AMDGPU.jl
AMDGPU.jl copied to clipboard
JuliaGPU
Allocating large (>= 4GB) with `AMDGPU.ones` fails
Questionnaire
- Does ROCm works for you outside of Julia, e.g. C/C++/Python?
yes
- Post output of
rocminfo.
output of `rocminfo`
$ rocminfo
ROCk module version 6.3.6 is loaded
=====================
HSA System Attributes
=====================
Runtime Version: 1.14
Runtime Ext Version: 1.6
System Timestamp Freq.: 1000.000000MHz
Sig. Max Wait Duration: 18446744073709551615 (0xFFFFFFFFFFFFFFFF) (timestamp count)
Machine Model: LARGE
System Endianness: LITTLE
Mwaitx: DISABLED
DMAbuf Support: YES
==========
HSA Agents
==========
*******
Agent 1
*******
Name: AMD EPYC 7A53 64-Core Processor
Uuid: CPU-XX
Marketing Name: AMD EPYC 7A53 64-Core Processor
Vendor Name: CPU
Feature: None specified
Profile: FULL_PROFILE
Float Round Mode: NEAR
Max Queue Number: 0(0x0)
Queue Min Size: 0(0x0)
Queue Max Size: 0(0x0)
Queue Type: MULTI
Node: 0
Device Type: CPU
Cache Info:
L1: 32768(0x8000) KB
Chip ID: 0(0x0)
ASIC Revision: 0(0x0)
Cacheline Size: 64(0x40)
Max Clock Freq. (MHz): 2000
BDFID: 0
Internal Node ID: 0
Compute Unit: 32
SIMDs per CU: 0
Shader Engines: 0
Shader Arrs. per Eng.: 0
WatchPts on Addr. Ranges:1
Memory Properties:
Features: None
Pool Info:
Pool 1
Segment: GLOBAL; FLAGS: FINE GRAINED
Size: 131342572(0x7d420ec) KB
Allocatable: TRUE
Alloc Granule: 4KB
Alloc Recommended Granule:4KB
Alloc Alignment: 4KB
Accessible by all: TRUE
Pool 2
Segment: GLOBAL; FLAGS: KERNARG, FINE GRAINED
Size: 131342572(0x7d420ec) KB
Allocatable: TRUE
Alloc Granule: 4KB
Alloc Recommended Granule:4KB
Alloc Alignment: 4KB
Accessible by all: TRUE
Pool 3
Segment: GLOBAL; FLAGS: COARSE GRAINED
Size: 131342572(0x7d420ec) KB
Allocatable: TRUE
Alloc Granule: 4KB
Alloc Recommended Granule:4KB
Alloc Alignment: 4KB
Accessible by all: TRUE
ISA Info:
*******
Agent 2
*******
Name: AMD EPYC 7A53 64-Core Processor
Uuid: CPU-XX
Marketing Name: AMD EPYC 7A53 64-Core Processor
Vendor Name: CPU
Feature: None specified
Profile: FULL_PROFILE
Float Round Mode: NEAR
Max Queue Number: 0(0x0)
Queue Min Size: 0(0x0)
Queue Max Size: 0(0x0)
Queue Type: MULTI
Node: 1
Device Type: CPU
Cache Info:
L1: 32768(0x8000) KB
Chip ID: 0(0x0)
ASIC Revision: 0(0x0)
Cacheline Size: 64(0x40)
Max Clock Freq. (MHz): 2000
BDFID: 0
Internal Node ID: 1
Compute Unit: 32
SIMDs per CU: 0
Shader Engines: 0
Shader Arrs. per Eng.: 0
WatchPts on Addr. Ranges:1
Memory Properties:
Features: None
Pool Info:
Pool 1
Segment: GLOBAL; FLAGS: FINE GRAINED
Size: 132111288(0x7dfdbb8) KB
Allocatable: TRUE
Alloc Granule: 4KB
Alloc Recommended Granule:4KB
Alloc Alignment: 4KB
Accessible by all: TRUE
Pool 2
Segment: GLOBAL; FLAGS: KERNARG, FINE GRAINED
Size: 132111288(0x7dfdbb8) KB
Allocatable: TRUE
Alloc Granule: 4KB
Alloc Recommended Granule:4KB
Alloc Alignment: 4KB
Accessible by all: TRUE
Pool 3
Segment: GLOBAL; FLAGS: COARSE GRAINED
Size: 132111288(0x7dfdbb8) KB
Allocatable: TRUE
Alloc Granule: 4KB
Alloc Recommended Granule:4KB
Alloc Alignment: 4KB
Accessible by all: TRUE
ISA Info:
*******
Agent 3
*******
Name: AMD EPYC 7A53 64-Core Processor
Uuid: CPU-XX
Marketing Name: AMD EPYC 7A53 64-Core Processor
Vendor Name: CPU
Feature: None specified
Profile: FULL_PROFILE
Float Round Mode: NEAR
Max Queue Number: 0(0x0)
Queue Min Size: 0(0x0)
Queue Max Size: 0(0x0)
Queue Type: MULTI
Node: 2
Device Type: CPU
Cache Info:
L1: 32768(0x8000) KB
Chip ID: 0(0x0)
ASIC Revision: 0(0x0)
Cacheline Size: 64(0x40)
Max Clock Freq. (MHz): 2000
BDFID: 0
Internal Node ID: 2
Compute Unit: 32
SIMDs per CU: 0
Shader Engines: 0
Shader Arrs. per Eng.: 0
WatchPts on Addr. Ranges:1
Memory Properties:
Features: None
Pool Info:
Pool 1
Segment: GLOBAL; FLAGS: FINE GRAINED
Size: 132111292(0x7dfdbbc) KB
Allocatable: TRUE
Alloc Granule: 4KB
Alloc Recommended Granule:4KB
Alloc Alignment: 4KB
Accessible by all: TRUE
Pool 2
Segment: GLOBAL; FLAGS: KERNARG, FINE GRAINED
Size: 132111292(0x7dfdbbc) KB
Allocatable: TRUE
Alloc Granule: 4KB
Alloc Recommended Granule:4KB
Alloc Alignment: 4KB
Accessible by all: TRUE
Pool 3
Segment: GLOBAL; FLAGS: COARSE GRAINED
Size: 132111292(0x7dfdbbc) KB
Allocatable: TRUE
Alloc Granule: 4KB
Alloc Recommended Granule:4KB
Alloc Alignment: 4KB
Accessible by all: TRUE
ISA Info:
*******
Agent 4
*******
Name: AMD EPYC 7A53 64-Core Processor
Uuid: CPU-XX
Marketing Name: AMD EPYC 7A53 64-Core Processor
Vendor Name: CPU
Feature: None specified
Profile: FULL_PROFILE
Float Round Mode: NEAR
Max Queue Number: 0(0x0)
Queue Min Size: 0(0x0)
Queue Max Size: 0(0x0)
Queue Type: MULTI
Node: 3
Device Type: CPU
Cache Info:
L1: 32768(0x8000) KB
Chip ID: 0(0x0)
ASIC Revision: 0(0x0)
Cacheline Size: 64(0x40)
Max Clock Freq. (MHz): 2000
BDFID: 0
Internal Node ID: 3
Compute Unit: 32
SIMDs per CU: 0
Shader Engines: 0
Shader Arrs. per Eng.: 0
WatchPts on Addr. Ranges:1
Memory Properties:
Features: None
Pool Info:
Pool 1
Segment: GLOBAL; FLAGS: FINE GRAINED
Size: 132008820(0x7de4b74) KB
Allocatable: TRUE
Alloc Granule: 4KB
Alloc Recommended Granule:4KB
Alloc Alignment: 4KB
Accessible by all: TRUE
Pool 2
Segment: GLOBAL; FLAGS: KERNARG, FINE GRAINED
Size: 132008820(0x7de4b74) KB
Allocatable: TRUE
Alloc Granule: 4KB
Alloc Recommended Granule:4KB
Alloc Alignment: 4KB
Accessible by all: TRUE
Pool 3
Segment: GLOBAL; FLAGS: COARSE GRAINED
Size: 132008820(0x7de4b74) KB
Allocatable: TRUE
Alloc Granule: 4KB
Alloc Recommended Granule:4KB
Alloc Alignment: 4KB
Accessible by all: TRUE
ISA Info:
*******
Agent 5
*******
Name: gfx90a
Uuid: GPU-23e41c33cd99e2ca
Marketing Name: AMD Instinct MI250X
Vendor Name: AMD
Feature: KERNEL_DISPATCH
Profile: BASE_PROFILE
Float Round Mode: NEAR
Max Queue Number: 128(0x80)
Queue Min Size: 64(0x40)
Queue Max Size: 131072(0x20000)
Queue Type: MULTI
Node: 4
Device Type: GPU
Cache Info:
L1: 16(0x10) KB
L2: 8192(0x2000) KB
Chip ID: 29704(0x7408)
ASIC Revision: 1(0x1)
Cacheline Size: 64(0x40)
Max Clock Freq. (MHz): 1700
BDFID: 55552
Internal Node ID: 4
Compute Unit: 110
SIMDs per CU: 4
Shader Engines: 8
Shader Arrs. per Eng.: 1
WatchPts on Addr. Ranges:4
Coherent Host Access: TRUE
Memory Properties:
Features: KERNEL_DISPATCH
Fast F16 Operation: TRUE
Wavefront Size: 64(0x40)
Workgroup Max Size: 1024(0x400)
Workgroup Max Size per Dimension:
x 1024(0x400)
y 1024(0x400)
z 1024(0x400)
Max Waves Per CU: 32(0x20)
Max Work-item Per CU: 2048(0x800)
Grid Max Size: 4294967295(0xffffffff)
Grid Max Size per Dimension:
x 4294967295(0xffffffff)
y 4294967295(0xffffffff)
z 4294967295(0xffffffff)
Max fbarriers/Workgrp: 32
Packet Processor uCode:: 78
SDMA engine uCode:: 8
IOMMU Support:: None
Pool Info:
Pool 1
Segment: GLOBAL; FLAGS: COARSE GRAINED
Size: 67092480(0x3ffc000) KB
Allocatable: TRUE
Alloc Granule: 4KB
Alloc Recommended Granule:2048KB
Alloc Alignment: 4KB
Accessible by all: FALSE
Pool 2
Segment: GLOBAL; FLAGS: EXTENDED FINE GRAINED
Size: 67092480(0x3ffc000) KB
Allocatable: TRUE
Alloc Granule: 4KB
Alloc Recommended Granule:2048KB
Alloc Alignment: 4KB
Accessible by all: FALSE
Pool 3
Segment: GLOBAL; FLAGS: FINE GRAINED
Size: 67092480(0x3ffc000) KB
Allocatable: TRUE
Alloc Granule: 4KB
Alloc Recommended Granule:2048KB
Alloc Alignment: 4KB
Accessible by all: FALSE
Pool 4
Segment: GROUP
Size: 64(0x40) KB
Allocatable: FALSE
Alloc Granule: 0KB
Alloc Recommended Granule:0KB
Alloc Alignment: 0KB
Accessible by all: FALSE
ISA Info:
ISA 1
Name: amdgcn-amd-amdhsa--gfx90a:sramecc+:xnack-
Machine Models: HSA_MACHINE_MODEL_LARGE
Profiles: HSA_PROFILE_BASE
Default Rounding Mode: NEAR
Default Rounding Mode: NEAR
Fast f16: TRUE
Workgroup Max Size: 1024(0x400)
Workgroup Max Size per Dimension:
x 1024(0x400)
y 1024(0x400)
z 1024(0x400)
Grid Max Size: 4294967295(0xffffffff)
Grid Max Size per Dimension:
x 4294967295(0xffffffff)
y 4294967295(0xffffffff)
z 4294967295(0xffffffff)
FBarrier Max Size: 32
*** Done ***
- Post output of
AMDGPU.versioninfo()if possible.
Info: AMDGPU versioninfo
┌───────────┬──────────────────┬───────────┬──────────────────────────────────────────────────────────────────────────────────────────┐
│ Available │ Name │ Version │ Path │
├───────────┼──────────────────┼───────────┼──────────────────────────────────────────────────────────────────────────────────────────┤
│ + │ LLD │ - │ /appl/lumi/SW/LUMI-24.03/G/EB/rocm/6.2.2/lib/llvm/bin/ld.lld │
│ + │ Device Libraries │ - │ /users/barthale/.julia/artifacts/b46ab46ef568406312e5f500efb677511199c2f9/amdgcn/bitcode │
│ + │ HIP │ 6.2.41134 │ /appl/lumi/SW/LUMI-24.03/G/EB/rocm/6.2.2/lib/libamdhip64.so │
│ + │ rocBLAS │ 4.2.1 │ /appl/lumi/SW/LUMI-24.03/G/EB/rocm/6.2.2/lib/librocblas.so │
│ + │ rocSOLVER │ 3.26.0 │ /appl/lumi/SW/LUMI-24.03/G/EB/rocm/6.2.2/lib/librocsolver.so │
│ + │ rocSPARSE │ 3.2.0 │ /appl/lumi/SW/LUMI-24.03/G/EB/rocm/6.2.2/lib/librocsparse.so │
│ + │ rocRAND │ 2.10.5 │ /appl/lumi/SW/LUMI-24.03/G/EB/rocm/6.2.2/lib/librocrand.so │
│ + │ rocFFT │ 1.0.29 │ /appl/lumi/SW/LUMI-24.03/G/EB/rocm/6.2.2/lib/librocfft.so │
│ + │ MIOpen │ 3.2.0 │ /appl/lumi/SW/LUMI-24.03/G/EB/rocm/6.2.2/lib/libMIOpen.so │
└───────────┴──────────────────┴───────────┴──────────────────────────────────────────────────────────────────────────────────────────┘
Reproducing the bug
- Describe what's not working.
While experimenting for issue https://github.com/JuliaGPU/AMDGPU.jl/issues/844 , I found that allocating large (>= 4GB) with AMDGPU.ones fails. The GPU should have 64 GB available.
- Provide MWE to reproduce it (if possible).
This is the first command in a julia session:
julia> using AMDGPU; a= AMDGPU.ones(UInt8,4*1024^3);
ERROR: HIPError(code hipErrorInvalidConfiguration, invalid configuration)
Stacktrace:
[1] check
@ ~/.julia/packages/AMDGPU/TqRG0/src/hip/error.jl:145 [inlined]
[2] macro expansion
@ ~/.julia/packages/AMDGPU/TqRG0/src/utils.jl:123 [inlined]
[3] hipModuleLaunchKernel(f::AMDGPU.HIP.HIPFunction, gridDimX::UInt32, gridDimY::UInt32, gridDimZ::UInt32, blockDimX::UInt32, blockDimY::UInt32, blockDimZ::UInt32, sharedMemBytes::Int64, stream::HIPStream, kernelParams::Vector{…}, extra::Ptr{…})
@ AMDGPU.HIP ~/.julia/packages/AMDGPU/TqRG0/src/hip/libhip.jl:5120
[4] #33
@ ~/.julia/packages/AMDGPU/TqRG0/src/runtime/hip-execution.jl:130 [inlined]
[5] macro expansion
@ ~/.julia/packages/AMDGPU/TqRG0/src/runtime/hip-execution.jl:110 [inlined]
[6] macro expansion
@ ./none:0 [inlined]
[7] pack_arguments(::AMDGPU.Runtime.var"#33#34"{…}, ::AMDGPU.KernelState, ::KernelAbstractions.CompilerMetadata{…}, ::AMDGPU.Device.ROCDeviceVector{…}, ::UInt8)
@ AMDGPU.Runtime ./none:0
[8] #launch#31
@ ~/.julia/packages/AMDGPU/TqRG0/src/runtime/hip-execution.jl:124 [inlined]
[9] launch
@ ~/.julia/packages/AMDGPU/TqRG0/src/runtime/hip-execution.jl:116 [inlined]
[10] #25
@ ~/.julia/packages/AMDGPU/TqRG0/src/runtime/hip-execution.jl:85 [inlined]
[11] macro expansion
@ ~/.julia/packages/AMDGPU/TqRG0/src/runtime/hip-execution.jl:78 [inlined]
[12] macro expansion
@ ./none:0 [inlined]
[13] convert_arguments
@ ./none:0 [inlined]
[14] #roccall#23
@ ~/.julia/packages/AMDGPU/TqRG0/src/runtime/hip-execution.jl:86 [inlined]
[15] roccall
@ ~/.julia/packages/AMDGPU/TqRG0/src/runtime/hip-execution.jl:84 [inlined]
[16] macro expansion
@ ~/.julia/packages/AMDGPU/TqRG0/src/runtime/hip-execution.jl:50 [inlined]
[17] macro expansion
@ ./none:0 [inlined]
[18] #call#1
@ ./none:0 [inlined]
[19] call
@ ./none:-1 [inlined]
[20] (::AMDGPU.Runtime.HIPKernel{…})(::KernelAbstractions.CompilerMetadata{…}, ::ROCArray{…}, ::UInt8; stream::HIPStream, call_kwargs::@Kwargs{…})
@ AMDGPU.Runtime ~/.julia/packages/AMDGPU/TqRG0/src/runtime/hip-execution.jl:59
[21] (::KernelAbstractions.Kernel{…})(::ROCArray{…}, ::Vararg{…}; ndrange::Int64, workgroupsize::Nothing)
@ AMDGPU.ROCKernels ~/.julia/packages/AMDGPU/TqRG0/src/ROCKernels.jl:113
[22] Kernel
@ ~/.julia/packages/AMDGPU/TqRG0/src/ROCKernels.jl:93 [inlined]
[23] fill!(A::ROCArray{UInt8, 1, AMDGPU.Runtime.Mem.HIPBuffer}, x::UInt8)
@ GPUArrays ~/.julia/packages/GPUArrays/w335n/src/host/construction.jl:22
[24] ones(T::Type, dims::Int64)
@ AMDGPU ~/.julia/packages/AMDGPU/TqRG0/src/array.jl:253
[25] top-level scope
@ REPL[1]:1
Some type information was truncated. Use `show(err)` to see complete types.
The allocation per-se seem to work, but the error is raised in the assignement.
julia> using AMDGPU; a = ROCVector{UInt8}(undef,4*1024^3);
julia> a .= 1;
ERROR: HIPError(code hipErrorInvalidConfiguration, invalid configuration)
Stacktrace:
[1] check
@ ~/.julia/packages/AMDGPU/TqRG0/src/hip/error.jl:145 [inlined]
[2] macro expansion
@ ~/.julia/packages/AMDGPU/TqRG0/src/utils.jl:123 [inlined]
[3] hipModuleLaunchKernel(f::AMDGPU.HIP.HIPFunction, gridDimX::UInt32, gridDimY::UInt32, gridDimZ::UInt32, blockDimX::UInt32, blockDimY::UInt32, blockDimZ::UInt32, sharedMemBytes::Int64, stream::HIPStream, kernelParams::Vector{…}, extra::Ptr{…})
@ AMDGPU.HIP ~/.julia/packages/AMDGPU/TqRG0/src/hip/libhip.jl:5120
[4] #33
@ ~/.julia/packages/AMDGPU/TqRG0/src/runtime/hip-execution.jl:130 [inlined]
[5] macro expansion
@ ~/.julia/packages/AMDGPU/TqRG0/src/runtime/hip-execution.jl:110 [inlined]
[6] macro expansion
@ ./none:0 [inlined]
[7] pack_arguments(::AMDGPU.Runtime.var"#33#34"{…}, ::AMDGPU.KernelState, ::KernelAbstractions.CompilerMetadata{…}, ::AMDGPU.Device.ROCDeviceVector{…}, ::Base.Broadcast.Broadcasted{…})
@ AMDGPU.Runtime ./none:0
[8] #launch#31
@ ~/.julia/packages/AMDGPU/TqRG0/src/runtime/hip-execution.jl:124 [inlined]
[9] launch
@ ~/.julia/packages/AMDGPU/TqRG0/src/runtime/hip-execution.jl:116 [inlined]
[10] #25
@ ~/.julia/packages/AMDGPU/TqRG0/src/runtime/hip-execution.jl:85 [inlined]
[11] macro expansion
@ ~/.julia/packages/AMDGPU/TqRG0/src/runtime/hip-execution.jl:78 [inlined]
[12] macro expansion
@ ./none:0 [inlined]
[13] convert_arguments
@ ./none:0 [inlined]
[14] #roccall#23
@ ~/.julia/packages/AMDGPU/TqRG0/src/runtime/hip-execution.jl:86 [inlined]
[15] roccall
@ ~/.julia/packages/AMDGPU/TqRG0/src/runtime/hip-execution.jl:84 [inlined]
[16] macro expansion
@ ~/.julia/packages/AMDGPU/TqRG0/src/runtime/hip-execution.jl:50 [inlined]
[17] macro expansion
@ ./none:0 [inlined]
[18] #call#1
@ ./none:0 [inlined]
[19] call
@ ./none:-1 [inlined]
[20] (::AMDGPU.Runtime.HIPKernel{…})(::KernelAbstractions.CompilerMetadata{…}, ::ROCArray{…}, ::Base.Broadcast.Broadcasted{…}; stream::HIPStream, call_kwargs::@Kwargs{…})
@ AMDGPU.Runtime ~/.julia/packages/AMDGPU/TqRG0/src/runtime/hip-execution.jl:59
[21] (::KernelAbstractions.Kernel{…})(::ROCArray{…}, ::Vararg{…}; ndrange::Tuple{…}, workgroupsize::Nothing)
@ AMDGPU.ROCKernels ~/.julia/packages/AMDGPU/TqRG0/src/ROCKernels.jl:113
[22] Kernel
@ ~/.julia/packages/AMDGPU/TqRG0/src/ROCKernels.jl:93 [inlined]
[23] _copyto!
@ ~/.julia/packages/GPUArrays/w335n/src/host/broadcast.jl:71 [inlined]
[24] materialize!
@ ~/.julia/packages/GPUArrays/w335n/src/host/broadcast.jl:38 [inlined]
[25] materialize!(dest::ROCArray{…}, bc::Base.Broadcast.Broadcasted{…})
@ Base.Broadcast ./broadcast.jl:902
[26] top-level scope
@ REPL[2]:1
Some type information was truncated. Use `show(err)` to see complete types.
(@lux-dev) pkg> st
Status `/pfs/lustrep4/users/barthale/.julia/environments/lux-dev/Project.toml`
[21141c5a] AMDGPU v2.1.2
⌃ [6e4b80f9] BenchmarkTools v1.6.0
[63c18a36] KernelAbstractions v0.9.39
⌃ [b2108857] Lux v1.24.0
[872c559c] NNlib v0.9.31
[90137ffa] StaticArrays v1.9.15
[36e64239] Unroll v0.1.0
Info Packages marked with ⌃ have new versions available and may be upgradable.
Is that a known issue?
The largest array that I can allocate with ones seems to be 4*1024^3-1024:
julia> using AMDGPU; a = AMDGPU.ones(UInt8,4*1024^3-1024);
julia> using AMDGPU; a = AMDGPU.ones(UInt8,4*1024^3-1023);
ERROR: HIPError(code hipErrorInvalidConfiguration, invalid configuration)
Stacktrace:
[1] check
@ ~/.julia/packages/AMDGPU/TqRG0/src/hip/error.jl:145 [inlined]
[2] macro expansion
@ ~/.julia/packages/AMDGPU/TqRG0/src/utils.jl:123 [inlined]
[3] hipModuleLaunchKernel(f::AMDGPU.HIP.HIPFunction, gridDimX::UInt32, gridDimY::UInt32, gridDimZ::UInt32, blockDimX::UInt32, blockDimY::UInt32, blockDimZ::UInt32, sharedMemBytes::Int64, stream::HIPStream, kernelParams::Vector{…}, extra::Ptr{…})
@ AMDGPU.HIP ~/.julia/packages/AMDGPU/TqRG0/src/hip/libhip.jl:5120
[4] #33
@ ~/.julia/packages/AMDGPU/TqRG0/src/runtime/hip-execution.jl:130 [inlined]
[5] macro expansion
@ ~/.julia/packages/AMDGPU/TqRG0/src/runtime/hip-execution.jl:110 [inlined]
[6] macro expansion
@ ./none:0 [inlined]
[7] pack_arguments(::AMDGPU.Runtime.var"#33#34"{…}, ::AMDGPU.KernelState, ::KernelAbstractions.CompilerMetadata{…}, ::AMDGPU.Device.ROCDeviceVector{…}, ::UInt8)
@ AMDGPU.Runtime ./none:0
[8] #launch#31
@ ~/.julia/packages/AMDGPU/TqRG0/src/runtime/hip-execution.jl:124 [inlined]
[9] launch
@ ~/.julia/packages/AMDGPU/TqRG0/src/runtime/hip-execution.jl:116 [inlined]
[10] #25
@ ~/.julia/packages/AMDGPU/TqRG0/src/runtime/hip-execution.jl:85 [inlined]
[11] macro expansion
@ ~/.julia/packages/AMDGPU/TqRG0/src/runtime/hip-execution.jl:78 [inlined]
[12] macro expansion
@ ./none:0 [inlined]
[13] convert_arguments
@ ./none:0 [inlined]
[14] #roccall#23
@ ~/.julia/packages/AMDGPU/TqRG0/src/runtime/hip-execution.jl:86 [inlined]
[15] roccall
@ ~/.julia/packages/AMDGPU/TqRG0/src/runtime/hip-execution.jl:84 [inlined]
[16] macro expansion
@ ~/.julia/packages/AMDGPU/TqRG0/src/runtime/hip-execution.jl:50 [inlined]
[17] macro expansion
@ ./none:0 [inlined]
[18] #call#1
@ ./none:0 [inlined]
[19] call
@ ./none:-1 [inlined]
[20] (::AMDGPU.Runtime.HIPKernel{…})(::KernelAbstractions.CompilerMetadata{…}, ::ROCArray{…}, ::UInt8; stream::HIPStream, call_kwargs::@Kwargs{…})
@ AMDGPU.Runtime ~/.julia/packages/AMDGPU/TqRG0/src/runtime/hip-execution.jl:59
[21] (::KernelAbstractions.Kernel{…})(::ROCArray{…}, ::Vararg{…}; ndrange::Int64, workgroupsize::Nothing)
@ AMDGPU.ROCKernels ~/.julia/packages/AMDGPU/TqRG0/src/ROCKernels.jl:113
[22] Kernel
@ ~/.julia/packages/AMDGPU/TqRG0/src/ROCKernels.jl:93 [inlined]
[23] fill!(A::ROCArray{UInt8, 1, AMDGPU.Runtime.Mem.HIPBuffer}, x::UInt8)
@ GPUArrays ~/.julia/packages/GPUArrays/w335n/src/host/construction.jl:22
[24] ones(T::Type, dims::Int64)
@ AMDGPU ~/.julia/packages/AMDGPU/TqRG0/src/array.jl:253
[25] top-level scope
@ REPL[36]:1
Some type information was truncated. Use `show(err)` to see complete types.
julia>
The element type does not seem to be important:
julia> using AMDGPU; a = AMDGPU.ones(UInt16,4*1024^3-1024);
julia> using AMDGPU; a = AMDGPU.ones(UInt16,4*1024^3-1023);
ERROR: HIPError(code hipErrorInvalidConfiguration, invalid configuration)
Seems the issue raises when gridsize > 4194303. Running this MWE fails
using AMDGPU
A = ROCVector{UInt8}(undef, 4 * 1024^3 - 1023)
s = UInt8(1)
function myfill!(a, b)
i = workitemIdx().x + (workgroupIdx().x - 1) * workgroupDim().x
if i ≤ length(a)
a[i] = b
end
return
end
groupsize = 1024
gridsize = cld(length(A), groupsize) # 4194304
@roc groupsize = groupsize gridsize = gridsize myfill!(A, s)
AMDGPU.synchronize()
but would work if gridsize = 4194303.
A HIP reproducer works though. I am unsure if HIP actually maps the 1D launch config to 2D under the hood to avoid failure. Making this in Julia seems to solve the issue:
using AMDGPU
A = ROCVector{UInt8}(undef, 4 * 1024^3 - 1023)
s = UInt8(1)
function myfill!(a, b)
i = workitemIdx().x +
(workgroupIdx().x - 1) * workgroupDim().x +
(workgroupIdx().y - 1) * workgroupDim().x * gridGroupDim().x
if i ≤ length(a)
a[i] = b
end
return
end
groupsize = 1024
blocks = cld(length(A), groupsize)
grid_x = min(blocks, 4194303)
grid_y = cld(blocks, 4194303)
@roc groupsize = groupsize gridsize = (grid_x, grid_y) myfill!(A, s)
AMDGPU.synchronize()
EDIT: HIP C++ example suffers from the same issue as AMDGPU one, suggesting it could be an issue or limitation on HIP.
Note that if reducing the groupsize from 1024 to 1023 (or lower 512 or 256), it also works without issues (and without the need of going 2D).
For info, AMDGPU.ones calls into fill! which calls into KernelAbstractions which launches in this case a 1D kernel with groupsize = 1024 and triggers the failure.
So the one question is, why does AMDGPU fail to launch the kernel for groupsize=1024 in this case of large amounts of blocks (or gridsize)?
A HIP reproducer also fails (using hipModuleLaunchKernel to use same kernel launch approach as in AMDGPU):
- kernel code (
my kernel.hip):
#include <hip/hip_runtime.h>
extern "C" __global__ void myfill(uint8_t *a, size_t N, uint8_t b)
{
size_t i = blockIdx.x * blockDim.x + threadIdx.x;
if (i < N)
a[i] = b;
}
- host side (
hip_host.cpp):
#include <hip/hip_runtime.h>
#include <stdint.h>
#include <stdio.h>
#define CHECK(cmd) do { \
hipError_t e = cmd; \
if (e != hipSuccess) { \
fprintf(stderr, \
"HIP ERROR %s:%d: %s\n", \
__FILE__, __LINE__, hipGetErrorString(e));\
return 1; \
} \
} while (0)
int main()
{
size_t N = 4ULL * 1024ULL * 1024ULL * 1024ULL - 1023ULL;
uint8_t b = 1;
// load module
hipModule_t mod;
hipFunction_t func;
CHECK( hipModuleLoad(&mod, "mykernel.hsaco") );
CHECK( hipModuleGetFunction(&func, mod, "myfill") );
// allocate device memory
uint8_t *A;
CHECK( hipMalloc((void**)&A, N) );
// launch parameters
int groupsize = 1024;
size_t gridsize = (N + groupsize - 1) / groupsize;
void *args[] = {
&A,
&N,
&b
};
CHECK( hipModuleLaunchKernel(func,
(unsigned int)gridsize, 1, 1,
(unsigned int)groupsize, 1, 1,
0, 0,
args, NULL) );
CHECK( hipDeviceSynchronize() );
CHECK( hipFree(A) );
CHECK( hipModuleUnload(mod) );
return 0;
}
To compile and run the above:
hipcc --genco mykernel.hip -o mykernel.hsaco
hipcc hip_host.cpp -O2
./a.out
HIP ERROR hip_host.cpp:44: invalid configuration argument
Thanks or suggesting this fix from Metal.jl @christiangnrd . I checked and indeed casting the global index as Int (-> Int64) solves the issue in the explicit AMDGPU case, i.e. the following executes without error:
using AMDGPU
A = ROCVector{UInt8}(undef, 4 * 1024^3 - 1023)
s = UInt8(1)
function myfill!(a, b)
i = Int(workitemIdx().x + (workgroupIdx().x - 1) * workgroupDim().x)
if i ≤ length(a)
a[i] = b
end
return
end
groupsize = 1024
gridsize = cld(length(A), groupsize) # 4194304
@roc groupsize = groupsize gridsize = gridsize myfill!(A, s)
AMDGPU.synchronize()
Now, about the original issue arising when using fill!, which calls into GPUArrays KA backed function https://github.com/JuliaGPU/GPUArrays.jl/blob/5a83c70ae885146e0e4a29116c45445d824e8e0b/src/host/construction.jl#L12-L24, I do not see how to fix things here https://github.com/JuliaGPU/AMDGPU.jl/blob/0c3073c4349b240c8a670b08b9400561a6747b3d/src/ROCKernels.jl#L140-L144 to make it work for ROCKernels in a similar fashion.
And broadcasting will certainly suffer from a similar issue given that it calls into GPUArrays which then uses KA on ROCBackend in e.g. https://github.com/JuliaGPU/GPUArrays.jl/blob/5a83c70ae885146e0e4a29116c45445d824e8e0b/src/host/broadcast.jl#L53
Hopefully addressed by #868
From the HIP documentation:
Please note, HIP does not support kernel launch with total work items defined in dimension with size gridDim x blockDim >= 2^32. So gridDim.x * blockDim.x, gridDim.y * blockDim.y and gridDim.z * blockDim.z are always less than 2^32.
This is also the case with Metal.
From the HIP documentation:
Please note, HIP does not support kernel launch with total work items defined in dimension with size gridDim x blockDim >= 2^32. So gridDim.x * blockDim.x, gridDim.y * blockDim.y and gridDim.z * blockDim.z are always less than 2^32.
This is also the case with Metal.
So we are indeed hitting this.
As reported by @Alexander-Barth , the largest array one can allocate is of size (4 * 1024^3 - 1024) which is (2^32 - 1024). Given that KA uses 1024 threads per block in 1D in the present case for e.g. fill!, having (4 * 1024^3 - 1023) would lead to launching one more block of 1024 threads and thus hitting the size gridDim x blockDim >= 2^32 case.