diff --git a/server/compositor/compositor.cpp b/server/compositor/compositor.cpp index 587da82e..f23b1ded 100644 --- a/server/compositor/compositor.cpp +++ b/server/compositor/compositor.cpp @@ -434,7 +434,7 @@ xrt_result_t compositor::layer_commit(xrt_graphics_sync_handle_t sync_handle) { if (encoder->stream_idx == 2 and not view_info.alpha) continue; - else if (encoder->need_transfer or encoder->target_queue == vk.queue_family_index) + else if (encoder->need_transfer or encoder->target_queue == vk.queue.family_index) { image_barriers.push_back( vk::ImageMemoryBarrier2{ @@ -452,7 +452,7 @@ xrt_result_t compositor::layer_commit(xrt_graphics_sync_handle_t sync_handle) // transition the encoder needs. .oldLayout = vk::ImageLayout::eGeneral, .newLayout = encoder->target_layout, - .srcQueueFamilyIndex = vk.queue_family_index, + .srcQueueFamilyIndex = vk.queue.family_index, .dstQueueFamilyIndex = encoder->target_queue, .image = images[i].image, .subresourceRange = {.aspectMask = vk::ImageAspectFlagBits::eColor, @@ -482,8 +482,8 @@ xrt_result_t compositor::layer_commit(xrt_graphics_sync_handle_t sync_handle) vk::CommandBufferSubmitInfo cmd_info{ .commandBuffer = cmd, }; - std::unique_lock lock{vk.queue_mutex}; - vk.queue.submit2(vk::SubmitInfo2{ + std::unique_lock lock{vk.queue.mutex}; + vk.queue.queue.submit2(vk::SubmitInfo2{ .commandBufferInfoCount = 1, .pCommandBufferInfos = &cmd_info, .signalSemaphoreInfoCount = 1, @@ -682,7 +682,7 @@ compositor::compositor(wivrn_session & session) : session(session), cmd_pool(vk.device, vk::CommandPoolCreateInfo{ .flags = vk::CommandPoolCreateFlagBits::eTransient, - .queueFamilyIndex = vk.queue_family_index, + .queueFamilyIndex = vk.queue.family_index, }), query_pool(vk.device, vk::QueryPoolCreateInfo{ .queryType = vk::QueryType::eTimestamp, @@ -706,7 +706,7 @@ compositor::compositor(wivrn_session & session) : *vk.instance, *vk.physical_device, *vk.device, - vk.queue_family_index, + vk.queue.family_index, 0, vk.has_device_ext(VK_KHR_EXTERNAL_FENCE_FD_EXTENSION_NAME), vk.has_device_ext(VK_KHR_EXTERNAL_SEMAPHORE_FD_EXTENSION_NAME), @@ -720,9 +720,7 @@ compositor::compositor(wivrn_session & session) : c_base->vk.graphics_queue = nullptr; if (c_base->vk.main_queue) - c_base->vk.main_queue->mutex = copy_mutex(vk.queue_mutex); - if (c_base->vk.encode_queue) - c_base->vk.encode_queue->mutex = copy_mutex(vk.encode_queue_mutex); + c_base->vk.main_queue->mutex = copy_mutex(vk.queue.mutex); { comp_vulkan_formats formats{}; diff --git a/server/encoder/encoder_settings.cpp b/server/encoder/encoder_settings.cpp index a487a148..cb02470c 100644 --- a/server/encoder/encoder_settings.cpp +++ b/server/encoder/encoder_settings.cpp @@ -181,7 +181,7 @@ class prober bool has_vk(video_codec codec) { - if (*vk.encode_queue == VK_NULL_HANDLE) + if (vk.encode_queues.empty()) return false; if (not std::get(vk.feat).videoMaintenance1) { @@ -190,7 +190,7 @@ class prober } auto prop = vk.physical_device.getQueueFamilyProperties2>(); assert(vk.encode_queue_family_index < prop.size()); - const auto flags = prop.at(vk.encode_queue_family_index).get().videoCodecOperations; + const auto flags = prop.at(vk.encode_queues[0].family_index).get().videoCodecOperations; switch (codec) { case h264: { diff --git a/server/encoder/ffmpeg/video_encoder_ffmpeg.h b/server/encoder/ffmpeg/video_encoder_ffmpeg.h index 5f9f7dfd..7bcbdeef 100644 --- a/server/encoder/ffmpeg/video_encoder_ffmpeg.h +++ b/server/encoder/ffmpeg/video_encoder_ffmpeg.h @@ -39,7 +39,7 @@ public: protected: video_encoder_ffmpeg(vk_bundle & vk, uint8_t stream_idx, const encoder_settings & settings) : - wivrn::video_encoder(vk, stream_idx, vk.queue_family_index, settings, std::make_unique(), true) {} + wivrn::video_encoder(vk, stream_idx, vk.queue.family_index, settings, std::make_unique(), true) {} virtual void push_frame(bool idr, uint8_t slot) = 0; diff --git a/server/encoder/ffmpeg/video_encoder_va.cpp b/server/encoder/ffmpeg/video_encoder_va.cpp index f31ba573..cc860af3 100644 --- a/server/encoder/ffmpeg/video_encoder_va.cpp +++ b/server/encoder/ffmpeg/video_encoder_va.cpp @@ -159,7 +159,7 @@ vk::raii::CommandPool make_cmd_pool(wivrn::vk_bundle & vk, uint8_t stream_idx) auto res = vk.device.createCommandPool(vk::CommandPoolCreateInfo{ .flags = vk::CommandPoolCreateFlagBits::eResetCommandBuffer | vk::CommandPoolCreateFlagBits::eTransient, - .queueFamilyIndex = vk.queue_family_index, + .queueFamilyIndex = vk.queue.family_index, }); vk.name(res, std::format("vaapi encoder {} command pool", stream_idx)); return res; @@ -536,17 +536,17 @@ void video_encoder_va::present_image(vk::Image y_cbcr, vk::SemaphoreSubmitInfo, cmd.end(); - std::unique_lock lock(vk.queue_mutex); + std::unique_lock lock(vk.queue.mutex); vk::CommandBufferSubmitInfo cmd_info{ .commandBuffer = *cmd, }; vk.device.resetFences(*in[slot].fence); - vk.queue.submit2(vk::SubmitInfo2{ - .commandBufferInfoCount = 1, - .pCommandBufferInfos = &cmd_info, - }, - *in[slot].fence); + vk.queue.queue.submit2(vk::SubmitInfo2{ + .commandBufferInfoCount = 1, + .pCommandBufferInfos = &cmd_info, + }, + *in[slot].fence); } void video_encoder_va::push_frame(bool idr, uint8_t slot) diff --git a/server/encoder/video_encoder.cpp b/server/encoder/video_encoder.cpp index 97696cf8..fc5904b6 100644 --- a/server/encoder/video_encoder.cpp +++ b/server/encoder/video_encoder.cpp @@ -193,7 +193,7 @@ video_encoder::video_encoder(vk_bundle & vk, bool async_send) : stream_idx(stream_idx), target_queue(target_queue), - need_transfer(not vk.optimal_transfer(vk.queue_family_index, target_queue)), + need_transfer(not vk.optimal_transfer(vk.queue.family_index, target_queue)), bitrate_multiplier(settings.bitrate_multiplier), shared_sender(async_send ? sender::get() : nullptr), idr(std::move(idr)), diff --git a/server/encoder/video_encoder_nvenc.cpp b/server/encoder/video_encoder_nvenc.cpp index 04600fc3..1ff1830d 100644 --- a/server/encoder/video_encoder_nvenc.cpp +++ b/server/encoder/video_encoder_nvenc.cpp @@ -153,7 +153,7 @@ static vk::raii::CommandPool make_cmd_pool(wivrn::vk_bundle & vk, uint8_t stream auto res = vk.device.createCommandPool(vk::CommandPoolCreateInfo{ .flags = vk::CommandPoolCreateFlagBits::eResetCommandBuffer | vk::CommandPoolCreateFlagBits::eTransient, - .queueFamilyIndex = vk.queue_family_index, + .queueFamilyIndex = vk.queue.family_index, }); vk.name(res, std::format("nvenc encoder {} command pool", stream_idx)); return res; @@ -163,7 +163,7 @@ video_encoder_nvenc::video_encoder_nvenc( wivrn::vk_bundle & vk, const encoder_settings & settings, uint8_t stream_idx) : - video_encoder(vk, stream_idx, vk.queue_family_index, settings, std::make_unique(), true), + video_encoder(vk, stream_idx, vk.queue.family_index, settings, std::make_unique(), true), vk(vk), cmd_pool{make_cmd_pool(vk, stream_idx)}, shared_state(video_encoder_nvenc_shared_state::get()), @@ -461,17 +461,17 @@ void video_encoder_nvenc::present_image(vk::Image y_cbcr, vk::SemaphoreSubmitInf }}); cmd.end(); - std::unique_lock lock(vk.queue_mutex); + std::unique_lock lock(vk.queue.mutex); vk::CommandBufferSubmitInfo cmd_info{ .commandBuffer = *cmd, }; vk.device.resetFences(*in[slot].fence); - vk.queue.submit2(vk::SubmitInfo2{ - .commandBufferInfoCount = 1, - .pCommandBufferInfos = &cmd_info, - }, - *in[slot].fence); + vk.queue.queue.submit2(vk::SubmitInfo2{ + .commandBufferInfoCount = 1, + .pCommandBufferInfos = &cmd_info, + }, + *in[slot].fence); } std::optional video_encoder_nvenc::encode(uint8_t slot, uint64_t frame_index) diff --git a/server/encoder/video_encoder_raw.cpp b/server/encoder/video_encoder_raw.cpp index 563c84b3..6e5ec78d 100644 --- a/server/encoder/video_encoder_raw.cpp +++ b/server/encoder/video_encoder_raw.cpp @@ -39,7 +39,7 @@ vk::raii::CommandPool make_cmd_pool(wivrn::vk_bundle & vk, uint8_t stream_idx) auto res = vk.device.createCommandPool(vk::CommandPoolCreateInfo{ .flags = vk::CommandPoolCreateFlagBits::eResetCommandBuffer | vk::CommandPoolCreateFlagBits::eTransient, - .queueFamilyIndex = *vk.transfer_queue ? vk.transfer_queue_family_index : vk.queue_family_index, + .queueFamilyIndex = vk.transfer_queue ? vk.transfer_queue.family_index : vk.queue.family_index, }); vk.name(res, std::format("raw encoder {} command pool", stream_idx)); return res; @@ -52,7 +52,7 @@ wivrn::video_encoder_raw::video_encoder_raw( uint8_t stream_idx) : video_encoder(vk, stream_idx, - *vk.transfer_queue ? vk.transfer_queue_family_index : vk.queue_family_index, + vk.transfer_queue ? vk.transfer_queue.family_index : vk.queue.family_index, settings, std::make_unique(), true), @@ -105,7 +105,7 @@ void wivrn::video_encoder_raw::present_image(vk::Image y_cbcr, vk::SemaphoreSubm vk::ImageMemoryBarrier2 barrier{ .dstStageMask = vk::PipelineStageFlagBits2KHR::eTransfer, .dstAccessMask = vk::AccessFlagBits2::eTransferRead, - .srcQueueFamilyIndex = vk.queue_family_index, + .srcQueueFamilyIndex = vk.queue.family_index, .dstQueueFamilyIndex = target_queue, .image = y_cbcr, .subresourceRange = {.aspectMask = vk::ImageAspectFlagBits::eColor, @@ -156,21 +156,21 @@ void wivrn::video_encoder_raw::present_image(vk::Image y_cbcr, vk::SemaphoreSubm cmd.end(); - std::unique_lock lock(*vk.transfer_queue ? vk.transfer_queue_mutex : vk.queue_mutex); + std::unique_lock lock(vk.transfer_queue ? vk.transfer_queue.mutex : vk.queue.mutex); vk::CommandBufferSubmitInfo cmd_info{ .commandBuffer = *cmd, }; compositor_sem.stageMask = vk::PipelineStageFlagBits2::eTransfer; vk.device.resetFences(*in[slot].fence); - (*vk.transfer_queue ? vk.transfer_queue : vk.queue) - .submit2(vk::SubmitInfo2{ - .waitSemaphoreInfoCount = 1, - .pWaitSemaphoreInfos = &compositor_sem, - .commandBufferInfoCount = 1, - .pCommandBufferInfos = &cmd_info, - }, - *in[slot].fence); + (vk.transfer_queue ? vk.transfer_queue : vk.queue) + .queue.submit2(vk::SubmitInfo2{ + .waitSemaphoreInfoCount = 1, + .pWaitSemaphoreInfos = &compositor_sem, + .commandBufferInfoCount = 1, + .pCommandBufferInfos = &cmd_info, + }, + *in[slot].fence); } std::optional wivrn::video_encoder_raw::encode(uint8_t slot, uint64_t frame_id) diff --git a/server/encoder/video_encoder_vulkan.cpp b/server/encoder/video_encoder_vulkan.cpp index 0f3a5ab0..ba2ef244 100644 --- a/server/encoder/video_encoder_vulkan.cpp +++ b/server/encoder/video_encoder_vulkan.cpp @@ -164,8 +164,9 @@ wivrn::video_encoder_vulkan::video_encoder_vulkan( const vk::VideoEncodeCapabilitiesKHR & in_encode_caps, uint8_t stream_idx, const encoder_settings & settings) : - video_encoder(vk, stream_idx, vk.encode_queue_family_index, settings, std::make_unique(), true), + video_encoder(vk, stream_idx, vk.encode_queues[stream_idx % vk.encode_queues.size()].family_index, settings, std::make_unique(), true), vk(vk), + encode_queue(vk.encode_queues[stream_idx % vk.encode_queues.size()]), encode_caps(patch_capabilities(in_encode_caps)), num_dpb_slots(std::min(video_caps.maxDpbSlots, 16u)) { @@ -270,7 +271,7 @@ void wivrn::video_encoder_vulkan::init(const vk::VideoCapabilitiesKHR & video_ca // NVIDIA has this flag while AMD has mutable flag on images, we don't have to // care for transfer on main queue - if (not(vk.physical_device.getQueueFamilyProperties().at(vk.encode_queue_family_index).queueFlags & vk::QueueFlagBits::eTransfer)) + if (not(vk.physical_device.getQueueFamilyProperties().at(encode_queue.family_index).queueFlags & vk::QueueFlagBits::eTransfer)) throw std::runtime_error("Vulkan encode queue is missing transfer capability"); tmp_image = image_allocation( @@ -428,7 +429,7 @@ void wivrn::video_encoder_vulkan::init(const vk::VideoCapabilitiesKHR & video_ca video_session = vk.device.createVideoSessionKHR(vk::VideoSessionCreateInfoKHR{ .pNext = video_session_create_next, - .queueFamilyIndex = vk.encode_queue_family_index, + .queueFamilyIndex = encode_queue.family_index, //.flags = vk::VideoSessionCreateFlagBitsKHR::eAllowEncodeParameterOptimizations, .pVideoProfile = &video_profile, .pictureFormat = picture_format.format, @@ -569,7 +570,7 @@ void wivrn::video_encoder_vulkan::init(const vk::VideoCapabilitiesKHR & video_ca video_command_pool = vk.device.createCommandPool( vk::CommandPoolCreateInfo{ .flags = vk::CommandPoolCreateFlagBits::eResetCommandBuffer | vk::CommandPoolCreateFlagBits::eTransient, - .queueFamilyIndex = vk.encode_queue_family_index, + .queueFamilyIndex = encode_queue.family_index, }); vk.name(video_command_pool, std::format("vulkan encoder {} video command pool", stream_idx)); @@ -585,12 +586,12 @@ void wivrn::video_encoder_vulkan::init(const vk::VideoCapabilitiesKHR & video_ca if (slot_data[0].host_buffer) { - if (*vk.transfer_queue) + if (vk.transfer_queue) { transfer_command_pool = vk.device.createCommandPool( vk::CommandPoolCreateInfo{ .flags = vk::CommandPoolCreateFlagBits::eResetCommandBuffer | vk::CommandPoolCreateFlagBits::eTransient, - .queueFamilyIndex = vk.transfer_queue_family_index, + .queueFamilyIndex = vk.transfer_queue.family_index, }); vk.name(transfer_command_pool, std::format("vulkan encoder {} transfer command pool", stream_idx)); @@ -605,7 +606,7 @@ void wivrn::video_encoder_vulkan::init(const vk::VideoCapabilitiesKHR & video_ca } else { - auto properties = vk.physical_device.getQueueFamilyProperties().at(vk.encode_queue_family_index); + auto properties = vk.physical_device.getQueueFamilyProperties().at(encode_queue.family_index); if (not(properties.queueFlags & vk::QueueFlagBits::eTransfer)) throw std::runtime_error("Incorrect vulkan queues for video encode support"); } @@ -656,7 +657,7 @@ void wivrn::video_encoder_vulkan::present_image(vk::Image y_cbcr, vk::SemaphoreS .dstAccessMask = vk::AccessFlagBits2::eTransferRead, .oldLayout = vk::ImageLayout::eGeneral, .newLayout = target_layout, - .srcQueueFamilyIndex = vk.queue_family_index, + .srcQueueFamilyIndex = vk.queue.family_index, .dstQueueFamilyIndex = target_queue, .image = y_cbcr, .subresourceRange = {.aspectMask = vk::ImageAspectFlagBits::eColor, @@ -759,7 +760,7 @@ void wivrn::video_encoder_vulkan::present_image(vk::Image y_cbcr, vk::SemaphoreS .dstAccessMask = vk::AccessFlagBits2::eVideoEncodeReadKHR, .oldLayout = vk::ImageLayout::eGeneral, .newLayout = target_layout, - .srcQueueFamilyIndex = need_transfer ? vk.queue_family_index : vk::QueueFamilyIgnored, + .srcQueueFamilyIndex = need_transfer ? vk.queue.family_index : vk::QueueFamilyIgnored, .dstQueueFamilyIndex = need_transfer ? target_queue : vk::QueueFamilyIgnored, .image = y_cbcr, .subresourceRange = {.aspectMask = vk::ImageAspectFlagBits::eColor, @@ -929,8 +930,8 @@ void wivrn::video_encoder_vulkan::present_image(vk::Image y_cbcr, vk::SemaphoreS .srcAccessMask = vk::AccessFlagBits2::eMemoryWrite, .dstStageMask = vk::PipelineStageFlagBits2::eTransfer, .dstAccessMask = vk::AccessFlagBits2::eMemoryRead, - .srcQueueFamilyIndex = vk.encode_queue_family_index, - .dstQueueFamilyIndex = vk.transfer_queue_family_index, + .srcQueueFamilyIndex = encode_queue.family_index, + .dstQueueFamilyIndex = vk.transfer_queue.family_index, .buffer = slot_item.output_buffer, .size = slot_item.copy_size, }; @@ -988,8 +989,8 @@ void wivrn::video_encoder_vulkan::present_image(vk::Image y_cbcr, vk::SemaphoreS .stageMask = vk::PipelineStageFlagBits2::eVideoEncodeKHR, }; - std::unique_lock lock(vk.encode_queue_mutex); - vk.encode_queue.submit2( + std::unique_lock lock(encode_queue.mutex); + encode_queue.queue.submit2( vk::SubmitInfo2{ .waitSemaphoreInfoCount = 1, .pWaitSemaphoreInfos = &compositor_sem, @@ -1013,14 +1014,14 @@ void wivrn::video_encoder_vulkan::present_image(vk::Image y_cbcr, vk::SemaphoreS .stageMask = vk::PipelineStageFlagBits2::eTransfer, }; - std::unique_lock lock(vk.transfer_queue_mutex); - vk.transfer_queue.submit2(vk::SubmitInfo2{ - .waitSemaphoreInfoCount = 1, - .pWaitSemaphoreInfos = &wait_info, - .commandBufferInfoCount = 1, - .pCommandBufferInfos = &cmd_info, - }, - *slot_item.fence); + std::unique_lock lock(vk.transfer_queue.mutex); + vk.transfer_queue.queue.submit2(vk::SubmitInfo2{ + .waitSemaphoreInfoCount = 1, + .pWaitSemaphoreInfos = &wait_info, + .commandBufferInfoCount = 1, + .pCommandBufferInfos = &cmd_info, + }, + *slot_item.fence); } } @@ -1065,8 +1066,8 @@ std::optional wivrn::video_encoder_vulkan::encode(ui cmd.end(); { - std::unique_lock lock(transfer_queue ? vk.transfer_queue_mutex : vk.encode_queue_mutex); - auto & queue = transfer_queue ? vk.transfer_queue : vk.encode_queue; + std::unique_lock lock(transfer_queue ? vk.transfer_queue.mutex : encode_queue.mutex); + auto & queue = (transfer_queue ? vk.transfer_queue : encode_queue).queue; vk::CommandBufferSubmitInfo cmd_info{ .commandBuffer = *cmd, diff --git a/server/encoder/video_encoder_vulkan.h b/server/encoder/video_encoder_vulkan.h index 2c36648d..e93d0642 100644 --- a/server/encoder/video_encoder_vulkan.h +++ b/server/encoder/video_encoder_vulkan.h @@ -22,6 +22,7 @@ #include #include +#include "utils/wivrn_vk_bundle.h" #include "video_encoder.h" #include "vk/allocation.h" @@ -30,6 +31,7 @@ namespace wivrn class video_encoder_vulkan : public video_encoder { wivrn::vk_bundle & vk; + wivrn::vk_bundle::queue_data & encode_queue; const vk::VideoEncodeCapabilitiesKHR encode_caps; #ifdef VK_KHR_video_encode_intra_refresh vk::VideoEncodeIntraRefreshInfoKHR intra_info{}; diff --git a/server/encoder/video_encoder_x264.cpp b/server/encoder/video_encoder_x264.cpp index afc154d5..b4be1e17 100644 --- a/server/encoder/video_encoder_x264.cpp +++ b/server/encoder/video_encoder_x264.cpp @@ -93,7 +93,7 @@ vk::raii::CommandPool make_cmd_pool(wivrn::vk_bundle & vk, uint8_t stream_idx) auto res = vk.device.createCommandPool(vk::CommandPoolCreateInfo{ .flags = vk::CommandPoolCreateFlagBits::eResetCommandBuffer | vk::CommandPoolCreateFlagBits::eTransient, - .queueFamilyIndex = *vk.transfer_queue ? vk.transfer_queue_family_index : vk.queue_family_index, + .queueFamilyIndex = vk.transfer_queue ? vk.transfer_queue.family_index : vk.queue.family_index, }); vk.name(res, std::format("x264 encoder {} command pool", stream_idx)); return res; @@ -106,7 +106,7 @@ video_encoder_x264::video_encoder_x264( uint8_t stream_idx) : video_encoder(vk, stream_idx, - *vk.transfer_queue ? vk.transfer_queue_family_index : vk.queue_family_index, + vk.transfer_queue ? vk.transfer_queue.family_index : vk.queue.family_index, settings, std::make_unique(), false), @@ -224,7 +224,7 @@ void video_encoder_x264::present_image(vk::Image y_cbcr, vk::SemaphoreSubmitInfo vk::ImageMemoryBarrier2 barrier{ .dstStageMask = vk::PipelineStageFlagBits2KHR::eTransfer, .dstAccessMask = vk::AccessFlagBits2::eTransferRead, - .srcQueueFamilyIndex = vk.queue_family_index, + .srcQueueFamilyIndex = vk.queue.family_index, .dstQueueFamilyIndex = target_queue, .image = y_cbcr, .subresourceRange = {.aspectMask = vk::ImageAspectFlagBits::eColor, @@ -274,21 +274,20 @@ void video_encoder_x264::present_image(vk::Image y_cbcr, vk::SemaphoreSubmitInfo cmd.end(); - std::unique_lock lock(*vk.transfer_queue ? vk.transfer_queue_mutex : vk.queue_mutex); + std::unique_lock lock(vk.transfer_queue ? vk.transfer_queue.mutex : vk.queue.mutex); vk::CommandBufferSubmitInfo cmd_info{ .commandBuffer = *cmd, }; compositor_sem.stageMask = vk::PipelineStageFlagBits2::eTransfer; vk.device.resetFences(*in[slot].fence); - (*vk.transfer_queue ? vk.transfer_queue : vk.queue) - .submit2(vk::SubmitInfo2{ - .waitSemaphoreInfoCount = 1, - .pWaitSemaphoreInfos = &compositor_sem, - .commandBufferInfoCount = 1, - .pCommandBufferInfos = &cmd_info, - }, - *in[slot].fence); + (vk.transfer_queue ? vk.transfer_queue : vk.queue).queue.submit2(vk::SubmitInfo2{ + .waitSemaphoreInfoCount = 1, + .pWaitSemaphoreInfos = &compositor_sem, + .commandBufferInfoCount = 1, + .pCommandBufferInfos = &cmd_info, + }, + *in[slot].fence); } std::optional video_encoder_x264::encode(uint8_t slot, uint64_t frame_index) diff --git a/server/utils/wivrn_vk_bundle.cpp b/server/utils/wivrn_vk_bundle.cpp index be0694dd..08aa27b0 100644 --- a/server/utils/wivrn_vk_bundle.cpp +++ b/server/utils/wivrn_vk_bundle.cpp @@ -126,7 +126,8 @@ int get_queue_index(const std::vector & queues, std:: return info.queueCount++; } } - static std::array prios{1.0, 1.0, 1.0}; + // worst case: compute, transfer and 3x encode are all of the same family + static std::array prios{1.0, 1.0, 1.0, 1.0, 1.0}; infos.push_back(vk::DeviceQueueCreateInfo{ .queueFamilyIndex = family_index, .queueCount = 1, @@ -140,12 +141,6 @@ wivrn::vk_bundle::vk_bundle() : instance(nullptr), physical_device(nullptr), device(nullptr), - queue(nullptr), - queue_family_index(vk::QueueFamilyIgnored), - transfer_queue(nullptr), - transfer_queue_family_index(vk::QueueFamilyIgnored), - encode_queue(nullptr), - encode_queue_family_index(vk::QueueFamilyIgnored), debug(nullptr) { // Create instance @@ -218,14 +213,15 @@ wivrn::vk_bundle::vk_bundle() : // Select queue families auto queues = physical_device.getQueueFamilyProperties(); + uint32_t encode_queue_family_index; { - queue_family_index = select_queue(queues, vk::QueueFlagBits::eCompute); - transfer_queue_family_index = select_queue(queues, vk::QueueFlagBits::eTransfer); + queue.family_index = select_queue(queues, vk::QueueFlagBits::eCompute); + transfer_queue.family_index = select_queue(queues, vk::QueueFlagBits::eTransfer); #if WIVRN_USE_VULKAN_ENCODE encode_queue_family_index = select_queue(queues, vk::QueueFlagBits::eVideoEncodeKHR); #endif // Technically allowed to have a device with only encode or decode capabilities - if (queue_family_index == vk::QueueFamilyIgnored) + if (queue.family_index == vk::QueueFamilyIgnored) throw std::runtime_error("GPU does not support vulkan compute"); } @@ -295,13 +291,21 @@ wivrn::vk_bundle::vk_bundle() : float prio = 1.0; std::vector queues_info; - int queue_index = get_queue_index(queues, queues_info, queue_family_index); + int queue_index = get_queue_index(queues, queues_info, queue.family_index); U_LOG_D("queue index: %d", queue_index); - int transfer_queue_index = get_queue_index(queues, queues_info, transfer_queue_family_index); + int transfer_queue_index = get_queue_index(queues, queues_info, transfer_queue.family_index); U_LOG_D("transfer queue index: %d", transfer_queue_index); #if WIVRN_USE_VULKAN_ENCODE - int encode_queue_index = get_queue_index(queues, queues_info, encode_queue_family_index); - U_LOG_D("encode queue index: %d", encode_queue_index); + std::vector encode_queue_indices; + // Get up to 3 encode queues (left, right and alpha encoders) + for (int i = 0; i < 3; ++i) + { + int encode_queue_index = get_queue_index(queues, queues_info, encode_queue_family_index); + if (encode_queue_index < 0) + break; + U_LOG_D("encode queue index: %d", encode_queue_index); + encode_queue_indices.push_back(encode_queue_index); + } #ifdef VK_KHR_video_maintenance1 if (has_device_ext(VK_KHR_VIDEO_MAINTENANCE_1_EXTENSION_NAME)) std::get(feat).videoMaintenance1 = @@ -353,19 +357,22 @@ wivrn::vk_bundle::vk_bundle() : .ppEnabledExtensionNames = device_extensions.data(), }); - queue = device.getQueue(queue_family_index, queue_index); - name(queue, "compute queue"); + queue.queue = device.getQueue(queue.family_index, queue_index); + name(queue.queue, "compute queue"); if (transfer_queue_index >= 0) { - transfer_queue = device.getQueue(transfer_queue_family_index, transfer_queue_index); - name(transfer_queue, "transfer queue"); + transfer_queue.queue = device.getQueue(transfer_queue.family_index, transfer_queue_index); + name(transfer_queue.queue, "transfer queue"); } #if WIVRN_USE_VULKAN_ENCODE - if (encode_queue_index >= 0) + for (auto encode_queue_index: encode_queue_indices) { - encode_queue = device.getQueue(encode_queue_family_index, encode_queue_index); - name(encode_queue, "encode queue"); + auto & queue = encode_queues.emplace_back(); + queue.queue = device.getQueue(encode_queue_family_index, encode_queue_index); + queue.family_index = encode_queue_family_index; + name(queue.queue, "encode queue"); } + U_LOG_D("Using %d vulkan encode queue(s)", int(encode_queues.size())); #endif } @@ -382,9 +389,9 @@ wivrn::vk_bundle::vk_bundle() : "\tGPU: %s\n" "\tqueue families: %d %d %d (main, encode, transfer)\n", prop.deviceName.data(), - int32_t(queue_family_index), + int32_t(queue.family_index), int32_t(encode_queue_family_index), - int32_t(transfer_queue_family_index)); + int32_t(transfer_queue.family_index)); } uint32_t wivrn::vk_bundle::get_memory_type(uint32_t type_bits, vk::MemoryPropertyFlags memory_props) diff --git a/server/utils/wivrn_vk_bundle.h b/server/utils/wivrn_vk_bundle.h index 98e4949f..572bec6c 100644 --- a/server/utils/wivrn_vk_bundle.h +++ b/server/utils/wivrn_vk_bundle.h @@ -21,6 +21,7 @@ #include "vk/vk_allocator.h" #include +#include #include #include #include @@ -61,23 +62,25 @@ uint64_t vk_handle(const T & handle) struct vk_bundle { + struct queue_data + { + std::mutex mutex; + vk::raii::Queue queue = nullptr; + uint32_t family_index = vk::QueueFamilyIgnored; + operator bool() const + { + return *queue != VK_NULL_HANDLE; + } + }; vk::raii::Context vk_ctx; vk::raii::Instance instance; vk::raii::PhysicalDevice physical_device; vk::raii::Device device; std::optional allocator; - std::mutex queue_mutex; - vk::raii::Queue queue; - uint32_t queue_family_index; - - std::mutex transfer_queue_mutex; - vk::raii::Queue transfer_queue; - uint32_t transfer_queue_family_index; - - std::mutex encode_queue_mutex; - vk::raii::Queue encode_queue; - uint32_t encode_queue_family_index; + queue_data queue; + queue_data transfer_queue; + beman::inplace_vector::inplace_vector encode_queues; vk::raii::DebugUtilsMessengerEXT debug;