Combine MISA and Winograd configs

nod-ai · Jun 19, 2024 · e5422a1 · e5422a1
1 parent 57aaf1d
commit e5422a1
Show file tree

Hide file tree

Showing 5 changed files with 40 additions and 47 deletions.
diff --git a/benchmark-scheduled-unet-misa.sh b/benchmark-scheduled-unet-misa.sh
diff --git a/compile-scheduled-unet-tk.sh b/compile-scheduled-unet-tk.sh
@@ -1,6 +1,6 @@
 #!/bin/bash
 
-# Usage: PATH=/path/to/iree/build/tools:$PATH ./compile-scheduled-unet.sh <target-chip> <default|winograd|misa> [extra flags]
+# Usage: PATH=/path/to/iree/build/tools:$PATH ./compile-scheduled-unet.sh <target-chip> <default|winograd|misa|hybrid> [extra flags]
 
 set -euo pipefail
 

diff --git a/compile-unet-base.sh b/compile-unet-base.sh
@@ -19,10 +19,13 @@ readonly CHIP="$2"
 readonly MODE="$3"
 USE_WINOGRAD=0
 USE_MISA=0
+USE_WINOGRAD_AND_MISA=0
 if [[ $MODE =~ "winograd" ]] ; then
   USE_WINOGRAD=1
 elif [[ $MODE =~ "misa" ]] ; then
   USE_MISA=1
+elif [[ $MODE =~ "hybrid" ]] ; then
+  USE_WINOGRAD_AND_MISA=1
 fi
 
 readonly ATTENTION_SPEC="$(realpath "$4")"
@@ -67,11 +70,21 @@ readonly MISA_FLAGS=(
   "--iree-preprocessing-transform-spec-filename=${SPEC_DIR}/misa_unet_spec.mlir"
 )
 
+readonly WINOGRAD_AND_MISA_FLAGS=(
+  "--iree-opt-const-expr-max-size-increase-threshold=1000000000000000"
+  "--iree-preprocessing-pass-pipeline=${WINOGRAD_PIPELINE[*]}"
+  "--iree-hal-executable-object-search-path=${SPEC_DIR}"
+  "--iree-preprocessing-transform-spec-filename=${SPEC_DIR}/misa_unet_spec.mlir"
+  "--iree-preprocessing-pass-pipeline=${WINOGRAD_PIPELINE[*]}"
+)
+
 declare -a FLAGS=("${DEFAULT_FLAGS[*]}")
 if [ "$USE_WINOGRAD" = 1 ] ; then
   FLAGS=("${WINOGRAD_FLAGS[@]}")
 elif [ "$USE_MISA" = 1 ] ; then
   FLAGS=("${MISA_FLAGS[@]}")
+elif [ "$USE_WINOGRAD_AND_MISA" = 1 ] ; then
+  FLAGS=("${WINOGRAD_AND_MISA_FLAGS[@]}")
 fi
 
 set -x

diff --git a/specs/misa_unet_spec.mlir b/specs/misa_unet_spec.mlir
@@ -857,15 +857,16 @@ module attributes {transform.with_named_sequence} {
     transform.foreach %funcs : !transform.any_op {
       ^bb1(%func: !transform.any_op):
         transform.foreach_match in %func
-            @match_conv_k1 -> @cast_and_call_dag_k1,
-            @match_conv_k2 -> @cast_and_call_dag_k2,
-            @match_conv_k3 -> @cast_and_call_dag_k3,
-            @match_conv_k4 -> @cast_and_call_dag_k4,
-            @match_conv_k5 -> @cast_and_call_dag_k5,
-            @match_conv_k6 -> @cast_and_call_dag_k6,
-            @match_conv_k7 -> @cast_and_call_dag_k7,
-            @match_conv_k8 -> @cast_and_call_dag_k8,
-            @match_conv_k9 -> @cast_and_call_dag_k9
+        // base - 63.5 ms
+            @match_conv_k1 -> @cast_and_call_dag_k1, // -1.4 ms
+            // @match_conv_k2 -> @cast_and_call_dag_k2, // -0.1 ms
+            // @match_conv_k3 -> @cast_and_call_dag_k3, // -0.1 ms
+            // @match_conv_k4 -> @cast_and_call_dag_k4, // 0 ms
+            // @match_conv_k5 -> @cast_and_call_dag_k5, // -0.1 ms
+            // @match_conv_k6 -> @cast_and_call_dag_k6, // -0.1 ms
+            @match_conv_k7 -> @cast_and_call_dag_k7 // -0.5 ms
+            // @match_conv_k8 -> @cast_and_call_dag_k8, // -0.1 ms
+            // @match_conv_k9 -> @cast_and_call_dag_k9 // -0.2 ms
           : (!transform.any_op) -> (!transform.any_op)
     }
     transform.apply_dce to %module : !transform.any_op

diff --git a/specs/winograd_conv_spec.mlir b/specs/winograd_conv_spec.mlir
@@ -453,24 +453,24 @@ module attributes { transform.with_named_sequence } {
 
 
   transform.named_sequence @__transform_main(%func: !transform.any_op {transform.consumed}) {
-    transform.foreach_match in %func // Base: 69.4ms // Best: 67.3ms
+    transform.foreach_match in %func // Base: 63.5ms // Best: 60.5ms
         // @match_conv2x4x128x128x3x3x320 -> @annotate_op, // fail to compile
 
-        @match_conv2x1280x64x64x3x3x1280 -> @annotate_op, // 68.8ms -0.6
-        @match_conv2x640x128x128x3x3x320 -> @annotate_op, // 68.9ms -0.5
-        @match_conv2x1920x64x64x3x3x640 -> @annotate_op, // 69.0ms -0.4
-        @match_conv2x960x128x128x3x3x320 -> @annotate_op, // 69.1ms -0.3
-        @match_conv2x640x128x128x3x3x640 -> @annotate_op, // 69.1ms -0.3
-        @match_conv2x1280x64x64x3x3x640 -> @annotate_op, // 69.2ms -0.2
-        @match_conv2x960x64x64x3x3x640 -> @annotate_op, // 69.2ms -0.2
-        // @match_conv2x320x64x64x3x3x640 -> @annotate_op, // 69.5ms -0.1
-        // @match_conv2x640x64x64x3x3x640 -> @annotate_op, // 69.3ms -0.1
-        // @match_conv2x2560x32x32x3x3x1280 -> @annotate_op, // 69.3ms -0.1
-        // @match_conv2x1920x32x32x3x3x1280 -> @annotate_op, // 69.4ms +0.0
-        // @match_conv2x640x32x32x3x3x1280 -> @annotate_op, // 69.5ms +0.1
-        // @match_conv2x320x128x128x3x3x320 -> @annotate_op, // 69.6ms +0.2
-        // @match_conv2x1280x32x32x3x3x1280 -> @annotate_op, // 69.7ms +0.3
-        // @match_conv2x320x128x128x3x3x4 -> @annotate_op, // 69.7ms +0.3
+        @match_conv2x1280x64x64x3x3x1280 -> @annotate_op, // 63.0ms -0.5
+        @match_conv2x640x128x128x3x3x320 -> @annotate_op, // 63.0ms -0.5
+        @match_conv2x1920x64x64x3x3x640 -> @annotate_op, // 63.2ms -0.3
+        @match_conv2x960x128x128x3x3x320 -> @annotate_op, // 63.1ms -0.4
+        @match_conv2x640x128x128x3x3x640 -> @annotate_op, // 63.0ms -0.5
+        @match_conv2x1280x64x64x3x3x640 -> @annotate_op, // 63.2ms -0.3
+        @match_conv2x960x64x64x3x3x640 -> @annotate_op, // 63.2ms -0.3
+        // @match_conv2x320x64x64x3x3x640 -> @annotate_op,
+        // @match_conv2x640x64x64x3x3x640 -> @annotate_op,
+        // @match_conv2x2560x32x32x3x3x1280 -> @annotate_op,
+        // @match_conv2x1920x32x32x3x3x1280 -> @annotate_op,
+        // @match_conv2x640x32x32x3x3x1280 -> @annotate_op,
+        // @match_conv2x320x128x128x3x3x320 -> @annotate_op,
+        // @match_conv2x1280x32x32x3x3x1280 -> @annotate_op,
+        // @match_conv2x320x128x128x3x3x4 -> @annotate_op,
         @placeholder -> @annotate_op
       : (!transform.any_op) -> (!transform.any_op)
     transform.yield