ncnn-android-vulkan vulkan acceleration is not work on snapdragon 8+ gen1

[superduelist]

android sdk 36
ndk 28
running on android 13

when open vulkan acceleration by
net.opt.use_vulkan_compute = true;

the infer time consumption is far more than not open(pure cpu infer)

may i ask why? pls.

[dosubot]

Hi @superduelist! I'm Dosu and I’m helping the ncnn team.

This is a known issue: on Snapdragon 8+ Gen1 (Adreno GPU), ncnn's Vulkan backend is often much slower than CPU inference. The main reasons are immature Vulkan drivers and lack of shader optimization on Adreno, plus ncnn's Vulkan backend is not as optimized as its ARM CPU backend. This is widely reported by users and confirmed by maintainers—Vulkan can be 2-4x slower than CPU on mobile devices, especially on Qualcomm/Adreno hardware (source, source).

Other factors that can make Vulkan slower include: some layers or operations (like certain padding or bias layers) are not implemented for Vulkan and will fallback to CPU, causing extra CPU-GPU data transfer overhead; preprocessing/postprocessing always runs on CPU; and OpenMP is still used for some CPU-side work even with Vulkan enabled (source).

You can try these troubleshooting steps, but they may not fully resolve the performance gap:

• Make sure you select the correct Vulkan device before loading the model:

  net.set_vulkan_device(0);

• Use proper Vulkan allocators:

  ncnn::VkAllocator* blob_vkallocator = vkdev.acquire_blob_allocator();
  ncnn::VkAllocator* staging_vkallocator = vkdev.acquire_blob_allocator();
  net.opt.blob_vkallocator = blob_vkallocator;
  net.opt.workspace_vkallocator = blob_vkallocator;
  net.opt.staging_vkallocator = staging_vkallocator;

• If your model has simple layers between CPU-only layers, force those to run on CPU to avoid transfer overhead by adding 31=16 to the layer in your param file (docs).
• If you use Winograd convolution, try disabling it for some layers with 31=64 to reduce memory usage and sometimes improve Vulkan performance (docs).
• For debugging, you can enable Vulkan validation layers by setting ENABLE_VALIDATION_LAYER to 1 in gpu.cpp (docs).

Unfortunately, there are no device-specific workarounds or fixes for Adreno 8xx/Gen1 in ncnn at this time. This is a limitation of current Adreno Vulkan drivers and ncnn's Vulkan backend (source, source).

If you need maximum performance, CPU inference is currently the best option on this hardware.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}

  

[superduelist]

thank you for your answer.
i wonder will this issue be solved in the future?
and by what method?

Hi @superduelist! I'm Dosu and I’m helping the ncnn team.

This is a known issue: on Snapdragon 8+ Gen1 (Adreno GPU), ncnn's Vulkan backend is often much slower than CPU inference. The main reasons are immature Vulkan drivers and lack of shader optimization on Adreno, plus ncnn's Vulkan backend is not as optimized as its ARM CPU backend. This is widely reported by users and confirmed by maintainers—Vulkan can be 2-4x slower than CPU on mobile devices, especially on Qualcomm/Adreno hardware (source, source).

Other factors that can make Vulkan slower include: some layers or operations (like certain padding or bias layers) are not implemented for Vulkan and will fallback to CPU, causing extra CPU-GPU data transfer overhead; preprocessing/postprocessing always runs on CPU; and OpenMP is still used for some CPU-side work even with Vulkan enabled (source).

You can try these troubleshooting steps, but they may not fully resolve the performance gap:

• Make sure you select the correct Vulkan device before loading the model:

  net.set_vulkan_device(0);

• Use proper Vulkan allocators:

  ncnn::VkAllocator* blob_vkallocator = vkdev.acquire_blob_allocator();
  ncnn::VkAllocator* staging_vkallocator = vkdev.acquire_blob_allocator();
  net.opt.blob_vkallocator = blob_vkallocator;
  net.opt.workspace_vkallocator = blob_vkallocator;
  net.opt.staging_vkallocator = staging_vkallocator;

• If your model has simple layers between CPU-only layers, force those to run on CPU to avoid transfer overhead by adding 31=16 to the layer in your param file (docs).
• If you use Winograd convolution, try disabling it for some layers with 31=64 to reduce memory usage and sometimes improve Vulkan performance (docs).
• For debugging, you can enable Vulkan validation layers by setting ENABLE_VALIDATION_LAYER to 1 in gpu.cpp (docs).

Unfortunately, there are no device-specific workarounds or fixes for Adreno 8xx/Gen1 in ncnn at this time. This is a limitation of current Adreno Vulkan drivers and ncnn's Vulkan backend (source, source).

If you need maximum performance, CPU inference is currently the best option on this hardware.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}

  

[dosubot]

There’s currently no official fix or roadmap for solving the slow Vulkan performance on Snapdragon 8+ Gen1 (Adreno GPU) in ncnn. The main bottleneck is the immaturity of Adreno Vulkan drivers and lack of shader optimization, which are outside ncnn’s direct control. Recent ncnn updates have added some Vulkan extensions for robustness and flexibility, but these do not address the core performance issues on Adreno 8+ Gen1 GPUs. Any future improvement would likely depend on Qualcomm releasing better Vulkan drivers, or ncnn maintainers finding feasible Adreno-specific optimizations—though there’s no indication of such work in progress right now. The best approach is to monitor ncnn and Adreno driver updates for any changes in support or performance improvements going forward. source source

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}

  

[superduelist]

Not considering the certain version like Adreno 730(snapdragon8+gen1)
Is the latest version of snapdragon like snapdragon 8 gen5 elite have the same problem?

There’s currently no official fix or roadmap for solving the slow Vulkan performance on Snapdragon 8+ Gen1 (Adreno GPU) in ncnn. The main bottleneck is the immaturity of Adreno Vulkan drivers and lack of shader optimization, which are outside ncnn’s direct control. Recent ncnn updates have added some Vulkan extensions for robustness and flexibility, but these do not address the core performance issues on Adreno 8+ Gen1 GPUs. Any future improvement would likely depend on Qualcomm releasing better Vulkan drivers, or ncnn maintainers finding feasible Adreno-specific optimizations—though there’s no indication of such work in progress right now. The best approach is to monitor ncnn and Adreno driver updates for any changes in support or performance improvements going forward. source source

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}

  

[dosubot]

Yes, the latest Snapdragon devices (like Snapdragon 8 Gen50 Elite or any new Adreno GPUs) are very likely to have the same Vulkan performance and stability issues with ncnn as older generations. There are ongoing reports of crashes and poor performance on Snapdragon 8 Gen 2 (Adreno 840) with ncnn's Vulkan backend, and there are no device-specific fixes, optimizations, or blacklists for newer Adreno GPUs in the ncnn codebase. The Vulkan backend treats all recent Adreno devices generically, and the main bottleneck remains the quality of Qualcomm's Vulkan drivers and lack of shader optimization. No improvements or special handling for the latest Snapdragon/Adreno GPUs have been added to ncnn as of now. Monitoring future ncnn and Qualcomm driver updates is still the best approach for any changes in support or performance improvements going forward.
Details and discussion

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}

  

[superduelist]

How about MediaTek Dimensity series android devices?

Yes, the latest Snapdragon devices (like Snapdragon 8 Gen50 Elite or any new Adreno GPUs) are very likely to have the same Vulkan performance and stability issues with ncnn as older generations. There are ongoing reports of crashes and poor performance on Snapdragon 8 Gen 2 (Adreno 840) with ncnn's Vulkan backend, and there are no device-specific fixes, optimizations, or blacklists for newer Adreno GPUs in the ncnn codebase. The Vulkan backend treats all recent Adreno devices generically, and the main bottleneck remains the quality of Qualcomm's Vulkan drivers and lack of shader optimization. No improvements or special handling for the latest Snapdragon/Adreno GPUs have been added to ncnn as of now. Monitoring future ncnn and Qualcomm driver updates is still the best approach for any changes in support or performance improvements going forward.
Details and discussion

To reply, just mention @dosu.

--Tek
^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}

  

[dosubot]

ncnn supports Vulkan on Mali GPUs, which are used in MediaTek Dimensity series devices. The Vulkan backend has been tested on several Mali GPUs (like Mali-G71, G72, G76), and models such as squeezenet, mobilenetssd, and yolov3 run successfully on many of them, though some older Mali devices may have incomplete support for certain models see test results.

ncnn applies device-specific flags for Mali GPUs (for example, handling fp16 arithmetic and setting a default Vulkan subgroup size of 16), and enables Vulkan extensions based on detected GPU features. There are no MediaTek/Dimensity-specific blacklists or optimizations—Mali GPUs are handled as a group see code logic.

Unlike Adreno (Snapdragon), there are no widespread reports of severe Vulkan crashes or unusable performance on Mali/Dimensity devices, but there also aren't many public benchmarks or user reports confirming superior performance. Actual results will depend on the specific Dimensity model and its Vulkan driver quality. If you have access to a target device, it's best to test your model directly.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}

  

[superduelist]

All known about.
Thank u very much.

ncnn supports Vulkan on Mali GPUs, which are used in MediaTek Dimensity series devices. The Vulkan backend has been tested on several Mali GPUs (like Mali-G71, G72, G76), and models such as squeezenet, mobilenetssd, and yolov3 run successfully on many of them, though some older Mali devices may have incomplete support for certain models see test results.

ncnn applies device-specific flags for Mali GPUs (for example, handling fp16 arithmetic and setting a default Vulkan subgroup size of 16), and enables Vulkan extensions based on detected GPU features. There are no MediaTek/Dimensity-specific blacklists or optimizations—Mali GPUs are handled as a group see code logic.

Unlike Adreno (Snapdragon), there are no widespread reports of severe Vulkan crashes or unusable performance on Mali/Dimensity devices, but there also aren't many public benchmarks or user reports confirming superior performance. Actual results will depend on the specific Dimensity model and its Vulkan driver quality. If you have access to a target device, it's best to test your model directly.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}