"Merge Tagged PR 6598"

src/video_core/macro/macro_hle.cpp

@@ -4,6 +4,8 @@
 
 #include <array>
 #include <vector>
+#include "common/scope_exit.h"
+#include "video_core/dirty_flags.h"
 #include "video_core/engines/maxwell_3d.h"
 #include "video_core/macro/macro_hle.h"
 #include "video_core/rasterizer_interface.h"
@@ -56,6 +58,7 @@ void HLE_0217920100488FF7(Engines::Maxwell3D& maxwell3d, const std::vector<u32>&
     maxwell3d.regs.index_array.first = parameters[3];
     maxwell3d.regs.reg_array[0x446] = element_base; // vertex id base?
     maxwell3d.regs.index_array.count = parameters[1];
+    maxwell3d.dirty.flags[VideoCommon::Dirty::IndexBuffer] = true;
     maxwell3d.regs.vb_element_base = element_base;
     maxwell3d.regs.vb_base_instance = base_instance;
     maxwell3d.mme_draw.instance_count = instance_count;
@@ -77,12 +80,70 @@ void HLE_0217920100488FF7(Engines::Maxwell3D& maxwell3d, const std::vector<u32>&
     maxwell3d.CallMethodFromMME(0x8e5, 0x0);
     maxwell3d.mme_draw.current_mode = Engines::Maxwell3D::MMEDrawMode::Undefined;
 }
+
+// Multidraw Indirect
+void HLE_3f5e74b9c9a50164(Engines::Maxwell3D& maxwell3d, const std::vector<u32>& parameters) {
+    SCOPE_EXIT({
+        // Clean everything.
+        maxwell3d.regs.reg_array[0x446] = 0x0; // vertex id base?
+        maxwell3d.regs.index_array.count = 0;
+        maxwell3d.regs.vb_element_base = 0x0;
+        maxwell3d.regs.vb_base_instance = 0x0;
+        maxwell3d.mme_draw.instance_count = 0;
+        maxwell3d.CallMethodFromMME(0x8e3, 0x640);
+        maxwell3d.CallMethodFromMME(0x8e4, 0x0);
+        maxwell3d.CallMethodFromMME(0x8e5, 0x0);
+        maxwell3d.mme_draw.current_mode = Engines::Maxwell3D::MMEDrawMode::Undefined;
+        maxwell3d.dirty.flags[VideoCommon::Dirty::IndexBuffer] = true;
+    });
+    const u32 start_indirect = parameters[0];
+    const u32 end_indirect = parameters[1];
+    if (start_indirect >= end_indirect) {
+        // Nothing to do.
+        return;
+    }
+    const auto topology =
+        static_cast<Tegra::Engines::Maxwell3D::Regs::PrimitiveTopology>(parameters[2]);
+    maxwell3d.regs.draw.topology.Assign(topology);
+    const u32 padding = parameters[3];
+    const std::size_t max_draws = parameters[4];
+
+    const u32 indirect_words = 5 + padding;
+    const std::size_t first_draw = start_indirect;
+    const std::size_t effective_draws = end_indirect - start_indirect;
+    const std::size_t last_draw = start_indirect + std::min(effective_draws, max_draws);
+
+    for (std::size_t index = first_draw; index < last_draw; index++) {
+        const std::size_t base = index * indirect_words + 5;
+        const u32 num_vertices = parameters[base];
+        const u32 instance_count = parameters[base + 1];
+        const u32 first_index = parameters[base + 2];
+        const u32 base_vertex = parameters[base + 3];
+        const u32 base_instance = parameters[base + 4];
+        maxwell3d.regs.index_array.first = first_index;
+        maxwell3d.regs.reg_array[0x446] = base_vertex;
+        maxwell3d.regs.index_array.count = num_vertices;
+        maxwell3d.regs.vb_element_base = base_vertex;
+        maxwell3d.regs.vb_base_instance = base_instance;
+        maxwell3d.mme_draw.instance_count = instance_count;
+        maxwell3d.CallMethodFromMME(0x8e3, 0x640);
+        maxwell3d.CallMethodFromMME(0x8e4, base_vertex);
+        maxwell3d.CallMethodFromMME(0x8e5, base_instance);
+        maxwell3d.dirty.flags[VideoCommon::Dirty::IndexBuffer] = true;
+        if (maxwell3d.ShouldExecute()) {
+            maxwell3d.Rasterizer().Draw(true, true);
+        }
+        maxwell3d.mme_draw.current_mode = Engines::Maxwell3D::MMEDrawMode::Undefined;
+    }
+}
+
 } // Anonymous namespace
 
-constexpr std::array<std::pair<u64, HLEFunction>, 3> hle_funcs{{
+constexpr std::array<std::pair<u64, HLEFunction>, 4> hle_funcs{{
     {0x771BB18C62444DA0, &HLE_771BB18C62444DA0},
     {0x0D61FC9FAAC9FCAD, &HLE_0D61FC9FAAC9FCAD},
     {0x0217920100488FF7, &HLE_0217920100488FF7},
+    {0x3f5e74b9c9a50164, &HLE_3f5e74b9c9a50164},
 }};
 
 HLEMacro::HLEMacro(Engines::Maxwell3D& maxwell3d_) : maxwell3d{maxwell3d_} {}

src/video_core/texture_cache/texture_cache.h

@@ -1079,7 +1079,7 @@ ImageId TextureCache<P>::JoinImages(const ImageInfo& info, GPUVAddr gpu_addr, VA
 template <class P>
 typename TextureCache<P>::BlitImages TextureCache<P>::GetBlitImages(
     const Tegra::Engines::Fermi2D::Surface& dst, const Tegra::Engines::Fermi2D::Surface& src) {
-    static constexpr auto FIND_OPTIONS = RelaxedOptions::Format | RelaxedOptions::Samples;
+    static constexpr auto FIND_OPTIONS = RelaxedOptions::Samples;
     const GPUVAddr dst_addr = dst.Address();
     const GPUVAddr src_addr = src.Address();
     ImageInfo dst_info(dst);
@@ -1093,9 +1093,7 @@ typename TextureCache<P>::BlitImages TextureCache<P>::GetBlitImages(
         const ImageBase* const dst_image = dst_id ? &slot_images[dst_id] : nullptr;
         const ImageBase* const src_image = src_id ? &slot_images[src_id] : nullptr;
         DeduceBlitImages(dst_info, src_info, dst_image, src_image);
-        if (GetFormatType(dst_info.format) != GetFormatType(src_info.format)) {
-            continue;
-        }
+        ASSERT(GetFormatType(dst_info.format) == GetFormatType(src_info.format));
         RelaxedOptions find_options{};
         if (src_info.num_samples > 1) {
             // it's a resolve, we must enforce the same format.

src/video_core/texture_cache/util.cpp

@@ -1152,36 +1152,14 @@ bool IsSubresource(const ImageInfo& candidate, const ImageBase& image, GPUVAddr
 void DeduceBlitImages(ImageInfo& dst_info, ImageInfo& src_info, const ImageBase* dst,
                       const ImageBase* src) {
     bool is_resolve = false;
-    const auto original_src_format = src_info.format;
-    const auto original_dst_format = dst_info.format;
     if (src) {
-        if (GetFormatType(src->info.format) != SurfaceType::ColorTexture) {
-            src_info.format = src->info.format;
-        }
         is_resolve = src->info.num_samples > 1;
         src_info.num_samples = src->info.num_samples;
         src_info.size = src->info.size;
     }
-    if (dst && GetFormatType(dst->info.format) != SurfaceType::ColorTexture) {
-        dst_info.format = dst->info.format;
-    }
-    if (src && GetFormatType(src->info.format) != SurfaceType::ColorTexture) {
-        if (dst) {
-            if (GetFormatType(dst->info.format) == SurfaceType::ColorTexture) {
-                src_info.format = original_src_format;
-            }
-        } else {
-            dst_info.format = src->info.format;
-        }
-    }
-    if (dst && GetFormatType(dst->info.format) != SurfaceType::ColorTexture) {
-        if (src) {
-            if (GetFormatType(src->info.format) == SurfaceType::ColorTexture) {
-                dst_info.format = original_dst_format;
-            }
-        } else {
-            src_info.format = dst->info.format;
-        }
+    if (dst) {
+        dst_info.num_samples = dst->info.num_samples;
+        dst_info.size = dst->info.size;
     }
     ASSERT(!is_resolve || dst_info.format == src_info.format);
 }