diff --git a/ext-tuner/example/rccl_tuner_gfx950.conf b/ext-tuner/example/rccl_tuner_gfx950.conf new file mode 100644 index 0000000000..8cb0ac6e2b --- /dev/null +++ b/ext-tuner/example/rccl_tuner_gfx950.conf @@ -0,0 +1,132 @@ +# Optimized configurations generated by optimize_config.py +# Optimization metric: latency_us +# Format: collective_type,min_bytes,max_bytes,algorithm,protocol,channels,nNodes,nRanks,numPipeOps,regBuff +# This file contains configurations for allgather, allreduce, and reducescatter on gfx950 with 2,4,8 nodes. +# The config for allgather upto 4MB will be bypassed by rccl as rccl runtime +# forced direct implementation for msgs upto 4MB +allgather,0,1536,ring,ll128,32,2,16,-1,-1 +allgather,1537,3072,ring,ll128,48,2,16,-1,-1 +allgather,3073,6144,ring,ll,32,2,16,-1,-1 +allgather,6145,12288,ring,simple,24,2,16,-1,-1 +allgather,12289,49152,ring,ll128,56,2,16,-1,-1 +allgather,49153,98304,ring,ll128,32,2,16,-1,-1 +allgather,98305,196608,ring,ll,64,2,16,-1,-1 +allgather,196609,393216,ring,ll128,32,2,16,-1,-1 +allgather,393217,786432,ring,simple,16,2,16,-1,-1 +allgather,786433,1572864,ring,ll,48,2,16,-1,-1 +allgather,1572865,6291456,ring,simple,16,2,16,-1,-1 +allgather,6291457,12582912,ring,ll128,32,2,16,-1,-1 +allgather,12582913,25165824,ring,ll128,40,2,16,-1,-1 +allgather,25165825,100663296,ring,ll128,56,2,16,-1,-1 +allgather,100663297,201326592,ring,simple,48,2,16,-1,-1 +allgather,201326593,17179869184,ring,simple,64,2,16,-1,-1 +allgather,0,1536,ring,ll128,56,4,32,-1,-1 +allgather,1537,3072,ring,ll,32,4,32,-1,-1 +allgather,3073,6144,ring,simple,32,4,32,-1,-1 +allgather,6145,24576,ring,ll,32,4,32,-1,-1 +allgather,24577,49152,ring,ll128,40,4,32,-1,-1 +allgather,49153,98304,ring,ll,56,4,32,-1,-1 +allgather,98305,196608,ring,ll128,40,4,32,-1,-1 +allgather,196609,393216,ring,ll,32,4,32,-1,-1 +allgather,393217,786432,ring,ll128,40,4,32,-1,-1 +allgather,786433,1572864,ring,simple,24,4,32,-1,-1 +allgather,1572865,3145728,ring,ll,32,4,32,-1,-1 +allgather,3145729,6291456,ring,simple,24,4,32,-1,-1 +allgather,6291457,12582912,ring,ll,56,4,32,-1,-1 +allgather,12582913,25165824,ring,ll128,32,4,32,-1,-1 +allgather,25165825,50331648,ring,ll128,40,4,32,-1,-1 +allgather,50331649,100663296,ring,ll128,56,4,32,-1,-1 +allgather,100663297,201326592,ring,ll128,48,4,32,-1,-1 +allgather,201326593,402653184,ring,ll128,64,4,32,-1,-1 +allgather,402653185,17179869184,ring,simple,64,4,32,-1,-1 +allgather,0,3072,ring,ll128,48,8,64,-1,-1 +allgather,3073,12288,ring,ll,56,8,64,-1,-1 +allgather,12289,24576,ring,ll128,48,8,64,-1,-1 +allgather,24577,49152,ring,ll,56,8,64,-1,-1 +allgather,49153,196608,ring,ll,64,8,64,-1,-1 +allgather,196609,393216,ring,ll,56,8,64,-1,-1 +allgather,393217,786432,ring,simple,40,8,64,-1,-1 +allgather,786433,1572864,ring,ll,48,8,64,-1,-1 +allgather,1572865,3145728,ring,ll,56,8,64,-1,-1 +allgather,3145729,6291456,ring,simple,48,8,64,-1,-1 +allgather,6291457,25165824,ring,ll,48,8,64,-1,-1 +allgather,25165825,50331648,ring,ll128,32,8,64,-1,-1 +allgather,50331649,100663296,ring,ll128,40,8,64,-1,-1 +allgather,100663297,201326592,ring,ll128,56,8,64,-1,-1 +allgather,201326593,402653184,ring,ll128,48,8,64,-1,-1 +allgather,402653185,805306368,ring,ll128,64,8,64,-1,-1 +allgather,805306369,17179869184,ring,simple,64,8,64,-1,-1 +allreduce,0,3072,tree,ll,1,2,16,-1,-1 +allreduce,3073,6144,tree,ll,16,2,16,-1,-1 +allreduce,6145,12288,tree,ll,56,2,16,-1,-1 +allreduce,12289,24576,tree,ll,64,2,16,-1,-1 +allreduce,24577,49152,tree,ll,56,2,16,-1,-1 +allreduce,49153,98304,tree,ll,40,2,16,-1,-1 +allreduce,98305,393216,tree,ll,64,2,16,-1,-1 +allreduce,393217,12582912,tree,ll128,64,2,16,-1,-1 +allreduce,12582913,25165824,ring,ll128,40,2,16,-1,-1 +allreduce,25165825,50331648,ring,ll128,56,2,16,-1,-1 +allreduce,50331649,100663296,tree,simple,64,2,16,-1,-1 +allreduce,100663297,201326592,ring,simple,48,2,16,-1,-1 +allreduce,201326593,17179869184,ring,simple,64,2,16,-1,-1 +allreduce,0,6144,tree,ll,1,4,32,-1,-1 +allreduce,6145,12288,tree,ll,16,4,32,-1,-1 +allreduce,12289,24576,tree,ll,56,4,32,-1,-1 +allreduce,24577,49152,tree,ll,16,4,32,-1,-1 +allreduce,49153,98304,tree,ll,40,4,32,-1,-1 +allreduce,98305,196608,tree,ll,64,4,32,-1,-1 +allreduce,196609,393216,tree,ll128,32,4,32,-1,-1 +allreduce,393217,25165824,tree,ll128,64,4,32,-1,-1 +allreduce,25165825,50331648,ring,ll128,40,4,32,-1,-1 +allreduce,50331649,201326592,ring,ll128,56,4,32,-1,-1 +allreduce,201326593,402653184,ring,simple,40,4,32,-1,-1 +allreduce,402653185,17179869184,ring,simple,64,4,32,-1,-1 +allreduce,0,6144,tree,ll,1,8,64,-1,-1 +allreduce,6145,12288,tree,ll,4,8,64,-1,-1 +allreduce,12289,24576,tree,ll,56,8,64,-1,-1 +allreduce,24577,49152,tree,ll,40,8,64,-1,-1 +allreduce,49153,98304,tree,ll,56,8,64,-1,-1 +allreduce,98305,196608,tree,ll,64,8,64,-1,-1 +allreduce,196609,393216,tree,ll128,48,8,64,-1,-1 +allreduce,393217,50331648,tree,ll128,64,8,64,-1,-1 +allreduce,50331649,100663296,tree,simple,64,8,64,-1,-1 +allreduce,100663297,201326592,ring,ll128,56,8,64,-1,-1 +allreduce,201326593,402653184,tree,simple,48,8,64,-1,-1 +allreduce,402653185,805306368,tree,simple,64,8,64,-1,-1 +allreduce,805306369,17179869184,ring,simple,64,8,64,-1,-1 +reducescatter,0,24576,ring,ll,1,2,16,-1,-1 +reducescatter,24577,98304,ring,ll,4,2,16,-1,-1 +reducescatter,98305,196608,ring,ll,8,2,16,-1,-1 +reducescatter,196609,786432,ring,ll,16,2,16,-1,-1 +reducescatter,786433,1572864,ring,ll128,16,2,16,-1,-1 +reducescatter,1572865,6291456,ring,ll128,24,2,16,-1,-1 +reducescatter,6291457,25165824,ring,ll128,40,2,16,-1,-1 +reducescatter,25165825,50331648,ring,ll128,64,2,16,-1,-1 +reducescatter,50331649,100663296,ring,ll128,56,2,16,-1,-1 +reducescatter,100663297,201326592,ring,simple,40,2,16,-1,-1 +reducescatter,201326593,3221225472,ring,simple,64,2,16,-1,-1 +reducescatter,3221225473,6442450944,ring,simple,48,2,16,-1,-1 +reducescatter,6442450945,12884901888,ring,simple,56,2,16,-1,-1 +reducescatter,12884901889,17179869184,ring,simple,64,2,16,-1,-1 +reducescatter,0,1536,ring,ll,1,4,32,-1,-1 +reducescatter,1537,6144,ring,ll,8,4,32,-1,-1 +reducescatter,6145,12288,ring,ll,1,4,32,-1,-1 +reducescatter,12289,98304,ring,ll,4,4,32,-1,-1 +reducescatter,98305,3145728,ring,ll,8,4,32,-1,-1 +reducescatter,3145729,100663296,ring,simple,16,4,32,-1,-1 +reducescatter,100663297,201326592,ring,simple,32,4,32,-1,-1 +reducescatter,201326593,402653184,ring,simple,40,4,32,-1,-1 +reducescatter,402653185,805306368,ring,simple,56,4,32,-1,-1 +reducescatter,805306369,12884901888,ring,simple,64,4,32,-1,-1 +reducescatter,12884901889,17179869184,ring,simple,56,4,32,-1,-1 +reducescatter,0,49152,ring,ll,1,8,64,-1,-1 +reducescatter,49153,196608,ring,ll,4,8,64,-1,-1 +reducescatter,196609,393216,ring,ll,8,8,64,-1,-1 +reducescatter,393217,3145728,ring,ll,16,8,64,-1,-1 +reducescatter,3145729,25165824,ring,ll128,24,8,64,-1,-1 +reducescatter,25165825,100663296,ring,ll128,40,8,64,-1,-1 +reducescatter,100663297,201326592,ring,ll128,64,8,64,-1,-1 +reducescatter,201326593,402653184,ring,ll128,56,8,64,-1,-1 +reducescatter,402653185,805306368,ring,simple,40,8,64,-1,-1 +reducescatter,805306369,1610612736,ring,simple,56,8,64,-1,-1 +reducescatter,1610612737,17179869184,ring,simple,64,8,64,-1,-1