path: root/builtin-programs/gpu/gpu.folk

# gpu.folk --
#
#     Sets up the GPU device.

if {[info exists this] && $::tcl_platform(os) eq "darwin"} {
    # We hard-code gpu.folk into thread 0, so we should abort if not
    # running that way.
    return
}

fn defineVulkanHandleType {cc type} {
    $cc argtype $type [format {
#ifdef VK_USE_64_BIT_PTR_DEFINES
        %s $argname; sscanf(Jim_String($obj), "(%s) 0x%%p", &$argname);
#else
        %s $argname; sscanf(Jim_String($obj), "(%s) 0x%%llx", &$argname);
#endif
    } $type $type $type $type]
    # Tcl_ObjPrintf doesn't work with %lld/%llx for some reason,
    # so we do it by hand.
    $cc rtype $type [format {
#ifdef VK_USE_64_BIT_PTR_DEFINES
        $robj = Jim_ObjPrintf("(%s) 0x%%" PRIxPTR, (uintptr_t) $rvalue);
#else
        {
          char buf[100]; snprintf(buf, 100, "(%s) 0x%%llx", $rvalue);
          $robj = Jim_NewStringObj(buf, -1);
        }
#endif
    } $type $type]
}
Claim the GPU Vulkan handle type definer is [fn defineVulkanHandleType]

When $::thisNode has had displays enumerated &\
     /someone/ wishes $::thisNode uses display /display/ with /...any/ {

set useGlfw $($display eq "glfw")

# First, we build the C GPU library, which can make direct calls into
# Vulkan -- this library then exposes functions that we can call from
# Tcl.
set gpuc [C]
$gpuc cflags -I./vendor
$gpuc include <pthread.h>
$gpuc code {
    #define VOLK_IMPLEMENTATION
    #include "volk/volk.h"
}
if {$useGlfw} {
    $gpuc code {
        // This must be included _after_ volk.h (Vulkan).
        #include <GLFW/glfw3.h>
    }
}

set macos [expr {$::tcl_platform(os) eq "darwin"}]
if {$macos} {
    $gpuc cflags -I/opt/homebrew/include -L/opt/homebrew/lib
}

if {$useGlfw} {
    $gpuc endcflags -lglfw
} else {
    foreach renderFile [glob -nocomplain "/dev/dri/render*"] {
        if {![file readable $renderFile]} {
            puts stderr "Gpu: Warning: $renderFile is not readable by current user; Vulkan device may not appear.
Try doing `sudo chmod 666 $renderFile`."
        }
    }
}

$gpuc argtype VkResult { long $argname; __ENSURE_OK(Jim_GetLong(interp, $obj, &$argname)); }
$gpuc proc VkResultToString {VkResult res} char* {
            switch (res) {
    #define CASE(x) case VK_##x: return #x;
            CASE(SUCCESS)                       CASE(NOT_READY)
            CASE(TIMEOUT)                       CASE(EVENT_SET)
            CASE(EVENT_RESET)                   CASE(INCOMPLETE)
            CASE(ERROR_OUT_OF_HOST_MEMORY)      CASE(ERROR_OUT_OF_DEVICE_MEMORY)
            CASE(ERROR_INITIALIZATION_FAILED)   CASE(ERROR_DEVICE_LOST)
            CASE(ERROR_MEMORY_MAP_FAILED)       CASE(ERROR_LAYER_NOT_PRESENT)
            CASE(ERROR_EXTENSION_NOT_PRESENT)   CASE(ERROR_FEATURE_NOT_PRESENT)
            CASE(ERROR_INCOMPATIBLE_DRIVER)     CASE(ERROR_TOO_MANY_OBJECTS)
            CASE(ERROR_FORMAT_NOT_SUPPORTED)    CASE(ERROR_FRAGMENTED_POOL)
            CASE(ERROR_UNKNOWN)                 CASE(ERROR_OUT_OF_POOL_MEMORY)
            CASE(ERROR_INVALID_EXTERNAL_HANDLE) CASE(ERROR_FRAGMENTATION)
            CASE(ERROR_INVALID_OPAQUE_CAPTURE_ADDRESS)
            CASE(PIPELINE_COMPILE_REQUIRED)      CASE(ERROR_SURFACE_LOST_KHR)
            CASE(ERROR_NATIVE_WINDOW_IN_USE_KHR) CASE(SUBOPTIMAL_KHR)
            CASE(ERROR_OUT_OF_DATE_KHR)          CASE(ERROR_INCOMPATIBLE_DISPLAY_KHR)
            CASE(ERROR_VALIDATION_FAILED_EXT)    CASE(ERROR_INVALID_SHADER_NV)
    #ifdef VK_ENABLE_BETA_EXTENSIONS
            CASE(ERROR_IMAGE_USAGE_NOT_SUPPORTED_KHR)
            CASE(ERROR_VIDEO_PICTURE_LAYOUT_NOT_SUPPORTED_KHR)
            CASE(ERROR_VIDEO_PROFILE_OPERATION_NOT_SUPPORTED_KHR)
            CASE(ERROR_VIDEO_PROFILE_FORMAT_NOT_SUPPORTED_KHR)
            CASE(ERROR_VIDEO_PROFILE_CODEC_NOT_SUPPORTED_KHR)
            CASE(ERROR_VIDEO_STD_VERSION_NOT_SUPPORTED_KHR)
    #endif
            CASE(ERROR_INVALID_DRM_FORMAT_MODIFIER_PLANE_LAYOUT_EXT)
            CASE(ERROR_NOT_PERMITTED_KHR)
            CASE(ERROR_FULL_SCREEN_EXCLUSIVE_MODE_LOST_EXT)
            CASE(THREAD_IDLE_KHR)        CASE(THREAD_DONE_KHR)
            CASE(OPERATION_DEFERRED_KHR) CASE(OPERATION_NOT_DEFERRED_KHR)
            default: return "unknown";
            }
    #undef CASE
}
local proc vktry {call} { string map {\n " "} [csubst {{
    VkResult res = $call;
    if (res != VK_SUCCESS) {
        fprintf(stderr, "Failed $call: %s (%d)\n",
                VkResultToString(res), res); exit(1);
    }
}}] }

$gpuc define {
    VkInstance instance;
    VkPhysicalDevice physicalDevice;
    VkPhysicalDeviceFeatures physicalDeviceFeatures;
    VkPhysicalDeviceProperties physicalDeviceProperties;

    VkDevice device;
    uint32_t computeQueueFamilyIndex;
    uint32_t graphicsQueueFamilyIndex = UINT32_MAX;

    pthread_mutex_t graphicsQueueMutex;
    VkQueue graphicsQueue;
    VkQueue presentQueue;
    VkQueue computeQueue;
}
defineVulkanHandleType $gpuc VkPhysicalDevice
defineVulkanHandleType $gpuc VkDevice
$gpuc proc getPhysicalDevice {} VkPhysicalDevice { return physicalDevice; }
$gpuc proc getDevice {} VkDevice { return device; }

$gpuc proc init {bool useGlfw} void {
    $[vktry volkInitialize()]
    $[if {$useGlfw} { expr {"glfwInit();"} }]

    // Set up VkInstance instance:
    {
        VkInstanceCreateInfo createInfo = {0};
        createInfo.sType = VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO;

        const char* validationLayers[] = {
           "VK_LAYER_KHRONOS_validation"
        };
        createInfo.enabledLayerCount = sizeof(validationLayers)/sizeof(validationLayers[0]);
        createInfo.ppEnabledLayerNames = validationLayers;

        $[if {$macos} {subst -nocommands {
            const char* enabledExtensions[] = {
                VK_KHR_PORTABILITY_ENUMERATION_EXTENSION_NAME,
                VK_KHR_SURFACE_EXTENSION_NAME,
                "VK_EXT_metal_surface",
                VK_KHR_GET_PHYSICAL_DEVICE_PROPERTIES_2_EXTENSION_NAME,
            };
            createInfo.enabledExtensionCount = sizeof(enabledExtensions)/sizeof(enabledExtensions[0]);

        }} elseif {$useGlfw} {subst -nocommands {
            const char** enabledExtensions = glfwGetRequiredInstanceExtensions(&createInfo.enabledExtensionCount);

        }} else {subst -nocommands {
            const char* enabledExtensions[] = {
                // 2 extensions for non-X11/Wayland display
                VK_KHR_SURFACE_EXTENSION_NAME,
                VK_KHR_DISPLAY_EXTENSION_NAME,
                VK_KHR_GET_PHYSICAL_DEVICE_PROPERTIES_2_EXTENSION_NAME,
            };
            createInfo.enabledExtensionCount = sizeof(enabledExtensions)/sizeof(enabledExtensions[0]);
        }}]
        createInfo.ppEnabledExtensionNames = enabledExtensions;

        $[if {$macos} { expr {{
            createInfo.flags = VK_INSTANCE_CREATE_ENUMERATE_PORTABILITY_BIT_KHR;
        }} }]
        VkResult res = vkCreateInstance(&createInfo, NULL, &instance);
        if (res != VK_SUCCESS) {
            FOLK_ERROR("Failed vkCreateInstance: %s (%d)\n",
                    VkResultToString(res), res);
            if (res == VK_ERROR_LAYER_NOT_PRESENT) {
                FOLK_ERROR("It looks like a required layer is missing.\n"
                           "Did you install `vulkan-validationlayers`?\n");
            }
        }
    }
    volkLoadInstance(instance);

    // Set up VkPhysicalDevice physicalDevice
    {
        uint32_t physicalDeviceCount = 0;
        vkEnumeratePhysicalDevices(instance, &physicalDeviceCount, NULL);
        if (physicalDeviceCount == 0) {
            FOLK_ERROR("Failed to find Vulkan physical device\n");
        }
        printf("gpu: Found %d Vulkan devices\n", physicalDeviceCount);
        VkPhysicalDevice physicalDevices[physicalDeviceCount];
        vkEnumeratePhysicalDevices(instance, &physicalDeviceCount, physicalDevices);

        physicalDevice = physicalDevices[0];
        vkGetPhysicalDeviceFeatures(physicalDevice, &physicalDeviceFeatures);
        vkGetPhysicalDeviceProperties(physicalDevice, &physicalDeviceProperties);
    }

    computeQueueFamilyIndex = UINT32_MAX; {
        uint32_t queueFamilyCount = 0;
        vkGetPhysicalDeviceQueueFamilyProperties(physicalDevice, &queueFamilyCount, NULL);
        VkQueueFamilyProperties queueFamilies[queueFamilyCount];
        vkGetPhysicalDeviceQueueFamilyProperties(physicalDevice, &queueFamilyCount, queueFamilies);
        for (int i = 0; i < queueFamilyCount; i++) {
            if (queueFamilies[i].queueFlags & VK_QUEUE_COMPUTE_BIT) {
                computeQueueFamilyIndex = i;
            }
            if (queueFamilies[i].queueFlags & VK_QUEUE_GRAPHICS_BIT) {
                graphicsQueueFamilyIndex = i;
                break;
            }
        }
        if (graphicsQueueFamilyIndex == UINT32_MAX) {
            fprintf(stderr, "Failed to find a Vulkan graphics queue family\n"); exit(1);
        }
        if (computeQueueFamilyIndex == UINT32_MAX) {
            fprintf(stderr, "Failed to find a Vulkan compute queue family\n"); exit(1);
        }
    }

    // Set up VkDevice device
    {
        VkDeviceQueueCreateInfo queueCreateInfo = {0};
        queueCreateInfo.sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO;
        queueCreateInfo.queueFamilyIndex = graphicsQueueFamilyIndex;
        queueCreateInfo.queueCount = 1;
        float queuePriority = 1.0f;
        queueCreateInfo.pQueuePriorities = &queuePriority;

        VkPhysicalDeviceFeatures deviceFeatures = {0};

        $[if {$macos} {subst -nocommands {
            const char *deviceExtensions[] = {
                VK_KHR_SWAPCHAIN_EXTENSION_NAME,
                "VK_KHR_portability_subset",
                VK_KHR_MAINTENANCE3_EXTENSION_NAME
            };
        }} elseif {$useGlfw} {subst -nocommands {
            const char *deviceExtensions[] = {
                VK_KHR_SWAPCHAIN_EXTENSION_NAME,
            };
        }} else {subst -nocommands {
            const char *deviceExtensions[] = {
                VK_KHR_SWAPCHAIN_EXTENSION_NAME,
                VK_KHR_MAINTENANCE3_EXTENSION_NAME
            };
        }}]

        VkDeviceCreateInfo createInfo = {0};
        createInfo.sType = VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO;
        createInfo.pQueueCreateInfos = &queueCreateInfo;
        createInfo.queueCreateInfoCount = 1;
        createInfo.pEnabledFeatures = &deviceFeatures;
        createInfo.enabledLayerCount = 0;
        createInfo.enabledExtensionCount = sizeof(deviceExtensions)/sizeof(deviceExtensions[0]);
        createInfo.ppEnabledExtensionNames = deviceExtensions;

        /* VkPhysicalDeviceDescriptorIndexingFeatures descriptorIndexingFeatures = {0}; */
        /* descriptorIndexingFeatures.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES; */
        /* descriptorIndexingFeatures.descriptorBindingPartiallyBound = VK_TRUE; */
        /* // TODO: Do we need more descriptor indexing features? */
        /* createInfo.pNext = &descriptorIndexingFeatures; */

        $[vktry {vkCreateDevice(physicalDevice, &createInfo, NULL, &device)}]
    }

    // Set up VkQueue graphicsQueue and VkQueue presentQueue and VkQueue computeQueue
    {
        pthread_mutex_init(&graphicsQueueMutex, NULL);
        vkGetDeviceQueue(device, graphicsQueueFamilyIndex, 0, &graphicsQueue);
        presentQueue = graphicsQueue;
        computeQueue = graphicsQueue;
    }
}
$gpuc proc getDoesSupportShaderSampledImageArrayDynamicIndexing {} bool {
    return physicalDeviceFeatures.shaderSampledImageArrayDynamicIndexing;
}
$gpuc proc getMaxTextures {} int {
    int maxTextures = physicalDeviceProperties.limits.maxPerStageDescriptorSamplers;
    // COMBINED_IMAGE_SAMPLER counts against both sampler and sampled image limits.
    int maxSampledImages = physicalDeviceProperties.limits.maxPerStageDescriptorSampledImages;
    if (maxSampledImages < maxTextures) { maxTextures = maxSampledImages; }
    // If it's actually a low cap (often 16), stick with that.
    if (maxTextures < 128) { return maxTextures; }
    // HACK: for now, I don't want to deal with all the other descriptor caps etc.
    return 128;
}

set gpuLib [$gpuc compile]

$gpuLib init $useGlfw
Claim the GPU library is $gpuLib

Claim the GPU physical device is [$gpuLib getPhysicalDevice]
Claim the GPU device is [$gpuLib getDevice]

}