# ============================================================================
# gpu.ludic — the seam between the renderer and a graphics API.
#
# render3d was written straight against OpenGL, so there was nowhere to put a second
# API. This file is where that stops: the renderer asks for what it wants here, and
# the backend decides how to say it. OpenGL is the default and the fallback; Vulkan
# is the second backend (docs: maroon-lake docs/plan/37-vulkan.md).
#
# The port is incremental. What has moved behind the seam so far:
#   - the choice of backend (R3D_GFX=gl|vk, or gpu_request before r3d_init)
#   - the fixed-function render state: depth test/func/write, blending, face
#     culling, colour writes, alpha-to-coverage, depth bias, scissor
#   - uniforms: looked up by program and name (gpu_uniform), set by the u_* setters
#   - vertex data: meshes, their attribute layouts, instance and stream buffers, draws
#   - textures: creation, upload, sampler state, mipmaps, units, read-back, freeing
#   - render targets and passes: framebuffers, attachments, blits, viewports, clears, the screen
#
# Render state is cached. A pipeline API bakes this state into an object picked by
# key; OpenGL gets the same effect by only telling the driver what changed. The
# cache is only right if NOTHING else in the renderer touches that state, so no
# gl_enable / gl_disable / gl_depth_func / gl_depth_mask / gl_blend_func /
# gl_cull_face / gl_color_mask / gl_scissor / gl_polygon_offset may appear outside
# this file, and neither may any gl_uniform* call.
# ============================================================================

const GPU_GL: int = 1
const GPU_VK: int = 2

var gpu_kind: int = 0                 # the backend running; 0 until gpu_select
var gpu_wanted: int = 0               # what was asked for (setting or R3D_GFX)
var gpu_fallback_reason: string = null # why the wanted backend is not the one running

# Ask for a backend before r3d_init ("gl", "opengl", "vk", "vulkan"). The environment
# (R3D_GFX) wins over it, so a test or a take can force one whatever the setting says.
function gpu_request(name: string) -> void { gpu_wanted = gpu_kind_of(name) }

function gpu_kind_of(name: string) -> int {
  if name == "vk" or name == "vulkan" { return GPU_VK }
  if name == "gl" or name == "opengl" { return GPU_GL }
  return GPU_GL
}
function gpu_name(kind: int) -> string {
  if kind == GPU_VK { return "vulkan" }
  return "opengl"
}

# Settle the backend. Anything that cannot run lands on OpenGL with a reason a
# caller can show the player (gpu_fallback_reason).
function gpu_select() -> int {
  if gpu_kind != 0 { return gpu_kind }
  if Os.has_env("R3D_GFX") { gpu_wanted = gpu_kind_of(Os.env("R3D_GFX")) }
  if gpu_wanted == 0 { gpu_wanted = GPU_GL }
  gpu_kind = GPU_GL
  if gpu_wanted == GPU_VK {
    # a window needs a surface, and only the Win32 one is built
    gvk_want_surface = is_windowed()
    if is_windowed() and Os.platform() != "windows" { gpu_fallback_reason = "the Vulkan window is built for Windows only" }
    else if not gvk_init() { gpu_fallback_reason = gvk_why }
    else if not gvk_manifest() { gpu_fallback_reason = "the renderer's SPIR-V manifest is missing" }
    else { gpu_kind = GPU_VK }
    if gpu_kind == GPU_GL { print(`r3d: vulkan requested: {gpu_fallback_reason}; using opengl`) }
  }
  return gpu_kind
}
function gpu_backend() -> string { return gpu_name(gpu_kind) }
function gpu_is_gl() -> bool { return gpu_kind != GPU_VK }

# ---- render state ----------------------------------------------------------------
# -1 = not known yet: the first set always reaches the driver, so the cache never
# assumes a default the context might not have.
var gpu_s_depth_test: int = -1
var gpu_s_depth_func: int = -1
var gpu_s_depth_write: int = -1
var gpu_s_blend: int = -1
var gpu_s_blend_src: int = -1
var gpu_s_blend_dst: int = -1
var gpu_s_cull: int = -1
var gpu_s_cull_face: int = -1
var gpu_s_color_write: int = -1
var gpu_s_a2c: int = -1

function gpu_b(on: bool) -> int { if on { return 1 };  return 0 }

# Forget the cache: after anything outside the renderer may have changed GL state
# (a context rebuilt, a foreign library drawing into the frame).
function gpu_state_forget() -> void {
  gpu_s_depth_test = -1;  gpu_s_depth_func = -1;  gpu_s_depth_write = -1
  gpu_s_blend = -1;  gpu_s_blend_src = -1;  gpu_s_blend_dst = -1
  gpu_s_cull = -1;  gpu_s_cull_face = -1;  gpu_s_color_write = -1;  gpu_s_a2c = -1
}

function gpu_gl_cap(cap: int, on: int) -> void { if gpu_kind == GPU_VK { return };  if on == 1 { gl_enable(cap) } else { gl_disable(cap) } }

function gpu_depth_test(on: bool) -> void {
  let v = gpu_b(on)
  if v == gpu_s_depth_test { return }
  gpu_s_depth_test = v
  gpu_gl_cap(GL_DEPTH_TEST, v)
}
# GL_LESS, GL_LEQUAL, GL_EQUAL, GL_ALWAYS, ... (the comparison names are the same in every API)
function gpu_depth_func(f: int) -> void {
  if f == gpu_s_depth_func { return }
  gpu_s_depth_func = f
  if gpu_kind != GPU_VK { gl_depth_func(f) }
}
function gpu_depth_write(on: bool) -> void {
  let v = gpu_b(on)
  if v == gpu_s_depth_write { return }
  gpu_s_depth_write = v
  if gpu_kind != GPU_VK { gl_depth_mask(v) }
}
function gpu_blend(on: bool) -> void {
  let v = gpu_b(on)
  if v == gpu_s_blend { return }
  gpu_s_blend = v
  gpu_gl_cap(GL_BLEND, v)
}
function gpu_blend_func(src: int, dst: int) -> void {
  if src == gpu_s_blend_src and dst == gpu_s_blend_dst { return }
  gpu_s_blend_src = src;  gpu_s_blend_dst = dst
  if gpu_kind != GPU_VK { gl_blend_func(src, dst) }
}
function gpu_cull(on: bool) -> void {
  let v = gpu_b(on)
  if v == gpu_s_cull { return }
  gpu_s_cull = v
  gpu_gl_cap(GL_CULL_FACE, v)
}
# GL_BACK or GL_FRONT
function gpu_cull_face(face: int) -> void {
  if face == gpu_s_cull_face { return }
  gpu_s_cull_face = face
  if gpu_kind != GPU_VK { gl_cull_face(face) }
}
# all four channels together: nothing in the renderer writes a partial mask
function gpu_color_write(on: bool) -> void {
  let v = gpu_b(on)
  if v == gpu_s_color_write { return }
  gpu_s_color_write = v
  if gpu_kind != GPU_VK { gl_color_mask(v, v, v, v) }
}
function gpu_alpha_to_coverage(on: bool) -> void {
  let v = gpu_b(on)
  if v == gpu_s_a2c { return }
  gpu_s_a2c = v
  gpu_gl_cap(GL_SAMPLE_ALPHA_TO_COVERAGE, v)
}

# Depth bias for the shadow casters; (0, 0) turns it off. factor/units are fixed, as
# gl_polygon_offset takes them. A pipeline API bakes the bias into the pipeline.
var gpu_s_bias: int = -1
var gpu_s_bias_f: int = 0         # float bits, for a backend that bakes the bias into a pipeline
var gpu_s_bias_u: int = 0
function gpu_depth_bias(factor: fixed, units: fixed) -> void {
  var v = 1
  if factor == 0.0 and units == 0.0 { v = 0 }
  if v != gpu_s_bias { gpu_s_bias = v;  gpu_gl_cap(GL_POLYGON_OFFSET_FILL, v) }
  gpu_s_bias_f = fx_to_f32(factor);  gpu_s_bias_u = fx_to_f32(units)
  if v == 1 and gpu_kind != GPU_VK { gl_polygon_offset(factor, units) }
}

# A scissor rectangle in top-down pixels (x, y from the top-left of the drawable), or
# off. Every API but OpenGL counts rows from the top; the GL backend flips it.
var gpu_s_scissor: int = -1
function gpu_scissor(x: int, y_top: int, w: int, h: int) -> void {
  if gpu_kind == GPU_VK { gpu_s_scissor = 1;  gvk_scissor(x, gl_h - y_top - h, w, h);  return }
  if gpu_s_scissor != 1 { gpu_s_scissor = 1;  gl_enable(GL_SCISSOR_TEST) }
  gl_scissor(x, gl_h - y_top - h, w, h)
  gpu_glcheck_after("scissor")
}
function gpu_scissor_off() -> void {
  if gpu_kind == GPU_VK { gpu_s_scissor = 0;  gvk_scissor_off();  return }
  if gpu_s_scissor == 0 { return }
  gpu_s_scissor = 0
  gl_disable(GL_SCISSOR_TEST)
  gpu_glcheck_after("scissor off")
}

# ---- uniforms --------------------------------------------------------------------
# A uniform is found by its program and its name, and set through a handle. On OpenGL
# the handle is the location. On a pipeline API it will name a slot in the program's
# uniform block, so the u_* setters below are the only code that knows which. Arrays
# are looked up by their first element ("u_bones[0]"), as every GL driver accepts.
function gpu_uniform(prog: int, name: string) -> int { if gpu_kind == GPU_VK { return gvk_uniform(prog, name) };  return gl_get_uniform_location(prog, name) }

# ---- programs ----------------------------------------------------------------------
# A program remembers the variant it was built from - vertex file, fragment file and the
# defines on one line, the key the SPIR-V manifest uses - which is how a backend that cannot
# compile shaders at run time finds its pipeline for the same handle.
var gpu_prog_ids: []int = null
var gpu_prog_keys: []string = null
var gpu_prog_cur: int = 0
function gpu_program(vs_src: string, fs_src: string, vs: string, fs: string, defines: string) -> int {
  if gpu_kind == GPU_VK { return gvk_program_new(vs, fs, defines) }
  let p = gl_program(vs_src, fs_src)
  if p == 0 { return 0 }
  if gpu_prog_ids == null { gpu_prog_ids = new []int;  gpu_prog_keys = new []string }
  push(gpu_prog_ids, p)
  push(gpu_prog_keys, `{vs}|{fs}|{Text.replace(defines, "\n", ";")}`)
  return p
}
# the manifest key a program was built from; "" for one this layer did not build
function gpu_program_key(p: int) -> string {
  if gpu_prog_ids == null { return "" }
  for i in 0 .. len(gpu_prog_ids) { if gpu_prog_ids[i] == p { return gpu_prog_keys[i] } }
  return ""
}
function gpu_use_program(p: int) -> void { ds_program_change(gpu_prog_cur, p);  if gpu_kind == GPU_VK { gpu_prog_cur = p;  return };  gl_use_program(p);  gpu_prog_cur = p;  gpu_glcheck_after("use program") }
function gpu_program_free(p: int) -> void {
  if gpu_kind == GPU_VK { return }
  if p == 0 { return }
  gl_delete_program(p)
  if gpu_prog_cur == p { gpu_prog_cur = 0 }
  if gpu_prog_ids != null { for i in 0 .. len(gpu_prog_ids) { if gpu_prog_ids[i] == p { gpu_prog_ids[i] = 0;  gpu_prog_keys[i] = "" } } }
  gpu_glcheck_after("program free")
}

# ---- GPU timers (R3D_PROF) ----------------------------------------------------------
function gpu_query_new(n: int, ids: words) -> void { if gpu_kind == GPU_VK { gvk_query_new(n, ids);  return };  gl_gen_queries(n, ids) }
function gpu_query_begin(id: int) -> void { if gpu_kind == GPU_VK { gvk_query_begin(id);  return };  gl_begin_query(GL_TIME_ELAPSED, id) }
function gpu_query_end() -> void { if gpu_kind == GPU_VK { gvk_query_end();  return };  gl_end_query(GL_TIME_ELAPSED) }
# true once the query has its result; the nanoseconds (low 32 bits) are then in out[0]
function gpu_query_result(id: int, out: words) -> bool {
  if gpu_kind == GPU_VK { return gvk_query_result(id, out) }
  gl_get_query_objectiv(id, GL_QUERY_RESULT_AVAILABLE, out)
  if out[0] == 0 { return false }
  gl_get_query_objectui64v(id, GL_QUERY_RESULT, out)
  return true
}

# ---- the context --------------------------------------------------------------------
function gpu_open(w: int, h: int, title: string) -> bool { if gpu_kind == GPU_VK { return gvk_open(w, h, title) };  return gl_open(w, h, title) }
function gpu_vsync(on: int) -> void { if gpu_kind == GPU_VK { gvk_vsync = on != 0;  if gvk_swap != 0 { gvk_swap_stale = true };  return };  gl_vsync(on) }
function gpu_renderer_name() -> string { if gpu_kind == GPU_VK { return `{gvk_device_name} (Vulkan)` };  return gl_get_string(GL_RENDERER) }
function gpu_resize_check() -> bool { if gpu_kind == GPU_VK { return gvk_resize_check() };  let r = gl_resize_check();  gpu_glcheck_after("the resize check");  return r }
# the finished frame: presented to the window, or (headless) the GPU's work finished
function gpu_present() -> void { if gpu_kind == GPU_VK { gvk_present();  return };  Gl.swap() }
# the frame as it will be presented, to a binary PPM with the top row first; before gpu_present
function gpu_screenshot(path: string) -> bool { if gpu_kind == GPU_VK { return gvk_screenshot(path) };  return Gl.screenshot(path: path) }

var gpu_u_tmp: words = null
function gpu_tmp() -> words { if gpu_u_tmp == null { gpu_u_tmp = words(4) };  return gpu_u_tmp }
# float bits (IEEE singles in an int), like every other number in the renderer
function u_f(loc: int, v: int) -> void { if gpu_kind == GPU_VK { let t = gpu_tmp();  t[0] = v;  gvk_u_set(loc, t, 4, 1);  return };  let t = gpu_tmp();  t[0] = v;  gl_uniform1fv(loc, 1, t) }
function u_f2(loc: int, x: int, y: int) -> void { if gpu_kind == GPU_VK { let t = gpu_tmp();  t[0] = x;  t[1] = y;  gvk_u_set(loc, t, 8, 1);  return };  let t = gpu_tmp();  t[0] = x;  t[1] = y;  gl_uniform2fv(loc, 1, t) }
function u_f3(loc: int, x: int, y: int, z: int) -> void { if gpu_kind == GPU_VK { let t = gpu_tmp();  t[0] = x;  t[1] = y;  t[2] = z;  gvk_u_set(loc, t, 12, 1);  return };  let t = gpu_tmp();  t[0] = x;  t[1] = y;  t[2] = z;  gl_uniform3fv(loc, 1, t) }
function u_f4(loc: int, x: int, y: int, z: int, w: int) -> void { if gpu_kind == GPU_VK { let t = gpu_tmp();  t[0] = x;  t[1] = y;  t[2] = z;  t[3] = w;  gvk_u_set(loc, t, 16, 1);  return };  let t = gpu_tmp();  t[0] = x;  t[1] = y;  t[2] = z;  t[3] = w;  gl_uniform4fv(loc, 1, t) }
function u_v3(loc: int, v: words) -> void { if gpu_kind == GPU_VK { gvk_u_set(loc, v, 12, 1);  return };  gl_uniform3fv(loc, 1, v) }
function u_fv(loc: int, n: int, v: words) -> void { if gpu_kind == GPU_VK { gvk_u_set(loc, v, 4, n);  return };  gl_uniform1fv(loc, n, v) }
# n vec4s from 4n float bits. Not u_fv with 4n: on Vulkan an array element is copied at the size
# given and placed at the array's stride, so a vec4 array fed floats got one float per element -
# which drew the chunked grass with every tile at a nonsense corner and zero blades a cell.
function u_f4v(loc: int, n: int, v: words) -> void { if gpu_kind == GPU_VK { gvk_u_set(loc, v, 16, n);  return };  gl_uniform4fv(loc, n, v) }
function u_mat4(loc: int, m: words) -> void { if gpu_kind == GPU_VK { gvk_u_set(loc, m, 64, 1);  return };  gl_uniform_matrix4fv(loc, 1, 0, m) }
function u_mat4n(loc: int, n: int, m: words) -> void { if gpu_kind == GPU_VK { gvk_u_set(loc, m, 64, n);  return };  gl_uniform_matrix4fv(loc, n, 0, m) }
function u_i(loc: int, v: int) -> void { if gpu_kind == GPU_VK { let t = gpu_tmp();  t[0] = v;  gvk_u_set(loc, t, 4, 1);  return };  gl_uniform1i(loc, v) }

# ---- what this machine can do ----------------------------------------------------
# The advanced graphics features are Windows features: the Vulkan renderer, ray tracing,
# DLSS, Reflex, HDR output, mesh-shader ground cover. gpu_caps_probe() asks Vulkan what
# the GPU offers - on Windows only - and a game's settings screen greys out whatever this
# machine cannot use, with the most specific reason. Detected every start, never saved:
# a settings file carried to another machine must not carry a stale "supported".
const GF_VULKAN: int = 0
const GF_RT_SHADOWS: int = 1
const GF_RT_REFLECTIONS: int = 2
const GF_RT_AO: int = 3
const GF_DLSS: int = 4
const GF_DLSS_RR: int = 5
const GF_DLSS_FG: int = 6
const GF_DLSS5: int = 7
const GF_REFLEX: int = 8
const GF_HDR: int = 9
const GF_MESH_GRASS: int = 10
const GF_COUNT: int = 11

var gpu_cap_probed: bool = false
var gpu_cap_windows: bool = false      # the advanced features exist on this platform at all
var gpu_cap_vulkan: bool = false       # a Vulkan loader, an instance and a device
var gpu_cap_floor: bool = false        # Vulkan 1.3 with everything the renderer's floor needs
var gpu_cap_rt: bool = false           # ray query + acceleration structures
var gpu_cap_mesh: bool = false         # VK_EXT_mesh_shader
var gpu_cap_nvidia: bool = false
var gpu_cap_rtx: int = 0               # the RTX generation (20, 30, 40, 50); 0 = not RTX
var gpu_cap_reflex: bool = false       # VK_NV_low_latency2
var gpu_cap_hdr: bool = false          # the instance offers HDR colour spaces
var gpu_cap_device: string = ""

# 20 for "NVIDIA GeForce RTX 2080", 50 for "RTX 5090"; 30 for a workstation "RTX A4000"
function gpu_rtx_generation(name: string) -> int {
  let at = Text.index_of(name, "RTX ")
  if at < 0 { return 0 }
  let p: pointer = name
  let c = p[at + 4]
  if c >= '0' and c <= '9' { return (c - '0') * 10 }
  return 30
}

function gpu_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
}

# R3D_CAPS=rtx50|rtx40|rtx30|amd|intel|none pretends to be a Windows machine with that GPU,
# so the settings screen can be shot and tested anywhere
function gpu_caps_fake(kind: string) -> void {
  gpu_cap_windows = true
  if kind == "none" { return }
  gpu_cap_vulkan = true;  gpu_cap_floor = true;  gpu_cap_hdr = true
  if kind == "amd" or kind == "intel" { gpu_cap_rt = true;  gpu_cap_mesh = true;  gpu_cap_device = `test {kind} GPU`;  return }
  gpu_cap_nvidia = true;  gpu_cap_rt = true;  gpu_cap_mesh = true;  gpu_cap_reflex = true
  gpu_cap_rtx = gpu_rtx_generation(`RTX {kind[3 .. 5]}`)
  gpu_cap_device = `test NVIDIA GeForce RTX {kind[3 .. 5]}`
}

function gpu_caps_probe() -> void {
  if gpu_cap_probed { return }
  gpu_cap_probed = true
  if Os.has_env("R3D_CAPS") { gpu_caps_fake(Os.env("R3D_CAPS"));  return }
  gpu_cap_windows = Os.platform() == "windows"
  if not gpu_cap_windows { return }
  if Vk.open() == 0 { return }
  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)
  gpu_cap_hdr = gpu_ext_in(iexts, nie, VK_EXT_SWAPCHAIN_COLOR_SPACE_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_i32(app, VkApplicationInfo_apiVersion, (1 << 22) | (3 << 12))
  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)
  # With the Vulkan renderer running, ask its own instance: the renderer's loader is NVIDIA Streamline's
  # interposer, and a second instance made and destroyed under it is what this avoids.
  var inst: pointer = null
  var own = false
  if gpu_kind == GPU_VK and gvk_ready {
    inst = gvk_inst
  } else {
    let out = bytes(8)
    if Vk.create_instance(ici, null, out) != VK_SUCCESS { return }
    inst = Vk.get_ptr(out, 0)
    own = true
  }
  Vk.put_i32(cnt, 0, 0)
  Vk.enumerate_physical_devices(inst, cnt, null)
  let nd = Vk.get_i32(cnt, 0)
  let devs = bytes(nd * 8 + 8)
  Vk.enumerate_physical_devices(inst, cnt, devs)
  let props = bytes(VkPhysicalDeviceProperties_sizeof)
  # the renderer runs on the first discrete GPU, else the first one listed
  var pick = -1
  for d in 0 .. nd {
    Vk.get_physical_device_properties(Vk.get_ptr(devs, d * 8), props)
    if pick < 0 and Vk.get_i32(props, VkPhysicalDeviceProperties_deviceType) == VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU { pick = d }
  }
  if pick < 0 and nd > 0 { pick = 0 }
  if pick >= 0 {
    let pd = Vk.get_ptr(devs, pick * 8)
    gpu_cap_vulkan = true
    Vk.get_physical_device_properties(pd, props)
    gpu_cap_device = string(Vk.at(props, VkPhysicalDeviceProperties_deviceName))
    gpu_cap_nvidia = Vk.get_i32(props, VkPhysicalDeviceProperties_vendorID) == 4318
    if gpu_cap_nvidia { gpu_cap_rtx = gpu_rtx_generation(gpu_cap_device) }
    let api = Vk.get_i32(props, VkPhysicalDeviceProperties_apiVersion)
    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(pd, f2)
    let major = (api >> 22) & 127
    let minor = (api >> 12) & 1023
    gpu_cap_floor = (major > 1 or (major == 1 and minor >= 3)) and Vk.get_i32(f13, VkPhysicalDeviceVulkan13Features_dynamicRendering) == 1 and Vk.get_i32(f13, VkPhysicalDeviceVulkan13Features_synchronization2) == 1 and Vk.get_i32(f12, VkPhysicalDeviceVulkan12Features_descriptorIndexing) == 1 and Vk.get_i32(f12, VkPhysicalDeviceVulkan12Features_bufferDeviceAddress) == 1 and Vk.get_i32(f12, VkPhysicalDeviceVulkan12Features_timelineSemaphore) == 1
    Vk.put_i32(cnt, 0, 0)
    Vk.enumerate_device_extension_properties(pd, null, cnt, null)
    let ne = Vk.get_i32(cnt, 0)
    let dexts = bytes(ne * VkExtensionProperties_sizeof + 8)
    Vk.enumerate_device_extension_properties(pd, null, cnt, dexts)
    gpu_cap_rt = gpu_ext_in(dexts, ne, VK_KHR_RAY_QUERY_EXTENSION_NAME) and gpu_ext_in(dexts, ne, VK_KHR_ACCELERATION_STRUCTURE_EXTENSION_NAME)
    gpu_cap_mesh = gpu_ext_in(dexts, ne, VK_EXT_MESH_SHADER_EXTENSION_NAME)
    gpu_cap_reflex = gpu_ext_in(dexts, ne, VK_NV_LOW_LATENCY_2_EXTENSION_NAME)
  }
  if own { Vk.destroy_instance(inst, null) }
  print(`r3d: gpu caps: {gpu_cap_device} vulkan {gpu_cap_vulkan} floor {gpu_cap_floor} rt {gpu_cap_rt} mesh {gpu_cap_mesh} rtx {gpu_cap_rtx} reflex {gpu_cap_reflex} hdr {gpu_cap_hdr}`)
}

# Whether the renderer actually draws a feature yet. Until a feature lands, choosing it is saved and
# shown, and says it takes effect later. The Vulkan renderer draws the whole game (phase 37-38), and
# DLSS super resolution and Reflex run through NVIDIA Streamline (streamline.ludic), and HDR output is an
# HDR10 swapchain (gpu_vk_draw.ludic). Mesh-shader grass draws but is not yet counted: at 4K on an
# RTX 3070 Ti its grass pass took 5.0 ms against the chunked path's 1.7. Whether this
# machine can use one is the caps' question, not this one.
function gpu_feature_implemented(f: int) -> bool { return f == GF_VULKAN or f == GF_DLSS or f == GF_REFLEX or f == GF_HDR }

# ---- vertex data --------------------------------------------------------------------
# A Mesh is built through these and records what it is made of - which buffer feeds which
# attribute, at what stride and offset, per vertex or per instance - so a backend that bakes
# vertex input into a pipeline (Vulkan) can read the layout back. On OpenGL each call is the
# GL it replaces, in the same order: a vertex array object per mesh, bound while it is built.
const GPU_F32: int = 1
const GPU_U8: int = 2
const GPU_U16: int = 3
const GPU_STATIC: int = 0
const GPU_DYNAMIC: int = 1
const GPU_STREAM: int = 2
const GPU_MAX_ATTRS: int = 8
const GPU_ATTR_W: int = 7         # per attribute index: buffer, comps, type, stride, offset, normalized, per instance
const GPU_MAX_VBUFS: int = 8

function gpu_gl_type(t: int) -> int {
  if t == GPU_U8 { return GL_UNSIGNED_BYTE }
  if t == GPU_U16 { return GL_UNSIGNED_SHORT }
  return GL_FLOAT
}
function gpu_type_bytes(t: int) -> int {
  if t == GPU_U8 { return 1 }
  if t == GPU_U16 { return 2 }
  return 4
}
function gpu_gl_usage(u: int) -> int {
  if u == GPU_DYNAMIC { return GL_DYNAMIC_DRAW }
  if u == GPU_STREAM { return GL_STREAM_DRAW }
  return GL_STATIC_DRAW
}

# a new mesh, its vertex array bound: the vertex, attribute and index calls below describe it
function gpu_mesh_new() -> Mesh {
  let m = new Mesh
  m.attrs = words(GPU_MAX_ATTRS * GPU_ATTR_W)
  for i in 0 .. GPU_MAX_ATTRS * GPU_ATTR_W { m.attrs[i] = 0 }
  m.vbufs = words(GPU_MAX_VBUFS)
  if gpu_kind != GPU_VK { m.vao = gl_vao() }
  return m
}
# a vertex buffer for the mesh being built (data may be null: storage only); returns it
function gpu_mesh_vertices(m: Mesh, data: pointer, nbytes: int, usage: int) -> int {
  var b = 0
  if gpu_kind == GPU_VK { b = gvk_buf_new();  gvk_buf_upload(b, nbytes, data) }
  else {
    b = gl_buffer()
    gl_bind_buffer(GL_ARRAY_BUFFER, b)
    gl_buffer_data(GL_ARRAY_BUFFER, nbytes, data, gpu_gl_usage(usage))
  }
  if m.vbo == 0 { m.vbo = b }
  if m.n_vbufs < GPU_MAX_VBUFS { m.vbufs[m.n_vbufs] = b;  m.n_vbufs += 1 }
  m.cur_buf = b
  return b
}
function gpu_mesh_record(m: Mesh, index: int, comps: int, type: int, stride: int, offset: int, normalized: bool, inst: bool) -> void {
  if index < 0 or index >= GPU_MAX_ATTRS { return }
  let o = index * GPU_ATTR_W
  var st = stride
  if st == 0 { st = comps * gpu_type_bytes(type) }
  # re-pointing an attribute at another buffer keeps the layout; changing its shape does not
  if m.attrs[o + 1] != comps or m.attrs[o + 2] != type or m.attrs[o + 3] != st or m.attrs[o + 4] != offset or m.attrs[o + 5] != gpu_b(normalized) or m.attrs[o + 6] != gpu_b(inst) or (m.attrs[o] == m.attrs[0]) != (m.cur_buf == m.attrs[0]) { m.vk_layout = 0 }
  m.attrs[o] = m.cur_buf;  m.attrs[o + 1] = comps;  m.attrs[o + 2] = type;  m.attrs[o + 3] = st
  m.attrs[o + 4] = offset;  m.attrs[o + 5] = gpu_b(normalized);  m.attrs[o + 6] = gpu_b(inst)
  if index + 1 > m.n_attrs { m.n_attrs = index + 1 }
}
# attribute `index` read from the last vertex buffer (stride 0: tightly packed)
function gpu_mesh_attr(m: Mesh, index: int, comps: int, type: int, stride: int, offset: int, normalized: bool) -> void {
  if gpu_kind != GPU_VK {
    gl_enable_vertex_attrib_array(index)
    gl_vertex_attrib_pointer(index, comps, gpu_gl_type(type), gpu_b(normalized), stride, gl_ptr(null, offset))
  }
  gpu_mesh_record(m, index, comps, type, stride, offset, normalized, false)
}
# the index buffer: 4-byte or 2-byte indices
function gpu_mesh_indices(m: Mesh, data: pointer, nbytes: int, index_bytes: int) -> void {
  m.itype = GL_UNSIGNED_INT
  if index_bytes == 2 { m.itype = GL_UNSIGNED_SHORT }
  if gpu_kind == GPU_VK { m.ebo = gvk_buf_new();  gvk_buf_upload(m.ebo, nbytes, data);  return }
  m.ebo = gl_buffer()
  gl_bind_buffer(GL_ELEMENT_ARRAY_BUFFER, m.ebo)
  gl_buffer_data(GL_ELEMENT_ARRAY_BUFFER, nbytes, data, GL_STATIC_DRAW)
}
# finished describing: nothing else is bound to it by accident
function gpu_mesh_done(m: Mesh) -> void { if gpu_kind == GPU_VK { return };  gl_bind_vertex_array(0) }

# Per-instance data: `buf` feeds the attributes named next, one element per instance. A mesh
# drawn from different instance buffers (the scatter layers' LOD buckets) is re-pointed here
# before each draw; on Vulkan that is a vertex-buffer binding, not a change of layout.
function gpu_mesh_bind_instances(m: Mesh, buf: int) -> void {
  if gpu_kind != GPU_VK {
    gl_bind_vertex_array(m.vao)
    gl_bind_buffer(GL_ARRAY_BUFFER, buf)
  }
  m.cur_buf = buf
  m.ibuf = buf
}
function gpu_mesh_attr_inst(m: Mesh, index: int, comps: int, type: int, stride: int, offset: int) -> void {
  if gpu_kind != GPU_VK {
    gl_enable_vertex_attrib_array(index)
    gl_vertex_attrib_pointer(index, comps, gpu_gl_type(type), 0, stride, gl_ptr(null, offset))
    gl_vertex_attrib_divisor(index, 1)
  }
  gpu_mesh_record(m, index, comps, type, stride, offset, false, true)
}

# a buffer on its own (instances, a stream): made, filled whole, freed
function gpu_buffer_new() -> int { if gpu_kind == GPU_VK { return gvk_buf_new() };  return gl_buffer() }
function gpu_buffer_upload(buf: int, nbytes: int, data: pointer, usage: int) -> void {
  if gpu_kind == GPU_VK { gvk_buf_upload(buf, nbytes, data);  return }
  gl_bind_buffer(GL_ARRAY_BUFFER, buf)
  gl_buffer_data(GL_ARRAY_BUFFER, nbytes, data, gpu_gl_usage(usage))
}
function gpu_buffer_free(buf: int) -> void {
  if gpu_kind == GPU_VK { if buf > 0 { gvk_buf_release(buf) };  return }
  if buf == 0 { return }
  let ids = gpu_tmp()
  ids[0] = buf
  gl_delete_buffers(1, ids)
}

# ---- compute and indirect draws (Vulkan) ----------------------------------------------------
# The GPU-driven path: a compute program writes instance lists and draw commands into buffers
# the draws then read. OpenGL here is 4.1 (macOS) with no compute, so gpu_compute is 0 there and
# a caller keeps its CPU path. A compute program's binding 0 is its parameter block (params,
# copied at the dispatch); bindings 1.. are `bufs`, gpu buffers.
function gpu_has_compute() -> bool { return gpu_kind == GPU_VK }
# several records in one indirect draw, each with its own firstInstance
function gpu_has_mdi() -> bool { return gpu_kind == GPU_VK and gvk_has_mdi }
# mesh shaders (VK_EXT_mesh_shader): a *.mesh program drawn with gpu_draw_mesh_tasks
function gpu_has_mesh() -> bool { return gpu_kind == GPU_VK and gvk_has_mesh }
function gpu_draw_mesh_tasks(x: int, y: int, z: int) -> void { if gpu_kind == GPU_VK { gvk_draw_mesh_tasks_now(x, y, z) } }
function gpu_compute(name: string, n_bufs: int) -> int {
  if gpu_kind != GPU_VK { return 0 }
  return gvk_compute_new(name, n_bufs)
}
function gpu_dispatch(c: int, params: pointer, n_params: int, bufs: words, groups: int) -> void {
  if gpu_kind == GPU_VK and c > 0 { gvk_dispatch(c, params, n_params, bufs, groups, 1, 1) }
}
# a buffer a compute pass writes (never reallocated under a draw that reads it)
function gpu_buffer_gpu_owned(buf: int) -> void { if gpu_kind == GPU_VK { gvk_buf_gpu_owned(buf) } }
# the host-visible contents of a buffer, for a readback after gpu_finish; null on OpenGL
function gpu_buffer_map(buf: int) -> pointer {
  if gpu_kind != GPU_VK or buf <= 0 { return null }
  return gvk_buf_map[buf]
}
function gpu_finish() -> void { if gpu_kind == GPU_VK { gvk_flush() } }
# n indexed draws of mesh m from VkDrawIndexedIndirectCommand records in buffer cmds at offset
# (bytes); each record's firstInstance selects its instances out of the bound instance buffer.
# With count_buf > 0 the GPU's own count (a uint at count_off) is used, up to n.
function gpu_draw_mesh_indirect(m: Mesh, cmds: int, offset: int, n: int, count_buf: int, count_off: int) -> void {
  if m != null { ds_draw(DS_INDIRECT, n, m.count) }
  if gpu_kind == GPU_VK { gvk_draw_indirect_now(m, cmds, offset, n, count_buf, count_off) }
}

# drawing
function gpu_mesh_bind(m: Mesh) -> void { if gpu_kind == GPU_VK { return };  gl_bind_vertex_array(m.vao) }
function gpu_mesh_unbind() -> void { if gpu_kind == GPU_VK { return };  gl_bind_vertex_array(0) }
function gpu_draw_mesh(m: Mesh) -> void {
  ds_draw(DS_MESH, 1, m.count)
  if gpu_kind == GPU_VK { gvk_draw_now(m, 0, 0, 1);  return }
  gpu_glcheck_before("a draw")
  gl_bind_vertex_array(m.vao)
  if m.ebo != 0 { gl_draw_elements(m.mode, m.count, m.itype, null) }
  else { gl_draw_arrays(m.mode, 0, m.count) }
  gpu_glcheck_after("a draw")
}
function gpu_draw_mesh_instanced(m: Mesh, n: int) -> void {
  ds_draw(DS_INSTANCED, n, m.count * n)
  if gpu_kind == GPU_VK { gvk_draw_now(m, 0, 0, n);  return }
  gpu_glcheck_before("an instanced draw")
  gl_bind_vertex_array(m.vao)
  if m.ebo != 0 { gl_draw_elements_instanced(m.mode, m.count, m.itype, null, n) }
  else { gl_draw_arrays_instanced(m.mode, 0, m.count, n) }
  gpu_glcheck_after("an instanced draw")
}
# the bound mesh's indices again (a patch mesh drawn once per terrain node)
function gpu_draw_bound_elements(m: Mesh) -> void {
  ds_draw(DS_PATCH, 1, m.count)
  if gpu_kind == GPU_VK { gvk_draw_now(m, 0, 0, 1);  return }
  gpu_glcheck_before("a terrain patch")
  gl_draw_elements(m.mode, m.count, m.itype, null)
  gpu_glcheck_after("a terrain patch")
}
# vertices [first, first + count) of the bound mesh, as triangles (the overlay's ranges)
function gpu_draw_range(m: Mesh, first: int, count: int) -> void {
  ds_draw(DS_RANGE, 1, count)
  if gpu_kind == GPU_VK { gvk_draw_now(m, first, count, 1);  return }
  gpu_glcheck_before("an overlay draw")
  gl_draw_arrays(GL_TRIANGLES, first, count)
  gpu_glcheck_after("an overlay draw")
}

function gpu_mesh_free(m: Mesh) -> void {
  if gpu_kind == GPU_VK { gvk_mesh_free(m);  return }
  if m == null { return }
  let ids = gpu_tmp()
  if m.vbufs != null {
    for i in 0 .. m.n_vbufs { ids[0] = m.vbufs[i];  gl_delete_buffers(1, ids) }
    m.n_vbufs = 0
  } else if m.vbo != 0 { ids[0] = m.vbo;  gl_delete_buffers(1, ids) }
  m.vbo = 0
  if m.ebo != 0 { ids[0] = m.ebo;  gl_delete_buffers(1, ids);  m.ebo = 0 }
  if m.vao != 0 { ids[0] = m.vao;  gl_delete_vertex_arrays(1, ids);  m.vao = 0 }
}

# ---- textures -----------------------------------------------------------------------------
# A texture is a handle (on OpenGL, the texture name). Each call below is the one GL call it
# replaces, in the renderer's own order, so OpenGL draws exactly what it drew. What those calls
# say about a texture - its size, format, filters, wraps, comparison, mipmaps, anisotropy - is
# recorded per handle as it is set, because a backend with immutable images and separate
# sampler objects (Vulkan) creates both from exactly that. Parameters apply to the texture last
# bound for that kind, which is how every call site here already works.
const GPU_TEX2D: int = 1
const GPU_TEX2D_ARRAY: int = 2
const GPU_TX_W: int = 12          # per handle: kind, w, h, layers, ifmt, min, mag, wrap s, wrap t, compare, mips, aniso
var gpu_tx: words = null
var gpu_tx_cap: int = 0
var gpu_bound_2d: int = 0
var gpu_bound_array: int = 0

function gpu_gl_target(kind: int) -> int { if kind == GPU_TEX2D_ARRAY { return GL_TEXTURE_2D_ARRAY };  return GL_TEXTURE_2D }
# the record for a handle, growing the table when a new name is larger than it
function gpu_tx_at(tex: int) -> int {
  if tex <= 0 { return -1 }
  if tex >= gpu_tx_cap {
    var cap = gpu_tx_cap * 2
    if cap < 256 { cap = 256 }
    while cap <= tex { cap = cap * 2 }
    let t = words(cap * GPU_TX_W)
    for i in 0 .. cap * GPU_TX_W { t[i] = 0 }
    if gpu_tx != null { mem_copy(t, gpu_tx, gpu_tx_cap * GPU_TX_W * 4);  free(gpu_tx) }
    gpu_tx = t
    gpu_tx_cap = cap
  }
  return tex * GPU_TX_W
}
function gpu_bound(kind: int) -> int { if kind == GPU_TEX2D_ARRAY { return gpu_bound_array };  return gpu_bound_2d }

function gpu_tex_new() -> int { if gpu_kind == GPU_VK { return gvk_tex_new() };  return gl_texture() }
# does texture `tex` have an image behind it? Always on OpenGL; on Vulkan an image whose memory could
# not be had is never made, and a caller that can fall back (a smaller shadow map) asks here.
function gpu_tex_ok(tex: int) -> bool {
  if gpu_kind != GPU_VK { return tex > 0 }
  return tex > 0 and tex < len(gvk_tex_image) and gvk_tex_image[tex] != 0
}
# what GL has on each unit's 2D target, for R3D_GLCHECK: deleting a texture unbinds it everywhere
var gpu_unit_2d: words = null
var gpu_unit_cur: int = 0
function gpu_tex_unit(unit: int) -> void { if gpu_kind == GPU_VK { gpu_unit_cur = unit;  return };  gl_active_texture(GL_TEXTURE0 + unit);  gpu_unit_cur = unit }
function gpu_tex_bind(kind: int, tex: int) -> void {
  ds_tex_bind()
  if gpu_kind != GPU_VK { gl_bind_texture(gpu_gl_target(kind), tex) }
  if kind == GPU_TEX2D_ARRAY { gpu_bound_array = tex } else { gpu_bound_2d = tex }
  if kind != GPU_TEX2D_ARRAY and gpu_unit_cur < 32 {
    if gpu_unit_2d == null { gpu_unit_2d = words(32);  for i in 0 .. 32 { gpu_unit_2d[i] = -1 } }
    gpu_unit_2d[gpu_unit_cur] = tex
  }
}
# pixel transfer packing (alignment, byte swap) for the uploads and read-backs that follow
function gpu_pixel_store(pname: int, value: int) -> void { if gpu_kind == GPU_VK { if pname == GL_UNPACK_SWAP_BYTES { gvk_unpack_swap = value == 1 };  return };  gl_pixel_storei(pname, value);  gpu_glcheck_after("pixel store") }
function gpu_tex_image2d(ifmt: int, w: int, h: int, fmt: int, ty: int, data: pointer) -> void {
  if gpu_kind == GPU_VK {
    gvk_flush()
    gvk_tex_storage(gpu_bound_2d, false, ifmt, w, h, 1, data != null)
    if data != null { gvk_tex_upload(gpu_bound_2d, ifmt, w, h, 1, fmt, ty, data) }
  } else { gl_tex_image2d(GL_TEXTURE_2D, 0, ifmt, w, h, 0, fmt, ty, data) }
  gpu_glcheck_after(`a {w}x{h} texture upload (format {ifmt})`)
  let o = gpu_tx_at(gpu_bound_2d)
  if o >= 0 { gpu_tx[o] = GPU_TEX2D;  gpu_tx[o + 1] = w;  gpu_tx[o + 2] = h;  gpu_tx[o + 3] = 1;  gpu_tx[o + 4] = ifmt }
}
function gpu_tex_image3d(ifmt: int, w: int, h: int, layers: int, fmt: int, ty: int, data: pointer) -> void {
  if gpu_kind == GPU_VK {
    gvk_flush()
    gvk_tex_storage(gpu_bound_array, true, ifmt, w, h, layers, data != null)
    if data != null { gvk_tex_upload(gpu_bound_array, ifmt, w, h, layers, fmt, ty, data) }
  } else { gl_tex_image3d(GL_TEXTURE_2D_ARRAY, 0, ifmt, w, h, layers, 0, fmt, ty, data) }
  gpu_glcheck_after(`a {w}x{h}x{layers} array upload (format {ifmt})`)
  let o = gpu_tx_at(gpu_bound_array)
  if o >= 0 { gpu_tx[o] = GPU_TEX2D_ARRAY;  gpu_tx[o + 1] = w;  gpu_tx[o + 2] = h;  gpu_tx[o + 3] = layers;  gpu_tx[o + 4] = ifmt }
}
function gpu_tex_param(kind: int, pname: int, value: int) -> void {
  if gpu_kind != GPU_VK { gl_tex_parameteri(gpu_gl_target(kind), pname, value) }
  let o = gpu_tx_at(gpu_bound(kind))
  if o < 0 { return }
  if pname == GL_TEXTURE_MIN_FILTER { gpu_tx[o + 5] = value }
  if pname == GL_TEXTURE_MAG_FILTER { gpu_tx[o + 6] = value }
  if pname == GL_TEXTURE_WRAP_S { gpu_tx[o + 7] = value }
  if pname == GL_TEXTURE_WRAP_T { gpu_tx[o + 8] = value }
  if pname == GL_TEXTURE_COMPARE_MODE { if value == GL_NONE { gpu_tx[o + 9] = 0 } }
  if pname == GL_TEXTURE_COMPARE_FUNC { gpu_tx[o + 9] = value }
  gpu_glcheck_after("tex param")
}
# a float parameter (fixed, as Gl.* takes it): anisotropy is the one the renderer sets
function gpu_tex_paramf(kind: int, pname: int, value: fixed) -> void {
  if gpu_kind != GPU_VK { gl_tex_parameterf(gpu_gl_target(kind), pname, value) }
  let o = gpu_tx_at(gpu_bound(kind))
  if o >= 0 and pname == 0x84FE { gpu_tx[o + 11] = fx_to_f32(value) }
  gpu_glcheck_after("tex paramf")
}
# the border colour clamp-to-border reads (four fixed values in `rgba`)
function gpu_tex_border(kind: int, rgba: pointer) -> void { if gpu_kind == GPU_VK { return };  gl_tex_parameterfv(gpu_gl_target(kind), GL_TEXTURE_BORDER_COLOR, rgba) }
function gpu_tex_mips(kind: int) -> void {
  if gpu_kind == GPU_VK {
    let mt = gpu_bound(kind)
    let mo = gpu_tx_at(mt)
    if mo >= 0 { gvk_mips_now(mt, gpu_tx[mo + 1], gpu_tx[mo + 2]) }
  } else { gl_generate_mipmap(gpu_gl_target(kind)) }
  gpu_glcheck_after(`mipmaps for texture {gpu_bound(kind)}`)
  let o = gpu_tx_at(gpu_bound(kind))
  if o >= 0 { gpu_tx[o + 10] = 1 }
}
# level 0 of the bound texture into `out`
function gpu_tex_read(kind: int, fmt: int, ty: int, out: pointer) -> void {
  if gpu_kind == GPU_VK { gvk_flush();  let rt = gpu_bound(kind);  let ro = gpu_tx_at(rt);  if ro >= 0 { gvk_tex_read(rt, gpu_tx[ro + 4], gpu_tx[ro + 1], gpu_tx[ro + 2], fmt, ty, out) };  return }
  gl_get_tex_image(gpu_gl_target(kind), 0, fmt, ty, out)
  gpu_glcheck_after(`a read-back of texture {gpu_bound(kind)}`)
}
function gpu_tex_free(tex: int) -> void {
  if tex == 0 { return }
  let ids = gpu_tmp()
  ids[0] = tex
  if gpu_kind == GPU_VK { gvk_flush();  if tex < len(gvk_tex_image) { gvk_tex_release(tex) } } else { gl_delete_textures(1, ids) }
  if gpu_unit_2d != null { for i in 0 .. 32 { if gpu_unit_2d[i] == tex { gpu_unit_2d[i] = 0;  if gpu_glcheck_on() { gpu_glcheck_say(`gpu: texture {tex} freed while bound on unit {i}`) } } } }
  if gpu_bound_2d == tex { gpu_bound_2d = 0 }
  let o = gpu_tx_at(tex)
  if o >= 0 { for i in 0 .. GPU_TX_W { gpu_tx[o + i] = 0 } }
}
# a texture on a unit for a program's sampler, by the sampler's name
function gpu_bind_sampler(prog: int, name: string, unit: int, kind: int, tex: int) -> void {
  if gpu_kind == GPU_VK { gvk_bind_texture(prog, name, tex);  return }
  gpu_tex_unit(unit)
  gpu_tex_bind(kind, tex)
  u_i(gpu_uniform(prog, name), unit)
  if gpu_glcheck_on() {
    # a sampler reading a texture nobody made (0 or freed) is the "unloadable" the driver warns of
    let o = gpu_tx_at(tex)
    if tex == 0 or o < 0 or gpu_tx[o] == 0 { gpu_glcheck_say(`gpu: {name} on unit {unit} of program {prog} samples texture {tex}, which has no image`) }
    else {
      # incomplete: a min filter that reads mipmaps (GL's default does, when none was set) on a
      # texture that never had them generated - the driver samples zero ("unloadable")
      let mn = gpu_tx[o + 5]
      let wants_mips = mn == 0 or mn == 0x2700 or mn == 0x2701 or mn == 0x2702 or mn == 0x2703
      if wants_mips and gpu_tx[o + 10] == 0 {
        var why = "a mipmap filter"
        if mn == 0 { why = "no min filter set (GL's default reads mipmaps)" }
        gpu_glcheck_say(`gpu: {name} on unit {unit} of program {prog} samples texture {tex} ({gpu_tx[o + 1]}x{gpu_tx[o + 2]}, format {gpu_tx[o + 4]}) with {why} but no mipmaps`)
      }
      # compare mode on: only a shadow sampler may read it; a plain one reads zero ("unloadable")
      if gpu_tx[o + 9] != 0 and not Text.contains(name, "shadow") { gpu_glcheck_say(`gpu: {name} on unit {unit} of program {prog} samples texture {tex} ({gpu_tx[o + 1]}x{gpu_tx[o + 2]}, format {gpu_tx[o + 4]}) with depth compare on`) }

    }
    gpu_glcheck_after(`binding {name} (texture {tex}) on unit {unit}`)
  }
}

# ---- render targets and passes -------------------------------------------------------------
# A framebuffer is a handle (OpenGL's name). What is attached to it - colour textures in slots,
# a depth texture or one layer of an array, renderbuffers and their sample count - is recorded as
# it is attached, which is what a backend with render passes and image views builds from. Each
# call is the GL call it replaces, in the renderer's order.
#
# gpu_check(tag) reports a pending error under a name, as gl_check did. R3D_GLCHECK=1 adds a
# completeness check of the bound framebuffer whenever a viewport is set, naming the handle, so
# a pass drawing into an incomplete target says which one; unset, it calls nothing.
const GPU_FB_W: int = 8           # per handle: colour 0, colour 1, depth texture, depth layer + 1, colour rb, depth rb, samples, colour layer + 1
var gpu_fb: words = null
var gpu_fb_cap: int = 0
var gpu_fb_cur: int = 0
var gpu_glcheck: int = -1
var gpu_rb_samples: int = 0
var gpu_drawbufs: words = null

function gpu_fb_at(fb: int) -> int {
  if fb <= 0 { return -1 }
  if fb >= gpu_fb_cap {
    var cap = gpu_fb_cap * 2
    if cap < 64 { cap = 64 }
    while cap <= fb { cap = cap * 2 }
    let t = words(cap * GPU_FB_W)
    for i in 0 .. cap * GPU_FB_W { t[i] = 0 }
    if gpu_fb != null { mem_copy(t, gpu_fb, gpu_fb_cap * GPU_FB_W * 4);  free(gpu_fb) }
    gpu_fb = t
    gpu_fb_cap = cap
  }
  return fb * GPU_FB_W
}
function gpu_glcheck_on() -> bool {
  if gpu_glcheck < 0 { gpu_glcheck = 0;  if Os.has_env("R3D_GLCHECK") { gpu_glcheck = 1 } }
  return gpu_glcheck == 1
}
function gpu_check(tag: string) -> int { if gpu_kind == GPU_VK { return 0 };  return gl_check(tag) }
# a named checkpoint that costs nothing unless R3D_GLCHECK is set
function gpu_debug_check(tag: string) -> void { if gpu_kind == GPU_VK { return };  if gpu_glcheck_on() { gl_check(tag) } }

function gpu_fb_new() -> int { if gpu_kind == GPU_VK { gvk_fb_counter += 1;  return gvk_fb_counter };  return gl_framebuffer() }
function gpu_fb_bind(fb: int) -> void {
  if gpu_kind == GPU_VK { gvk_rebind(fb) } else { gl_bind_framebuffer(GL_FRAMEBUFFER, fb) }
  gpu_fb_cur = fb
  gpu_glcheck_after("fb bind")
}
function gpu_fb_bind_read(fb: int) -> void { if gpu_kind == GPU_VK { gvk_fb_read = fb;  return };  gl_bind_framebuffer(GL_READ_FRAMEBUFFER, fb);  gpu_glcheck_after("fb bind read") }
function gpu_fb_bind_draw(fb: int) -> void { if gpu_kind == GPU_VK { gvk_fb_draw = fb;  return };  gl_bind_framebuffer(GL_DRAW_FRAMEBUFFER, fb);  gpu_glcheck_after("fb bind draw") }
function gpu_fb_color(slot: int, tex: int) -> void {
  if gpu_kind == GPU_VK { gvk_pass_end() } else { gl_framebuffer_texture2d(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0 + slot, GL_TEXTURE_2D, tex, 0) }
  let o = gpu_fb_at(gpu_fb_cur)
  if o >= 0 and slot < 2 { gpu_fb[o + slot] = tex;  if slot == 0 { gpu_fb[o + 7] = 0 } }
  gpu_glcheck_after("attaching to {gpu_fb_describe(gpu_fb_cur)}")
}
function gpu_fb_color_layer(slot: int, tex: int, layer: int) -> void {
  if gpu_kind == GPU_VK { gvk_pass_end() } else { gl_framebuffer_texture_layer(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0 + slot, tex, 0, layer) }
  let o = gpu_fb_at(gpu_fb_cur)
  if o >= 0 and slot < 2 { gpu_fb[o + slot] = tex;  if slot == 0 { gpu_fb[o + 7] = layer + 1 } }
  gpu_glcheck_after("attaching to {gpu_fb_describe(gpu_fb_cur)}")
}
function gpu_fb_depth(tex: int) -> void {
  if gpu_kind == GPU_VK { gvk_pass_end() } else { gl_framebuffer_texture2d(GL_FRAMEBUFFER, GL_DEPTH_ATTACHMENT, GL_TEXTURE_2D, tex, 0) }
  let o = gpu_fb_at(gpu_fb_cur)
  if o >= 0 { gpu_fb[o + 2] = tex;  gpu_fb[o + 3] = 0 }
  gpu_glcheck_after("attaching to {gpu_fb_describe(gpu_fb_cur)}")
}
function gpu_fb_depth_layer(tex: int, layer: int) -> void {
  if gpu_kind == GPU_VK { gvk_pass_end() } else { gl_framebuffer_texture_layer(GL_FRAMEBUFFER, GL_DEPTH_ATTACHMENT, tex, 0, layer) }
  let o = gpu_fb_at(gpu_fb_cur)
  if o >= 0 { gpu_fb[o + 2] = tex;  gpu_fb[o + 3] = layer + 1 }
  gpu_glcheck_after("attaching to {gpu_fb_describe(gpu_fb_cur)}")
}
function gpu_rb_new() -> int {
  if gpu_kind == GPU_VK { return gvk_tex_new() }
  let ids = gpu_tmp()
  gl_gen_renderbuffers(1, ids)
  return ids[0]
}
# storage for a renderbuffer: samples > 0 makes it multisampled
function gpu_rb_storage(rb: int, ifmt: int, w: int, h: int, samples: int) -> void {
  if gpu_kind == GPU_VK {
    gvk_flush();  gpu_rb_samples = samples
    gvk_storage_samples = samples
    gvk_tex_storage(rb, false, ifmt, w, h, 1, false)
    gvk_storage_samples = 1
    return
  }
  gl_bind_renderbuffer(GL_RENDERBUFFER, rb)
  if samples > 0 { gl_renderbuffer_storage_multisample(GL_RENDERBUFFER, samples, ifmt, w, h) }
  else { gl_renderbuffer_storage(GL_RENDERBUFFER, ifmt, w, h) }
  gpu_rb_samples = samples
  gpu_glcheck_after("rb storage")
}
function gpu_fb_color_rb(slot: int, rb: int) -> void {
  if gpu_kind == GPU_VK { gvk_pass_end();  let co = gpu_fb_at(gpu_fb_cur);  if co >= 0 and slot < 2 { gpu_fb[co + slot] = rb };  return }
  gl_framebuffer_renderbuffer(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0 + slot, GL_RENDERBUFFER, rb)
  let o = gpu_fb_at(gpu_fb_cur)
  if o >= 0 { gpu_fb[o + 4] = rb;  gpu_fb[o + 6] = gpu_rb_samples }
}
function gpu_fb_depth_rb(rb: int) -> void {
  if gpu_kind == GPU_VK { gvk_pass_end();  let dop = gpu_fb_at(gpu_fb_cur);  if dop >= 0 { gpu_fb[dop + 2] = rb;  gpu_fb[dop + 3] = 0 };  return }
  gl_framebuffer_renderbuffer(GL_FRAMEBUFFER, GL_DEPTH_ATTACHMENT, GL_RENDERBUFFER, rb)
  let o = gpu_fb_at(gpu_fb_cur)
  if o >= 0 { gpu_fb[o + 5] = rb;  gpu_fb[o + 6] = gpu_rb_samples }
}
# colour slots 0 .. n-1 are drawn into (several: an MRT bake)
function gpu_fb_draw_buffers(n: int) -> void {
  if gpu_kind == GPU_VK { gvk_fb_colors(gpu_fb_cur, n);  return }
  if gpu_drawbufs == null { gpu_drawbufs = words(8) }
  for i in 0 .. n { gpu_drawbufs[i] = GL_COLOR_ATTACHMENT0 + i }
  gl_draw_buffers(n, gpu_drawbufs)
  gpu_glcheck_after("fb draw buffers")
}
# a depth-only target: no colour is drawn or read
function gpu_fb_no_color() -> void {
  if gpu_kind == GPU_VK { gvk_fb_colors(gpu_fb_cur, 0);  return }
  gl_draw_buffer(GL_NONE)
  gl_read_buffer(GL_NONE)
  gpu_glcheck_after("fb no color")
}
function gpu_fb_status() -> int { if gpu_kind == GPU_VK { return GL_FRAMEBUFFER_COMPLETE };  return gl_check_framebuffer_status(GL_FRAMEBUFFER) }
function gpu_fb_free(fb: int) -> void {
  if gpu_kind == GPU_VK { gvk_fb_forget(fb);  let fo = gpu_fb_at(fb);  if fo >= 0 { for i in 0 .. GPU_FB_W { gpu_fb[fo + i] = 0 } };  return }
  if fb == 0 { return }
  let ids = gpu_tmp()
  ids[0] = fb
  gl_delete_framebuffers(1, ids)
  let o = gpu_fb_at(fb)
  if o >= 0 { for i in 0 .. GPU_FB_W { gpu_fb[o + i] = 0 } }
  gpu_glcheck_after("fb free")
}
function gpu_rb_free(rb: int) -> void {
  if gpu_kind == GPU_VK { gvk_flush();  if rb > 0 and rb < len(gvk_tex_image) { gvk_tex_release(rb) };  return }
  if rb == 0 { return }
  let ids = gpu_tmp()
  ids[0] = rb
  gl_delete_renderbuffers(1, ids)
  gpu_glcheck_after("rb free")
}
function gpu_viewport(x: int, y: int, w: int, h: int) -> void { if gpu_kind == GPU_VK { gvk_viewport(x, y, w, h);  return };  gl_viewport(x, y, w, h);  gpu_glcheck_after("viewport") }

# R3D_GLCHECK: before a draw or a clear, the bound framebuffer must be complete; after it, no
# error may be pending. Each report names the framebuffer and what was being done, once per
# distinct message, so the first bad pass is found without a flood. (Checking when a viewport
# is set reported the shadow pass, which sets its viewport before attaching a cascade.)
var gpu_glcheck_seen: []string = null
function gpu_glcheck_say(msg: string) -> void {
  if gpu_glcheck_seen == null { gpu_glcheck_seen = new []string }
  for i in 0 .. len(gpu_glcheck_seen) { if gpu_glcheck_seen[i] == msg { return } }
  push(gpu_glcheck_seen, msg)
  print(msg)
}
# a framebuffer as a person reads it: its handle and its colour (or depth) attachment's size and format
function gpu_fb_describe(fb: int) -> string {
  if fb == 0 { return "the default framebuffer" }
  let o = gpu_fb_at(fb)
  if o < 0 { return `framebuffer {fb}` }
  var tex = gpu_fb[o]
  var what = "colour"
  if tex == 0 { tex = gpu_fb[o + 2];  what = "depth" }
  let t = gpu_tx_at(tex)
  if tex == 0 or t < 0 { return `framebuffer {fb} (nothing recorded attached)` }
  return `framebuffer {fb} ({what} texture {tex}, {gpu_tx[t + 1]}x{gpu_tx[t + 2]}, format {gpu_tx[t + 4]})`
}
function gpu_glcheck_before(what: string) -> void {
  if gpu_kind == GPU_VK { return }
  if not gpu_glcheck_on() { return }
  let pending = gl_get_error()
  if pending != 0 { gpu_glcheck_say(`gpu: error {pending} pending before {what} into {gpu_fb_describe(gpu_fb_cur)}`) }
  if gpu_unit_2d != null and gpu_unit_2d[0] == 0 { gpu_glcheck_say(`gpu: {what} into {gpu_fb_describe(gpu_fb_cur)} with unit 0's 2D texture deleted`) }
  let st = gl_check_framebuffer_status(GL_FRAMEBUFFER)
  if st != GL_FRAMEBUFFER_COMPLETE { gpu_glcheck_say(`gpu: {gpu_fb_describe(gpu_fb_cur)} incomplete ({st}) at {what}`) }
}
function gpu_glcheck_after(what: string) -> void {
  if gpu_kind == GPU_VK { return }
  if not gpu_glcheck_on() { return }
  let e = gl_get_error()
  if e != 0 { gpu_glcheck_say(`gpu: error {e} from {what} into {gpu_fb_describe(gpu_fb_cur)}`) }
}
function gpu_clear_color(r: fixed, g: fixed, b: fixed, a: fixed) -> void { if gpu_kind == GPU_VK { gvk_clear_color(fx_to_f32(r), fx_to_f32(g), fx_to_f32(b), fx_to_f32(a));  return };  gl_clear_color(r, g, b, a) }
function gpu_clear(mask: int) -> void {
  if gpu_kind == GPU_VK { gvk_clear(mask, gpu_fb, gpu_fb_at(gvk_fb_cur));  return }
  gpu_glcheck_before("a clear")
  gl_clear(mask)
  gpu_glcheck_after("a clear")
}
# the bound read framebuffer's [0, w) x [0, h) into the bound draw framebuffer's, unscaled
function gpu_blit(w: int, h: int, mask: int) -> void {
  if gpu_kind == GPU_VK { gvk_blit(w, h, mask);  return }
  gl_blit_framebuffer(0, 0, w, h, 0, 0, w, h, mask, GL_NEAREST)
  gpu_glcheck_after("a blit")
}
# the framebuffer the finished frame is presented from (an offscreen one, headless)
function gpu_screen_fb() -> int { if gpu_kind == GPU_VK { return 0 };  return gl_screen }
function gpu_multisample(on: bool) -> void { if gpu_kind == GPU_VK { return };  gpu_gl_cap(GL_MULTISAMPLE, gpu_b(on));  gpu_glcheck_after("multisample") }
# the most samples the scene may be drawn with: the device's colour-and-depth limit on Vulkan (0 before
# it is open), 4 on OpenGL, which has always asked for up to that
function gpu_msaa_max() -> int { if gpu_kind == GPU_VK { return gvk_msaa_max };  return 4 }
function gpu_wireframe(on: bool) -> void {
  if gpu_kind == GPU_VK { gvk_wireframe = gpu_b(on);  return }
  if on { gl_polygon_mode(GL_FRONT_AND_BACK, GL_LINE) } else { gl_polygon_mode(GL_FRONT_AND_BACK, GL_FILL) }
  gpu_glcheck_after("wireframe")
}
# the presented frame as RGB8, bottom row first (a photograph)
function gpu_read_screen(w: int, h: int, out: pointer) -> void {
  if gpu_kind == GPU_VK { gvk_read_screen(w, h, out);  return }
  gl_bind_framebuffer(GL_READ_FRAMEBUFFER, gl_screen)
  gl_pixel_storei(GL_PACK_ALIGNMENT, 1)
  gl_read_pixels(0, 0, w, h, GL_RGB, GL_UNSIGNED_BYTE, out)
  gpu_glcheck_after(`a {w}x{h} read of the screen`)
}