RPCS3/rpcs3
1
0
Fork 0
mirror of https://github.com/RPCS3/rpcs3.git synced 2025-12-06 07:12:28 +01:00
Code Issues Actions 1 Packages Projects Releases Wiki Activity

vk: Use BAR memory pool for low-latency resource uploads

Browse source
This commit is contained in:
kd-11 2025-08-24 20:58:55 +03:00 committed by kd-11
parent 21809731e2
commit 20a7297ac4
4 changed files with 33 additions and 20 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

26
rpcs3/Emu/RSX/VK/VKGSRender.cpp
View file

@ -509,17 +509,17 @@ VKGSRender::VKGSRender(utils::serial* ar) noexcept : GSRender(ar)
semaphore_info.sType = VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO;
// VRAM allocation
m_attrib_ring_info.create(VK_BUFFER_USAGE_UNIFORM_TEXEL_BUFFER_BIT, VK_ATTRIB_RING_BUFFER_SIZE_M * 0x100000, "attrib buffer", 0x400000, VK_TRUE);
m_fragment_env_ring_info.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, VK_UBO_RING_BUFFER_SIZE_M * 0x100000, "fragment env buffer");
m_vertex_env_ring_info.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, VK_UBO_RING_BUFFER_SIZE_M * 0x100000, "vertex env buffer");
m_fragment_texture_params_ring_info.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, VK_UBO_RING_BUFFER_SIZE_M * 0x100000, "fragment texture params buffer");
m_vertex_layout_ring_info.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, VK_UBO_RING_BUFFER_SIZE_M * 0x100000, "vertex layout buffer", 0x10000, VK_TRUE);
m_fragment_constants_ring_info.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, VK_UBO_RING_BUFFER_SIZE_M * 0x100000, "fragment constants buffer");
m_transform_constants_ring_info.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, VK_TRANSFORM_CONSTANTS_BUFFER_SIZE_M * 0x100000, "transform constants buffer");
m_index_buffer_ring_info.create(VK_BUFFER_USAGE_INDEX_BUFFER_BIT, VK_INDEX_RING_BUFFER_SIZE_M * 0x100000, "index buffer");
m_texture_upload_buffer_ring_info.create(VK_BUFFER_USAGE_TRANSFER_SRC_BIT, VK_TEXTURE_UPLOAD_RING_BUFFER_SIZE_M * 0x100000, "texture upload buffer", 32 * 0x100000);
m_raster_env_ring_info.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, VK_UBO_RING_BUFFER_SIZE_M * 0x100000, "raster env buffer");
m_instancing_buffer_ring_info.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, VK_TRANSFORM_CONSTANTS_BUFFER_SIZE_M * 0x100000, "instancing data buffer");
m_attrib_ring_info.create(VK_BUFFER_USAGE_UNIFORM_TEXEL_BUFFER_BIT, VK_ATTRIB_RING_BUFFER_SIZE_M * 0x100000, vk::heap_pool_default, "attrib buffer", 0x400000, VK_TRUE);
m_fragment_env_ring_info.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, VK_UBO_RING_BUFFER_SIZE_M * 0x100000, vk::heap_pool_low_latency, "fragment env buffer");
m_vertex_env_ring_info.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, VK_UBO_RING_BUFFER_SIZE_M * 0x100000, vk::heap_pool_low_latency, "vertex env buffer");
m_fragment_texture_params_ring_info.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, VK_UBO_RING_BUFFER_SIZE_M * 0x100000, vk::heap_pool_low_latency, "fragment texture params buffer");
m_vertex_layout_ring_info.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, VK_UBO_RING_BUFFER_SIZE_M * 0x100000, vk::heap_pool_low_latency, "vertex layout buffer", 0x10000, VK_TRUE);
m_fragment_constants_ring_info.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, VK_UBO_RING_BUFFER_SIZE_M * 0x100000, vk::heap_pool_low_latency, "fragment constants buffer");
m_transform_constants_ring_info.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, VK_TRANSFORM_CONSTANTS_BUFFER_SIZE_M * 0x100000, vk::heap_pool_low_latency, "transform constants buffer");
m_index_buffer_ring_info.create(VK_BUFFER_USAGE_INDEX_BUFFER_BIT, VK_INDEX_RING_BUFFER_SIZE_M * 0x100000, vk::heap_pool_default, "index buffer");
m_texture_upload_buffer_ring_info.create(VK_BUFFER_USAGE_TRANSFER_SRC_BIT, VK_TEXTURE_UPLOAD_RING_BUFFER_SIZE_M * 0x100000, vk::heap_pool_default, "texture upload buffer", 32 * 0x100000);
m_raster_env_ring_info.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, VK_UBO_RING_BUFFER_SIZE_M * 0x100000, vk::heap_pool_low_latency, "raster env buffer");
m_instancing_buffer_ring_info.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, VK_TRANSFORM_CONSTANTS_BUFFER_SIZE_M * 0x100000, vk::heap_pool_low_latency, "instancing data buffer");
vk::data_heap_manager::register_ring_buffers
({
@ -540,8 +540,8 @@ VKGSRender::VKGSRender(utils::serial* ar) noexcept : GSRender(ar)
if (shadermode == shader_mode::async_with_interpreter || shadermode == shader_mode::interpreter_only)
{
m_vertex_instructions_buffer.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, 64 * 0x100000, "vertex instructions buffer", 512 * 16);
m_fragment_instructions_buffer.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, 64 * 0x100000, "fragment instructions buffer", 2048);
m_vertex_instructions_buffer.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, 64 * 0x100000, vk::heap_pool_low_latency, "vertex instructions buffer", 512 * 16);
m_fragment_instructions_buffer.create(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, 64 * 0x100000, vk::heap_pool_low_latency, "fragment instructions buffer", 2048);
vk::data_heap_manager::register_ring_buffers
({

4
rpcs3/Emu/RSX/VK/VKOverlays.cpp
View file

@ -42,12 +42,12 @@ namespace vk
{
if (!m_vao.heap)
{
m_vao.create(VK_BUFFER_USAGE_VERTEX_BUFFER_BIT, 1 * 0x100000, "overlays VAO", 128);
m_vao.create(VK_BUFFER_USAGE_VERTEX_BUFFER_BIT, 1 * 0x100000, vk::heap_pool_default, "overlays VAO", 128);
}
if (!m_ubo.heap && m_num_uniform_buffers > 0)
{
m_ubo.create(VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT, 8 * 0x100000, "overlays UBO", 128);
m_ubo.create(VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT, 8 * 0x100000, vk::heap_pool_default, "overlays UBO", 128);
}
}

14
rpcs3/Emu/RSX/VK/vkutils/data_heap.cpp
View file

@ -13,14 +13,20 @@ namespace vk
{
data_heap g_upload_heap;
void data_heap::create(VkBufferUsageFlags usage, usz size, const char* name, usz guard, VkBool32 notify)
void data_heap::create(VkBufferUsageFlags usage, usz size, rsx::flags32_t flags, const char* name, usz guard, VkBool32 notify)
{
::data_heap::init(size, name, guard);
const auto& memory_map = g_render_device->get_memory_mapping();
if (flags & heap_pool_low_latency)
{
// Prefer uploading to BAR if low latency is desired.
m_prefer_writethrough = memory_map.device_bar_total_bytes > (2048ull * 0x100000);
}
VkFlags memory_flags = VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT | VK_MEMORY_PROPERTY_HOST_COHERENT_BIT;
auto memory_index = memory_map.host_visible_coherent;
auto memory_index = m_prefer_writethrough ? memory_map.device_bar : memory_map.host_visible_coherent;
if (!(get_heap_compatible_buffer_types() & usage))
{
@ -75,7 +81,7 @@ namespace vk
const auto& memory_map = g_render_device->get_memory_mapping();
VkFlags memory_flags = VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT | VK_MEMORY_PROPERTY_HOST_COHERENT_BIT;
auto memory_index = memory_map.host_visible_coherent;
auto memory_index = m_prefer_writethrough ? memory_map.device_bar : memory_map.host_visible_coherent;
// Update heap information and reset the allocator
::data_heap::init(aligned_new_size, m_name, m_min_guard_size);
@ -163,7 +169,7 @@ namespace vk
{
if (!g_upload_heap.heap)
{
g_upload_heap.create(VK_BUFFER_USAGE_TRANSFER_SRC_BIT, 64 * 0x100000, "auxilliary upload heap", 0x100000);
g_upload_heap.create(VK_BUFFER_USAGE_TRANSFER_SRC_BIT, 64 * 0x100000, vk::heap_pool_default, "auxilliary upload heap", 0x100000);
}
return &g_upload_heap;

9
rpcs3/Emu/RSX/VK/vkutils/data_heap.h
View file

@ -11,6 +11,12 @@
namespace vk
{
enum data_heap_pool_flags
{
heap_pool_default = 0,
heap_pool_low_latency = 1,
};
class data_heap : public ::data_heap
{
private:
@ -19,6 +25,7 @@ namespace vk
void* _ptr = nullptr;
bool notify_on_grow = false;
bool m_prefer_writethrough = false;
std::unique_ptr<buffer> shadow;
std::vector<VkBufferCopy> dirty_ranges;
@ -33,7 +40,7 @@ namespace vk
// Avoid mapping/unmapping to keep these drivers from stalling
// NOTE2: HOST_CACHED flag does not keep the mapped ptr around in the driver either
void create(VkBufferUsageFlags usage, usz size, const char* name, usz guard = 0x10000, VkBool32 notify = VK_FALSE);
void create(VkBufferUsageFlags usage, usz size, rsx::flags32_t flags, const char* name, usz guard = 0x10000, VkBool32 notify = VK_FALSE);
void destroy();
void* map(usz offset, usz size);