buffer_cache: Fix copy based uniform bindings tracking

master
ReinUsesLisp 2021-06-10 02:24:12 +07:00 committed by ameerj
parent 15bdd27cac
commit 60a96c49e5
2 changed files with 22 additions and 9 deletions

@ -680,6 +680,9 @@ void BufferCache<P>::SetUniformBuffersState(const std::array<u32, NUM_STAGES>& m
const UniformBufferSizes* sizes) {
if constexpr (HAS_PERSISTENT_UNIFORM_BUFFER_BINDINGS) {
if (enabled_uniform_buffer_masks != mask) {
if constexpr (IS_OPENGL) {
fast_bound_uniform_buffers.fill(0);
}
dirty_uniform_buffers.fill(~u32{0});
}
}
@ -1020,6 +1023,7 @@ void BufferCache<P>::BindHostGraphicsUniformBuffer(size_t stage, u32 index, u32
// Fast path for Nvidia
if (!HasFastUniformBufferBound(stage, binding_index)) {
// We only have to bind when the currently bound buffer is not the fast version
fast_bound_uniform_buffers[stage] |= 1U << binding_index;
runtime.BindFastUniformBuffer(stage, binding_index, size);
}
const auto span = ImmediateBufferWithData(cpu_addr, size);
@ -1027,8 +1031,9 @@ void BufferCache<P>::BindHostGraphicsUniformBuffer(size_t stage, u32 index, u32
return;
}
}
fast_bound_uniform_buffers[stage] |= 1U << binding_index;
if constexpr (IS_OPENGL) {
fast_bound_uniform_buffers[stage] |= 1U << binding_index;
}
// Stream buffer path to avoid stalling on non-Nvidia drivers or Vulkan
const std::span<u8> span = runtime.BindMappedUniformBuffer(stage, binding_index, size);
cpu_memory.ReadBlockUnsafe(cpu_addr, span.data(), size);
@ -1046,9 +1051,15 @@ void BufferCache<P>::BindHostGraphicsUniformBuffer(size_t stage, u32 index, u32
// This exists to avoid instances where the fast buffer is bound and a GPU write happens
return;
}
fast_bound_uniform_buffers[stage] &= ~(1U << binding_index);
const u32 offset = buffer.Offset(cpu_addr);
if constexpr (IS_OPENGL) {
// Fast buffer will be unbound
fast_bound_uniform_buffers[stage] &= ~(1U << binding_index);
// Mark the index as dirty if offset doesn't match
const bool is_copy_bind = offset != 0 && !runtime.SupportsNonZeroUniformOffset();
dirty_uniform_buffers[stage] |= (is_copy_bind ? 1U : 0U) << index;
}
if constexpr (NEEDS_BIND_UNIFORM_INDEX) {
runtime.BindUniformBuffer(stage, binding_index, buffer, offset, size);
} else {

@ -92,16 +92,14 @@ public:
VideoCore::Surface::PixelFormat format);
void BindFastUniformBuffer(size_t stage, u32 binding_index, u32 size) {
const GLuint handle = fast_uniforms[stage][binding_index].handle;
const GLsizeiptr gl_size = static_cast<GLsizeiptr>(size);
if (use_assembly_shaders) {
const GLuint handle = fast_uniforms[stage][binding_index].handle;
const GLsizeiptr gl_size = static_cast<GLsizeiptr>(size);
glBindBufferRangeNV(PABO_LUT[stage], binding_index, handle, 0, gl_size);
} else {
const GLuint base_binding = graphics_base_uniform_bindings[stage];
const GLuint binding = base_binding + binding_index;
glBindBufferRange(GL_UNIFORM_BUFFER, binding,
fast_uniforms[stage][binding_index].handle, 0,
static_cast<GLsizeiptr>(size));
glBindBufferRange(GL_UNIFORM_BUFFER, binding, handle, 0, gl_size);
}
}
@ -134,6 +132,10 @@ public:
return has_fast_buffer_sub_data;
}
[[nodiscard]] bool SupportsNonZeroUniformOffset() const noexcept {
return !use_assembly_shaders;
}
void SetBaseUniformBindings(const std::array<GLuint, 5>& bindings) {
graphics_base_uniform_bindings = bindings;
}