diff --git a/GPU/GPUTracking/Definitions/GPUParameters.json b/GPU/GPUTracking/Definitions/GPUParameters.json index e8f1c24520813..d2987bb8edd89 100644 --- a/GPU/GPUTracking/Definitions/GPUParameters.json +++ b/GPU/GPUTracking/Definitions/GPUParameters.json @@ -1,62 +1,122 @@ { "CORE": { "WARP_SIZE": { - "default": 32, - "MI100": 64, - "VEGA": 64, - "AMPERE": 32, - "TURING": 32 + "default": 32, + "MI100": 64, + "VEGA": 64, + "AMPERE": 32, + "TURING": 32 }, "THREAD_COUNT_DEFAULT": { - "default": 256, - "MI100": 256, - "VEGA": 256, - "AMPERE": 512, - "TURING": 512 + "default": 256, + "MI100": 256, + "VEGA": 256, + "AMPERE": 512, + "TURING": 512 } }, "LB": { "GPUTPCCreateTrackingData": { - "default": 256, - "MI100": [256, 7], - "VEGA": [192, 2], - "AMPERE": 384, - "TURING": 256 + "default": 256, + "MI100": [ + 256, + 7 + ], + "VEGA": [ + 192, + 2 + ], + "AMPERE": 384, + "TURING": 256 }, "GPUTPCTrackletConstructor": { - "default": 256, - "MI100": [768, 8], - "VEGA": [512, 10], - "AMPERE": [256, 2], - "TURING": [256, 2] + "default": 256, + "MI100": [ + 768, + 8 + ], + "VEGA": [ + 512, + 10 + ], + "AMPERE": [ + 256, + 2 + ], + "TURING": [ + 256, + 2 + ] }, "GPUTPCTrackletSelector": { - "default": 256, - "MI100": [384, 5], - "VEGA": [192, 10], - "AMPERE": [192, 3], - "TURING": [192, 3] + "default": 256, + "MI100": [ + 384, + 5 + ], + "VEGA": [ + 192, + 10 + ], + "AMPERE": [ + 192, + 3 + ], + "TURING": [ + 192, + 3 + ] }, "GPUTPCNeighboursFinder": { - "default": 256, - "MI100": [192, 8], - "VEGA": [960, 8], - "AMPERE": [640, 1], - "TURING": [640, 1] + "default": 256, + "MI100": [ + 192, + 8 + ], + "VEGA": [ + 960, + 8 + ], + "AMPERE": [ + 640, + 1 + ], + "TURING": [ + 640, + 1 + ] }, "GPUTPCNeighboursCleaner": { - "default": 256, - "MI100": [128, 5], - "VEGA": [384, 9], - "AMPERE": 512, - "TURING": 512 + "default": 256, + "MI100": [ + 128, + 5 + ], + "VEGA": [ + 384, + 9 + ], + "AMPERE": 512, + "TURING": 512 }, "GPUTPCExtrapolationTracking": { - "default": 256, - "MI100": [256, 7], - "VEGA": [256, 2], - "AMPERE": [128, 4], - "TURING": [192, 2] + "default": 256, + "MI100": [ + 256, + 7 + ], + "VEGA": [ + 256, + 2 + ], + "AMPERE": [ + 128, + 4 + ], + "TURING": [ + 192, + 2 + ] }, "GPUTRDTrackerKernels_gpuVersion": { "default": 512 @@ -71,32 +131,77 @@ "default": 512 }, "GPUTPCCompressionKernels_step0attached": { - "default": 256, - "MI100": [128, 1], - "VEGA": [64, 2], - "AMPERE": [64, 2], - "TURING": 128 + "default": 256, + "MI100": [ + 128, + 1 + ], + "VEGA": [ + 64, + 2 + ], + "AMPERE": [ + 64, + 2 + ], + "TURING": 128 }, "GPUTPCCompressionKernels_step1unattached": { - "default": 256, - "MI100": [512, 2], - "VEGA": [512, 2], - "AMPERE": [512, 3], - "TURING": [512, 2] + "default": 256, + "MI100": [ + 512, + 2 + ], + "VEGA": [ + 512, + 2 + ], + "AMPERE": [ + 512, + 3 + ], + "TURING": [ + 512, + 2 + ] }, "GPUTPCDecompressionKernels_step0attached": { - "default": 256, - "MI100": [128, 2], - "VEGA": [128, 2], - "AMPERE": [32, 1], - "TURING": [32, 1] + "default": 256, + "MI100": [ + 128, + 2 + ], + "VEGA": [ + 128, + 2 + ], + "AMPERE": [ + 32, + 1 + ], + "TURING": [ + 32, + 1 + ] }, "GPUTPCDecompressionKernels_step1unattached": { - "default": 256, - "MI100": [64, 2], - "VEGA": [64, 2], - "AMPERE": [32, 1], - "TURING": [32, 1] + "default": 256, + "MI100": [ + 64, + 2 + ], + "VEGA": [ + 64, + 2 + ], + "AMPERE": [ + 32, + 1 + ], + "TURING": [ + 32, + 1 + ] }, "GPUTPCDecompressionUtilKernels_sortPerSectorRow": { "default": 256 @@ -108,205 +213,304 @@ "default": 256 }, "GPUTPCCFDecodeZS": { - "default": [128, 4], - "MI100": [64, 4], - "VEGA": [64, 1], - "AMPERE": [64, 10], - "TURING": [64, 8] + "default": [ + 128, + 4 + ], + "MI100": [ + 64, + 4 + ], + "VEGA": [ + 64, + 1 + ], + "AMPERE": [ + 64, + 10 + ], + "TURING": [ + 64, + 8 + ] }, "GPUTPCCFDecodeZSLink": { - "default": "GPUCA_WARP_SIZE", - "MI100": "GPUCA_WARP_SIZE", - "VEGA": "GPUCA_WARP_SIZE", - "AMPERE": "GPUCA_WARP_SIZE", - "TURING": "GPUCA_WARP_SIZE" + "default": "GPUCA_WARP_SIZE", + "MI100": "GPUCA_WARP_SIZE", + "VEGA": "GPUCA_WARP_SIZE", + "AMPERE": "GPUCA_WARP_SIZE", + "TURING": "GPUCA_WARP_SIZE" }, "GPUTPCCFDecodeZSDenseLink": { - "default": "GPUCA_WARP_SIZE", - "MI100": ["GPUCA_WARP_SIZE", 4], - "VEGA": ["GPUCA_WARP_SIZE", 14], - "AMPERE": "GPUCA_WARP_SIZE", - "TURING": "GPUCA_WARP_SIZE" + "default": "GPUCA_WARP_SIZE", + "MI100": [ + "GPUCA_WARP_SIZE", + 4 + ], + "VEGA": [ + "GPUCA_WARP_SIZE", + 14 + ], + "AMPERE": "GPUCA_WARP_SIZE", + "TURING": "GPUCA_WARP_SIZE" }, "GPUTPCCFGather": { - "default": [1024, 1], - "MI100": [1024, 5], - "VEGA": [1024, 1], - "AMPERE": [1024, 1], - "TURING": [1024, 1] + "default": [ + 1024, + 1 + ], + "MI100": [ + 1024, + 5 + ], + "VEGA": [ + 1024, + 1 + ], + "AMPERE": [ + 1024, + 1 + ], + "TURING": [ + 1024, + 1 + ] }, "COMPRESSION_GATHER": { - "default": 1024, - "MI100": 1024, - "VEGA": 1024, - "AMPERE": 1024, - "TURING": 1024 + "default": 1024, + "MI100": 1024, + "VEGA": 1024, + "AMPERE": 1024, + "TURING": 1024 }, "GPUTPCGMMergerTrackFit": { - "default": 256, - "MI100": [192, 2], - "VEGA": [64, 7], - "AMPERE": [64, 4], - "TURING": [32, 8] + "default": 256, + "MI100": [ + 192, + 2 + ], + "VEGA": [ + 64, + 7 + ], + "AMPERE": [ + 64, + 4 + ], + "TURING": [ + 32, + 8 + ] }, "GPUTPCGMMergerFollowLoopers": { - "default": 256, - "MI100": [256, 5], - "VEGA": [256, 4], - "AMPERE": [64, 12], - "TURING": [128, 4] + "default": 256, + "MI100": [ + 256, + 5 + ], + "VEGA": [ + 256, + 4 + ], + "AMPERE": [ + 64, + 12 + ], + "TURING": [ + 128, + 4 + ] }, "GPUTPCGMMergerSectorRefit": { - "default": 256, - "MI100": [64, 4], - "VEGA": [256, 2], - "AMPERE": [32, 6], - "TURING": [64, 5] + "default": 256, + "MI100": [ + 64, + 4 + ], + "VEGA": [ + 256, + 2 + ], + "AMPERE": [ + 32, + 6 + ], + "TURING": [ + 64, + 5 + ] }, "GPUTPCGMMergerUnpackResetIds": { - "default": 256, - "MI100": 256, - "VEGA": 256, - "AMPERE": 256, - "TURING": 256 + "default": 256, + "MI100": 256, + "VEGA": 256, + "AMPERE": 256, + "TURING": 256 }, "GPUTPCGMMergerUnpackGlobal": { - "default": 256, - "MI100": 256, - "VEGA": 256, - "AMPERE": 256, - "TURING": 256 + "default": 256, + "MI100": 256, + "VEGA": 256, + "AMPERE": 256, + "TURING": 256 }, "GPUTPCGMMergerResolve_step0": { - "default": 256, - "MI100": 512, - "VEGA": 256, - "AMPERE": 256, - "TURING": 256 + "default": 256, + "MI100": 512, + "VEGA": 256, + "AMPERE": 256, + "TURING": 256 }, "GPUTPCGMMergerResolve_step1": { - "default": 256, - "MI100": 512, - "VEGA": 256, - "AMPERE": 256, - "TURING": 256 + "default": 256, + "MI100": 512, + "VEGA": 256, + "AMPERE": 256, + "TURING": 256 }, "GPUTPCGMMergerResolve_step2": { - "default": 256, - "MI100": 512, - "VEGA": 256, - "AMPERE": 256, - "TURING": 256 + "default": 256, + "MI100": 512, + "VEGA": 256, + "AMPERE": 256, + "TURING": 256 }, "GPUTPCGMMergerResolve_step3": { - "default": 256, - "MI100": 512, - "VEGA": 256, - "AMPERE": 256, - "TURING": 256 + "default": 256, + "MI100": 512, + "VEGA": 256, + "AMPERE": 256, + "TURING": 256 }, "GPUTPCGMMergerResolve_step4": { - "default": 256, - "MI100": 512, - "VEGA": 256, - "AMPERE": [256, 4], - "TURING": [256, 4] + "default": 256, + "MI100": 512, + "VEGA": 256, + "AMPERE": [ + 256, + 4 + ], + "TURING": [ + 256, + 4 + ] }, "GPUTPCGMMergerClearLinks": { - "default": 256, - "MI100": 256, - "VEGA": 256, - "AMPERE": 256, - "TURING": 256 + "default": 256, + "MI100": 256, + "VEGA": 256, + "AMPERE": 256, + "TURING": 256 }, "GPUTPCGMMergerMergeWithinPrepare": { - "default": 256, - "MI100": 256, - "VEGA": 256, - "AMPERE": 256, - "TURING": 256 + "default": 256, + "MI100": 256, + "VEGA": 256, + "AMPERE": 256, + "TURING": 256 }, "GPUTPCGMMergerMergeSectorsPrepare": { - "default": 256, - "MI100": 256, - "VEGA": 256, - "AMPERE": [256, 2], - "TURING": [256, 2] + "default": 256, + "MI100": 256, + "VEGA": 256, + "AMPERE": [ + 256, + 2 + ], + "TURING": [ + 256, + 2 + ] }, "GPUTPCGMMergerMergeBorders_step0": { - "default": 256, - "MI100": 512, - "VEGA": 256, - "AMPERE": 192, - "TURING": 192 + "default": 256, + "MI100": 512, + "VEGA": 256, + "AMPERE": 192, + "TURING": 192 }, "GPUTPCGMMergerMergeBorders_step2": { - "default": 256, - "MI100": 512, - "VEGA": 256, - "AMPERE": [64, 2], - "TURING": 256 + "default": 256, + "MI100": 512, + "VEGA": 256, + "AMPERE": [ + 64, + 2 + ], + "TURING": 256 }, "GPUTPCGMMergerMergeCE": { - "default": 256, - "MI100": 512, - "VEGA": 256, - "AMPERE": 256, - "TURING": 256 + "default": 256, + "MI100": 512, + "VEGA": 256, + "AMPERE": 256, + "TURING": 256 }, "GPUTPCGMMergerLinkExtrapolatedTracks": { - "default": 256, - "MI100": 256, - "VEGA": 256, - "AMPERE": 256, - "TURING": 256 + "default": 256, + "MI100": 256, + "VEGA": 256, + "AMPERE": 256, + "TURING": 256 }, "GPUTPCGMMergerCollect": { - "default": 256, - "MI100": [768, 1], - "VEGA": [1024, 1], - "AMPERE": [256, 2], - "TURING": [128, 2] + "default": 256, + "MI100": [ + 768, + 1 + ], + "VEGA": [ + 1024, + 1 + ], + "AMPERE": [ + 256, + 2 + ], + "TURING": [ + 128, + 2 + ] }, "GPUTPCGMMergerSortTracksPrepare": { - "default": 256, - "MI100": 256, - "VEGA": 256, - "AMPERE": 256, - "TURING": 256 + "default": 256, + "MI100": 256, + "VEGA": 256, + "AMPERE": 256, + "TURING": 256 }, "GPUTPCGMMergerPrepareForFit_step0": { - "default": 256, - "MI100": 256, - "VEGA": 256, - "AMPERE": 256, - "TURING": 256 + "default": 256, + "MI100": 256, + "VEGA": 256, + "AMPERE": 256, + "TURING": 256 }, "GPUTPCGMMergerPrepareForFit_step1": { - "default": 256, - "MI100": 256, - "VEGA": 256, - "AMPERE": 256, - "TURING": 256 + "default": 256, + "MI100": 256, + "VEGA": 256, + "AMPERE": 256, + "TURING": 256 }, "GPUTPCGMMergerPrepareForFit_step2": { - "default": 256, - "MI100": 256, - "VEGA": 256, - "AMPERE": 256, - "TURING": 256 + "default": 256, + "MI100": 256, + "VEGA": 256, + "AMPERE": 256, + "TURING": 256 }, "GPUTPCGMMergerFinalize_step0": { - "default": 256, - "VEGA": 256 + "default": 256, + "VEGA": 256 }, "GPUTPCGMMergerFinalize_step1": { - "default": 256, - "VEGA": 256 + "default": 256, + "VEGA": 256 }, "GPUTPCGMMergerFinalize_step2": { - "default": 256, - "VEGA": 256 + "default": 256, + "VEGA": 256 }, "GPUTPCGMMergerMergeLoopers_step0": { "default": 256 @@ -324,66 +528,111 @@ "default": 256 }, "GPUTPCStartHitsFinder": { - "default": 256, - "MI100": [1024, 2], - "VEGA": [1024, 7], - "AMPERE": 512, - "TURING": 512 + "default": 256, + "MI100": [ + 1024, + 2 + ], + "VEGA": [ + 1024, + 7 + ], + "AMPERE": 512, + "TURING": 512 }, "GPUTPCStartHitsSorter": { - "default": 256, - "MI100": [1024, 5], - "VEGA": [512, 7], - "AMPERE": [512, 1], - "TURING": [512, 1] + "default": 256, + "MI100": [ + 1024, + 5 + ], + "VEGA": [ + 512, + 7 + ], + "AMPERE": [ + 512, + 1 + ], + "TURING": [ + 512, + 1 + ] }, "GPUTPCCFCheckPadBaseline": { - "default": 64, - "MI100": [64, 10], - "VEGA": [64, 2], - "AMPERE": [64, 8] + "default": 64, + "MI100": [ + 64, + 10 + ], + "VEGA": [ + 64, + 2 + ], + "AMPERE": [ + 64, + 8 + ] }, "GPUTPCCFChargeMapFiller_fillIndexMap": { - "default": 512, - "MI100": 512, - "VEGA": 512, - "AMPERE": 448 + "default": 512, + "MI100": 512, + "VEGA": 512, + "AMPERE": 448 }, "GPUTPCCFChargeMapFiller_fillFromDigits": { - "default": 512, - "MI100": 512, - "VEGA": 512, - "AMPERE": 448 + "default": 512, + "MI100": 512, + "VEGA": 512, + "AMPERE": 448 }, "GPUTPCCFChargeMapFiller_findFragmentStart": { - "default": 512, - "MI100": 512, - "VEGA": 512, - "AMPERE": 448 + "default": 512, + "MI100": 512, + "VEGA": 512, + "AMPERE": 448 }, "GPUTPCCFPeakFinder": { - "default": 512, - "MI100": [512, 9], - "VEGA": [512, 4], - "AMPERE": 128 + "default": 512, + "MI100": [ + 512, + 9 + ], + "VEGA": [ + 512, + 4 + ], + "AMPERE": 128 }, "GPUTPCCFNoiseSuppression": { - "default": 512, - "MI100": 512, - "VEGA": 512, - "AMPERE": 448 + "default": 512, + "MI100": 512, + "VEGA": 512, + "AMPERE": 448 }, "GPUTPCCFDeconvolution": { - "default": 512, - "MI100": [512, 5], - "VEGA": [512, 5], - "AMPERE": 384 + "default": 512, + "MI100": [ + 512, + 5 + ], + "VEGA": [ + 512, + 5 + ], + "AMPERE": 384 }, "GPUTPCCFClusterizer": { - "default": 512, - "MI100": [448, 3], - "VEGA": [512, 2], - "AMPERE": 448 + "default": 512, + "MI100": [ + 448, + 3 + ], + "VEGA": [ + 512, + 2 + ], + "AMPERE": 448 }, "GPUTPCNNClusterizerKernels": { "default": 512 @@ -395,10 +644,16 @@ "default": 256 }, "GPUMemClean16": { - "default": ["GPUCA_THREAD_COUNT_DEFAULT", 1] + "default": [ + "GPUCA_THREAD_COUNT_DEFAULT", + 1 + ] }, "GPUitoa": { - "default": ["GPUCA_THREAD_COUNT_DEFAULT", 1] + "default": [ + "GPUCA_THREAD_COUNT_DEFAULT", + 1 + ] }, "GPUTPCCFNoiseSuppression_noiseSuppression": { "default": "GPUCA_LB_GPUTPCCFNoiseSuppression" @@ -461,122 +716,122 @@ "default": "GPUCA_LB_COMPRESSION_GATHER" }, "GPUTPCGMMergerFinalize_0": { - "default": 256, - "MI100": 256, - "AMPERE": 256, - "TURING": 256 + "default": 256, + "MI100": 256, + "AMPERE": 256, + "TURING": 256 }, "GPUTPCGMMergerFinalize_1": { - "default": 256, - "MI100": 256, - "AMPERE": 256, - "TURING": 256 + "default": 256, + "MI100": 256, + "AMPERE": 256, + "TURING": 256 }, "GPUTPCGMMergerFinalize_2": { - "default": 256, - "MI100": 256, - "AMPERE": 256, - "TURING": 256 + "default": 256, + "MI100": 256, + "AMPERE": 256, + "TURING": 256 } }, "PAR": { "AMD_EUS_PER_CU": { - "default": 0, - "default_cpu": 0, - "MI100": 4, - "VEGA": 4 + "default": 0, + "default_cpu": 0, + "MI100": 4, + "VEGA": 4 }, "SORT_STARTHITS": { - "default": 1, - "default_cpu": 0 + "default": 1, + "default_cpu": 0 }, "NEIGHBOURS_FINDER_MAX_NNEIGHUP": { - "default": 6, - "default_cpu": 0, - "MI100": 10, - "VEGA": 4, - "AMPERE": 4, - "TURING": 4 + "default": 6, + "default_cpu": 0, + "MI100": 10, + "VEGA": 4, + "AMPERE": 4, + "TURING": 4 }, "NEIGHBOURS_FINDER_UNROLL_GLOBAL": { - "default": 4, - "default_cpu": 0, - "MI100": 4, - "VEGA": 2 + "default": 4, + "default_cpu": 0, + "MI100": 4, + "VEGA": 2 }, "NEIGHBOURS_FINDER_UNROLL_SHARED": { - "default": 1, - "default_cpu": 0, - "MI100": 0, - "VEGA": 0 + "default": 1, + "default_cpu": 0, + "MI100": 0, + "VEGA": 0 }, "TRACKLET_SELECTOR_HITS_REG_SIZE": { - "default": 12, - "default_cpu": 0, - "MI100": 9, - "VEGA": 27, - "AMPERE": 20, - "TURING": 20 + "default": 12, + "default_cpu": 0, + "MI100": 9, + "VEGA": 27, + "AMPERE": 20, + "TURING": 20 }, "ALTERNATE_BORDER_SORT": { - "default": 0, - "default_cpu": 0, - "MI100": 1, - "VEGA": 1, - "AMPERE": 1, - "TURING": 1 + "default": 0, + "default_cpu": 0, + "MI100": 1, + "VEGA": 1, + "AMPERE": 1, + "TURING": 1 }, "SORT_BEFORE_FIT": { - "default": 0, - "default_cpu": 0, - "MI100": 1, - "VEGA": 1, - "AMPERE": 1, - "TURING": 1 + "default": 0, + "default_cpu": 0, + "MI100": 1, + "VEGA": 1, + "AMPERE": 1, + "TURING": 1 }, "NO_ATOMIC_PRECHECK": { - "default": 0, - "default_cpu": 0, - "MI100": 1, - "VEGA": 1, - "AMPERE": 1, - "TURING": 1 + "default": 0, + "default_cpu": 0, + "MI100": 1, + "VEGA": 1, + "AMPERE": 1, + "TURING": 1 }, "DEDX_STORAGE_TYPE": { - "default": "float", - "default_cpu": "float", - "MI100": "uint16_t", - "VEGA": "uint16_t", - "AMPERE": "uint16_t", - "TURING": "uint16_t" + "default": "float", + "default_cpu": "float", + "MI100": "uint16_t", + "VEGA": "uint16_t", + "AMPERE": "uint16_t", + "TURING": "uint16_t" }, "MERGER_INTERPOLATION_ERROR_TYPE": { - "default": "float", - "default_cpu": "float", - "MI100": "half", - "VEGA": "half", - "AMPERE": "half", - "TURING": "half" + "default": "float", + "default_cpu": "float", + "MI100": "half", + "VEGA": "half", + "AMPERE": "half", + "TURING": "half" }, "COMP_GATHER_KERNEL": { - "default": 0, - "default_cpu": 0, - "MI100": 4, - "VEGA": 4, - "AMPERE": 4, - "TURING": 4 + "default": 0, + "default_cpu": 0, + "MI100": 4, + "VEGA": 4, + "AMPERE": 4, + "TURING": 4 }, "COMP_GATHER_MODE": { - "default": 2, - "default_cpu": 0, - "MI100": 3, - "VEGA": 3, - "AMPERE": 3, - "TURING": 3 + "default": 2, + "default_cpu": 0, + "MI100": 3, + "VEGA": 3, + "AMPERE": 3, + "TURING": 3 }, "CF_SCAN_WORKGROUP_SIZE": { - "default": 512, - "default_cpu": 0 + "default": 512, + "default_cpu": 0 } } }