, tc%d);\r\n", i,
+                                       i, i, i);
+                    } else {
+                        len += sprintf(buf + len, "        float4 blendVal%d = float4(0, 0, 0, 0);\r\n", i);
+                    }
+                    len += sprintf(buf + len,
+                                   "        texVal%d = lerp(texVal%d, blendVal%d, "
+                                   "g_textureMask%d.Sample(g_sampler%d, tc%d).a);\r\n",
+                                   i, i, i, i, i, i);
+                }
                 len += sprintf(buf + len, "    }\r\n");
             } else {
                 len +=

src/graphic/Fast3D/gfx_metal_shader.cpp

@@ -125,7 +125,7 @@ static void append_formula(char* buf, size_t* len, const uint8_t c[2][4], bool d
 
 // MARK: - Public Methods
 
-MTL::VertexDescriptor* gfx_metal_build_shader(char buf[4096], size_t& num_floats, const CCFeatures& cc_features,
+MTL::VertexDescriptor* gfx_metal_build_shader(char buf[8192], size_t& num_floats, const CCFeatures& cc_features,
                                               bool three_point_filtering) {
 
     size_t len = 0;

src/graphic/Fast3D/gfx_metal_shader.h

@@ -13,7 +13,7 @@
 #include <stdio.h>
 #include "gfx_cc.h"
 
-MTL::VertexDescriptor* gfx_metal_build_shader(char buf[4096], size_t& num_floats, const CCFeatures& cc_features,
+MTL::VertexDescriptor* gfx_metal_build_shader(char buf[8192], size_t& num_floats, const CCFeatures& cc_features,
                                               bool three_point_filtering);
 
 #endif /* GFX_METAL_SHADER_H */

src/graphic/Fast3D/gfx_pc.cpp

@@ -901,10 +901,10 @@ static void import_texture_raw(int tile, bool importReplacement) {
     // if texture type is CI4 or CI8 we need to apply tlut to it
     switch (type) {
         case LUS::TextureType::Palette4bpp:
-            import_texture_ci4(tile, false);
+            import_texture_ci4(tile, importReplacement);
             return;
         case LUS::TextureType::Palette8bpp:
-            import_texture_ci8(tile, false);
+            import_texture_ci8(tile, importReplacement);
             return;
         default:
             break;
@@ -1565,6 +1565,12 @@ static void gfx_sp_tri1(uint8_t vtx1_idx, uint8_t vtx2_idx, uint8_t vtx3_idx, bo
             if (linear_filter != rendering_state.textures[i]->second.linear_filter ||
                 cms != rendering_state.textures[i]->second.cms || cmt != rendering_state.textures[i]->second.cmt) {
                 gfx_flush();
+
+                // Set the same sampler params on the blended texture. Needed for opengl.
+                if (rdp.loaded_texture[i].blended) {
+                    gfx_rapi->set_sampler_parameters(SHADER_FIRST_REPLACEMENT_TEXTURE + i, linear_filter, cms, cmt);
+                }
+
                 gfx_rapi->set_sampler_parameters(i, linear_filter, cms, cmt);
                 rendering_state.textures[i]->second.linear_filter = linear_filter;
                 rendering_state.textures[i]->second.cms = cms;