gl_device: Expose NV_vertex_buffer_unified_memory except on Turing

Expose NV_vertex_buffer_unified_memory when the driver supports it.

This commit adds a function the determine if a GL_RENDERER is a Turing
GPU. This is required because on Turing GPUs Nvidia's driver crashes
when the buffer is marked as resident or on DeleteBuffers. Without a
synchronous debug output (single threaded driver), it's likely that
the driver will crash in the first blocking call.
This commit is contained in:
ReinUsesLisp 2020-06-18 03:54:13 -03:00
parent 00c66a7289
commit 73fb3a304b
2 changed files with 30 additions and 1 deletions

View file

@ -188,16 +188,32 @@ bool IsASTCSupported() {
return true;
}
/// @brief Returns true when a GL_RENDERER is a Turing GPU
/// @param renderer GL_RENDERER string
bool IsTuring(std::string_view renderer) {
static constexpr std::array<std::string_view, 12> TURING_GPUS = {
"GTX 1650", "GTX 1660", "RTX 2060", "RTX 2070",
"RTX 2080", "TITAN RTX", "Quadro RTX 3000", "Quadro RTX 4000",
"Quadro RTX 5000", "Quadro RTX 6000", "Quadro RTX 8000", "Tesla T4",
};
return std::any_of(TURING_GPUS.begin(), TURING_GPUS.end(),
[renderer](std::string_view candidate) {
return renderer.find(candidate) != std::string_view::npos;
});
}
} // Anonymous namespace
Device::Device()
: max_uniform_buffers{BuildMaxUniformBuffers()}, base_bindings{BuildBaseBindings()} {
const std::string_view vendor = reinterpret_cast<const char*>(glGetString(GL_VENDOR));
const std::string_view renderer = reinterpret_cast<const char*>(glGetString(GL_RENDERER));
const std::string_view version = reinterpret_cast<const char*>(glGetString(GL_VERSION));
const std::vector extensions = GetExtensions();
const bool is_nvidia = vendor == "NVIDIA Corporation";
const bool is_amd = vendor == "ATI Technologies Inc.";
const bool is_turing = is_nvidia && IsTuring(renderer);
bool disable_fast_buffer_sub_data = false;
if (is_nvidia && version == "4.6.0 NVIDIA 443.24") {
@ -221,8 +237,16 @@ Device::Device()
has_variable_aoffi = TestVariableAoffi();
has_component_indexing_bug = is_amd;
has_precise_bug = TestPreciseBug();
has_fast_buffer_sub_data = is_nvidia && !disable_fast_buffer_sub_data;
has_nv_viewport_array2 = GLAD_GL_NV_viewport_array2;
// At the moment of writing this, only Nvidia's driver optimizes BufferSubData on exclusive
// uniform buffers as "push constants"
has_fast_buffer_sub_data = is_nvidia && !disable_fast_buffer_sub_data;
// Nvidia's driver on Turing GPUs randomly crashes when the buffer is made resident, or on
// DeleteBuffers. Disable unified memory on these devices.
has_vertex_buffer_unified_memory = GLAD_GL_NV_vertex_buffer_unified_memory && !is_turing;
use_assembly_shaders = Settings::values.use_assembly_shaders && GLAD_GL_NV_gpu_program5 &&
GLAD_GL_NV_compute_program5 && GLAD_GL_NV_transform_feedback &&
GLAD_GL_NV_transform_feedback2;

View file

@ -72,6 +72,10 @@ public:
return has_texture_shadow_lod;
}
bool HasVertexBufferUnifiedMemory() const {
return has_vertex_buffer_unified_memory;
}
bool HasASTC() const {
return has_astc;
}
@ -115,6 +119,7 @@ private:
bool has_vertex_viewport_layer{};
bool has_image_load_formatted{};
bool has_texture_shadow_lod{};
bool has_vertex_buffer_unified_memory{};
bool has_astc{};
bool has_variable_aoffi{};
bool has_component_indexing_bug{};