ludic/packages/ludic.render3d/gpu_vk.ludic

# gpu_vk.ludic — the Vulkan side of gpu.ludic: the device every other part of the backend
# draws with, the memory it allocates from and the one-shot command buffers that uploads,
# bakes and read-backs go through.
#
# gpu.ludic keeps the renderer's handles (a texture, a buffer, a framebuffer, a program as an
# int) and branches on the backend; on Vulkan those ints index tables of Vulkan objects kept
# here. Nothing in this file runs unless the Vulkan backend was chosen and came up.

# ---- the device ---------------------------------------------------------------------------
var gvk_ready: bool = false
var gvk_inst: pointer = null
var gvk_pd: pointer = null
var gvk_dev: pointer = null
var gvk_queue: pointer = null
var gvk_family: int = -1
var gvk_mp: bytes = null             # VkPhysicalDeviceMemoryProperties
var gvk_device_name: string = ""
var gvk_why: string = ""             # why the device did not come up, for the fallback notice
var gvk_max_aniso: int = 0x3F800000  # the device's anisotropy limit as float bits; 1.0 when it has none
var gvk_want_surface: bool = false   # set before gvk_init when the renderer will present to a window
var gvk_has_surface: bool = false    # the instance and device can make a surface and a swapchain

function gvk_fail(what: string, r: int) -> bool {
  gvk_why = `{what} failed (VkResult {r})`
  print(`r3d: vulkan: {gvk_why}`)
  return false
}
function gvk_handle(out: bytes) -> long { return Vk.get_i64(out, 0) }

function gvk_ext_in(props: bytes, n: int, want: string) -> bool {
  for i in 0 .. n {
    if string(Vk.at(props, i * VkExtensionProperties_sizeof + VkExtensionProperties_extensionName)) == want { return true }
  }
  return false
}

# Instance, the first discrete GPU (else the first listed), its first graphics queue, and a
# device with the Tier 1 floor switched on. False, with gvk_why set, if any of it is missing:
# the caller stays on OpenGL.
var gvk_has_mdi: bool = false         # multiDrawIndirect + drawIndirectFirstInstance
var gvk_has_dic: bool = false         # drawIndirectCount

function gvk_init() -> bool {
  if gvk_ready { return true }
  if Vk.open() == 0 { gvk_why = "no Vulkan loader";  return false }

  let cnt = bytes(4)
  Vk.put_i32(cnt, 0, 0)
  Vk.enumerate_instance_extension_properties(null, cnt, null)
  let nie = Vk.get_i32(cnt, 0)
  let iexts = bytes(nie * VkExtensionProperties_sizeof + 8)
  Vk.enumerate_instance_extension_properties(null, cnt, iexts)
  # MoltenVK is a portability driver, and is only listed to a program that says it knows
  let portability = gvk_ext_in(iexts, nie, VK_KHR_PORTABILITY_ENUMERATION_EXTENSION_NAME)
  let app = bytes(VkApplicationInfo_sizeof)
  Vk.zero(app, VkApplicationInfo_sizeof)
  Vk.put_i32(app, VkApplicationInfo_sType, VK_STRUCTURE_TYPE_APPLICATION_INFO)
  Vk.put_ptr(app, VkApplicationInfo_pApplicationName, "ludic.render3d")
  Vk.put_i32(app, VkApplicationInfo_apiVersion, (1 << 22) | (3 << 12))
  let iext_names = bytes(32)
  var n_iext = 0
  let ici = bytes(VkInstanceCreateInfo_sizeof)
  Vk.zero(ici, VkInstanceCreateInfo_sizeof)
  Vk.put_i32(ici, VkInstanceCreateInfo_sType, VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO)
  Vk.put_ptr(ici, VkInstanceCreateInfo_pApplicationInfo, app)
  if portability {
    Vk.put_ptr(iext_names, n_iext * 8, VK_KHR_PORTABILITY_ENUMERATION_EXTENSION_NAME);  n_iext += 1
    Vk.put_i32(ici, VkInstanceCreateInfo_flags, VK_INSTANCE_CREATE_ENUMERATE_PORTABILITY_BIT_KHR)
  }
  # a window's surface: only the Win32 one is built, so a window elsewhere stays on OpenGL
  gvk_has_surface = gvk_want_surface and gvk_ext_in(iexts, nie, VK_KHR_SURFACE_EXTENSION_NAME) and gvk_ext_in(iexts, nie, VK_KHR_WIN32_SURFACE_EXTENSION_NAME)
  if gvk_want_surface and not gvk_has_surface { gvk_why = "no Win32 surface in this Vulkan instance";  return false }
  if gvk_has_surface {
    Vk.put_ptr(iext_names, n_iext * 8, VK_KHR_SURFACE_EXTENSION_NAME);  n_iext += 1
    Vk.put_ptr(iext_names, n_iext * 8, VK_KHR_WIN32_SURFACE_EXTENSION_NAME);  n_iext += 1
  }
  if n_iext > 0 {
    Vk.put_i32(ici, VkInstanceCreateInfo_enabledExtensionCount, n_iext)
    Vk.put_ptr(ici, VkInstanceCreateInfo_ppEnabledExtensionNames, iext_names)
  }
  let out = bytes(8)
  var r = Vk.create_instance(ici, null, out)
  if r != VK_SUCCESS { return gvk_fail("vkCreateInstance", r) }
  gvk_inst = Vk.get_ptr(out, 0)

  Vk.put_i32(cnt, 0, 0)
  Vk.enumerate_physical_devices(gvk_inst, cnt, null)
  let nd = Vk.get_i32(cnt, 0)
  if nd == 0 { gvk_why = "no Vulkan device";  return false }
  let devs = bytes(nd * 8 + 8)
  Vk.enumerate_physical_devices(gvk_inst, cnt, devs)
  let props = bytes(VkPhysicalDeviceProperties_sizeof)
  var pick = 0
  for d in 0 .. nd {
    Vk.get_physical_device_properties(Vk.get_ptr(devs, d * 8), props)
    if Vk.get_i32(props, VkPhysicalDeviceProperties_deviceType) == VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU { pick = d;  break }
  }
  gvk_pd = Vk.get_ptr(devs, pick * 8)
  Vk.get_physical_device_properties(gvk_pd, props)
  gvk_device_name = string(Vk.at(props, VkPhysicalDeviceProperties_deviceName))

  Vk.put_i32(cnt, 0, 0)
  Vk.get_physical_device_queue_family_properties(gvk_pd, cnt, null)
  let nq = Vk.get_i32(cnt, 0)
  let qprops = bytes(nq * VkQueueFamilyProperties_sizeof + 8)
  Vk.get_physical_device_queue_family_properties(gvk_pd, cnt, qprops)
  for q in 0 .. nq {
    let flags = Vk.get_i32(qprops, q * VkQueueFamilyProperties_sizeof + VkQueueFamilyProperties_queueFlags)
    if gvk_family < 0 and (flags & VK_QUEUE_GRAPHICS_BIT) != 0 { gvk_family = q }
  }
  if gvk_family < 0 { gvk_why = `{gvk_device_name} has no graphics queue`;  return false }

  # what the device offers, then the same structs handed back asking for the floor
  let f13 = bytes(VkPhysicalDeviceVulkan13Features_sizeof)
  Vk.zero(f13, VkPhysicalDeviceVulkan13Features_sizeof)
  Vk.put_i32(f13, VkPhysicalDeviceVulkan13Features_sType, VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_3_FEATURES)
  let f12 = bytes(VkPhysicalDeviceVulkan12Features_sizeof)
  Vk.zero(f12, VkPhysicalDeviceVulkan12Features_sizeof)
  Vk.put_i32(f12, VkPhysicalDeviceVulkan12Features_sType, VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES)
  Vk.put_ptr(f12, VkPhysicalDeviceVulkan12Features_pNext, f13)
  let f2 = bytes(VkPhysicalDeviceFeatures2_sizeof)
  Vk.zero(f2, VkPhysicalDeviceFeatures2_sizeof)
  Vk.put_i32(f2, VkPhysicalDeviceFeatures2_sType, VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2)
  Vk.put_ptr(f2, VkPhysicalDeviceFeatures2_pNext, f12)
  Vk.get_physical_device_features2(gvk_pd, f2)
  var missing = ""
  if Vk.get_i32(f13, VkPhysicalDeviceVulkan13Features_dynamicRendering) != 1 { missing = missing + " dynamicRendering" }
  if Vk.get_i32(f13, VkPhysicalDeviceVulkan13Features_synchronization2) != 1 { missing = missing + " synchronization2" }
  if Vk.get_i32(f12, VkPhysicalDeviceVulkan12Features_descriptorIndexing) != 1 { missing = missing + " descriptorIndexing" }
  if Vk.get_i32(f12, VkPhysicalDeviceVulkan12Features_timelineSemaphore) != 1 { missing = missing + " timelineSemaphore" }
  if len(missing) > 0 { gvk_why = `{gvk_device_name} lacks{missing}`;  return false }
  # anisotropic filtering is a 1.0 feature every desktop GPU has; ask for it when it is there
  let aniso = Vk.get_i32(f2, VkPhysicalDeviceFeatures2_features + VkPhysicalDeviceFeatures_samplerAnisotropy) == 1
  if aniso { gvk_max_aniso = Vk.get_i32(props, VkPhysicalDeviceProperties_limits + VkPhysicalDeviceLimits_maxSamplerAnisotropy) }
  let want13 = bytes(VkPhysicalDeviceVulkan13Features_sizeof)
  Vk.zero(want13, VkPhysicalDeviceVulkan13Features_sizeof)
  Vk.put_i32(want13, VkPhysicalDeviceVulkan13Features_sType, VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_3_FEATURES)
  Vk.put_i32(want13, VkPhysicalDeviceVulkan13Features_dynamicRendering, 1)
  Vk.put_i32(want13, VkPhysicalDeviceVulkan13Features_synchronization2, 1)
  let want12 = bytes(VkPhysicalDeviceVulkan12Features_sizeof)
  Vk.zero(want12, VkPhysicalDeviceVulkan12Features_sizeof)
  Vk.put_i32(want12, VkPhysicalDeviceVulkan12Features_sType, VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES)
  Vk.put_ptr(want12, VkPhysicalDeviceVulkan12Features_pNext, want13)
  Vk.put_i32(want12, VkPhysicalDeviceVulkan12Features_descriptorIndexing, 1)
  Vk.put_i32(want12, VkPhysicalDeviceVulkan12Features_timelineSemaphore, 1)
  let want2 = bytes(VkPhysicalDeviceFeatures2_sizeof)
  Vk.zero(want2, VkPhysicalDeviceFeatures2_sizeof)
  Vk.put_i32(want2, VkPhysicalDeviceFeatures2_sType, VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2)
  Vk.put_ptr(want2, VkPhysicalDeviceFeatures2_pNext, want12)
  if aniso { Vk.put_i32(want2, VkPhysicalDeviceFeatures2_features + VkPhysicalDeviceFeatures_samplerAnisotropy, 1) }
  # GPU-driven drawing: one indirect buffer holds a layer's draws, and a compute pass may write
  # how many of them there are. Asked for where the device has them; the renderer checks
  # gvk_has_mdi / gvk_has_dic before it takes that path.
  gvk_has_mdi = Vk.get_i32(f2, VkPhysicalDeviceFeatures2_features + VkPhysicalDeviceFeatures_multiDrawIndirect) == 1 and Vk.get_i32(f2, VkPhysicalDeviceFeatures2_features + VkPhysicalDeviceFeatures_drawIndirectFirstInstance) == 1
  if gvk_has_mdi {
    Vk.put_i32(want2, VkPhysicalDeviceFeatures2_features + VkPhysicalDeviceFeatures_multiDrawIndirect, 1)
    Vk.put_i32(want2, VkPhysicalDeviceFeatures2_features + VkPhysicalDeviceFeatures_drawIndirectFirstInstance, 1)
  }
  gvk_has_dic = Vk.get_i32(f12, VkPhysicalDeviceVulkan12Features_drawIndirectCount) == 1
  if gvk_has_dic { Vk.put_i32(want12, VkPhysicalDeviceVulkan12Features_drawIndirectCount, 1) }

  Vk.put_i32(cnt, 0, 0)
  Vk.enumerate_device_extension_properties(gvk_pd, null, cnt, null)
  let nde = Vk.get_i32(cnt, 0)
  let dexts = bytes(nde * VkExtensionProperties_sizeof + 8)
  Vk.enumerate_device_extension_properties(gvk_pd, null, cnt, dexts)
  let prio = bytes(4)
  Vk.put_i32(prio, 0, 0x3F800000)
  let qci = bytes(VkDeviceQueueCreateInfo_sizeof)
  Vk.zero(qci, VkDeviceQueueCreateInfo_sizeof)
  Vk.put_i32(qci, VkDeviceQueueCreateInfo_sType, VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO)
  Vk.put_i32(qci, VkDeviceQueueCreateInfo_queueFamilyIndex, gvk_family)
  Vk.put_i32(qci, VkDeviceQueueCreateInfo_queueCount, 1)
  Vk.put_ptr(qci, VkDeviceQueueCreateInfo_pQueuePriorities, prio)
  let dext_names = bytes(32)
  var n_dext = 0
  let dci = bytes(VkDeviceCreateInfo_sizeof)
  Vk.zero(dci, VkDeviceCreateInfo_sizeof)
  Vk.put_i32(dci, VkDeviceCreateInfo_sType, VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO)
  Vk.put_ptr(dci, VkDeviceCreateInfo_pNext, want2)
  Vk.put_i32(dci, VkDeviceCreateInfo_queueCreateInfoCount, 1)
  Vk.put_ptr(dci, VkDeviceCreateInfo_pQueueCreateInfos, qci)
  if gvk_ext_in(dexts, nde, "VK_KHR_portability_subset") { Vk.put_ptr(dext_names, n_dext * 8, "VK_KHR_portability_subset");  n_dext += 1 }
  if gvk_has_surface {
    if not gvk_ext_in(dexts, nde, VK_KHR_SWAPCHAIN_EXTENSION_NAME) { gvk_why = `{gvk_device_name} has no swapchain`;  return false }
    Vk.put_ptr(dext_names, n_dext * 8, VK_KHR_SWAPCHAIN_EXTENSION_NAME);  n_dext += 1
  }
  if n_dext > 0 {
    Vk.put_i32(dci, VkDeviceCreateInfo_enabledExtensionCount, n_dext)
    Vk.put_ptr(dci, VkDeviceCreateInfo_ppEnabledExtensionNames, dext_names)
  }
  r = Vk.create_device(gvk_pd, dci, null, out)
  if r != VK_SUCCESS { return gvk_fail("vkCreateDevice", r) }
  gvk_dev = Vk.get_ptr(out, 0)
  Vk.get_device_queue(gvk_dev, gvk_family, 0, out)
  gvk_queue = Vk.get_ptr(out, 0)
  gvk_mp = bytes(VkPhysicalDeviceMemoryProperties_sizeof)
  Vk.get_physical_device_memory_properties(gvk_pd, gvk_mp)

  if not gvk_cmd_init() { return false }
  gvk_ready = true
  print(`r3d: vulkan on {gvk_device_name}, anisotropy up to {gvk_aniso_x(gvk_max_aniso)}x, multi-draw indirect {gvk_has_mdi}, indirect count {gvk_has_dic}`)
  return true
}

# a positive float, as its bits, to a whole number (16.0 -> 16) for a message
function gvk_aniso_x(bits: int) -> int {
  let e = ((bits >> 23) & 255) - 127
  if e < 0 { return 0 }
  return ((bits & 0x7FFFFF) | 0x800000) >> (23 - e)
}

# ---- memory -------------------------------------------------------------------------------
# The first memory type a resource allows with every property wanted; -1 if there is none.
function gvk_mem_type(allowed: int, want: int) -> int {
  for t in 0 .. Vk.get_i32(gvk_mp, VkPhysicalDeviceMemoryProperties_memoryTypeCount) {
    let pf = Vk.get_i32(gvk_mp, VkPhysicalDeviceMemoryProperties_memoryTypes + t * VkMemoryType_sizeof + VkMemoryType_propertyFlags)
    if ((allowed >> t) & 1) == 1 and (pf & want) == want { return t }
  }
  return -1
}
# Memory for one resource, from its requirements (a VkMemoryRequirements). One allocation per
# resource while the backend comes up; the block allocator with sub-allocation replaces it
# before the forest and the streams are on this backend, which allocate thousands.
var gvk_n_allocs: int = 0
function gvk_alloc(req: bytes, want: int) -> long {
  let allowed = Vk.get_i32(req, VkMemoryRequirements_memoryTypeBits)
  var t = gvk_mem_type(allowed, want)
  # device-local is a preference; host-visible is a need
  if t < 0 and want == VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT { t = gvk_mem_type(allowed, 0) }
  let zero: long = 0
  if t < 0 { print(`r3d: vulkan: no memory type for properties {want}`);  return zero }
  let mai = bytes(VkMemoryAllocateInfo_sizeof)
  Vk.zero(mai, VkMemoryAllocateInfo_sizeof)
  Vk.put_i32(mai, VkMemoryAllocateInfo_sType, VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO)
  Vk.put_i64(mai, VkMemoryAllocateInfo_allocationSize, Vk.get_i64(req, VkMemoryRequirements_size))
  Vk.put_i32(mai, VkMemoryAllocateInfo_memoryTypeIndex, t)
  let out = bytes(8)
  let r = Vk.allocate_memory(gvk_dev, mai, null, out)
  if r != VK_SUCCESS { print(`r3d: vulkan: vkAllocateMemory failed (VkResult {r})`);  return zero }
  gvk_n_allocs += 1
  return gvk_handle(out)
}

# ---- one-shot commands --------------------------------------------------------------------
# Uploads, bakes and read-backs record into a command buffer, submit it and wait. The frame
# itself does not go through here.
var gvk_pool: long = 0
var gvk_fence: bytes = null
function gvk_cmd_init() -> bool {
  let cpi = bytes(VkCommandPoolCreateInfo_sizeof)
  Vk.zero(cpi, VkCommandPoolCreateInfo_sizeof)
  Vk.put_i32(cpi, VkCommandPoolCreateInfo_sType, VK_STRUCTURE_TYPE_COMMAND_POOL_CREATE_INFO)
  Vk.put_i32(cpi, VkCommandPoolCreateInfo_flags, VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT)
  Vk.put_i32(cpi, VkCommandPoolCreateInfo_queueFamilyIndex, gvk_family)
  let out = bytes(8)
  var r = Vk.create_command_pool(gvk_dev, cpi, null, out)
  if r != VK_SUCCESS { return gvk_fail("vkCreateCommandPool", r) }
  gvk_pool = gvk_handle(out)
  let fci = bytes(VkFenceCreateInfo_sizeof)
  Vk.zero(fci, VkFenceCreateInfo_sizeof)
  Vk.put_i32(fci, VkFenceCreateInfo_sType, VK_STRUCTURE_TYPE_FENCE_CREATE_INFO)
  gvk_fence = bytes(8)
  r = Vk.create_fence(gvk_dev, fci, null, gvk_fence)
  if r != VK_SUCCESS { return gvk_fail("vkCreateFence", r) }
  return true
}
# a command buffer, begun
function gvk_once_begin() -> pointer {
  let cbai = bytes(VkCommandBufferAllocateInfo_sizeof)
  Vk.zero(cbai, VkCommandBufferAllocateInfo_sizeof)
  Vk.put_i32(cbai, VkCommandBufferAllocateInfo_sType, VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO)
  Vk.put_i64(cbai, VkCommandBufferAllocateInfo_commandPool, gvk_pool)
  Vk.put_i32(cbai, VkCommandBufferAllocateInfo_level, VK_COMMAND_BUFFER_LEVEL_PRIMARY)
  Vk.put_i32(cbai, VkCommandBufferAllocateInfo_commandBufferCount, 1)
  let cbs = bytes(8)
  if Vk.allocate_command_buffers(gvk_dev, cbai, cbs) != VK_SUCCESS { return null }
  let cb = Vk.get_ptr(cbs, 0)
  let cbbi = bytes(VkCommandBufferBeginInfo_sizeof)
  Vk.zero(cbbi, VkCommandBufferBeginInfo_sizeof)
  Vk.put_i32(cbbi, VkCommandBufferBeginInfo_sType, VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO)
  Vk.put_i32(cbbi, VkCommandBufferBeginInfo_flags, VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT)
  Vk.begin_command_buffer(cb, cbbi)
  return cb
}
# end it, submit it, wait for it, free it
function gvk_once_end(cb: pointer) -> bool {
  var r = Vk.end_command_buffer(cb)
  if r != VK_SUCCESS { return gvk_fail("vkEndCommandBuffer", r) }
  let cbs = bytes(8)
  Vk.put_ptr(cbs, 0, cb)
  Vk.reset_fences(gvk_dev, 1, gvk_fence)
  let si = bytes(VkSubmitInfo_sizeof)
  Vk.zero(si, VkSubmitInfo_sizeof)
  Vk.put_i32(si, VkSubmitInfo_sType, VK_STRUCTURE_TYPE_SUBMIT_INFO)
  Vk.put_i32(si, VkSubmitInfo_commandBufferCount, 1)
  Vk.put_ptr(si, VkSubmitInfo_pCommandBuffers, cbs)
  r = Vk.queue_submit(gvk_queue, 1, si, Vk.get_i64(gvk_fence, 0))
  if r != VK_SUCCESS { return gvk_fail("vkQueueSubmit", r) }
  let forever: long = -1
  r = Vk.wait_for_fences(gvk_dev, 1, gvk_fence, 1, forever)
  Vk.free_command_buffers(gvk_dev, gvk_pool, 1, cbs)
  if r != VK_SUCCESS { return gvk_fail("vkWaitForFences", r) }
  return true
}

# ---- teardown -----------------------------------------------------------------------------
function gvk_shutdown() -> void {
  if not gvk_ready { return }
  Vk.device_wait_idle(gvk_dev)
  Vk.destroy_fence(gvk_dev, Vk.get_i64(gvk_fence, 0), null)
  Vk.destroy_command_pool(gvk_dev, gvk_pool, null)
  Vk.destroy_device(gvk_dev, null)
  Vk.destroy_instance(gvk_inst, null)
  gvk_ready = false
}