llnl · amagela · Apr 14, 2026 · Mar 31, 2026 · Apr 2, 2026 · Apr 9, 2026
diff --git a/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_DIFFUSION3DPA_bandwidth_raw.csv b/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_DIFFUSION3DPA_bandwidth_raw.csv
@@ -0,0 +1,12 @@
+Problem size,Base_CUDA-block_64,Base_Seq-default,RAJA_CUDA-block_64
+405,12.0974,8.79032,12.8804
+1566,46.2209,9.0806,49.3488
+6183,181.159,9.20569,189.324
+12339,342.937,9.06822,356.197
+24678,626.709,8.08964,656.577
+49329,971.099,7.8621,1058.14
+98631,1417.29,8.33524,1371.79
+197262,1662.67,7.19463,1655.15
+394524,1623.27,5.15081,1670.4
+577935,1724.22,5.04846,1773.47
+1155843,1887.69,4.83564,1913.84
diff --git a/...3_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_DIFFUSION3DPA_bandwidth_smoothed.csv b/...3_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_DIFFUSION3DPA_bandwidth_smoothed.csv
@@ -0,0 +1,12 @@
+Problem size,Base_CUDA-block_64,Base_Seq-default,RAJA_CUDA-block_64
+405,46.2209,9.0806,49.3488
+1566,113.68995,9.07441,119.3364
+6183,181.159,9.06822,189.324
+12339,342.937,9.06822,356.197
+24678,626.709,8.33524,656.577
+49329,971.099,8.08964,1058.14
+98631,1417.29,7.8621,1371.79
+197262,1623.27,7.19463,1655.15
+394524,1662.67,5.15081,1670.4
+577935,1693.4450000000002,5.099635,1721.935
+1155843,1724.22,5.04846,1773.47
diff --git a/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_DIFFUSION3DPA_flops_raw.csv b/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_DIFFUSION3DPA_flops_raw.csv
@@ -0,0 +1,12 @@
+Problem size,Base_CUDA-block_64,Base_Seq-default,RAJA_CUDA-block_64
+405,18.0915,13.1459,19.2625
+1566,69.2993,13.6146,73.989
+6183,271.794,13.8113,284.043
+12339,514.567,13.6066,534.463
+24678,940.413,12.139,985.231
+49329,1457.23,11.7979,1587.84
+98631,2126.82,12.508,2058.54
+197262,2495.05,10.7965,2483.77
+394524,2435.93,7.72949,2506.66
+577935,2587.42,7.57591,2661.33
+1155843,2832.74,7.25656,2871.99
diff --git a/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_DIFFUSION3DPA_flops_smoothed.csv b/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_DIFFUSION3DPA_flops_smoothed.csv
@@ -0,0 +1,12 @@
+Problem size,Base_CUDA-block_64,Base_Seq-default,RAJA_CUDA-block_64
+405,69.2993,13.6146,73.989
+1566,170.54665,13.6106,179.01600000000002
+6183,271.794,13.6066,284.043
+12339,514.567,13.6066,534.463
+24678,940.413,12.508,985.231
+49329,1457.23,12.139,1587.84
+98631,2126.82,11.7979,2058.54
+197262,2435.93,10.7965,2483.77
+394524,2495.05,7.72949,2506.66
+577935,2541.235,7.6527,2583.995
+1155843,2587.42,7.57591,2661.33
diff --git a/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_EDGE3D_bandwidth_raw.csv b/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_EDGE3D_bandwidth_raw.csv
@@ -0,0 +1,12 @@
+Problem size,Base_CUDA-block_256,Base_Seq-default,RAJA_CUDA-block_256
+512,0.197104,0.0229243,0.196309
+3375,0.802097,0.0221669,0.801571
+17576,2.64385,0.0217854,2.65177
+35937,2.7385,0.021533,2.74046
+79507,3.30983,0.021455,3.29378
+166375,3.43745,0.0213676,3.42995
+343000,3.568,0.0213174,3.55459
+704969,3.6228,0.0213042,3.60819
+1442897,3.65749,0.0211735,3.64609
+2146689,3.67155,0.0212504,3.66075
+4410944,3.67202,0.0211402,3.65886
diff --git a/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_EDGE3D_bandwidth_smoothed.csv b/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_EDGE3D_bandwidth_smoothed.csv
@@ -0,0 +1,12 @@
+Problem size,Base_CUDA-block_256,Base_Seq-default,RAJA_CUDA-block_256
+512,0.802097,0.0221669,0.801571
+3375,1.7229735,0.02197615,1.7266705
+17576,2.64385,0.0217854,2.65177
+35937,2.7385,0.021533,2.74046
+79507,3.30983,0.021455,3.29378
+166375,3.43745,0.0213676,3.42995
+343000,3.568,0.0213174,3.55459
+704969,3.6228,0.0213042,3.60819
+1442897,3.65749,0.0212504,3.64609
+2146689,3.66452,0.02121195,3.652475
+4410944,3.67155,0.0211735,3.65886
diff --git a/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_EDGE3D_flops_raw.csv b/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_EDGE3D_flops_raw.csv
@@ -0,0 +1,12 @@
+Problem size,Base_CUDA-block_256,Base_Seq-default,RAJA_CUDA-block_256
+512,80.6755,9.38302,80.35
+3375,339.719,9.38853,339.496
+17576,1140.57,9.39835,1143.99
+35937,1188.01,9.34143,1188.86
+79507,1442.95,9.35348,1435.95
+166375,1503.98,9.34894,1500.7
+343000,1565.47,9.35308,1559.59
+704969,1593.03,9.36795,1586.61
+1442897,1611.09,9.32671,1606.07
+2146689,1618.58,9.36815,1613.82
+4410944,1620.76,9.33091,1614.96
diff --git a/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_EDGE3D_flops_smoothed.csv b/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_EDGE3D_flops_smoothed.csv
@@ -0,0 +1,12 @@
+Problem size,Base_CUDA-block_256,Base_Seq-default,RAJA_CUDA-block_256
+512,339.719,9.38853,339.496
+3375,740.1445,9.385774999999999,741.7429999999999
+17576,1140.57,9.38302,1143.99
+35937,1188.01,9.35348,1188.86
+79507,1442.95,9.35308,1435.95
+166375,1503.98,9.35308,1500.7
+343000,1565.47,9.35308,1559.59
+704969,1593.03,9.35308,1586.61
+1442897,1611.09,9.35308,1606.07
+2146689,1614.835,9.34943,1609.945
+4410944,1618.58,9.33091,1613.82
diff --git a/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_ENERGY_bandwidth_raw.csv b/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_ENERGY_bandwidth_raw.csv
@@ -0,0 +1,12 @@
+Problem size,Base_CUDA-block_256,Base_Seq-default,RAJA_CUDA-block_256
+417,6.49675,48.7956,6.50062
+1667,25.3733,47.0544,25.255
+6667,100.275,47.0702,99.7526
+13334,199.931,47.2522,199.262
+26667,391.454,48.6333,390.714
+53334,727.103,37.0917,724.576
+106667,1325.45,33.7378,1326.25
+213334,2240.46,32.8315,2251.23
+426667,3105.64,30.7464,3097.01
+625000,2943.18,26.281,2934.85
+1250000,2643.02,18.7451,2646.71
diff --git a/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_ENERGY_bandwidth_smoothed.csv b/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_ENERGY_bandwidth_smoothed.csv
@@ -0,0 +1,12 @@
+Problem size,Base_CUDA-block_256,Base_Seq-default,RAJA_CUDA-block_256
+417,25.3733,47.0702,25.255
+1667,62.82415,47.1612,62.5038
+6667,100.275,47.2522,99.7526
+13334,199.931,47.0702,199.262
+26667,391.454,47.0702,390.714
+53334,727.103,37.0917,724.576
+106667,1325.45,33.7378,1326.25
+213334,2240.46,32.8315,2251.23
+426667,2643.02,30.7464,2646.71
+625000,2793.1,28.5137,2790.7799999999997
+1250000,2943.18,26.281,2934.85
diff --git a/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_ENERGY_flops_raw.csv b/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_ENERGY_flops_raw.csv
@@ -0,0 +1,12 @@
+Problem size,Base_CUDA-block_256,Base_Seq-default,RAJA_CUDA-block_256
+417,1.77625,13.341,1.77731
+1667,6.93723,12.865,6.90487
+6667,27.4158,12.8693,27.273
+13334,54.6624,12.9191,54.4796
+26667,107.026,13.2967,106.824
+53334,198.795,10.1411,198.104
+106667,362.386,9.22412,362.606
+213334,612.557,8.97634,615.5
+426667,849.103,8.40628,846.742
+625000,804.683,7.18539,802.407
+1250000,722.619,5.12504,723.629
diff --git a/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_ENERGY_flops_smoothed.csv b/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_ENERGY_flops_smoothed.csv
@@ -0,0 +1,12 @@
+Problem size,Base_CUDA-block_256,Base_Seq-default,RAJA_CUDA-block_256
+417,6.93723,12.8693,6.90487
+1667,17.176515000000002,12.894200000000001,17.088935
+6667,27.4158,12.9191,27.273
+13334,54.6624,12.8693,54.4796
+26667,107.026,12.8693,106.824
+53334,198.795,10.1411,198.104
+106667,362.386,9.22412,362.606
+213334,612.557,8.97634,615.5
+426667,722.619,8.40628,723.629
+625000,763.6510000000001,7.795835,763.018
+1250000,804.683,7.18539,802.407
diff --git a/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_FEMSWEEP_bandwidth_raw.csv b/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_FEMSWEEP_bandwidth_raw.csv
@@ -0,0 +1,11 @@
+Problem size,Base_CUDA-block_128,Base_Seq-default,RAJA_CUDA-block_128
+31104,14.928049999999999,0.6055889999999999,10.06305
+73728,24.5073,0.5725,16.3838
+144000,34.1906,0.578163,23.3096
+248832,44.7641,0.599649,30.9269
+395136,46.0821,0.602929,36.1339
+589824,53.0155,0.594489,42.4744
+1152000,58.4488,0.566533,48.8716
+1533312,61.8105,0.521593,52.502
+1990656,65.9083,0.509787,55.1352
+2530944,66.4358,0.498932,57.3794
diff --git a/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_FEMSWEEP_bandwidth_smoothed.csv b/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_FEMSWEEP_bandwidth_smoothed.csv
@@ -0,0 +1,11 @@
+Problem size,Base_CUDA-block_128,Base_Seq-default,RAJA_CUDA-block_128
+31104,19.745,0.586811,13.2291
+73728,24.5073,0.595459,16.3838
+144000,34.1906,0.599649,23.3096
+248832,44.7641,0.594489,30.9269
+395136,46.0821,0.594489,36.1339
+589824,53.0155,0.594489,42.4744
+1152000,58.4488,0.566533,48.8716
+1533312,61.8105,0.521593,52.502
+1990656,63.859399999999994,0.51569,53.8186
+2530944,65.9083,0.509787,55.1352
diff --git a/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_FEMSWEEP_flops_raw.csv b/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_FEMSWEEP_flops_raw.csv
@@ -0,0 +1,11 @@
+Problem size,Base_CUDA-block_128,Base_Seq-default,RAJA_CUDA-block_128
+31104,83.12595,3.37219,56.03545
+73728,132.791,3.10205,88.7746
+144000,182.372,3.08391,124.333
+248832,236.347,3.16604,163.289
+395136,241.569,3.16064,189.419
+589824,276.445,3.09991,221.479
+1152000,302.549,2.93255,252.975
+1533312,319.107,2.69281,271.05
+1990656,339.519,2.62611,284.022
+2530944,341.606,2.56545,295.038
diff --git a/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_FEMSWEEP_flops_smoothed.csv b/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_FEMSWEEP_flops_smoothed.csv
@@ -0,0 +1,11 @@
+Problem size,Base_CUDA-block_128,Base_Seq-default,RAJA_CUDA-block_128
+31104,108.1106,3.208915,72.43655000000001
+73728,132.791,3.16604,88.7746
+144000,182.372,3.16064,124.333
+248832,236.347,3.10205,163.289
+395136,241.569,3.09991,189.419
+589824,276.445,3.09991,221.479
+1152000,302.549,2.93255,252.975
+1533312,319.107,2.69281,271.05
+1990656,329.313,2.65946,277.536
+2530944,339.519,2.62611,284.022
diff --git a/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_INTSC_HEXRECT_bandwidth_raw.csv b/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_INTSC_HEXRECT_bandwidth_raw.csv
@@ -0,0 +1,12 @@
+Problem size,Base_CUDA-block_64,Base_Seq-default,RAJA_CUDA-block_64
+1728,0.511948,0.0515405,0.517215
+5832,1.71116,0.0532159,1.72196
+17576,4.83834,0.0532582,4.94421
+39304,9.85649,0.0530228,10.1052
+74088,15.1717,0.0530315,11.2695
+140608,15.8649,0.0529853,13.8482
+287496,17.2826,0.0527919,14.9171
+551368,18.8845,0.0527725,15.7543
+1124864,19.8885,0.0527279,16.5051
+1643032,20.0555,0.0526587,16.6066
+3241792,20.4644,0.0526191,16.9192
diff --git a/...3_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_INTSC_HEXRECT_bandwidth_smoothed.csv b/...3_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_INTSC_HEXRECT_bandwidth_smoothed.csv
@@ -0,0 +1,12 @@
+Problem size,Base_CUDA-block_64,Base_Seq-default,RAJA_CUDA-block_64
+1728,1.71116,0.0532159,1.72196
+5832,3.27475,0.053119349999999996,3.333085
+17576,4.83834,0.0530315,4.94421
+39304,9.85649,0.0530315,10.1052
+74088,15.1717,0.0530228,11.2695
+140608,15.8649,0.0529853,13.8482
+287496,17.2826,0.0527919,14.9171
+551368,18.8845,0.0527725,15.7543
+1124864,19.8885,0.0527279,16.5051
+1643032,19.972,0.0526933,16.55585
+3241792,20.0555,0.0526587,16.6066
diff --git a/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_INTSC_HEXRECT_flops_raw.csv b/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_INTSC_HEXRECT_flops_raw.csv
@@ -0,0 +1,12 @@
+Problem size,Base_CUDA-block_64,Base_Seq-default,RAJA_CUDA-block_64
+1728,40.5816,4.08557,40.9991
+5832,137.471,4.27525,138.338
+17576,391.7,4.31165,400.271
+39304,801.073,4.30936,821.286
+74088,1235.97,4.32023,918.07
+140608,1294.89,4.32463,1130.28
+287496,1412.94,4.31601,1219.55
+551368,1545.74,4.31956,1289.52
+1124864,1629.6,4.32034,1352.37
+1643032,1644.02,4.31662,1361.3
+3241792,1678.67,4.31628,1387.86
diff --git a/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_INTSC_HEXRECT_flops_smoothed.csv b/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_INTSC_HEXRECT_flops_smoothed.csv
@@ -0,0 +1,12 @@
+Problem size,Base_CUDA-block_64,Base_Seq-default,RAJA_CUDA-block_64
+1728,137.471,4.27525,138.338
+5832,264.5855,4.292305,269.3045
+17576,391.7,4.30936,400.271
+39304,801.073,4.31165,821.286
+74088,1235.97,4.31601,918.07
+140608,1294.89,4.31956,1130.28
+287496,1412.94,4.32023,1219.55
+551368,1545.74,4.31956,1289.52
+1124864,1629.6,4.31662,1352.37
+1643032,1636.81,4.31809,1356.835
+3241792,1644.02,4.31662,1361.3
diff --git a/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_MASS3DEA_bandwidth_raw.csv b/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_MASS3DEA_bandwidth_raw.csv
@@ -0,0 +1,12 @@
+Problem size,Base_CUDA-compile_time_block_stride_loop_64,Base_Seq-default,RAJA_CUDA-cached_block_stride_loop_64,RAJA_CUDA-compile_time_block_stride_loop_64
+196608,13.5704,0.189198,19.1632,20.2873
+389120,26.8231,0.189853,38.096,40.0042
+778240,53.0688,0.192069,75.3143,79.797
+1556480,94.9269,0.191844,108.856,117.855
+3108864,162.839,0.187071,152.047,164.577
+6213632,179.945,0.186147,175.507,192.851
+9101312,188.677,0.185918,180.851,198.776
+18198528,218.752,0.185614,201.752,224.806
+24260608,220.49,0.184607,201.658,224.953
+30326784,222.162,0.186821,202.225,225.791
+36392960,228.994,0.184956,208.098,214.629
diff --git a/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_MASS3DEA_bandwidth_smoothed.csv b/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_MASS3DEA_bandwidth_smoothed.csv
@@ -0,0 +1,12 @@
+Problem size,Base_CUDA-compile_time_block_stride_loop_64,Base_Seq-default,RAJA_CUDA-cached_block_stride_loop_64,RAJA_CUDA-compile_time_block_stride_loop_64
+196608,26.8231,0.189853,38.096,40.0042
+389120,39.94595,0.19084849999999998,56.70515,59.9006
+778240,53.0688,0.189853,75.3143,79.797
+1556480,94.9269,0.189853,108.856,117.855
+3108864,162.839,0.187071,152.047,164.577
+6213632,179.945,0.186147,175.507,192.851
+9101312,188.677,0.185918,180.851,198.776
+18198528,218.752,0.185918,201.658,224.806
+24260608,220.49,0.185614,201.752,224.806
+30326784,221.32600000000002,0.185285,201.9885,224.8795
+36392960,222.162,0.184956,202.225,224.953
diff --git a/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_MASS3DEA_flops_raw.csv b/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_MASS3DEA_flops_raw.csv
@@ -0,0 +1,12 @@
+Problem size,Base_CUDA-compile_time_block_stride_loop_64,Base_Seq-default,RAJA_CUDA-cached_block_stride_loop_64,RAJA_CUDA-compile_time_block_stride_loop_64
+196608,12.371,0.172475,17.4694,18.4941
+389120,24.4534,0.173081,34.7305,36.4701
+778240,48.3817,0.175105,68.6625,72.7492
+1556480,86.5439,0.174902,99.2427,107.447
+3108864,148.459,0.170552,138.621,150.044
+6213632,164.056,0.16971,160.009,175.822
+9101312,172.016,0.169502,164.882,181.224
+18198528,199.437,0.169224,183.938,204.956
+24260608,201.02,0.168307,183.852,205.09
+30326784,202.545,0.170325,184.368,205.854
+36392960,208.774,0.168625,189.723,195.678
diff --git a/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_MASS3DEA_flops_smoothed.csv b/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_MASS3DEA_flops_smoothed.csv
@@ -0,0 +1,12 @@
+Problem size,Base_CUDA-compile_time_block_stride_loop_64,Base_Seq-default,RAJA_CUDA-cached_block_stride_loop_64,RAJA_CUDA-compile_time_block_stride_loop_64
+196608,24.4534,0.173081,34.7305,36.4701
+389120,36.41755,0.17399150000000002,51.6965,54.60965
+778240,48.3817,0.173081,68.6625,72.7492
+1556480,86.5439,0.173081,99.2427,107.447
+3108864,148.459,0.170552,138.621,150.044
+6213632,164.056,0.16971,160.009,175.822
+9101312,172.016,0.169502,164.882,181.224
+18198528,199.437,0.169502,183.852,204.956
+24260608,201.02,0.169224,183.938,204.956
+30326784,201.7825,0.1689245,184.153,205.023
+36392960,202.545,0.168625,184.368,205.09
diff --git a/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_MASS3DPA_ATOMIC_bandwidth_raw.csv b/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_MASS3DPA_ATOMIC_bandwidth_raw.csv
@@ -0,0 +1,12 @@
+Problem size,Base_CUDA-block_64,Base_Seq-default,RAJA_CUDA-block_64
+729,15.994,5.63123,15.9327
+3375,78.6741,5.57301,78.0764
+9261,195.668,5.48866,198.452
+19683,342.273,5.42675,339.859
+35937,493.318,5.50448,492.121
+68921,646.287,5.32245,641.236
+132651,753.564,5.48991,756.865
+250047,788.718,5.13006,788.566
+531441,823.891,4.66743,823.152
+753571,852.829,5.21751,852.44
+1442897,854.982,4.43834,854.334
diff --git a/...rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_MASS3DPA_ATOMIC_bandwidth_smoothed.csv b/...rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_MASS3DPA_ATOMIC_bandwidth_smoothed.csv
@@ -0,0 +1,12 @@
+Problem size,Base_CUDA-block_64,Base_Seq-default,RAJA_CUDA-block_64
+729,78.6741,5.57301,78.0764
+3375,137.17105,5.530835,138.26420000000002
+9261,195.668,5.50448,198.452
+19683,342.273,5.48866,339.859
+35937,493.318,5.48866,492.121
+68921,646.287,5.42675,641.236
+132651,753.564,5.32245,756.865
+250047,788.718,5.21751,788.566
+531441,823.891,5.13006,823.152
+753571,838.3599999999999,4.898745,837.796
+1442897,852.829,4.66743,852.44
diff --git a/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_MASS3DPA_ATOMIC_flops_raw.csv b/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_MASS3DPA_ATOMIC_flops_raw.csv
@@ -0,0 +1,12 @@
+Problem size,Base_CUDA-block_64,Base_Seq-default,RAJA_CUDA-block_64
+729,31.9709,11.2564,31.8483
+3375,163.533,11.5841,162.29
+9261,412.739,11.5777,418.612
+19683,727.574,11.5357,722.443
+35937,1053.65,11.7567,1051.09
+68921,1386.0,11.4143,1375.17
+132651,1621.29,11.8115,1628.39
+250047,1701.15,11.0648,1700.82
+531441,1781.14,10.0903,1779.54
+753571,1845.33,11.2895,1844.49
+1442897,1852.55,9.61689,1851.15
diff --git a/.../13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_MASS3DPA_ATOMIC_flops_smoothed.csv b/.../13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_MASS3DPA_ATOMIC_flops_smoothed.csv
@@ -0,0 +1,12 @@
+Problem size,Base_CUDA-block_64,Base_Seq-default,RAJA_CUDA-block_64
+729,163.533,11.5777,162.29
+3375,288.13599999999997,11.5567,290.451
+9261,412.739,11.5777,418.612
+19683,727.574,11.5777,722.443
+35937,1053.65,11.5777,1051.09
+68921,1386.0,11.5357,1375.17
+132651,1621.29,11.4143,1628.39
+250047,1701.15,11.2895,1700.82
+531441,1781.14,11.0648,1779.54
+753571,1813.2350000000001,10.577549999999999,1812.0149999999999
+1442897,1845.33,10.0903,1844.49
diff --git a/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_MASSVEC3DPA_bandwidth_raw.csv b/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_MASSVEC3DPA_bandwidth_raw.csv
@@ -0,0 +1,12 @@
+Problem size,Base_CUDA-direct_64,Base_CUDA-runtime_block_stride_loop_64,Base_Seq-default,RAJA_CUDA-cached_block_stride_loop_64,RAJA_CUDA-direct_64,RAJA_CUDA-runtime_block_stride_loop_64
+5376,10.3695,6.36613,3.89184,6.69729,10.9591,6.75047
+21312,40.5998,25.2277,3.97645,26.1748,41.7224,26.2766
+85056,159.182,96.3617,4.09511,101.046,166.765,101.436
+169920,279.378,178.801,4.1078,186.69,291.695,187.242
+339840,430.166,304.323,4.06186,312.599,446.023,312.635
+679680,568.529,380.403,4.04509,380.057,557.927,379.635
+1359360,681.084,458.687,4.14675,454.106,672.94,453.847
+2718720,729.4,500.232,4.08355,488.989,724.725,489.954
+5437248,758.443,521.457,3.72476,492.28,749.481,492.59
+7964736,776.721,533.27,3.74583,502.632,767.326,502.798
+15929280,797.352,547.173,3.68644,514.449,788.629,514.867
diff --git a/.../13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_MASSVEC3DPA_bandwidth_smoothed.csv b/.../13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_MASSVEC3DPA_bandwidth_smoothed.csv
@@ -0,0 +1,12 @@
+Problem size,Base_CUDA-direct_64,Base_CUDA-runtime_block_stride_loop_64,Base_Seq-default,RAJA_CUDA-cached_block_stride_loop_64,RAJA_CUDA-direct_64,RAJA_CUDA-runtime_block_stride_loop_64
+5376,40.5998,25.2277,3.97645,26.1748,41.7224,26.2766
+21312,99.89089999999999,60.7947,4.03578,63.610400000000006,104.24369999999999,63.856300000000005
+85056,159.182,96.3617,4.06186,101.046,166.765,101.436
+169920,279.378,178.801,4.06186,186.69,291.695,187.242
+339840,430.166,304.323,4.09511,312.599,446.023,312.635
+679680,568.529,380.403,4.08355,380.057,557.927,379.635
+1359360,681.084,458.687,4.06186,454.106,672.94,453.847
+2718720,729.4,500.232,4.04509,488.989,724.725,489.954
+5437248,758.443,521.457,3.74583,492.28,749.481,492.59
+7964736,767.582,527.3634999999999,3.735295,497.456,758.4035,497.69399999999996
+15929280,776.721,533.27,3.72476,502.632,767.326,502.798
diff --git a/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_MASSVEC3DPA_flops_raw.csv b/docs/13_rajaperf/baseline_data/RPBenchmark_H100_tier1/Apps_MASSVEC3DPA_flops_raw.csv
@@ -0,0 +1,12 @@
+Problem size,Base_CUDA-direct_64,Base_CUDA-runtime_block_stride_loop_64,Base_Seq-default,RAJA_CUDA-cached_block_stride_loop_64,RAJA_CUDA-direct_64,RAJA_CUDA-runtime_block_stride_loop_64
+5376,33.9294,20.8302,12.7342,21.9138,35.8586,22.0878
+21312,133.032,82.6628,13.0295,85.7661,136.71,86.0997
+85056,521.775,315.859,13.4231,331.214,546.63,332.491
+169920,915.812,586.117,13.4656,611.976,956.187,613.788
+339840,1410.14,997.613,13.3154,1024.74,1462.13,1024.86
+679680,1863.74,1247.03,13.2606,1245.9,1828.99,1244.51
+1359360,2232.74,1503.67,13.5939,1488.66,2206.04,1487.81
+2718720,2391.14,1639.87,13.3868,1603.02,2375.81,1606.18
+5437248,2486.35,1709.46,12.2106,1613.81,2456.97,1614.83
+7964736,2546.27,1748.18,12.2797,1647.75,2515.47,1648.29
+15929280,2613.91,1793.76,12.085,1686.48,2585.31,1687.86