[Flint] adding binary semaphores

README.md

@@ -521,7 +521,7 @@ vkCreatePipelineLayout                           | ⚙️ WIP
 vkCreateQueryPool                                | ⚙️ WIP
 vkCreateRenderPass                               | ⚙️ WIP
 vkCreateSampler                                  | ⚙️ WIP
-vkCreateSemaphore                                | ⚙️ WIP
+vkCreateSemaphore                                | ✅ Implemented
 vkCreateShaderModule                             | ⚙️ WIP
 vkCreateSwapchainKHR                             | ⚙️ WIP
 vkCreateWaylandSurfaceKHR                        | ⚙️ WIP
@@ -546,7 +546,7 @@ vkDestroyPipelineLayout                          | ⚙️ WIP
 vkDestroyQueryPool                               | ⚙️ WIP
 vkDestroyRenderPass                              | ⚙️ WIP
 vkDestroySampler                                 | ⚙️ WIP
-vkDestroySemaphore                               | ⚙️ WIP
+vkDestroySemaphore                               | ✅ Implemented
 vkDestroyShaderModule                            | ⚙️ WIP
 vkDestroySurfaceKHR                              | ⚙️ WIP
 vkDestroySwapchainKHR                            | ⚙️ WIP

src/intel/FlintBinarySemaphore.zig

@@ -1,20 +1,71 @@
 const std = @import("std");
 const vk = @import("vulkan");
 const base = @import("base");
+const kmd = @import("kmd.zig");
 
 const VkError = base.VkError;
 const Device = base.Device;
+const FlintDevice = @import("FlintDevice.zig");
+
+const drm_syncobj_create = 0xbf;
+const drm_syncobj_destroy = 0xc0;
+const drm_syncobj_wait = 0xc3;
+const drm_syncobj_reset = 0xc4;
+const drm_syncobj_signal = 0xc5;
+
+const syncobj_wait_all: u32 = 1 << 0;
+const syncobj_wait_for_submit: u32 = 1 << 1;
+
+const SyncObjCreate = extern struct {
+    handle: u32,
+    flags: u32,
+};
+
+const SyncObjDestroy = extern struct {
+    handle: u32,
+    pad: u32,
+};
+
+const SyncObjWait = extern struct {
+    handles: u64,
+    timeout_nsec: i64,
+    count_handles: u32,
+    flags: u32,
+    first_signaled: u32,
+    pad: u32,
+    deadline_nsec: u64,
+};
+
+const SyncObjArray = extern struct {
+    handles: u64,
+    count_handles: u32,
+    pad: u32,
+};
 
 const Self = @This();
 pub const Interface = base.BinarySemaphore;
 
 interface: Interface,
+handle: u32,
 
 pub fn create(device: *base.Device, allocator: std.mem.Allocator, info: *const vk.SemaphoreCreateInfo) VkError!*Self {
     const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
     errdefer allocator.destroy(self);
 
     var interface = try Interface.init(device, allocator, info);
+    const flint_device: *FlintDevice = @alignCast(@fieldParentPtr("interface", device));
+    var create_info = SyncObjCreate{
+        .handle = 0,
+        .flags = 0,
+    };
+
+    base.utils.ioctl(
+        try flint_device.kmd.file(),
+        device.io(),
+        kmd.drmIoctlIowr(drm_syncobj_create, SyncObjCreate),
+        &create_info,
+    ) catch return VkError.DeviceLost;
+    errdefer destroyHandle(flint_device, device.io(), create_info.handle);
 
     interface.vtable = &.{
         .destroy = destroy,
@@ -24,21 +75,89 @@ pub fn create(device: *base.Device, allocator: std.mem.Allocator, info: *const v
 
     self.* = .{
         .interface = interface,
+        .handle = create_info.handle,
     };
     return self;
 }
 
 pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) void {
     const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    const device: *FlintDevice = @alignCast(@fieldParentPtr("interface", interface.owner));
+    destroyHandle(device, interface.owner.io(), self.handle);
     allocator.destroy(self);
 }
 
 pub fn signal(interface: *Interface) VkError!void {
     const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
-    _ = self;
+    const device: *FlintDevice = @alignCast(@fieldParentPtr("interface", interface.owner));
+    var handles = [_]u32{self.handle};
+    var signal_info = SyncObjArray{
+        .handles = @intFromPtr(&handles),
+        .count_handles = handles.len,
+        .pad = 0,
+    };
+    base.utils.ioctl(
+        try device.kmd.file(),
+        interface.owner.io(),
+        kmd.drmIoctlIowr(drm_syncobj_signal, SyncObjArray),
+        &signal_info,
+    ) catch return VkError.DeviceLost;
 }
 
 pub fn wait(interface: *Interface) VkError!void {
     const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
-    _ = self;
+    const device: *FlintDevice = @alignCast(@fieldParentPtr("interface", interface.owner));
+    var handles = [_]u32{self.handle};
+    var wait_info = SyncObjWait{
+        .handles = @intFromPtr(&handles),
+        .timeout_nsec = std.math.maxInt(i64),
+        .count_handles = handles.len,
+        .flags = syncobj_wait_all | syncobj_wait_for_submit,
+        .first_signaled = 0,
+        .pad = 0,
+        .deadline_nsec = 0,
+    };
+
+    const errno = base.utils.ioctlErrno(
+        try device.kmd.file(),
+        interface.owner.io(),
+        kmd.drmIoctlIowr(drm_syncobj_wait, SyncObjWait),
+        &wait_info,
+    ) catch return VkError.DeviceLost;
+
+    switch (errno) {
+        .SUCCESS => {},
+        else => return VkError.DeviceLost,
+    }
+    try reset(interface);
+}
+
+pub fn reset(interface: *Interface) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    const device: *FlintDevice = @alignCast(@fieldParentPtr("interface", interface.owner));
+    var handles = [_]u32{self.handle};
+    var reset_info = SyncObjArray{
+        .handles = @intFromPtr(&handles),
+        .count_handles = handles.len,
+        .pad = 0,
+    };
+    base.utils.ioctl(
+        try device.kmd.file(),
+        interface.owner.io(),
+        kmd.drmIoctlIowr(drm_syncobj_reset, SyncObjArray),
+        &reset_info,
+    ) catch return VkError.DeviceLost;
+}
+
+fn destroyHandle(device: *FlintDevice, io: std.Io, handle: u32) void {
+    var destroy_info = SyncObjDestroy{
+        .handle = handle,
+        .pad = 0,
+    };
+    base.utils.ioctl(
+        device.kmd.file() catch return,
+        io,
+        kmd.drmIoctlIowr(drm_syncobj_destroy, SyncObjDestroy),
+        &destroy_info,
+    ) catch {};
 }

src/intel/FlintQueue.zig

@@ -2,6 +2,7 @@ const std = @import("std");
 const vk = @import("vulkan");
 const base = @import("base");
 
+const FlintBinarySemaphore = @import("FlintBinarySemaphore.zig");
 const FlintCommandBuffer = @import("FlintCommandBuffer.zig");
 const FlintDevice = @import("FlintDevice.zig");
 const FlintFence = @import("FlintFence.zig");
@@ -55,51 +56,67 @@ pub fn bindSparse(interface: *Interface, info: []const vk.BindSparseInfo, fence:
 pub fn submit(interface: *Interface, infos: []Interface.SubmitInfo, fence: ?*base.Fence) VkError!void {
     const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
     const device: *FlintDevice = @alignCast(@fieldParentPtr("interface", interface.owner));
-
-    var remaining_batches: usize = 0;
-    for (infos) |info| {
-        for (info.command_buffers.items) |command_buffer| {
-            const intel_command_buffer: *FlintCommandBuffer = @alignCast(@fieldParentPtr("interface", command_buffer));
-            if (intel_command_buffer.batch.items.len != 0) remaining_batches += 1;
-        }
-    }
-    const batch_count = remaining_batches;
+    const allocator = interface.host_allocator.allocator();
 
     try self.completion.interface.reset();
 
-    for (infos) |info| {
-        for (info.wait_semaphores.items) |semaphore| {
-            try semaphore.wait();
-        }
+    for (infos, 0..) |info, info_index| {
+        const last_info = info_index + 1 == infos.len;
+        const request_count = @max(info.command_buffers.items.len, 1);
 
-        for (info.command_buffers.items) |command_buffer| {
-            const intel_command_buffer: *FlintCommandBuffer = @alignCast(@fieldParentPtr("interface", command_buffer));
-            if (intel_command_buffer.batch.items.len == 0) {
-                try intel_command_buffer.submitGpuBatch(&.{});
-                continue;
-            }
+        for (0..request_count) |request_index| {
+            const first_request = request_index == 0;
+            const last_request = request_index + 1 == request_count;
 
-            remaining_batches -= 1;
-            var syncs: [2]kmd.SyncDependency = undefined;
-            var sync_count: usize = 0;
-            if (remaining_batches == 0) {
-                syncs[sync_count] = .{ .handle = self.completion.handle, .signal = true };
-                sync_count += 1;
-                if (fence) |base_fence| {
-                    const flint_fence: *FlintFence = @alignCast(@fieldParentPtr("interface", base_fence));
-                    syncs[sync_count] = .{ .handle = flint_fence.handle, .signal = true };
-                    sync_count += 1;
+            var syncs = std.ArrayList(kmd.SyncDependency).empty;
+            defer syncs.deinit(allocator);
+
+            if (first_request) {
+                for (info.wait_semaphores.items) |base_semaphore| {
+                    const semaphore: *FlintBinarySemaphore = @alignCast(@fieldParentPtr("interface", base_semaphore));
+                    syncs.append(allocator, .{ .handle = semaphore.handle, .wait = true }) catch return VkError.OutOfHostMemory;
                 }
             }
-            try intel_command_buffer.submitGpuBatch(syncs[0..sync_count]);
-        }
 
-        for (info.signal_semaphores.items) |semaphore| {
-            try semaphore.signal();
+            if (last_request) {
+                for (info.signal_semaphores.items) |base_semaphore| {
+                    const semaphore: *FlintBinarySemaphore = @alignCast(@fieldParentPtr("interface", base_semaphore));
+                    syncs.append(allocator, .{ .handle = semaphore.handle, .signal = true }) catch return VkError.OutOfHostMemory;
+                }
+
+                if (last_info) {
+                    syncs.append(allocator, .{ .handle = self.completion.handle, .signal = true }) catch return VkError.OutOfHostMemory;
+                    if (fence) |base_fence| {
+                        const flint_fence: *FlintFence = @alignCast(@fieldParentPtr("interface", base_fence));
+                        syncs.append(allocator, .{ .handle = flint_fence.handle, .signal = true }) catch return VkError.OutOfHostMemory;
+                    }
+                }
+            }
+
+            if (info.command_buffers.items.len == 0) {
+                try device.kmd.submitBatch(
+                    interface.owner.io(),
+                    allocator,
+                    &.{},
+                    &.{},
+                    syncs.items,
+                );
+            } else {
+                const command_buffer = info.command_buffers.items[request_index];
+                const intel_command_buffer: *FlintCommandBuffer = @alignCast(@fieldParentPtr("interface", command_buffer));
+                try intel_command_buffer.submitGpuBatch(syncs.items);
+            }
+
+            if (first_request) {
+                for (info.wait_semaphores.items) |base_semaphore| {
+                    const semaphore: *FlintBinarySemaphore = @alignCast(@fieldParentPtr("interface", base_semaphore));
+                    try FlintBinarySemaphore.reset(&semaphore.interface);
+                }
+            }
         }
     }
 
-    if (batch_count == 0) {
+    if (infos.len == 0) {
         var syncs: [2]kmd.SyncDependency = undefined;
         var sync_count: usize = 1;
         syncs[0] = .{ .handle = self.completion.handle, .signal = true };
@@ -109,11 +126,9 @@ pub fn submit(interface: *Interface, infos: []Interface.SubmitInfo, fence: ?*bas
             sync_count += 1;
         }
 
-        // A real no-op request preserves queue order: its output fence cannot
-        // signal ahead of an earlier request that is still running.
         try device.kmd.submitBatch(
             interface.owner.io(),
-            interface.host_allocator.allocator(),
+            allocator,
             &.{},
             &.{},
             syncs[0..sync_count],