[ET-VK] Experimental FP Linear Implementation with NV cooperative matrix 2 extension#17581
Open
HarryHu-art wants to merge 1 commit intogh/HarryHu-art/4/basefrom
Open
[ET-VK] Experimental FP Linear Implementation with NV cooperative matrix 2 extension#17581HarryHu-art wants to merge 1 commit intogh/HarryHu-art/4/basefrom
HarryHu-art wants to merge 1 commit intogh/HarryHu-art/4/basefrom
Conversation
…rix 2 extension Experimental FP Linear Implementation with NV cooperativate matrix 2 ``` buck run @fbcode/mode/win //xplat/executorch/backends/vulkan/test/custom_ops:test_fp_linear >> File changed: fbcode//executorch/backends/vulkan/runtime/gen_vulkan_spv.py File changed: fbcode//executorch/backends/vulkan/runtime/graph/ops/glsl/pack_fp_linear_weight.yaml File changed: fbcode//executorch/backends/vulkan/runtime/graph/ops/impl/LinearExperimental.cpp 15 additional file change events Buck UI: https://www.internalfb.com/buck2/34f0710d-d349-4cba-9e35-10926968dd39 Network: Up: 0B Down: 0B Command: run. Time elapsed: 19.0s BUILD SUCCEEDED - starting your binary === Compute Shader Performance Benchmark === FP32/FP16 Linear Layer Benchmark ---------------------------------------------------------------------- === Cooperative Matrix Properties === Loader Message 0 Inserted device layer "VK_LAYER_KHRONOS_validation" (C:\VulkanSDK\1.4.321.1\Bin\.\VkLayer_khronos_validation.dll) Loader Message 0 Inserted device layer "VK_LAYER_NV_present" (C:\Windows\System32\DriverStore\FileRepository\nvlei.inf_amd64_28d22cc4fdec4a49\.\nvoglv64.dll) Loader Message 0 Inserted device layer "VK_LAYER_NV_optimus" (C:\Windows\System32\DriverStore\FileRepository\nvlei.inf_amd64_28d22cc4fdec4a49\.\nvoglv64.dll) Loader Message 0 vkCreateDevice layer callstack setup to: Loader Message 0 <Application> Loader Message 0 || Loader Message 0 <Loader> Loader Message 0 || Loader Message 0 VK_LAYER_NV_optimus Loader Message 0 Type: Implicit Loader Message 0 Enabled By: Implicit Layer Loader Message 0 Disable Env Var: DISABLE_LAYER_NV_OPTIMUS_1 Loader Message 0 Manifest: C:\Windows\System32\DriverStore\FileRepository\nvlei.inf_amd64_28d22cc4fdec4a49\nv-vk64.json Loader Message 0 Library: C:\Windows\System32\DriverStore\FileRepository\nvlei.inf_amd64_28d22cc4fdec4a49\.\nvoglv64.dll Loader Message 0 || Loader Message 0 VK_LAYER_NV_present Loader Message 0 Type: Implicit Loader Message 0 Enabled By: Implicit Layer Loader Message 0 Disable Env Var: DISABLE_LAYER_NV_PRESENT_1 Loader Message 0 Manifest: C:\Windows\System32\DriverStore\FileRepository\nvlei.inf_amd64_28d22cc4fdec4a49\nv-vk64.json Loader Message 0 Library: C:\Windows\System32\DriverStore\FileRepository\nvlei.inf_amd64_28d22cc4fdec4a49\.\nvoglv64.dll Loader Message 0 || Loader Message 0 VK_LAYER_KHRONOS_validation Loader Message 0 Type: Explicit Loader Message 0 Enabled By: By the Application Loader Message 0 Manifest: C:\VulkanSDK\1.4.321.1\Bin\VkLayer_khronos_validation.json Loader Message 0 Library: C:\VulkanSDK\1.4.321.1\Bin\.\VkLayer_khronos_validation.dll Loader Message 0 || Loader Message 0 <Device> Loader Message 0 Using "NVIDIA GeForce RTX 5080" with driver: "C:\Windows\System32\DriverStore\FileRepository\nvlei.inf_amd64_28d22cc4fdec4a49\.\nvoglv64.dll" Validation 0 vkCreateImage(): The following VkImageCreateInfo returned VK_ERROR_FORMAT_NOT_SUPPORTED when calling vkGetPhysicalDeviceImageFormatProperties2 format (VK_FORMAT_R32G32B32A32_SFLOAT) type (VK_IMAGE_TYPE_3D) tiling (VK_IMAGE_TILING_LINEAR) usage (VK_IMAGE_USAGE_SAMPLED_BIT|VK_IMAGE_USAGE_STORAGE_BIT) flags (VkImageCreateFlags(0)) VkImageCreateInfo::pNext is NULL. The Vulkan spec states: Each of the following values (as described in Image Creation Limits) must not be undefined : imageCreateMaxMipLevels, imageCreateMaxArrayLayers, imageCreateMaxExtent, and imageCreateSampleCounts (https://vulkan.lunarg.com/doc/view/1.4.321.1/windows/antora/spec/latest/chapters/resources.html#VUID-VkImageCreateInfo-imageCreateMaxMipLevels-02251) Found 15 cooperative matrix configurations: ---------------------------------------------------------------------- # | M | N | K | A Type | B Type | C Type | R Type | Scope ---------------------------------------------------------------------- 0 | 16 | 16 | 16 | float16 | float16 | float16 | float16 | Subgroup 1 | 16 | 8 | 16 | float16 | float16 | float16 | float16 | Subgroup 2 | 16 | 8 | 8 | float16 | float16 | float16 | float16 | Subgroup 3 | 16 | 16 | 16 | float16 | float16 | float32 | float32 | Subgroup 4 | 16 | 8 | 16 | float16 | float16 | float32 | float32 | Subgroup 5 | 16 | 8 | 8 | float16 | float16 | float32 | float32 | Subgroup 6 | 16 | 16 | 32 | uint8 | uint8 | uint32 | uint32 | Subgroup 7 | 16 | 16 | 32 | int8 | int8 | int32 | int32 | Subgroup 8 | 16 | 8 | 32 | uint8 | uint8 | uint32 | uint32 | Subgroup 9 | 16 | 8 | 32 | int8 | int8 | int32 | int32 | Subgroup 10 | 16 | 16 | 16 | unknown | unknown | float32 | float32 | Subgroup 11 | 16 | 16 | 32 | unknown | unknown | float16 | float16 | Subgroup 12 | 16 | 16 | 32 | unknown | unknown | float32 | float32 | Subgroup 13 | 16 | 16 | 32 | unknown | unknown | float16 | float16 | Subgroup 14 | 16 | 16 | 32 | unknown | unknown | float32 | float32 | Subgroup ---------------------------------------------------------------------- Configurations with float32 A, B, C types: Configurations with float16 A/B, float32 C (mixed precision): M=16, N=16, K=16, Scope=Subgroup M=16, N=8, K=16, Scope=Subgroup M=16, N=8, K=8, Scope=Subgroup Test: ACCU B=4 I=128 O=128 Buf fp16+bias L input_tensor Data: Type: ValueSpec(type=Tensor, sizes=[4, 128], dtype=half, memory_layout=WidthPacked, storage_type=Buffer, data_gen=RANDOM) Total elements: 512 Data (first 64 elements): [-0.250732, 0.592773, 0.901367, -0.632812, 0.463867, 0.559082, 0.197266, 0.193604, -0.687500, -0.108276, -0.687988, -0.799805, -0.883789, -0.081482, 0.731934, -0.332520, 0.202148, -0.713867, 0.416016, 0.301758, -0.958496, -0.886719, 0.939453, 0.443848, 0.664551, 0.876953, -0.575195, -0.998047, -0.636230, 0.984375, -0.632812, 0.234863, -0.391357, 0.223267, 0.049500, -0.985840, -0.136108, -0.953613, -0.417480, 0.049530, 0.223633, -0.200195, -0.720703, -0.906250, -0.415527, 0.947266, -0.267090, -0.534180, -0.087830, -0.818359, 0.570312, 0.236694, -0.600586, -0.234985, 0.028458, 0.966309, 0.184814, -0.066467, -0.906738, 0.719727, 0.215088, 0.360596, -0.658691, -0.098999, ... (448 more)] Statistics: min=0.229682, max=1.468703, mean=0.922048, sum=472.088684 weight_tensor Data: Type: ValueSpec(type=Tensor, sizes=[128, 128], dtype=half, memory_layout=WidthPacked, storage_type=Buffer, data_gen=RANDOM) Total elements: 16384 Data (first 64 elements): [-0.769531, -0.006275, 0.218018, -0.793945, -0.732910, -0.702148, -0.518555, -0.655273, -0.345703, 0.622070, 0.718262, -0.850098, 0.332031, -0.909668, 0.082275, 0.224243, -0.941895, 0.189575, 0.467285, -0.498535, -0.210083, -0.705078, 0.604004, -0.977051, -0.490967, -0.063721, -0.885742, 0.909180, 0.732910, 0.111511, -0.557617, 0.299316, -0.189941, 0.161743, -0.367676, -0.662109, -0.846191, -0.168213, 0.686035, 0.305908, 0.697754, 0.241089, 0.942871, -0.192383, -0.229126, 0.747070, 0.908691, 0.110107, -0.108459, -0.142090, 0.339355, -0.720215, -0.834961, -0.539062, 0.793945, -0.605469, -0.403809, 0.596680, -0.475342, -0.335205, -0.989258, -0.157715, 0.086365, -0.110352, ... (16320 more)] Statistics: min=0.013550, max=1.468764, mean=0.913034, sum=14959.146484 bias_tensor Data: Type: ValueSpec(type=Tensor, sizes=[128], dtype=half, memory_layout=WidthPacked, storage_type=Buffer, data_gen=RANDOM) Total elements: 128 Data (first 64 elements): [0.669434, -0.134888, -0.790039, -0.936035, 0.489258, 0.255615, -0.278809, -0.630859, -0.281250, -0.407227, 0.218384, -0.485840, -0.212402, 0.359619, -0.181763, -0.434814, 0.019791, 0.833496, 0.420166, -0.583496, 0.920898, 0.262939, -0.086731, 0.597168, -0.144653, 0.065247, -0.772949, -0.650391, -0.563965, -0.645020, 0.914551, 0.467285, 0.886230, -0.200317, 0.763184, 0.024109, 0.292725, -0.917969, -0.572266, -0.051605, 0.273438, -0.978027, -0.721680, 0.602051, -0.082581, -0.718262, 0.747559, -0.410889, -0.482910, 0.708496, 0.329590, 0.599609, 0.725098, -0.226318, -0.702148, 0.677734, 0.125854, -0.276367, -0.681641, 0.816406, -0.653809, -0.542480, -0.791504, -0.032104, ... (64 more)] Statistics: min=0.289590, max=1.467422, mean=0.940261, sum=120.353394 Executing 1 test cases for FPLinear ---------------------------------------------------------------------- ==================== Shared Object List ==================== idx sizes users 0 16 [7,] ==================== Value List ============================ idx type sizes node_type storage_bytes so_idx 0 TENSOR [4, 128] INPUT 1024 1 STAGING 2 TENSORREF [128, 128] PREPACK 3 TENSORREF [128] PREPACK 4 INT 5 TENSOR [4, 128] OUTPUT 1024 6 TENSOR [128, 128] PREPACK 32768 7 TENSOR [] 16 0 8 TENSOR [128] PREPACK 256 9 STAGING ==================== Prepack Node List ===================== idx shader_name tref packed 0pack_fp_linear_weight_buffer_half 2 6 1 nchw_to_buffer_half_half 3 8 ==================== Execute Node List ===================== idx shader_name READ_arg WRITE_arg 0 nchw_to_buffer_half_float [1,] [0,] 1 linear_tiled_nv_cm2_half [0,6,8,] [5,] 2 buffer_to_nchw_half_float [5,] [9,] Output[0] Data: Type: ValueSpec(type=Tensor, sizes=[4, 128], dtype=half, memory_layout=WidthPacked, storage_type=Buffer, data_gen=ZEROS) Total elements: 512 Data (first 20 elements): [-1.498047, -3.060547, -0.001218, -2.798828, 1.934570, 2.558594, 3.414062, 5.820312, -2.109375, -3.806641, 3.628906, 0.022491, -1.992188, -1.054688, 0.149658, -8.593750, -0.514648, 8.828125, 3.468750, -1.464844, ... (492 more)] Statistics: min=0.101169, max=1.581347, mean=1.030442, sum=527.586243 Output[0] (ref) Data: Type: ValueSpec(type=Tensor, sizes=[4, 128], dtype=half, memory_layout=WidthPacked, storage_type=Buffer, data_gen=ZEROS) Total elements: 512 Data (first 20 elements): [-1.499553, -3.061874, -0.002150, -2.799306, 1.934955, 2.555415, 3.412702, 5.819950, -2.108622, -3.801892, 3.630069, 0.021923, -1.991505, -1.054845, 0.149873, -8.598531, -0.516323, 8.826857, 3.467027, -1.466808, ... (492 more)] Statistics: min=0.101169, max=1.581347, mean=1.030442, sum=527.586243 Correctness validation PASSED linear_tiled_nv_cm2_half ACCU B=4 I=128 O=128 Buf fp16+bias L [4x128] 5.920 ╬╝s 11.070 GFLOP/s PASSED ---------------------------------------------------------------------- ```` Differential Revision: [D91945037](https://our.internmc.facebook.com/intern/diff/D91945037/) [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/17581
Note: Links to docs will display an error until the docs builds have been completed. ❌ 2 New FailuresAs of commit ab6b3da with merge base 0c87468 ( NEW FAILURES - The following jobs have failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
This was referenced Feb 20, 2026
meta-cla
bot
added
the
CLA Signed
This PR needs a
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stack from ghstack (oldest at bottom):
Experimental FP Linear Implementation with NV cooperativate matrix 2
Differential Revision: D91945037