shader: Add support for native 16-bit floats

This commit is contained in:
ReinUsesLisp 2021-06-16 03:03:08 -03:00 committed by ameerj
parent 376aa94819
commit cbbca26d18
9 changed files with 50 additions and 14 deletions

View file

@ -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]};
Shader::IR::Program program_vb{TranslateProgram(pools.inst, pools.block, env, cfg)};
auto& program_va{programs[0]};
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) {

View file

@ -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;
};

View file

@ -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]};
Shader::IR::Program program_vb{TranslateProgram(pools.inst, pools.block, env, cfg)};
auto& program_va{programs[0]};
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)};

View file

@ -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;

View file

@ -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");
// is_float16_supported = false;
LOG_WARNING(Render_Vulkan, "Blacklisting Intel proprietary from float16 math");
is_float16_supported = false;
}
graphics_queue = logical.GetQueue(graphics_family);