From a843a8b3576fc7695f3394ff947ecf5f41be0ab4 Mon Sep 17 00:00:00 2001 From: w4yne Date: Wed, 11 Sep 2024 09:13:21 +0800 Subject: [PATCH] [KUNLUNXIN] add FlagGems blas op support & minor fix for env vars (#727) Co-authored-by: w4yne --- operation/benchmarks/addmm/kunlunxin/R300p/case_config.yaml | 5 +++++ operation/benchmarks/addmm/kunlunxin/R300p/env.sh | 5 +++++ operation/benchmarks/bmm/kunlunxin/R300p/case_config.yaml | 5 +++++ operation/benchmarks/bmm/kunlunxin/R300p/env.sh | 5 +++++ .../benchmarks/cross_entropy_loss/kunlunxin/R300p/env.sh | 2 +- operation/benchmarks/linear/kunlunxin/R300p/case_config.yaml | 5 +++++ operation/benchmarks/linear/kunlunxin/R300p/env.sh | 5 +++++ operation/benchmarks/max/kunlunxin/R300p/env.sh | 2 +- operation/benchmarks/min/kunlunxin/R300p/env.sh | 2 +- operation/benchmarks/mm/kunlunxin/R300p/case_config.yaml | 5 +++++ operation/benchmarks/mm/kunlunxin/R300p/env.sh | 5 +++++ 11 files changed, 43 insertions(+), 3 deletions(-) create mode 100644 operation/benchmarks/addmm/kunlunxin/R300p/case_config.yaml create mode 100644 operation/benchmarks/addmm/kunlunxin/R300p/env.sh create mode 100644 operation/benchmarks/bmm/kunlunxin/R300p/case_config.yaml create mode 100644 operation/benchmarks/bmm/kunlunxin/R300p/env.sh create mode 100644 operation/benchmarks/linear/kunlunxin/R300p/case_config.yaml create mode 100644 operation/benchmarks/linear/kunlunxin/R300p/env.sh create mode 100644 operation/benchmarks/mm/kunlunxin/R300p/case_config.yaml create mode 100644 operation/benchmarks/mm/kunlunxin/R300p/env.sh diff --git a/operation/benchmarks/addmm/kunlunxin/R300p/case_config.yaml b/operation/benchmarks/addmm/kunlunxin/R300p/case_config.yaml new file mode 100644 index 000000000..37f065e91 --- /dev/null +++ b/operation/benchmarks/addmm/kunlunxin/R300p/case_config.yaml @@ -0,0 +1,5 @@ +M : 16 +N : 40 +K : 256 +ITERS: 50 +SPECTFLOPS: 9999 diff --git a/operation/benchmarks/addmm/kunlunxin/R300p/env.sh b/operation/benchmarks/addmm/kunlunxin/R300p/env.sh new file mode 100644 index 000000000..6e1e159ef --- /dev/null +++ b/operation/benchmarks/addmm/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" diff --git a/operation/benchmarks/bmm/kunlunxin/R300p/case_config.yaml b/operation/benchmarks/bmm/kunlunxin/R300p/case_config.yaml new file mode 100644 index 000000000..37f065e91 --- /dev/null +++ b/operation/benchmarks/bmm/kunlunxin/R300p/case_config.yaml @@ -0,0 +1,5 @@ +M : 16 +N : 40 +K : 256 +ITERS: 50 +SPECTFLOPS: 9999 diff --git a/operation/benchmarks/bmm/kunlunxin/R300p/env.sh b/operation/benchmarks/bmm/kunlunxin/R300p/env.sh new file mode 100644 index 000000000..6e1e159ef --- /dev/null +++ b/operation/benchmarks/bmm/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" diff --git a/operation/benchmarks/cross_entropy_loss/kunlunxin/R300p/env.sh b/operation/benchmarks/cross_entropy_loss/kunlunxin/R300p/env.sh index 38a0db6a6..9dc60f91e 100644 --- a/operation/benchmarks/cross_entropy_loss/kunlunxin/R300p/env.sh +++ b/operation/benchmarks/cross_entropy_loss/kunlunxin/R300p/env.sh @@ -1,6 +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 +export TRITONXPU_BUFFER_SIZE=128 echo "KUNLUNXIN ENV.SH end" diff --git a/operation/benchmarks/linear/kunlunxin/R300p/case_config.yaml b/operation/benchmarks/linear/kunlunxin/R300p/case_config.yaml new file mode 100644 index 000000000..37f065e91 --- /dev/null +++ b/operation/benchmarks/linear/kunlunxin/R300p/case_config.yaml @@ -0,0 +1,5 @@ +M : 16 +N : 40 +K : 256 +ITERS: 50 +SPECTFLOPS: 9999 diff --git a/operation/benchmarks/linear/kunlunxin/R300p/env.sh b/operation/benchmarks/linear/kunlunxin/R300p/env.sh new file mode 100644 index 000000000..6e1e159ef --- /dev/null +++ b/operation/benchmarks/linear/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" diff --git a/operation/benchmarks/max/kunlunxin/R300p/env.sh b/operation/benchmarks/max/kunlunxin/R300p/env.sh index 176026da9..b4d3583bc 100644 --- a/operation/benchmarks/max/kunlunxin/R300p/env.sh +++ b/operation/benchmarks/max/kunlunxin/R300p/env.sh @@ -1,6 +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 +export INST_COMBINE_LOOP_THRESHOLD=1000 echo "KUNLUNXIN ENV.SH end" diff --git a/operation/benchmarks/min/kunlunxin/R300p/env.sh b/operation/benchmarks/min/kunlunxin/R300p/env.sh index 176026da9..b4d3583bc 100644 --- a/operation/benchmarks/min/kunlunxin/R300p/env.sh +++ b/operation/benchmarks/min/kunlunxin/R300p/env.sh @@ -1,6 +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 +export INST_COMBINE_LOOP_THRESHOLD=1000 echo "KUNLUNXIN ENV.SH end" diff --git a/operation/benchmarks/mm/kunlunxin/R300p/case_config.yaml b/operation/benchmarks/mm/kunlunxin/R300p/case_config.yaml new file mode 100644 index 000000000..37f065e91 --- /dev/null +++ b/operation/benchmarks/mm/kunlunxin/R300p/case_config.yaml @@ -0,0 +1,5 @@ +M : 16 +N : 40 +K : 256 +ITERS: 50 +SPECTFLOPS: 9999 diff --git a/operation/benchmarks/mm/kunlunxin/R300p/env.sh b/operation/benchmarks/mm/kunlunxin/R300p/env.sh new file mode 100644 index 000000000..6e1e159ef --- /dev/null +++ b/operation/benchmarks/mm/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"