shader: Add support for native 16-bit floats

src/shader_recompiler/CMakeLists.txt

@@ -211,6 +211,7 @@ add_library(shader_recompiler STATIC
     frontend/maxwell/translate/translate.h
     frontend/maxwell/translate_program.cpp
     frontend/maxwell/translate_program.h
+    host_translate_info.h
     ir_opt/collect_shader_info_pass.cpp
     ir_opt/constant_propagation_pass.cpp
     ir_opt/dead_code_elimination_pass.cpp

src/shader_recompiler/frontend/maxwell/translate_program.cpp

@@ -13,6 +13,7 @@
 #include "shader_recompiler/frontend/maxwell/structured_control_flow.h"
 #include "shader_recompiler/frontend/maxwell/translate/translate.h"
 #include "shader_recompiler/frontend/maxwell/translate_program.h"
+#include "shader_recompiler/host_translate_info.h"
 #include "shader_recompiler/ir_opt/passes.h"
 
 namespace Shader::Maxwell {
@@ -120,7 +121,7 @@ void AddNVNStorageBuffers(IR::Program& program) {
 } // Anonymous namespace
 
 IR::Program TranslateProgram(ObjectPool<IR::Inst>& inst_pool, ObjectPool<IR::Block>& block_pool,
-                             Environment& env, Flow::CFG& cfg) {
+                             Environment& env, Flow::CFG& cfg, const HostTranslateInfo& host_info) {
     IR::Program program;
     program.syntax_list = BuildASL(inst_pool, block_pool, env, cfg);
     program.blocks = GenerateBlocks(program.syntax_list);
@@ -150,8 +151,9 @@ IR::Program TranslateProgram(ObjectPool<IR::Inst>& inst_pool, ObjectPool<IR::Blo
     RemoveUnreachableBlocks(program);
 
     // Replace instructions before the SSA rewrite
+    if (!host_info.support_float16) {
         Optimization::LowerFp16ToFp32(program);
-
+    }
     Optimization::SsaRewritePass(program);
 
     Optimization::GlobalMemoryToStorageBufferPass(program);

src/shader_recompiler/frontend/maxwell/translate_program.h

@@ -8,13 +8,14 @@
 #include "shader_recompiler/frontend/ir/basic_block.h"
 #include "shader_recompiler/frontend/ir/program.h"
 #include "shader_recompiler/frontend/maxwell/control_flow.h"
+#include "shader_recompiler/host_translate_info.h"
 #include "shader_recompiler/object_pool.h"
 
 namespace Shader::Maxwell {
 
 [[nodiscard]] IR::Program TranslateProgram(ObjectPool<IR::Inst>& inst_pool,
                                            ObjectPool<IR::Block>& block_pool, Environment& env,
-                                           Flow::CFG& cfg);
+                                           Flow::CFG& cfg, const HostTranslateInfo& host_info);
 
 [[nodiscard]] IR::Program MergeDualVertexPrograms(IR::Program& vertex_a, IR::Program& vertex_b,
                                                   Environment& env_vertex_b);

src/shader_recompiler/host_translate_info.h

@@ -0,0 +1,18 @@
+// Copyright 2021 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+namespace Shader {
+
+// Try to keep entries here to a minimum
+// They can accidentally change the cached information in a shader
+
+/// Misc information about the host
+struct HostTranslateInfo {
+    bool support_float16{}; ///< True when the device supports 16-bit floats
+    bool support_int64{};   ///< True when the device supports 64-bit integers
+};
+
+} // namespace Shader

src/video_core/renderer_opengl/gl_shader_cache.cpp

@@ -201,6 +201,10 @@ ShaderCache::ShaderCache(RasterizerOpenGL& rasterizer_, Core::Frontend::EmuWindo
           .has_gl_component_indexing_bug = device.HasComponentIndexingBug(),
           .has_gl_precise_bug = device.HasPreciseBug(),
           .ignore_nan_fp_comparisons = true,
+      },
+      host_info{
+          .support_float16 = false,
+          .support_int64 = true,
       } {
     if (use_asynchronous_shaders) {
         workers = CreateWorkers();
@@ -373,15 +377,15 @@ std::unique_ptr<GraphicsPipeline> ShaderCache::CreateGraphicsPipeline(
         Shader::Maxwell::Flow::CFG cfg(env, pools.flow_block, cfg_offset, index == 0);
         if (!uses_vertex_a || index != 1) {
             // Normal path
-            programs[index] = TranslateProgram(pools.inst, pools.block, env, cfg);
+            programs[index] = TranslateProgram(pools.inst, pools.block, env, cfg, host_info);
 
             for (const auto& desc : programs[index].info.storage_buffers_descriptors) {
                 total_storage_buffers += desc.count;
             }
         } else {
             // VertexB path when VertexA is present.
-            Shader::IR::Program& program_va{programs[0]};
+            auto& program_va{programs[0]};
-            Shader::IR::Program program_vb{TranslateProgram(pools.inst, pools.block, env, cfg)};
+            auto program_vb{TranslateProgram(pools.inst, pools.block, env, cfg, host_info)};
             for (const auto& desc : program_vb.info.storage_buffers_descriptors) {
                 total_storage_buffers += desc.count;
             }
@@ -449,7 +453,7 @@ std::unique_ptr<ComputePipeline> ShaderCache::CreateComputePipeline(
     LOG_INFO(Render_OpenGL, "0x{:016x}", key.Hash());
 
     Shader::Maxwell::Flow::CFG cfg{env, pools.flow_block, env.StartAddress()};
-    Shader::IR::Program program{TranslateProgram(pools.inst, pools.block, env, cfg)};
+    auto program{TranslateProgram(pools.inst, pools.block, env, cfg, host_info)};
 
     u32 num_storage_buffers{};
     for (const auto& desc : program.info.storage_buffers_descriptors) {

src/video_core/renderer_opengl/gl_shader_cache.h

@@ -14,6 +14,7 @@
 #include "common/common_types.h"
 #include "common/thread_worker.h"
 #include "shader_recompiler/frontend/ir/value.h"
+#include "shader_recompiler/host_translate_info.h"
 #include "shader_recompiler/object_pool.h"
 #include "video_core/engines/shader_type.h"
 #include "video_core/renderer_opengl/gl_compute_pipeline.h"
@@ -82,6 +83,8 @@ private:
     std::unordered_map<ComputePipelineKey, std::unique_ptr<ComputePipeline>> compute_cache;
 
     Shader::Profile profile;
+    Shader::HostTranslateInfo host_info;
+
     std::filesystem::path shader_cache_filename;
     std::unique_ptr<ShaderWorker> workers;
 };

src/video_core/renderer_vulkan/vk_pipeline_cache.cpp

@@ -307,6 +307,10 @@ PipelineCache::PipelineCache(RasterizerVulkan& rasterizer_, Tegra::Engines::Maxw
         .has_broken_signed_operations = false,
         .ignore_nan_fp_comparisons = false,
     };
+    host_info = Shader::HostTranslateInfo{
+        .support_float16 = device.IsFloat16Supported(),
+        .support_int64 = true,
+    };
 }
 
 PipelineCache::~PipelineCache() = default;
@@ -484,11 +488,11 @@ std::unique_ptr<GraphicsPipeline> PipelineCache::CreateGraphicsPipeline(
         Shader::Maxwell::Flow::CFG cfg(env, pools.flow_block, cfg_offset, index == 0);
         if (!uses_vertex_a || index != 1) {
             // Normal path
-            programs[index] = TranslateProgram(pools.inst, pools.block, env, cfg);
+            programs[index] = TranslateProgram(pools.inst, pools.block, env, cfg, host_info);
         } else {
             // VertexB path when VertexA is present.
-            Shader::IR::Program& program_va{programs[0]};
+            auto& program_va{programs[0]};
-            Shader::IR::Program program_vb{TranslateProgram(pools.inst, pools.block, env, cfg)};
+            auto program_vb{TranslateProgram(pools.inst, pools.block, env, cfg, host_info)};
             programs[index] = MergeDualVertexPrograms(program_va, program_vb, env);
         }
     }
@@ -575,7 +579,7 @@ std::unique_ptr<ComputePipeline> PipelineCache::CreateComputePipeline(
     LOG_INFO(Render_Vulkan, "0x{:016x}", key.Hash());
 
     Shader::Maxwell::Flow::CFG cfg{env, pools.flow_block, env.StartAddress()};
-    Shader::IR::Program program{TranslateProgram(pools.inst, pools.block, env, cfg)};
+    auto program{TranslateProgram(pools.inst, pools.block, env, cfg, host_info)};
     const std::vector<u32> code{EmitSPIRV(profile, program)};
     device.SaveShader(code);
     vk::ShaderModule spv_module{BuildShader(device, code)};

src/video_core/renderer_vulkan/vk_pipeline_cache.h

@@ -19,6 +19,7 @@
 #include "shader_recompiler/frontend/ir/basic_block.h"
 #include "shader_recompiler/frontend/ir/value.h"
 #include "shader_recompiler/frontend/maxwell/control_flow.h"
+#include "shader_recompiler/host_translate_info.h"
 #include "shader_recompiler/object_pool.h"
 #include "shader_recompiler/profile.h"
 #include "video_core/engines/maxwell_3d.h"
@@ -157,6 +158,8 @@ private:
     ShaderPools main_pools;
 
     Shader::Profile profile;
+    Shader::HostTranslateInfo host_info;
+
     std::filesystem::path pipeline_cache_filename;
 
     Common::ThreadWorker workers;

src/video_core/vulkan_common/vulkan_device.cpp

@@ -497,8 +497,8 @@ Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR
     }
     if (is_float16_supported && driver_id == VK_DRIVER_ID_INTEL_PROPRIETARY_WINDOWS) {
         // Intel's compiler crashes when using fp16 on Astral Chain, disable it for the time being.
-        // LOG_WARNING(Render_Vulkan, "Blacklisting Intel proprietary from float16 math");
+        LOG_WARNING(Render_Vulkan, "Blacklisting Intel proprietary from float16 math");
-        // is_float16_supported = false;
+        is_float16_supported = false;
     }
 
     graphics_queue = logical.GetQueue(graphics_family);