CI: Optimize autotuning pipeline and inital the docs

.github/scripts/op_tune.sh

@@ -27,8 +27,22 @@ declare -a tune_jobs=(
 
 for job in "${tune_jobs[@]}"; do
     IFS=':' read -r shape dir test_path tune_cmd <<< "$job"
-    if [ -n "$shape_filter" ] && [ "$shape" != "$shape_filter" ]; then
-        continue
+    # If shape_filter is not empty, check if the current shape exists in the filter list.
+    # shape_filter is a comma-separated list, e.g. "ck_gemm_a8w8,ck_batched_gemm_a8w8"
+    if [ -n "$shape_filter" ]; then
+        # Remove all whitespace from the shape_filter string
+        shape_filter_no_space="${shape_filter//[[:space:]]/}"
+        IFS=',' read -ra filter_shapes <<< "$shape_filter_no_space"
+        found_match=false
+        for filter_shape in "${filter_shapes[@]}"; do
+            if [[ "$shape" == "$filter_shape" ]]; then
+                found_match=true
+                break
+            fi
+        done
+        if [ "$found_match" = false ]; then
+            continue
+        fi
     fi
     echo "============================================================"
     echo "🧪 Processing shape: $shape under directory: $dir"

.github/workflows/operators-tuning.yaml

@@ -1,13 +1,10 @@
 name: Operators Tuning
 
 on:
-  pull_request:
-    paths:
-      - 'aiter/configs/*untuned*.csv'
   workflow_dispatch:
     inputs:
       shapes:
-        description: 'Comma separated shape names to run (leave empty for all)'
+        description: 'Comma separated shape names to run, e.g. ck_batched_gemm_a8w8, ck_gemm_a8w8, ck_gemm_a8w8_blockscale, ck_gemm_a8w8_blockscale_bpreshuffle, ck_gemm_a8w8_bpreshuffle etc. (leave empty for all)'
         required: false
         default: ''
       arguments:
@@ -81,7 +78,7 @@ jobs:
           docker exec \
           -w /workspace \
           operators_tuning_test \
-          ./.github/scripts/op_tune.sh test "${{ github.event.inputs.shapes }}"
+          ./.github/scripts/op_tune.sh test "ck_batched_gemm_a8w8, ck_gemm_a8w8, ck_gemm_a8w8_blockscale"
 
       - name: Operators tuning Tests
         run: |
@@ -90,7 +87,7 @@ jobs:
           docker exec \
           -w /workspace \
           operators_tuning_test \
-          ./.github/scripts/op_tune.sh tune "${{ github.event.inputs.shapes }}" "${{ github.event.inputs.arguments }}"
+          ./.github/scripts/op_tune.sh tune "ck_batched_gemm_a8w8, ck_gemm_a8w8, ck_gemm_a8w8_blockscale" "${{ github.event.inputs.arguments }}"
 
       - name: Show the difference after tuning
         run: |
@@ -103,7 +100,7 @@ jobs:
           docker exec \
           -w /workspace \
           operators_tuning_test \
-          ./.github/scripts/op_tune.sh test "${{ github.event.inputs.shapes }}"
+          ./.github/scripts/op_tune.sh test "ck_batched_gemm_a8w8, ck_gemm_a8w8, ck_gemm_a8w8_blockscale"
 
       - name: Upload tuned CSVs
         uses: actions/upload-artifact@v4

docs/autotuning_pipeline.md

@@ -0,0 +1,17 @@
+# Autotuning Pipelines in Aiter CI
+
+## What is the tuning pipeline workflow?
+
+An automated tuning system that ingests and benchmarks a volume of inputs, then records the best operator for each input in a database based on test results, so that future identical inputs can directly return the optimal operator.
+
+## Implementation
+
+In the Aiter repository, there are tuning scripts designed for various shapes, such as `aiter/csrc/ck_batched_gemm_a8w8` (see: [ROCm/aiter](https://github.com/ROCm/aiter)).
+
+Running these scripts generates tuned results, which are stored in the `aiter/configs` directory, for example: `aiter/configs/a8w8_tuned_batched_gemm.csv`. These CSV files are compiled during the Aiter installation process and are referenced when using Aiter operators.
+
+Based on this, we provide CI pipelines to generate and use these tuned CSV files:
+
+- [Manual Pipeline](https://github.com/ROCm/aiter/actions/workflows/operators-tuning.yaml): Allows users to select specific shapes to tune and choose whether to upload the results to the Aiter repository.
+
+- Scheduled Pipeline: Runs nightly or weekly to generate all tuned CSV files and automatically upload the results to the Aiter repository.