Plexi09/eden-miror
1
0
Fork 0
mirror of https://git.eden-emu.dev/eden-emu/eden synced 2026-04-11 09:38:57 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 2

[vk] Fast UBO: fix tracking, resize heuristics, add debug guard (#2695)

Browse source 
Co-authored-by: Ribbit <ribbit@placeholder.com>
Reviewed-on: https://git.eden-emu.dev/eden-emu/eden/pulls/2695
Reviewed-by: CamilleLaVey <camillelavey99@gmail.com>
Co-authored-by: Ribbit <ribbit@eden-emu.dev>
Co-committed-by: Ribbit <ribbit@eden-emu.dev>
This commit is contained in:
Ribbit 2025-10-08 05:39:08 +02:00 • committed by crueter
parent db65f10768
commit 8078990b9b
No known key found for this signature in database
GPG key ID: 425ACD2D4830EBC6
6 changed files with 76 additions and 33 deletions
Download patch file Download diff file Expand all files Collapse all files

42
src/video_core/buffer_cache/buffer_cache.h
View file

@ -386,11 +386,10 @@ void BufferCache<P>::BindHostComputeBuffers() {
template <class P>
void BufferCache<P>::SetUniformBuffersState(const std::array<u32, NUM_STAGES>& mask,
const UniformBufferSizes* sizes) {
if constexpr (HAS_PERSISTENT_UNIFORM_BUFFER_BINDINGS) {
if (channel_state->enabled_uniform_buffer_masks != mask) {
if constexpr (IS_OPENGL) {
channel_state->fast_bound_uniform_buffers.fill(0);
}
const bool mask_changed = channel_state->enabled_uniform_buffer_masks != mask;
if (mask_changed) {
channel_state->fast_bound_uniform_buffers.fill(0);
if constexpr (HAS_PERSISTENT_UNIFORM_BUFFER_BINDINGS) {
channel_state->dirty_uniform_buffers.fill(~u32{0});
channel_state->uniform_buffer_binding_sizes.fill({});
}
@ -806,7 +805,7 @@ void BufferCache<P>::BindHostGraphicsUniformBuffer(size_t stage, u32 index, u32
channel_state->uniform_buffer_binding_sizes[stage][binding_index] != size;
if (should_fast_bind) {
// We only have to bind when the currently bound buffer is not the fast version
channel_state->fast_bound_uniform_buffers[stage] |= 1U << binding_index;
channel_state->fast_bound_uniform_buffers[stage] |= 1u << binding_index;
channel_state->uniform_buffer_binding_sizes[stage][binding_index] = size;
runtime.BindFastUniformBuffer(stage, binding_index, size);
}
@ -815,13 +814,22 @@ void BufferCache<P>::BindHostGraphicsUniformBuffer(size_t stage, u32 index, u32
return;
}
}
if constexpr (IS_OPENGL) {
channel_state->fast_bound_uniform_buffers[stage] |= 1U << binding_index;
channel_state->uniform_buffer_binding_sizes[stage][binding_index] = size;
}
channel_state->fast_bound_uniform_buffers[stage] |= 1u << binding_index;
channel_state->uniform_buffer_binding_sizes[stage][binding_index] = size;
// Stream buffer path to avoid stalling on non-Nvidia drivers or Vulkan
const std::span<u8> span = runtime.BindMappedUniformBuffer(stage, binding_index, size);
#ifdef YUZU_DEBUG
ASSERT(binding_index < NUM_GRAPHICS_UNIFORM_BUFFERS);
ASSERT(span.size() >= size && "UBO stream span too small");
if (!device_memory.ReadBlockFastChecked(device_addr, span.data(), size)) {
LOG_CRITICAL(Render, "DeviceMemory OOB/unmapped: addr=0x{:x} size={}", device_addr, size);
channel_state->fast_bound_uniform_buffers[stage] &= ~(1u << binding_index);
ASSERT(false);
return;
}
#else
device_memory.ReadBlockUnsafe(device_addr, span.data(), size);
#endif
return;
}
// Classic cached path
@ -830,7 +838,8 @@ void BufferCache<P>::BindHostGraphicsUniformBuffer(size_t stage, u32 index, u32
}
// Skip binding if it's not needed and if the bound buffer is not the fast version
// This exists to avoid instances where the fast buffer is bound and a GPU write happens
needs_bind |= HasFastUniformBufferBound(stage, binding_index);
const bool was_fast_bound = HasFastUniformBufferBound(stage, binding_index);
needs_bind |= was_fast_bound;
if constexpr (HAS_PERSISTENT_UNIFORM_BUFFER_BINDINGS) {
needs_bind |= channel_state->uniform_buffer_binding_sizes[stage][binding_index] != size;
}
@ -839,9 +848,6 @@ void BufferCache<P>::BindHostGraphicsUniformBuffer(size_t stage, u32 index, u32
}
const u32 offset = buffer.Offset(device_addr);
if constexpr (IS_OPENGL) {
// Fast buffer will be unbound
channel_state->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();
channel_state->dirty_uniform_buffers[stage] |= (is_copy_bind ? 1U : 0U) << index;
@ -855,6 +861,7 @@ void BufferCache<P>::BindHostGraphicsUniformBuffer(size_t stage, u32 index, u32
} else {
runtime.BindUniformBuffer(buffer, offset, size);
}
channel_state->fast_bound_uniform_buffers[stage] &= ~(1u << binding_index);
}
template <class P>
@ -1789,12 +1796,7 @@ std::span<u8> BufferCache<P>::ImmediateBuffer(size_t wanted_capacity) {
template <class P>
bool BufferCache<P>::HasFastUniformBufferBound(size_t stage, u32 binding_index) const noexcept {
if constexpr (IS_OPENGL) {
return ((channel_state->fast_bound_uniform_buffers[stage] >> binding_index) & 1) != 0;
} else {
// Only OpenGL has fast uniform buffers
return false;
}
return ((channel_state->fast_bound_uniform_buffers[stage] >> binding_index) & 1u) != 0;
}
template <class P>

5
src/video_core/buffer_cache/buffer_cache_base.h
View file

@ -53,6 +53,7 @@ constexpr u32 NUM_COMPUTE_UNIFORM_BUFFERS = 8;
constexpr u32 NUM_STORAGE_BUFFERS = 16;
constexpr u32 NUM_TEXTURE_BUFFERS = 32;
constexpr u32 NUM_STAGES = 5;
static_assert(NUM_GRAPHICS_UNIFORM_BUFFERS <= 32, "fast bitmask must fit u32");
using UniformBufferSizes = std::array<std::array<u32, NUM_GRAPHICS_UNIFORM_BUFFERS>, NUM_STAGES>;
using ComputeUniformBufferSizes = std::array<u32, NUM_COMPUTE_UNIFORM_BUFFERS>;
@ -137,8 +138,8 @@ public:
u32 written_compute_texture_buffers = 0;
u32 image_compute_texture_buffers = 0;
std::array<u32, 16> uniform_cache_hits{};
std::array<u32, 16> uniform_cache_shots{};
std::array<u32, NUM_GRAPHICS_UNIFORM_BUFFERS> uniform_cache_hits{};
std::array<u32, NUM_GRAPHICS_UNIFORM_BUFFERS> uniform_cache_shots{};
u32 uniform_buffer_skip_cache_size = DEFAULT_SKIP_CACHE_SIZE;