Add an option to control on applying tensor.pad -> insert_slice pattern. (#10435)
This avoids having dispatches that only have single tensor.pad op. The tensor.pad dispatches are not supported. It makes the option more robust, and can be applied on all the tracking models.
diff --git a/benchmarks/TFLite/android-mali.cmake b/benchmarks/TFLite/android-mali.cmake
index 766b5ff..a849e0f 100644
--- a/benchmarks/TFLite/android-mali.cmake
+++ b/benchmarks/TFLite/android-mali.cmake
@@ -114,7 +114,7 @@
"GPU-Mali-Valhall"
COMPILATION_FLAGS
${ANDROID_MALI_GPU_COMPILATION_FLAGS}
- "--iree-flow-enable-fuse-padding-into-consumer-ops"
+ "--iree-flow-enable-fuse-padding-into-linalg-consumer-ops"
BENCHMARK_TOOL
iree-benchmark-module
CONFIG
@@ -140,7 +140,7 @@
"--iree-input-type=tosa"
"--iree-flow-demote-f32-to-f16"
"--iree-vulkan-target-triple=valhall-unknown-android31"
- "--iree-flow-enable-fuse-padding-into-consumer-ops"
+ "--iree-flow-enable-fuse-padding-into-linalg-consumer-ops"
BENCHMARK_TOOL
iree-benchmark-module
CONFIG
@@ -180,7 +180,7 @@
"GPU-Mali-Valhall"
COMPILATION_FLAGS
${ANDROID_MALI_GPU_COMPILATION_FLAGS}
- "--iree-flow-enable-fuse-padding-into-consumer-ops"
+ "--iree-flow-enable-fuse-padding-into-linalg-consumer-ops"
"--iree-hal-benchmark-dispatch-repeat-count=32"
BENCHMARK_TOOL
iree-benchmark-module
@@ -209,7 +209,7 @@
"--iree-input-type=tosa"
"--iree-flow-demote-f32-to-f16"
"--iree-vulkan-target-triple=valhall-unknown-android31"
- "--iree-flow-enable-fuse-padding-into-consumer-ops"
+ "--iree-flow-enable-fuse-padding-into-linalg-consumer-ops"
"--iree-hal-benchmark-dispatch-repeat-count=32"
BENCHMARK_TOOL
iree-benchmark-module