LuxRender Bug Tracker

View Issue Details Jump to Notes ] Issue History ] Print ]
IDProjectCategoryView StatusDate SubmittedLast Update
0001662LuxRenderCorepublic2017-10-24 13:532017-10-24 13:56
Reporterjamilbk 
Assigned To 
PrioritynormalSeveritymajorReproducibilityalways
StatusnewResolutionopen 
Platformamd64OSUbuntuOS Version16.04.3
Product Version 
Target VersionFixed in Version 
Summary0001662: Extremely high MultiGPU score in Luxmark Hotel scene on Vega FE
DescriptionHi there,

While I do enjoy being the king of hill for the MultiGPU Hotel scene, I can't help but think the score is wrong in some way. See http://luxmark.info/node/5018 [^]

Seeing as I only get around 2,000 on a single Vega FE, it seems hard to believe that four Vega FEs score over 75,000.

I'm not sure how to begin debugging this, but here are my system details in case it matters:

Hardware:
- Threadripper 1950x
- Gigabyte Gaming 7 X399
- 128 GB 2933 Mhz C14 DDR4
- 4x Vega FE (water cooled)

Software:
- ROCm 1.6.4
- Ubuntu 16.04.3

I've noticed the Kernel Compilation time is much, much higher than on Windows -- about 750,000 ms total for this scene. Maybe that's a clue?

Steps To ReproduceInstall AMD ROCm 1.6.4 on a multi-gpu Vega linux host and run Luxmark 3.1 Hotel benchmark.
Additional InformationOutput of clinfo:

Number of platforms: 1
  Platform Profile: FULL_PROFILE
  Platform Version: OpenCL 2.0 AMD-APP (2508.0)
  Platform Name: AMD Accelerated Parallel Processing
  Platform Vendor: Advanced Micro Devices, Inc.
  Platform Extensions: cl_khr_icd cl_amd_event_callback


  Platform Name: AMD Accelerated Parallel Processing
Number of devices: 4
  Device Type: CL_DEVICE_TYPE_GPU
  Vendor ID: 1002h
  Board name: Device 6863
  Device Topology: PCI[ B#10, D#0, F#0 ]
  Max compute units: 64
  Max work items dimensions: 3
    Max work items[0]: 1024
    Max work items[1]: 1024
    Max work items[2]: 1024
  Max work group size: 256
  Preferred vector width char: 4
  Preferred vector width short: 2
  Preferred vector width int: 1
  Preferred vector width long: 1
  Preferred vector width float: 1
  Preferred vector width double: 1
  Native vector width char: 4
  Native vector width short: 2
  Native vector width int: 1
  Native vector width long: 1
  Native vector width float: 1
  Native vector width double: 1
  Max clock frequency: 1600Mhz
  Address bits: 64
  Max memory allocation: 14588628172
  Image support: Yes
  Max number of images read arguments: 128
  Max number of images write arguments: 8
  Max image 2D width: 16384
  Max image 2D height: 16384
  Max image 3D width: 2048
  Max image 3D height: 2048
  Max image 3D depth: 2048
  Max samplers within kernel: 26723
  Max size of kernel argument: 1024
  Alignment (bits) of base address: 1024
  Minimum alignment (bytes) for any datatype: 128
  Single precision floating point capability
    Denorms: Yes
    Quiet NaNs: Yes
    Round to nearest even: Yes
    Round to zero: Yes
    Round to +ve and infinity: Yes
    IEEE754-2008 fused multiply-add: Yes
  Cache type: Read/Write
  Cache line size: 64
  Cache size: 16384
  Global memory size: 17163091968
  Constant buffer size: 14588628172
  Max number of constant args: 8
  Local memory type: Scratchpad
  Local memory size: 65536
  Max pipe arguments: 0
  Max pipe active reservations: 0
  Max pipe packet size: 0
  Max global variable size: 14588628172
  Max global variable preferred total size: 17163091968
  Max read/write image args: 64
  Max on device events: 0
  Queue on device max size: 0
  Max on device queues: 0
  Queue on device preferred size: 0
  SVM capabilities:
    Coarse grain buffer: Yes
    Fine grain buffer: Yes
    Fine grain system: No
    Atomics: No
  Preferred platform atomic alignment: 0
  Preferred global atomic alignment: 0
  Preferred local atomic alignment: 0
  Kernel Preferred work group size multiple: 64
  Error correction support: 0
  Unified memory for Host and Device: 0
  Profiling timer resolution: 1
  Device endianess: Little
  Available: Yes
  Compiler available: Yes
  Execution capabilities:
    Execute OpenCL kernels: Yes
    Execute native function: No
  Queue on Host properties:
    Out-of-Order: No
    Profiling : Yes
  Queue on Device properties:
    Out-of-Order: No
    Profiling : No
  Platform ID: 0x7f4ae8bbe110
  Name: gfx900
  Vendor: Advanced Micro Devices, Inc.
  Device OpenCL C version: OpenCL C 2.0
  Driver version: 1.1 (HSA,LC)
  Profile: FULL_PROFILE
  Version: OpenCL 1.2
  Extensions: cl_khr_fp64 cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_khr_int64_base_atomics cl_khr_int64_extended_atomics cl_khr_3d_image_writes cl_khr_byte_addressable_store cl_khr_fp16 cl_khr_gl_sharing cl_amd_media_ops cl_amd_media_ops2 cl_khr_subgroups cl_khr_depth_images cl_amd_liquid_flash cl_amd_copy_buffer_p2p


  Device Type: CL_DEVICE_TYPE_GPU
  Vendor ID: 1002h
  Board name: Device 6863
  Device Topology: PCI[ B#13, D#0, F#0 ]
  Max compute units: 64
  Max work items dimensions: 3
    Max work items[0]: 1024
    Max work items[1]: 1024
    Max work items[2]: 1024
  Max work group size: 256
  Preferred vector width char: 4
  Preferred vector width short: 2
  Preferred vector width int: 1
  Preferred vector width long: 1
  Preferred vector width float: 1
  Preferred vector width double: 1
  Native vector width char: 4
  Native vector width short: 2
  Native vector width int: 1
  Native vector width long: 1
  Native vector width float: 1
  Native vector width double: 1
  Max clock frequency: 1600Mhz
  Address bits: 64
  Max memory allocation: 14588628172
  Image support: Yes
  Max number of images read arguments: 128
  Max number of images write arguments: 8
  Max image 2D width: 16384
  Max image 2D height: 16384
  Max image 3D width: 2048
  Max image 3D height: 2048
  Max image 3D depth: 2048
  Max samplers within kernel: 26723
  Max size of kernel argument: 1024
  Alignment (bits) of base address: 1024
  Minimum alignment (bytes) for any datatype: 128
  Single precision floating point capability
    Denorms: Yes
    Quiet NaNs: Yes
    Round to nearest even: Yes
    Round to zero: Yes
    Round to +ve and infinity: Yes
    IEEE754-2008 fused multiply-add: Yes
  Cache type: Read/Write
  Cache line size: 64
  Cache size: 16384
  Global memory size: 17163091968
  Constant buffer size: 14588628172
  Max number of constant args: 8
  Local memory type: Scratchpad
  Local memory size: 65536
  Max pipe arguments: 0
  Max pipe active reservations: 0
  Max pipe packet size: 0
  Max global variable size: 14588628172
  Max global variable preferred total size: 17163091968
  Max read/write image args: 64
  Max on device events: 0
  Queue on device max size: 0
  Max on device queues: 0
  Queue on device preferred size: 0
  SVM capabilities:
    Coarse grain buffer: Yes
    Fine grain buffer: Yes
    Fine grain system: No
    Atomics: No
  Preferred platform atomic alignment: 0
  Preferred global atomic alignment: 0
  Preferred local atomic alignment: 0
  Kernel Preferred work group size multiple: 64
  Error correction support: 0
  Unified memory for Host and Device: 0
  Profiling timer resolution: 1
  Device endianess: Little
  Available: Yes
  Compiler available: Yes
  Execution capabilities:
    Execute OpenCL kernels: Yes
    Execute native function: No
  Queue on Host properties:
    Out-of-Order: No
    Profiling : Yes
  Queue on Device properties:
    Out-of-Order: No
    Profiling : No
  Platform ID: 0x7f4ae8bbe110
  Name: gfx900
  Vendor: Advanced Micro Devices, Inc.
  Device OpenCL C version: OpenCL C 2.0
  Driver version: 1.1 (HSA,LC)
  Profile: FULL_PROFILE
  Version: OpenCL 1.2
  Extensions: cl_khr_fp64 cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_khr_int64_base_atomics cl_khr_int64_extended_atomics cl_khr_3d_image_writes cl_khr_byte_addressable_store cl_khr_fp16 cl_khr_gl_sharing cl_amd_media_ops cl_amd_media_ops2 cl_khr_subgroups cl_khr_depth_images cl_amd_liquid_flash cl_amd_copy_buffer_p2p


  Device Type: CL_DEVICE_TYPE_GPU
  Vendor ID: 1002h
  Board name: Device 6863
  Device Topology: PCI[ B#69, D#0, F#0 ]
  Max compute units: 64
  Max work items dimensions: 3
    Max work items[0]: 1024
    Max work items[1]: 1024
    Max work items[2]: 1024
  Max work group size: 256
  Preferred vector width char: 4
  Preferred vector width short: 2
  Preferred vector width int: 1
  Preferred vector width long: 1
  Preferred vector width float: 1
  Preferred vector width double: 1
  Native vector width char: 4
  Native vector width short: 2
  Native vector width int: 1
  Native vector width long: 1
  Native vector width float: 1
  Native vector width double: 1
  Max clock frequency: 1600Mhz
  Address bits: 64
  Max memory allocation: 14588628172
  Image support: Yes
  Max number of images read arguments: 128
  Max number of images write arguments: 8
  Max image 2D width: 16384
  Max image 2D height: 16384
  Max image 3D width: 2048
  Max image 3D height: 2048
  Max image 3D depth: 2048
  Max samplers within kernel: 26723
  Max size of kernel argument: 1024
  Alignment (bits) of base address: 1024
  Minimum alignment (bytes) for any datatype: 128
  Single precision floating point capability
    Denorms: Yes
    Quiet NaNs: Yes
    Round to nearest even: Yes
    Round to zero: Yes
    Round to +ve and infinity: Yes
    IEEE754-2008 fused multiply-add: Yes
  Cache type: Read/Write
  Cache line size: 64
  Cache size: 16384
  Global memory size: 17163091968
  Constant buffer size: 14588628172
  Max number of constant args: 8
  Local memory type: Scratchpad
  Local memory size: 65536
  Max pipe arguments: 0
  Max pipe active reservations: 0
  Max pipe packet size: 0
  Max global variable size: 14588628172
  Max global variable preferred total size: 17163091968
  Max read/write image args: 64
  Max on device events: 0
  Queue on device max size: 0
  Max on device queues: 0
  Queue on device preferred size: 0
  SVM capabilities:
    Coarse grain buffer: Yes
    Fine grain buffer: Yes
    Fine grain system: No
    Atomics: No
  Preferred platform atomic alignment: 0
  Preferred global atomic alignment: 0
  Preferred local atomic alignment: 0
  Kernel Preferred work group size multiple: 64
  Error correction support: 0
  Unified memory for Host and Device: 0
  Profiling timer resolution: 1
  Device endianess: Little
  Available: Yes
  Compiler available: Yes
  Execution capabilities:
    Execute OpenCL kernels: Yes
    Execute native function: No
  Queue on Host properties:
    Out-of-Order: No
    Profiling : Yes
  Queue on Device properties:
    Out-of-Order: No
    Profiling : No
  Platform ID: 0x7f4ae8bbe110
  Name: gfx900
  Vendor: Advanced Micro Devices, Inc.
  Device OpenCL C version: OpenCL C 2.0
  Driver version: 1.1 (HSA,LC)
  Profile: FULL_PROFILE
  Version: OpenCL 1.2
  Extensions: cl_khr_fp64 cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_khr_int64_base_atomics cl_khr_int64_extended_atomics cl_khr_3d_image_writes cl_khr_byte_addressable_store cl_khr_fp16 cl_khr_gl_sharing cl_amd_media_ops cl_amd_media_ops2 cl_khr_subgroups cl_khr_depth_images cl_amd_liquid_flash cl_amd_copy_buffer_p2p


  Device Type: CL_DEVICE_TYPE_GPU
  Vendor ID: 1002h
  Board name: Device 6863
  Device Topology: PCI[ B#72, D#0, F#0 ]
  Max compute units: 64
  Max work items dimensions: 3
    Max work items[0]: 1024
    Max work items[1]: 1024
    Max work items[2]: 1024
  Max work group size: 256
  Preferred vector width char: 4
  Preferred vector width short: 2
  Preferred vector width int: 1
  Preferred vector width long: 1
  Preferred vector width float: 1
  Preferred vector width double: 1
  Native vector width char: 4
  Native vector width short: 2
  Native vector width int: 1
  Native vector width long: 1
  Native vector width float: 1
  Native vector width double: 1
  Max clock frequency: 1600Mhz
  Address bits: 64
  Max memory allocation: 14588628172
  Image support: Yes
  Max number of images read arguments: 128
  Max number of images write arguments: 8
  Max image 2D width: 16384
  Max image 2D height: 16384
  Max image 3D width: 2048
  Max image 3D height: 2048
  Max image 3D depth: 2048
  Max samplers within kernel: 26723
  Max size of kernel argument: 1024
  Alignment (bits) of base address: 1024
  Minimum alignment (bytes) for any datatype: 128
  Single precision floating point capability
    Denorms: Yes
    Quiet NaNs: Yes
    Round to nearest even: Yes
    Round to zero: Yes
    Round to +ve and infinity: Yes
    IEEE754-2008 fused multiply-add: Yes
  Cache type: Read/Write
  Cache line size: 64
  Cache size: 16384
  Global memory size: 17163091968
  Constant buffer size: 14588628172
  Max number of constant args: 8
  Local memory type: Scratchpad
  Local memory size: 65536
  Max pipe arguments: 0
  Max pipe active reservations: 0
  Max pipe packet size: 0
  Max global variable size: 14588628172
  Max global variable preferred total size: 17163091968
  Max read/write image args: 64
  Max on device events: 0
  Queue on device max size: 0
  Max on device queues: 0
  Queue on device preferred size: 0
  SVM capabilities:
    Coarse grain buffer: Yes
    Fine grain buffer: Yes
    Fine grain system: No
    Atomics: No
  Preferred platform atomic alignment: 0
  Preferred global atomic alignment: 0
  Preferred local atomic alignment: 0
  Kernel Preferred work group size multiple: 64
  Error correction support: 0
  Unified memory for Host and Device: 0
  Profiling timer resolution: 1
  Device endianess: Little
  Available: Yes
  Compiler available: Yes
  Execution capabilities:
    Execute OpenCL kernels: Yes
    Execute native function: No
  Queue on Host properties:
    Out-of-Order: No
    Profiling : Yes
  Queue on Device properties:
    Out-of-Order: No
    Profiling : No
  Platform ID: 0x7f4ae8bbe110
  Name: gfx900
  Vendor: Advanced Micro Devices, Inc.
  Device OpenCL C version: OpenCL C 2.0
  Driver version: 1.1 (HSA,LC)
  Profile: FULL_PROFILE
  Version: OpenCL 1.2
  Extensions: cl_khr_fp64 cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_khr_int64_base_atomics cl_khr_int64_extended_atomics cl_khr_3d_image_writes cl_khr_byte_addressable_store cl_khr_fp16 cl_khr_gl_sharing cl_amd_media_ops cl_amd_media_ops2 cl_khr_subgroups cl_khr_depth_images cl_amd_liquid_flash cl_amd_copy_buffer_p2p
TagsNo tags attached.
Mercurial Changeset #
Requires Documentation UpdateNo
Requires Exporter Update
Attached Files

- Relationships

-  Notes
(0004337)
jamilbk (reporter)
2017-10-24 13:56

If by chance this benchmark is now using half-floats (FP16) for computation, this was recently exposed in the ROCm 1.6.4 OpenCL stack so that may explain the high score. See https://www.anandtech.com/show/11717/the-amd-radeon-rx-vega-64-and-56-review/4 [^]

- Issue History
Date Modified Username Field Change
2017-10-24 13:53 jamilbk New Issue
2017-10-24 13:54 jamilbk Category LuxMax => Core
2017-10-24 13:55 jamilbk Description Updated View Revisions
2017-10-24 13:56 jamilbk Note Added: 0004337


Copyright © 2000 - 2018 MantisBT Team
Powered by Mantis Bugtracker