[D3D12] DXBC samplers

src/xenia/gpu/d3d12/d3d12_command_processor.cc

@@ -129,11 +129,11 @@ ID3D12RootSignature* D3D12CommandProcessor::GetRootSignature(
   uint32_t pixel_texture_count = 0, pixel_sampler_count = 0;
   if (pixel_shader != nullptr) {
     pixel_shader->GetTextureSRVs(pixel_texture_count);
-    pixel_shader->GetSamplerFetchConstants(pixel_sampler_count);
+    pixel_shader->GetSamplerBindings(pixel_sampler_count);
   }
   uint32_t vertex_texture_count, vertex_sampler_count;
   vertex_shader->GetTextureSRVs(vertex_texture_count);
-  vertex_shader->GetSamplerFetchConstants(vertex_sampler_count);
+  vertex_shader->GetSamplerBindings(vertex_sampler_count);
   // Max 96 textures (if all kinds of tfetch instructions are used for all fetch
   // registers) and 32 samplers (one sampler per used fetch), but different
   // shader stages have different texture sets.
@@ -319,11 +319,11 @@ uint32_t D3D12CommandProcessor::GetRootExtraParameterIndices(
   uint32_t pixel_texture_count = 0, pixel_sampler_count = 0;
   if (pixel_shader != nullptr) {
     pixel_shader->GetTextureSRVs(pixel_texture_count);
-    pixel_shader->GetSamplerFetchConstants(pixel_sampler_count);
+    pixel_shader->GetSamplerBindings(pixel_sampler_count);
   }
   uint32_t vertex_texture_count, vertex_sampler_count;
   vertex_shader->GetTextureSRVs(vertex_texture_count);
-  vertex_shader->GetSamplerFetchConstants(vertex_sampler_count);
+  vertex_shader->GetSamplerBindings(vertex_sampler_count);
 
   uint32_t index = kRootParameter_Count_Base;
   if (pixel_texture_count != 0) {
@@ -1525,11 +1525,10 @@ bool D3D12CommandProcessor::UpdateBindings(
   // Get used textures and samplers.
   uint32_t pixel_texture_count, pixel_sampler_count;
   const D3D12Shader::TextureSRV* pixel_textures;
-  const uint32_t* pixel_samplers;
+  const D3D12Shader::SamplerBinding* pixel_samplers;
   if (pixel_shader != nullptr) {
     pixel_textures = pixel_shader->GetTextureSRVs(pixel_texture_count);
-    pixel_samplers =
+    pixel_samplers = pixel_shader->GetSamplerBindings(pixel_sampler_count);
-        pixel_shader->GetSamplerFetchConstants(pixel_sampler_count);
   } else {
     pixel_textures = nullptr;
     pixel_texture_count = 0;
@@ -1539,8 +1538,8 @@ bool D3D12CommandProcessor::UpdateBindings(
   uint32_t vertex_texture_count, vertex_sampler_count;
   const D3D12Shader::TextureSRV* vertex_textures =
       vertex_shader->GetTextureSRVs(vertex_texture_count);
-  const uint32_t* vertex_samplers =
+  const D3D12Shader::SamplerBinding* vertex_samplers =
-      vertex_shader->GetSamplerFetchConstants(vertex_sampler_count);
+      vertex_shader->GetSamplerBindings(vertex_sampler_count);
   uint32_t texture_count = pixel_texture_count + vertex_texture_count;
   uint32_t sampler_count = pixel_sampler_count + vertex_sampler_count;
 
@@ -1785,7 +1784,10 @@ bool D3D12CommandProcessor::UpdateBindings(
                   RootExtraParameterIndices::kUnavailable);
       gpu_handle_pixel_samplers_ = sampler_gpu_handle;
       for (uint32_t i = 0; i < pixel_sampler_count; ++i) {
-        texture_cache_->WriteSampler(pixel_samplers[i], sampler_cpu_handle);
+        const D3D12Shader::SamplerBinding& sampler = pixel_samplers[i];
+        texture_cache_->WriteSampler(sampler.fetch_constant, sampler.mag_filter,
+                                     sampler.min_filter, sampler.mip_filter,
+                                     sampler.aniso_filter, sampler_cpu_handle);
         sampler_cpu_handle.ptr += descriptor_size_sampler;
         sampler_gpu_handle.ptr += descriptor_size_sampler;
       }
@@ -1797,7 +1799,10 @@ bool D3D12CommandProcessor::UpdateBindings(
                   RootExtraParameterIndices::kUnavailable);
       gpu_handle_vertex_samplers_ = sampler_gpu_handle;
       for (uint32_t i = 0; i < vertex_sampler_count; ++i) {
-        texture_cache_->WriteSampler(vertex_samplers[i], sampler_cpu_handle);
+        const D3D12Shader::SamplerBinding& sampler = vertex_samplers[i];
+        texture_cache_->WriteSampler(sampler.fetch_constant, sampler.mag_filter,
+                                     sampler.min_filter, sampler.mip_filter,
+                                     sampler.aniso_filter, sampler_cpu_handle);
         sampler_cpu_handle.ptr += descriptor_size_sampler;
         sampler_gpu_handle.ptr += descriptor_size_sampler;
       }

src/xenia/gpu/d3d12/d3d12_shader.cc

@@ -29,27 +29,33 @@ D3D12Shader::~D3D12Shader() {
 
 void D3D12Shader::SetTexturesAndSamplers(
     const DxbcShaderTranslator::TextureSRV* texture_srvs,
-    uint32_t texture_srv_count, const uint32_t* sampler_fetch_constants,
+    uint32_t texture_srv_count,
-    uint32_t sampler_count) {
+    const DxbcShaderTranslator::SamplerBinding* sampler_bindings,
+    uint32_t sampler_binding_count) {
+  texture_srvs_.clear();
+  texture_srvs_.reserve(texture_srv_count);
   used_texture_mask_ = 0;
   for (uint32_t i = 0; i < texture_srv_count; ++i) {
-    TextureSRV& srv = texture_srvs_[i];
+    TextureSRV srv;
     const DxbcShaderTranslator::TextureSRV& translator_srv = texture_srvs[i];
     srv.fetch_constant = translator_srv.fetch_constant;
     srv.dimension = translator_srv.dimension;
+    texture_srvs_.push_back(srv);
     used_texture_mask_ |= 1u << translator_srv.fetch_constant;
   }
-  texture_srv_count_ = texture_srv_count;
+  sampler_bindings_.clear();
-#if 0
+  sampler_bindings_.reserve(sampler_binding_count);
-  // If there's a texture, there's a sampler for it.
+  for (uint32_t i = 0; i < texture_srv_count; ++i) {
-  used_texture_mask_ = 0;
+    SamplerBinding sampler;
-  for (uint32_t i = 0; i < sampler_count; ++i) {
+    const DxbcShaderTranslator::SamplerBinding& translator_sampler =
-    uint32_t sampler_fetch_constant = sampler_fetch_constants[i];
+        sampler_bindings[i];
-    sampler_fetch_constants_[i] = sampler_fetch_constant;
+    sampler.fetch_constant = translator_sampler.fetch_constant;
-    used_texture_mask_ |= 1u << sampler_fetch_constant;
+    sampler.mag_filter = translator_sampler.mag_filter;
+    sampler.min_filter = translator_sampler.min_filter;
+    sampler.mip_filter = translator_sampler.mip_filter;
+    sampler.aniso_filter = translator_sampler.aniso_filter;
+    sampler_bindings_.push_back(sampler);
   }
-  sampler_count_ = sampler_count;
-#endif
 }
 
 bool D3D12Shader::DisassembleDXBC() {

src/xenia/gpu/d3d12/d3d12_shader.h

@@ -10,6 +10,8 @@
 #ifndef XENIA_GPU_D3D12_D3D12_SHADER_H_
 #define XENIA_GPU_D3D12_D3D12_SHADER_H_
 
+#include <vector>
+
 #include "xenia/gpu/dxbc_shader_translator.h"
 #include "xenia/gpu/shader.h"
 #include "xenia/ui/d3d12/d3d12_api.h"
@@ -26,34 +28,48 @@ class D3D12Shader : public Shader {
 
   void SetTexturesAndSamplers(
       const DxbcShaderTranslator::TextureSRV* texture_srvs,
-      uint32_t texture_srv_count, const uint32_t* sampler_fetch_constants,
+      uint32_t texture_srv_count,
-      uint32_t sampler_count);
+      const DxbcShaderTranslator::SamplerBinding* sampler_bindings,
+      uint32_t sampler_binding_count);
 
   bool DisassembleDXBC();
 
+  static constexpr uint32_t kMaxTextureSRVIndexBits =
+      DxbcShaderTranslator::kMaxTextureSRVIndexBits;
+  static constexpr uint32_t kMaxTextureSRVs =
+      DxbcShaderTranslator::kMaxTextureSRVs;
   struct TextureSRV {
     uint32_t fetch_constant;
     TextureDimension dimension;
   };
   const TextureSRV* GetTextureSRVs(uint32_t& count_out) const {
-    count_out = texture_srv_count_;
+    count_out = uint32_t(texture_srvs_.size());
-    return texture_srvs_;
+    return texture_srvs_.data();
-  }
-  const uint32_t* GetSamplerFetchConstants(uint32_t& count_out) const {
-    count_out = sampler_count_;
-    return sampler_fetch_constants_;
   }
   const uint32_t GetUsedTextureMask() const { return used_texture_mask_; }
 
+  static constexpr uint32_t kMaxSamplerBindingIndexBits =
+      DxbcShaderTranslator::kMaxSamplerBindingIndexBits;
+  static constexpr uint32_t kMaxSamplerBindings =
+      DxbcShaderTranslator::kMaxSamplerBindings;
+  struct SamplerBinding {
+    uint32_t fetch_constant;
+    TextureFilter mag_filter;
+    TextureFilter min_filter;
+    TextureFilter mip_filter;
+    AnisoFilter aniso_filter;
+  };
+  const SamplerBinding* GetSamplerBindings(uint32_t& count_out) const {
+    count_out = uint32_t(sampler_bindings_.size());
+    return sampler_bindings_.data();
+  }
+
  private:
   ID3DBlob* blob_ = nullptr;
 
-  // Up to 32 2D array textures, 32 3D textures and 32 cube textures.
+  std::vector<TextureSRV> texture_srvs_;
-  TextureSRV texture_srvs_[96];
-  uint32_t texture_srv_count_ = 0;
-  uint32_t sampler_fetch_constants_[32];
-  uint32_t sampler_count_ = 0;
   uint32_t used_texture_mask_ = 0;
+  std::vector<SamplerBinding> sampler_bindings_;
 };
 
 }  // namespace d3d12

src/xenia/gpu/d3d12/pipeline_cache.cc

@@ -196,17 +196,14 @@ bool PipelineCache::TranslateShader(D3D12Shader* shader,
     return false;
   }
 
-  // TODO(Triang3l): Re-enable this when the DXBC shader translators supports
-  // textures.
   uint32_t texture_srv_count;
   const DxbcShaderTranslator::TextureSRV* texture_srvs =
       shader_translator_->GetTextureSRVs(texture_srv_count);
-#if 0
+  uint32_t sampler_binding_count;
-  uint32_t sampler_count;
+  const DxbcShaderTranslator::SamplerBinding* sampler_bindings =
-  const uint32_t* sampler_fetch_constants =
+      shader_translator_->GetSamplerBindings(sampler_binding_count);
-      shader_translator_->GetSamplerFetchConstants(sampler_count);
+  shader->SetTexturesAndSamplers(texture_srvs, texture_srv_count,
-#endif
+                                 sampler_bindings, sampler_binding_count);
-  shader->SetTexturesAndSamplers(texture_srvs, texture_srv_count, nullptr, 0);
 
   if (shader->is_valid()) {
     XELOGGPU("Generated %s shader (%db) - hash %.16" PRIX64 ":\n%s\n",

src/xenia/gpu/d3d12/texture_cache.cc

@@ -445,34 +445,46 @@ void TextureCache::WriteTextureSRV(uint32_t fetch_constant,
 }
 
 void TextureCache::WriteSampler(uint32_t fetch_constant,
+                                TextureFilter mag_filter,
+                                TextureFilter min_filter,
+                                TextureFilter mip_filter,
+                                AnisoFilter aniso_filter,
                                 D3D12_CPU_DESCRIPTOR_HANDLE handle) {
   auto& regs = *register_file_;
   uint32_t r = XE_GPU_REG_SHADER_CONSTANT_FETCH_00_0 + fetch_constant * 6;
   auto group =
       reinterpret_cast<const xenos::xe_gpu_fetch_group_t*>(&regs.values[r]);
   auto& fetch = group->texture_fetch;
-  // TODO(Triang3l): Fetch shader instruction overrides.
+  if (mag_filter == TextureFilter::kUseFetchConst) {
+    mag_filter = TextureFilter(fetch.mag_filter);
+  }
+  if (min_filter == TextureFilter::kUseFetchConst) {
+    min_filter = TextureFilter(fetch.min_filter);
+  }
+  if (mip_filter == TextureFilter::kUseFetchConst) {
+    mip_filter = TextureFilter(fetch.mip_filter);
+  }
+  if (aniso_filter == AnisoFilter::kUseFetchConst) {
+    aniso_filter = AnisoFilter(fetch.aniso_filter);
+  }
   D3D12_SAMPLER_DESC desc;
   if (fetch.aniso_filter) {
     desc.Filter = D3D12_FILTER_ANISOTROPIC;
-    desc.MaxAnisotropy = std::min(1u << (fetch.aniso_filter - 1), 16u);
+    desc.MaxAnisotropy = std::min(1u << (uint32_t(aniso_filter) - 1), 16u);
   } else {
-    D3D12_FILTER_TYPE filter_min =
+    D3D12_FILTER_TYPE d3d_filter_min = min_filter == TextureFilter::kLinear
-        TextureFilter(fetch.min_filter) == TextureFilter::kLinear
+                                           ? D3D12_FILTER_TYPE_LINEAR
-            ? D3D12_FILTER_TYPE_LINEAR
+                                           : D3D12_FILTER_TYPE_POINT;
-            : D3D12_FILTER_TYPE_POINT;
+    D3D12_FILTER_TYPE d3d_filter_mag = mag_filter == TextureFilter::kLinear
-    D3D12_FILTER_TYPE filter_mag =
+                                           ? D3D12_FILTER_TYPE_LINEAR
-        TextureFilter(fetch.mag_filter) == TextureFilter::kLinear
+                                           : D3D12_FILTER_TYPE_POINT;
-            ? D3D12_FILTER_TYPE_LINEAR
+    D3D12_FILTER_TYPE d3d_filter_mip = mip_filter == TextureFilter::kLinear
-            : D3D12_FILTER_TYPE_POINT;
+                                           ? D3D12_FILTER_TYPE_LINEAR
-    D3D12_FILTER_TYPE filter_mip =
+                                           : D3D12_FILTER_TYPE_POINT;
-        TextureFilter(fetch.mip_filter) == TextureFilter::kLinear
-            ? D3D12_FILTER_TYPE_LINEAR
-            : D3D12_FILTER_TYPE_POINT;
     // TODO(Triang3l): Investigate mip_filter TextureFilter::kBaseMap.
-    desc.Filter =
+    desc.Filter = D3D12_ENCODE_BASIC_FILTER(
-        D3D12_ENCODE_BASIC_FILTER(filter_min, filter_mag, filter_mip,
+        d3d_filter_min, d3d_filter_mag, d3d_filter_mip,
-                                  D3D12_FILTER_REDUCTION_TYPE_STANDARD);
+        D3D12_FILTER_REDUCTION_TYPE_STANDARD);
     desc.MaxAnisotropy = 1;
   }
   // FIXME(Triang3l): Halfway and mirror clamp to border aren't mapped properly.

src/xenia/gpu/d3d12/texture_cache.h

@@ -76,7 +76,9 @@ class TextureCache {
   void WriteTextureSRV(uint32_t fetch_constant,
                        TextureDimension shader_dimension,
                        D3D12_CPU_DESCRIPTOR_HANDLE handle);
-  void WriteSampler(uint32_t fetch_constant,
+  void WriteSampler(uint32_t fetch_constant, TextureFilter mag_filter,
+                    TextureFilter min_filter, TextureFilter mip_filter,
+                    AnisoFilter aniso_filter,
                     D3D12_CPU_DESCRIPTOR_HANDLE handle);
 
   static DXGI_FORMAT GetResolveDXGIFormat(TextureFormat format);

src/xenia/gpu/dxbc_shader_translator.cc

@@ -13,6 +13,8 @@
 
 #include <algorithm>
 #include <cstring>
+#include <memory>
+#include <sstream>
 
 #include "third_party/dxbc/DXBCChecksum.h"
 #include "third_party/dxbc/d3d12TokenizedProgramFormat.hpp"
@@ -2870,6 +2872,10 @@ uint32_t DxbcShaderTranslator::FindOrAddTextureSRV(uint32_t fetch_constant,
       return 1 + i;
     }
   }
+  if (texture_srvs_.size() >= kMaxTextureSRVs) {
+    assert_always();
+    return 1 + (kMaxTextureSRVs - 1);
+  }
   TextureSRV new_texture_srv;
   new_texture_srv.fetch_constant = fetch_constant;
   new_texture_srv.dimension = dimension;
@@ -2887,10 +2893,63 @@ uint32_t DxbcShaderTranslator::FindOrAddTextureSRV(uint32_t fetch_constant,
   new_texture_srv.name =
       xe::format_string("xe_texture%u_%s", fetch_constant, dimension_name);
   uint32_t srv_register = 1 + uint32_t(texture_srvs_.size());
-  texture_srvs_.push_back(new_texture_srv);
+  texture_srvs_.emplace_back(std::move(new_texture_srv));
   return srv_register;
 }
 
+uint32_t DxbcShaderTranslator::FindOrAddSamplerBinding(
+    uint32_t fetch_constant, TextureFilter mag_filter, TextureFilter min_filter,
+    TextureFilter mip_filter, AnisoFilter aniso_filter) {
+  // In Direct3D 12, anisotropic filtering implies linear filtering.
+  if (aniso_filter != AnisoFilter::kDisabled &&
+      aniso_filter != AnisoFilter::kUseFetchConst) {
+    mag_filter = TextureFilter::kLinear;
+    min_filter = TextureFilter::kLinear;
+    mip_filter = TextureFilter::kLinear;
+    aniso_filter = std::min(aniso_filter, AnisoFilter::kMax_16_1);
+  }
+
+  for (uint32_t i = 0; i < uint32_t(sampler_bindings_.size()); ++i) {
+    const SamplerBinding& sampler_binding = sampler_bindings_[i];
+    if (sampler_binding.fetch_constant == fetch_constant &&
+        sampler_binding.mag_filter == mag_filter &&
+        sampler_binding.min_filter == min_filter &&
+        sampler_binding.mip_filter == mip_filter &&
+        sampler_binding.aniso_filter == aniso_filter) {
+      return i;
+    }
+  }
+
+  if (sampler_bindings_.size() >= kMaxSamplerBindings) {
+    assert_always();
+    return kMaxSamplerBindings - 1;
+  }
+
+  std::ostringstream name;
+  name << "xe_sampler" << fetch_constant;
+  if (aniso_filter != AnisoFilter::kUseFetchConst) {
+    name << "_a" << (1u << uint32_t(aniso_filter));
+  }
+  if (aniso_filter == AnisoFilter::kDisabled ||
+      aniso_filter == AnisoFilter::kUseFetchConst) {
+    static const char* kFilterSuffixes[] = {"p", "l", "b", "f"};
+    name << "_" << kFilterSuffixes[uint32_t(mag_filter)]
+         << kFilterSuffixes[uint32_t(min_filter)]
+         << kFilterSuffixes[uint32_t(mip_filter)];
+  }
+
+  SamplerBinding new_sampler_binding;
+  new_sampler_binding.fetch_constant = fetch_constant;
+  new_sampler_binding.mag_filter = mag_filter;
+  new_sampler_binding.min_filter = min_filter;
+  new_sampler_binding.mip_filter = mip_filter;
+  new_sampler_binding.aniso_filter = aniso_filter;
+  new_sampler_binding.name = name.str();
+  uint32_t sampler_register = 1 + uint32_t(sampler_bindings_.size());
+  sampler_bindings_.emplace_back(std::move(new_sampler_binding));
+  return sampler_register;
+}
+
 void DxbcShaderTranslator::ProcessTextureFetchInstruction(
     const ParsedTextureFetchInstruction& instr) {
   CheckPredicate(instr.is_predicated, instr.predicate_condition);
@@ -2927,7 +2986,11 @@ void DxbcShaderTranslator::ProcessTextureFetchInstruction(
       srv_register_3d = UINT32_MAX;
     }
 
-    // TODO(Triang3l): Sampler, actually sample instead of this stub.
+    uint32_t sampler_register = FindOrAddSamplerBinding(
+        tfetch_index, instr.attributes.mag_filter, instr.attributes.min_filter,
+        instr.attributes.mip_filter, instr.attributes.aniso_filter);
+
+    // TODO(Triang3l): Actually sample instead of this stub.
     shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_MOV) |
                            ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(8));
     shader_code_.push_back(

src/xenia/gpu/dxbc_shader_translator.h

@@ -62,6 +62,11 @@ class DxbcShaderTranslator : public ShaderTranslator {
     uint32_t color_output_map[4];
   };
 
+  // 96 textures at most because there are 32 fetch constants, and textures can
+  // be 2D array, 3D or cube.
+  static constexpr uint32_t kMaxTextureSRVIndexBits = 7;
+  static constexpr uint32_t kMaxTextureSRVs =
+      (1 << kMaxTextureSRVIndexBits) - 1;
   struct TextureSRV {
     uint32_t fetch_constant;
     TextureDimension dimension;
@@ -73,6 +78,17 @@ class DxbcShaderTranslator : public ShaderTranslator {
     return texture_srvs_.data();
   }
 
+  // Arbitrary limit - there can't be more than 2048 in a shader-visible
+  // descriptor heap, though some older hardware (tier 1 resource binding -
+  // Nvidia Fermi) doesn't support more than 16 samplers bound at once (we can't
+  // really do anything if a game uses more than 16), but just to have some
+  // limit so sampler count can easily be packed into 32-bit map keys (for
+  // instance, for root signatures). But shaders can specify overrides for
+  // filtering modes, and the number of possible combinations is huge - let's
+  // limit it to something sane.
+  static constexpr uint32_t kMaxSamplerBindingIndexBits = 7;
+  static constexpr uint32_t kMaxSamplerBindings =
+      (1 << kMaxSamplerBindingIndexBits) - 1;
   struct SamplerBinding {
     uint32_t fetch_constant;
     TextureFilter mag_filter;
@@ -341,6 +357,12 @@ class DxbcShaderTranslator : public ShaderTranslator {
   // Returns T#/t# index (they are the same in this translator).
   uint32_t FindOrAddTextureSRV(uint32_t fetch_constant,
                                TextureDimension dimension);
+  // Returns S#/s# index (they are the same in this translator).
+  uint32_t FindOrAddSamplerBinding(uint32_t fetch_constant,
+                                   TextureFilter mag_filter,
+                                   TextureFilter min_filter,
+                                   TextureFilter mip_filter,
+                                   AnisoFilter aniso_filter);
 
   void ProcessVectorAluInstruction(const ParsedAluInstruction& instr);
   void ProcessScalarAluInstruction(const ParsedAluInstruction& instr);