[KUNLUNXIN] FlagGems operation test: update dockerfile, config for ops

operation/benchmarks/abs/kunlunxin/R300p/case_config.yaml

@@ -0,0 +1,2 @@
+ITERS: 50
+SPECTFLOPS: 9999

operation/benchmarks/abs/kunlunxin/R300p/env.sh

@@ -0,0 +1,5 @@
+echo "KUNLUNXIN ENV.SH start"
+
+source /root/miniconda/etc/profile.d/conda.sh && conda activate python38_torch201_cuda
+
+echo "KUNLUNXIN ENV.SH end"

operation/benchmarks/add/kunlunxin/R300p/case_config.yaml

@@ -0,0 +1,2 @@
+ITERS: 50
+SPECTFLOPS: 9999

operation/benchmarks/add/kunlunxin/R300p/env.sh

@@ -0,0 +1,5 @@
+echo "KUNLUNXIN ENV.SH start"
+
+source /root/miniconda/etc/profile.d/conda.sh && conda activate python38_torch201_cuda
+
+echo "KUNLUNXIN ENV.SH end"

operation/benchmarks/all/kunlunxin/R300p/case_config.yaml

@@ -0,0 +1,3 @@
+Shape: [4096, 256]
+ITERS: 50
+SPECTFLOPS: 9999

operation/benchmarks/all/kunlunxin/R300p/env.sh

@@ -0,0 +1,5 @@
+echo "KUNLUNXIN ENV.SH start"
+
+source /root/miniconda/etc/profile.d/conda.sh && conda activate python38_torch201_cuda
+
+echo "KUNLUNXIN ENV.SH end"

operation/benchmarks/all/main.py

@@ -10,6 +10,7 @@
 import yaml
 import sys
 import subprocess
+import math
 
 sys.path.append("..")
 from drivers.utils import *
@@ -67,14 +68,21 @@ def main(config, case_config):
 
 
     m = case_config.Melements
+    # default shape: M * 1024 * 1024
+    shape = m * 1024 * 1024
 
+    if config.vendor == 'kunlunxin':
+        # if `Shape' specified in `case_config.yaml', use it
+        if case_config.__contains__('Shape') and case_config.Shape is not None:
+            shape = case_config.Shape
 
-    a = torch.arange(0, m * 1024 * 1024).to(0)
+    a = torch.arange(0, math.prod(shape)).to(0)
+    print(f'Shape for performance_test: {a.shape}')
 
     latency_nowarm, latency_warm, cputime, kerneltime = do_test(
         torch.all, (a, ), host_device_sync, config, case_config)
 
-    op2flops = lambda x: x * m * 1024 * 1024
+    op2flops = lambda x: x * math.prod(shape)
 
     perf_result = cal_perf(cputime, kerneltime, op2flops,
                            config.spectflops)
@@ -99,4 +107,4 @@ def main(config, case_config):
         print("Using flaggems")
     else:
         print("Using nativetorch")
-    main(config, case_config)
+    main(config, case_config)

operation/benchmarks/amax/kunlunxin/R300p/case_config.yaml

@@ -0,0 +1,3 @@
+Shape: [4096, 1]
+ITERS: 50
+SPECTFLOPS: 9999

operation/benchmarks/amax/kunlunxin/R300p/env.sh

@@ -0,0 +1,5 @@
+echo "KUNLUNXIN ENV.SH start"
+
+source /root/miniconda/etc/profile.d/conda.sh && conda activate python38_torch201_cuda
+
+echo "KUNLUNXIN ENV.SH end"

operation/benchmarks/amax/main.py

@@ -10,6 +10,7 @@
 import yaml
 import sys
 import subprocess
+import math
 
 sys.path.append("..")
 from drivers.utils import *
@@ -68,14 +69,21 @@ def main(config, case_config):
 
     m = case_config.M
     n = case_config.N
+    # default shape: (M*80, N*80)
+    shape = (m * 80, n * 80)
 
+    if config.vendor == 'kunlunxin':
+        # if `Shape' specified in `case_config.yaml', use it
+        if case_config.__contains__('Shape') and case_config.Shape is not None:
+            shape = case_config.Shape
 
-    a = torch.randn(m * 80 , n * 80 , dtype=dtype[config.dataformat]).to(0)
+    a = torch.randn(shape, dtype=dtype[config.dataformat]).to(0)
+    print(f'Shape for performance_test: {a.shape}')
 
     latency_nowarm, latency_warm, cputime, kerneltime = do_test(
         torch.amax, (a, 1), host_device_sync, config, case_config)
 
-    op2flops = lambda x: x * m * 80  * n * 80
+    op2flops = lambda x: x * math.prod(shape)
 
     perf_result = cal_perf(cputime, kerneltime, op2flops,
                            config.spectflops)
@@ -100,4 +108,4 @@ def main(config, case_config):
         print("Using flaggems")
     else:
         print("Using nativetorch")
-    main(config, case_config)
+    main(config, case_config)

operation/benchmarks/argmax/kunlunxin/R300p/case_config.yaml

@@ -0,0 +1,3 @@
+Shape: [4096, 1]
+ITERS: 50
+SPECTFLOPS: 9999

operation/benchmarks/argmax/kunlunxin/R300p/env.sh

@@ -0,0 +1,5 @@
+echo "KUNLUNXIN ENV.SH start"
+
+source /root/miniconda/etc/profile.d/conda.sh && conda activate python38_torch201_cuda
+
+echo "KUNLUNXIN ENV.SH end"

operation/benchmarks/argmax/main.py

@@ -10,6 +10,7 @@
 import yaml
 import sys
 import subprocess
+import math
 
 sys.path.append("..")
 from drivers.utils import *
@@ -68,14 +69,21 @@ def main(config, case_config):
 
     m = case_config.M
     n = case_config.N
+    # default shape: (M * 80, N * 80)
+    shape = (m * 80, n * 80)
 
+    if config.vendor == 'kunlunxin':
+        # if `Shape' specified in `case_config.yaml', use it
+        if case_config.__contains__('Shape') and case_config.Shape is not None:
+            shape = case_config.Shape
 
-    a = torch.randn(m * 80, n * 80, dtype=dtype[config.dataformat]).to(0)
+    a = torch.randn(shape, dtype=dtype[config.dataformat]).to(0)
+    print(f'Shape for performance_test: {a.shape}')
 
     latency_nowarm, latency_warm, cputime, kerneltime = do_test(
         torch.argmax, (a, 1), host_device_sync, config, case_config)
 
-    op2flops = lambda x: x * m * n * 80 * 80
+    op2flops = lambda x: x * math.prod(shape)
 
     perf_result = cal_perf(cputime, kerneltime, op2flops,
                            config.spectflops)
@@ -100,4 +108,4 @@ def main(config, case_config):
         print("Using flaggems")
     else:
         print("Using nativetorch")
-    main(config, case_config)
+    main(config, case_config)

operation/benchmarks/bitwise_and/kunlunxin/R300p/case_config.yaml

@@ -0,0 +1,2 @@
+ITERS: 50
+SPECTFLOPS: 9999

operation/benchmarks/bitwise_and/kunlunxin/R300p/env.sh

@@ -0,0 +1,5 @@
+echo "KUNLUNXIN ENV.SH start"
+
+source /root/miniconda/etc/profile.d/conda.sh && conda activate python38_torch201_cuda
+
+echo "KUNLUNXIN ENV.SH end"

operation/benchmarks/bitwise_not/kunlunxin/R300p/case_config.yaml

@@ -0,0 +1,2 @@
+ITERS: 50
+SPECTFLOPS: 9999

operation/benchmarks/bitwise_not/kunlunxin/R300p/env.sh

@@ -0,0 +1,5 @@
+echo "KUNLUNXIN ENV.SH start"
+
+source /root/miniconda/etc/profile.d/conda.sh && conda activate python38_torch201_cuda
+
+echo "KUNLUNXIN ENV.SH end"

operation/benchmarks/bitwise_or/kunlunxin/R300p/case_config.yaml

@@ -0,0 +1,2 @@
+ITERS: 50
+SPECTFLOPS: 9999

operation/benchmarks/bitwise_or/kunlunxin/R300p/env.sh

@@ -0,0 +1,5 @@
+echo "KUNLUNXIN ENV.SH start"
+
+source /root/miniconda/etc/profile.d/conda.sh && conda activate python38_torch201_cuda
+
+echo "KUNLUNXIN ENV.SH end"

operation/benchmarks/cos/kunlunxin/R300p/case_config.yaml

@@ -0,0 +1,2 @@
+ITERS: 50
+SPECTFLOPS: 9999

operation/benchmarks/cos/kunlunxin/R300p/env.sh

@@ -0,0 +1,6 @@
+echo "KUNLUNXIN ENV.SH start"
+
+source /root/miniconda/etc/profile.d/conda.sh && conda activate python38_torch201_cuda
+export XPU_enable_reorder=1
+
+echo "KUNLUNXIN ENV.SH end"

operation/benchmarks/cross_entropy_loss/kunlunxin/R300p/case_config.yaml

@@ -0,0 +1,4 @@
+bs: 4096
+elements: 256
+ITERS: 50
+SPECTFLOPS: 9999

operation/benchmarks/cross_entropy_loss/kunlunxin/R300p/env.sh

@@ -0,0 +1,6 @@
+echo "KUNLUNXIN ENV.SH start"
+
+source /root/miniconda/etc/profile.d/conda.sh && conda activate python38_torch201_cuda
+export XPU_enable_reorder=1
+
+echo "KUNLUNXIN ENV.SH end"

operation/benchmarks/div/kunlunxin/R300p/case_config.yaml

@@ -0,0 +1,2 @@
+ITERS: 50
+SPECTFLOPS: 9999

operation/benchmarks/div/kunlunxin/R300p/env.sh

@@ -0,0 +1,5 @@
+echo "KUNLUNXIN ENV.SH start"
+
+source /root/miniconda/etc/profile.d/conda.sh && conda activate python38_torch201_cuda
+
+echo "KUNLUNXIN ENV.SH end"

operation/benchmarks/dropout/kunlunxin/R300p/case_config.yaml

@@ -0,0 +1,3 @@
+Melements: 1
+ITERS: 50
+SPECTFLOPS: 9999

operation/benchmarks/dropout/kunlunxin/R300p/env.sh

@@ -0,0 +1,5 @@
+echo "KUNLUNXIN ENV.SH start"
+
+source /root/miniconda/etc/profile.d/conda.sh && conda activate python38_torch201_cuda
+
+echo "KUNLUNXIN ENV.SH end"

operation/benchmarks/eq/kunlunxin/R300p/case_config.yaml

@@ -0,0 +1,2 @@
+ITERS: 50
+SPECTFLOPS: 9999

operation/benchmarks/eq/kunlunxin/R300p/env.sh

@@ -0,0 +1,5 @@
+echo "KUNLUNXIN ENV.SH start"
+
+source /root/miniconda/etc/profile.d/conda.sh && conda activate python38_torch201_cuda
+
+echo "KUNLUNXIN ENV.SH end"

operation/benchmarks/exp/kunlunxin/R300p/env.sh

@@ -1,7 +1,5 @@
 echo "KUNLUNXIN ENV.SH start"
 
 source /root/miniconda/etc/profile.d/conda.sh && conda activate python38_torch201_cuda
-export TRITON_XPU_ARCH=3
-export CUDART_DUMMY_REGISTER=1
 
 echo "KUNLUNXIN ENV.SH end"

operation/benchmarks/ge/kunlunxin/R300p/case_config.yaml

@@ -0,0 +1,2 @@
+ITERS: 50
+SPECTFLOPS: 9999

operation/benchmarks/ge/kunlunxin/R300p/env.sh

@@ -0,0 +1,5 @@
+echo "KUNLUNXIN ENV.SH start"
+
+source /root/miniconda/etc/profile.d/conda.sh && conda activate python38_torch201_cuda
+
+echo "KUNLUNXIN ENV.SH end"

operation/benchmarks/gelu/kunlunxin/R300p/case_config.yaml

@@ -0,0 +1,3 @@
+Shape: [1024, 1024]
+ITERS: 50
+SPECTFLOPS: 9999

operation/benchmarks/gelu/kunlunxin/R300p/env.sh

@@ -0,0 +1,5 @@
+echo "KUNLUNXIN ENV.SH start"
+
+source /root/miniconda/etc/profile.d/conda.sh && conda activate python38_torch201_cuda
+
+echo "KUNLUNXIN ENV.SH end"

operation/benchmarks/gelu/main.py

@@ -10,6 +10,7 @@
 import yaml
 import sys
 import subprocess
+import math
 
 sys.path.append("..")
 from drivers.utils import *
@@ -66,14 +67,22 @@ def main(config, case_config):
     set_ieee_float32(config.vendor)
 
     m = case_config.Melements
-    f = torch.nn.GELU()
+    # default shape: (M, 1024, 1024)
+    shape = (m, 1024, 1024)
+
+    if config.vendor == 'kunlunxin':
+        # if `Shape' specified in `case_config.yaml', use it
+        if case_config.__contains__('Shape') and case_config.Shape is not None:
+            shape = case_config.Shape
 
-    a = torch.randn(m, 1024, 1024, dtype=dtype[config.dataformat]).to(0)
+    a = torch.randn(shape, dtype=dtype[config.dataformat]).to(0)
+    print(f'Shape for performance_test: {a.shape}')
 
+    f = torch.nn.GELU()
     latency_nowarm, latency_warm, cputime, kerneltime = do_test(
         f, (a, ), host_device_sync, config, case_config) # 调整为torch.sub
 
-    op2flops = lambda x: x * 9 * m * 1024 * 1024 
+    op2flops = lambda x: x * 9 * math.prod(shape)
 
     perf_result = cal_perf(cputime, kerneltime, op2flops,
                            config.spectflops)

operation/benchmarks/group_norm/kunlunxin/R300p/case_config.yaml

@@ -0,0 +1,5 @@
+bs: 1
+channel: 3
+hiddensize: 4
+ITERS: 50
+SPECTFLOPS: 9999

operation/benchmarks/group_norm/kunlunxin/R300p/env.sh

@@ -0,0 +1,5 @@
+echo "KUNLUNXIN ENV.SH start"
+
+source /root/miniconda/etc/profile.d/conda.sh && conda activate python38_torch201_cuda
+
+echo "KUNLUNXIN ENV.SH end"

operation/benchmarks/gt/kunlunxin/R300p/case_config.yaml

@@ -0,0 +1,2 @@
+ITERS: 50
+SPECTFLOPS: 9999

operation/benchmarks/gt/kunlunxin/R300p/env.sh

@@ -0,0 +1,5 @@
+echo "KUNLUNXIN ENV.SH start"
+
+source /root/miniconda/etc/profile.d/conda.sh && conda activate python38_torch201_cuda
+
+echo "KUNLUNXIN ENV.SH end"

operation/benchmarks/isinf/kunlunxin/R300p/case_config.yaml

@@ -0,0 +1,2 @@
+ITERS: 50
+SPECTFLOPS: 9999

operation/benchmarks/isinf/kunlunxin/R300p/env.sh

@@ -0,0 +1,5 @@
+echo "KUNLUNXIN ENV.SH start"
+
+source /root/miniconda/etc/profile.d/conda.sh && conda activate python38_torch201_cuda
+
+echo "KUNLUNXIN ENV.SH end"

operation/benchmarks/isnan/kunlunxin/R300p/case_config.yaml

@@ -0,0 +1,2 @@
+ITERS: 50
+SPECTFLOPS: 9999

operation/benchmarks/isnan/kunlunxin/R300p/env.sh

@@ -0,0 +1,5 @@
+echo "KUNLUNXIN ENV.SH start"
+
+source /root/miniconda/etc/profile.d/conda.sh && conda activate python38_torch201_cuda
+
+echo "KUNLUNXIN ENV.SH end"

operation/benchmarks/layer_norm/kunlunxin/R300p/case_config.yaml

@@ -0,0 +1,2 @@
+ITERS: 50
+SPECTFLOPS: 9999

operation/benchmarks/layer_norm/kunlunxin/R300p/env.sh

@@ -0,0 +1,5 @@
+echo "KUNLUNXIN ENV.SH start"
+
+source /root/miniconda/etc/profile.d/conda.sh && conda activate python38_torch201_cuda
+
+echo "KUNLUNXIN ENV.SH end"

operation/benchmarks/le/kunlunxin/R300p/case_config.yaml

@@ -0,0 +1,2 @@
+ITERS: 50
+SPECTFLOPS: 9999

operation/benchmarks/le/kunlunxin/R300p/env.sh

@@ -0,0 +1,5 @@
+echo "KUNLUNXIN ENV.SH start"
+
+source /root/miniconda/etc/profile.d/conda.sh && conda activate python38_torch201_cuda
+
+echo "KUNLUNXIN ENV.SH end"

operation/benchmarks/log_softmax/kunlunxin/R300p/env.sh

@@ -1,7 +1,5 @@
 echo "KUNLUNXIN ENV.SH start"
 
 source /root/miniconda/etc/profile.d/conda.sh && conda activate python38_torch201_cuda
-export TRITON_XPU_ARCH=3
-export CUDART_DUMMY_REGISTER=1
 
 echo "KUNLUNXIN ENV.SH end"

operation/benchmarks/lt/kunlunxin/R300p/case_config.yaml

@@ -0,0 +1,2 @@
+ITERS: 50
+SPECTFLOPS: 9999

operation/benchmarks/lt/kunlunxin/R300p/env.sh

@@ -0,0 +1,5 @@
+echo "KUNLUNXIN ENV.SH start"
+
+source /root/miniconda/etc/profile.d/conda.sh && conda activate python38_torch201_cuda
+
+echo "KUNLUNXIN ENV.SH end"

operation/benchmarks/max/kunlunxin/R300p/case_config.yaml

@@ -0,0 +1,3 @@
+Shape: [4096, 1]
+ITERS: 50
+SPECTFLOPS: 9999

operation/benchmarks/max/kunlunxin/R300p/env.sh

@@ -0,0 +1,6 @@
+echo "KUNLUNXIN ENV.SH start"
+
+source /root/miniconda/etc/profile.d/conda.sh && conda activate python38_torch201_cuda
+export Triton_big_instcombine=1000
+
+echo "KUNLUNXIN ENV.SH end"