video_core: Only enable sampler comparison for comparison fetches (#5020)

* video_core: Only enable sampler comparison for comparison fetches

* Use HashCombine for the sampler hash

src/shader_recompiler/ir/passes/resource_patching_pass.cpp

@@ -126,7 +126,8 @@ public:
     u32 Add(const SamplerResource& desc) {
         const u32 index{Add(sampler_resources, desc, [this, &desc](const auto& existing) {
             return desc.sharp_fetch == existing.sharp_fetch && desc.post_op == existing.post_op &&
-                   desc.post_op_tsharp_dw3_off == existing.post_op_tsharp_dw3_off;
+                   desc.post_op_tsharp_dw3_off == existing.post_op_tsharp_dw3_off &&
+                   desc.is_depth == existing.is_depth;
         })};
         return index;
     }
@@ -312,6 +313,7 @@ void PatchImageSharp(const ResourceDiscovery& resource, Info& info, Descriptors&
             .post_op = resource.sharps[1].post_op,
             .post_op_tsharp_dw3_off =
                 lod_prod.IsEmpty() ? UNKNOWN_LOCATION : SharpLocationFromSource(lod_prod.Inst()),
+            .is_depth = bool(inst_info.is_depth), // true for the _C (compare) opcodes
         });
         inst.SetArg(0, ir.Imm32(image_binding | sampler_binding << 16));
     } else {

src/shader_recompiler/resource.h

@@ -170,6 +170,7 @@ struct SamplerResource {
     SharpFetch<AmdGpu::Sampler> sharp_fetch{};
     SharpFetchPostOp post_op{};
     SharpLocation post_op_tsharp_dw3_off{};
+    bool is_depth{};
 
     constexpr AmdGpu::Sampler GetSharp(const auto& info) const noexcept {
         AmdGpu::Sampler sampler{};

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -950,7 +950,8 @@ void Rasterizer::BindTextures(const Shader::Info& stage, Shader::Backend::Bindin
 
     for (const auto& sampler : stage.samplers) {
         auto ssharp = sampler.GetSharp(stage);
-        const auto vk_sampler = texture_cache.GetSampler(ssharp, liverpool->regs.ta_bc_base);
+        const auto vk_sampler =
+            texture_cache.GetSampler(ssharp, liverpool->regs.ta_bc_base, sampler.is_depth);
         image_infos.emplace_back(vk_sampler, VK_NULL_HANDLE, vk::ImageLayout::eGeneral);
         auto& set_write = set_writes[set_write_index++];
         set_write.dstSet = VK_NULL_HANDLE;

src/video_core/texture_cache/sampler.cpp

@@ -9,7 +9,7 @@
 namespace VideoCore {
 
 Sampler::Sampler(const Vulkan::Instance& instance, const AmdGpu::Sampler& sampler,
-                 const AmdGpu::BorderColorBuffer border_color_base) {
+                 const AmdGpu::BorderColorBuffer border_color_base, const bool is_depth) {
     using namespace Vulkan;
     const bool anisotropy_enable = instance.IsAnisotropicFilteringSupported() &&
                                    (AmdGpu::IsAnisoFilter(sampler.xy_mag_filter) ||
@@ -54,7 +54,8 @@ Sampler::Sampler(const Vulkan::Instance& instance, const AmdGpu::Sampler& sample
         .mipLodBias = std::min(sampler.LodBias(), instance.MaxSamplerLodBias()),
         .anisotropyEnable = anisotropy_enable,
         .maxAnisotropy = max_anisotropy,
-        .compareEnable = sampler.depth_compare_func != AmdGpu::DepthCompare::Never,
+        // GCN compares per instruction; a plain read of a compare sampler is undefined in Vulkan.
+        .compareEnable = is_depth,
         .compareOp = LiverpoolToVK::DepthCompare(sampler.depth_compare_func),
         .minLod = sampler.MinLod(),
         .maxLod = sampler.MaxLod(),

src/video_core/texture_cache/sampler.h

@@ -16,7 +16,7 @@ namespace VideoCore {
 class Sampler {
 public:
     explicit Sampler(const Vulkan::Instance& instance, const AmdGpu::Sampler& sampler,
-                     const AmdGpu::BorderColorBuffer border_color_base);
+                     const AmdGpu::BorderColorBuffer border_color_base, bool is_depth);
     ~Sampler();
 
     Sampler(const Sampler&) = delete;

src/video_core/texture_cache/texture_cache.cpp

@@ -6,6 +6,7 @@
 #include "common/assert.h"
 #include "common/debug.h"
 #include "common/div_ceil.h"
+#include "common/hash.h"
 #include "common/scope_exit.h"
 #include "core/emulator_settings.h"
 #include "core/memory.h"
@@ -766,11 +767,14 @@ void TextureCache::RefreshImage(Image& image) {
 }
 
 vk::Sampler TextureCache::GetSampler(const AmdGpu::Sampler& sampler,
-                                     AmdGpu::BorderColorBuffer border_color_base) {
-    const u64 hash = XXH3_64bits(&sampler, sizeof(sampler));
+                                     AmdGpu::BorderColorBuffer border_color_base,
+                                     const bool is_depth) {
+    // Compare and plain uses of one S# need separate samplers.
+    const u64 hash = HashCombine(XXH3_64bits(&sampler, sizeof(sampler)), is_depth);
 
     std::scoped_lock lock{samplers_mutex};
-    const auto [it, new_sampler] = samplers.try_emplace(hash, instance, sampler, border_color_base);
+    const auto [it, new_sampler] =
+        samplers.try_emplace(hash, instance, sampler, border_color_base, is_depth);
     if (new_sampler) {
         samplers.at(hash).lru_id = sampler_lru_cache.Insert(hash, gc_tick);
     } else {

src/video_core/texture_cache/texture_cache.h

@@ -143,7 +143,8 @@ public:
 
     /// Retrieves the sampler that matches the provided S# descriptor.
     [[nodiscard]] vk::Sampler GetSampler(const AmdGpu::Sampler& sampler,
-                                         AmdGpu::BorderColorBuffer border_color_base);
+                                         AmdGpu::BorderColorBuffer border_color_base,
+                                         bool is_depth);
 
     /// Retrieves the image with the specified id.
     [[nodiscard]] Image& GetImage(ImageId id) {