[ET-VK] Miscellaneous fixes

Pull Request resolved: #14732

Collecting fixes for various models/ops in this diff/PR.

They have all been squashed into this single change to make it easier to cherry pick.

# Fixes

## Wav2Letter

Type: Output correctness failure

This is caused by a bug in swiftshader, and not reproducible on any other platform. Specifically, the issue is in the softmax shader; the exact cause of the issue is unknown, but it is related to using shared memory within shaders. The workaround for this issue is to use separate shared memory arrays for the shared max and shared sum.

## ConvNeXT

Type: Exception during runtime

This is caused by an incompatible memory layout being used for mean2d. More technically, the packed dimension of the tensor cannot be one of the dims being reduced. The current operator registry system did not have a way to select valid tensor representations based on the actual arguments of an op.

To fix, we have to introduce a mechanism for ops to specify valid representations once a node's arguments are known. Once the model is exported with supported memory layout, the model test passes.

## Inception_V3/ViT

Type: Exception during runtime

The root cause of this was an interaction betwen the fuse batch norm pass and how `vulkan_preprocess.py` was applying passes. Essentially, the fuse batch norm pass creates a new param node for the fused weight, which uses the original name of the tensor which may contain capital letters. However after re-tracing the graph, the node's name was being lowercased. `vulkan_preprocess` was using _copy_module to update the exported program's graph module in place, which was not updating the ep's graph signature with the new lowercase name after retracing.

The solution was to migrate vulkan_preprocess.py to use the _transform() API instead of using _copy_module.

There was also a small bug in Pool.cpp where `bool` was used to pass a UBO field that is received as an `int`.

## DenseNet 161 (w/ dynamic shapes)

Type: Output Mismatch

Cause: the native_batch_norm op doesn't support dynamic shapes. However, the backend test runner doesn't set the correct compile option to filter ops without dynamic shape support.

Since batch norm is easy to implement, fix by implementing resize for batch norm.
ghstack-source-id: 313794474

Differential Revision: [D83703496](https://our.internmc.facebook.com/intern/diff/D83703496/)

Author: ssjia

.github/workflows/pull.yml

@@ -970,11 +970,16 @@ jobs:
         PYTHON_EXECUTABLE=python bash backends/vulkan/test/scripts/test_model.sh --build
 
         # Test models serially
-        models="mv2 mv3 edsr resnet18 resnet50 dl3"
+        models="mv2 mv3 edsr resnet18 resnet50 dl3 w2l ic3 ic4 convnext_small vit"
         for model in $models; do
           python -m examples.vulkan.export --model_name=$model --test
         done
 
+        # For selected vision models, test with dynamic shapes
+        models="mv2 mv3 resnet18 resnet50 ic3 ic4 densenet161"
+        for model in $models; do
+          python -m examples.vulkan.export --model_name=$model --test -d
+        done
 
   test-vulkan-operators-linux:
     name: test-vulkan-operators-linux

backends/vulkan/_passes/fold_qdq.py

@@ -17,9 +17,8 @@ class FoldQDQPass(ExportPass):
     valid quant op patterns have already been fused before this pass.
     """
 
-    def __init__(self, edge_program: torch.export.ExportedProgram):
-        super(FoldQDQPass, self).__init__()
-        self.edge_program = edge_program
+    def __init__(self):
+        super().__init__()
 
     def call(self, graph_module: torch.fx.GraphModule):
         for node in graph_module.graph.nodes:

backends/vulkan/_passes/fuse_patterns.py

@@ -4,6 +4,8 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
+from typing import Optional
+
 import executorch.backends.vulkan.patterns as vk_patterns
 
 import torch
@@ -13,13 +15,15 @@
 
 
 class FusePatternsPass(ExportPass):
-    def __init__(self, exported_program: ExportedProgram) -> None:
+    def __init__(self) -> None:
         super().__init__()
-        self.program = exported_program
+        self._exported_program: Optional[ExportedProgram] = None
 
     def call(self, graph_module: torch.fx.GraphModule):
+        assert self._exported_program is not None
+
         total_replaced = vk_patterns.replace_all_fusable_subgraphs(
-            self.program, graph_module
+            self._exported_program, graph_module
         )
 
         if total_replaced > 0:

backends/vulkan/_passes/fuse_quantized_ops.py

@@ -211,18 +211,20 @@ def fuse_into_linear_qcnw_node(
 
 
 class FuseQuantizedOpsTransform(ExportPass):
-    def __init__(self, exported_program: ExportedProgram) -> None:
+    def __init__(self) -> None:
         super().__init__()
-        self.program = exported_program
+        self._exported_program: Optional[ExportedProgram] = None
 
     def call(self, graph_module: torch.fx.GraphModule) -> PassResult:
+        assert self._exported_program is not None
+
         for node in graph_module.graph.nodes:
             # Check for linear_qcnw pattern (weight-only quantization)
-            qcnw_details = matches_linear_qcnw_pattern(self.program, node)
+            qcnw_details = matches_linear_qcnw_pattern(self._exported_program, node)
             if qcnw_details is not None:
                 qcnw_method, qcnw_nbits = qcnw_details
                 fuse_into_linear_qcnw_node(
-                    self.program, graph_module, node, qcnw_method, qcnw_nbits
+                    self._exported_program, graph_module, node, qcnw_method, qcnw_nbits
                 )
                 continue
 

backends/vulkan/op_registry.py

@@ -16,8 +16,6 @@
 
 import torch
 
-from executorch.backends.vulkan.serialization.vulkan_graph_schema import VkMemoryLayout
-
 from executorch.exir.dialects._ops import ops as exir_ops
 
 from executorch.exir.dialects.edge._ops import EdgeOpOverload
@@ -48,6 +46,9 @@ class OpFeatures:
         # Optional check function used during partitioning to determine if a node's
         # inputs are supported by the operator implementation.
         "are_node_inputs_supported_fn",
+        # Optional function to determine valid representation sets for input and outputs
+        # once a node's actual inputs are known.
+        "pick_io_storage_fn",
     ]
 
     def __init__(
@@ -61,6 +62,7 @@ def __init__(
         supports_resize: bool = False,
         supports_prepacking: bool = False,
         are_node_inputs_supported_fn: Optional[Callable] = allow_node,
+        pick_io_storage_fn: Optional[Callable] = None,
     ):
         self.inputs_storage: utils.TensorRepSetList = utils.TensorRepSetList(
             inputs_storage if inputs_storage is not None else []
@@ -77,15 +79,21 @@ def __init__(
         self.supports_prepacking = supports_prepacking
 
         self.are_node_inputs_supported_fn = are_node_inputs_supported_fn
+        self.pick_io_storage_fn = pick_io_storage_fn
 
     def make_op_repsets(
         self,
         op_node: torch.fx.Node,
         texture_limits: utils.ImageExtents = utils.DEFAULT_TEXTURE_LIMITS,
     ) -> utils.OpRepSets:
-        return utils.OpRepSets(
-            self.inputs_storage, self.outputs_storage, op_node, texture_limits
-        )
+        inputs_storage = self.inputs_storage
+        outputs_storage = self.outputs_storage
+        if self.pick_io_storage_fn is not None:
+            i_storage, o_storage = self.pick_io_storage_fn(op_node)
+            inputs_storage = utils.TensorRepSetList(i_storage)
+            outputs_storage = utils.TensorRepSetList(o_storage)
+
+        return utils.OpRepSets(inputs_storage, outputs_storage, op_node, texture_limits)
 
 
 #######################
@@ -410,28 +418,16 @@ def register_softmax_op():
 )
 def register_reduce_op():
     def check_reduce_node(node: torch.fx.Node) -> bool:
+        # Only one argument implies that the reduction is over the entire tensor, which
+        # is not supported yet.
+        if len(node.args) == 1:
+            return False
+
         dim_list = node.args[1]
+        # Only 1D and 2D reductions are supported at the moment.
         if isinstance(dim_list, list) and len(dim_list) > 2:
             return False
 
-        if isinstance(dim_list, list) and len(dim_list) == 2:
-            # Try to get the memory layout for this node
-            try:
-                memory_layout = utils.get_node_memory_layout(node)
-
-                # If we have memory layout information, check if any dimension in dim_list corresponds to a packed dimension
-                if (
-                    memory_layout is not None
-                    and memory_layout != VkMemoryLayout.DEFAULT_LAYOUT
-                ):
-                    # For now only default layout is supported for 2D reduction.
-                    # Because we can't determine if the input is NCHW or NHWC here,
-                    # assume the reduction dimension is packed so we cannot support it.
-                    return False
-            except (AssertionError, KeyError, AttributeError):
-                # If we can't get memory layout information, we'll assume the dims aren't packed
-                pass
-
         def try_find_keepdim_arg(node: torch.fx.Node) -> bool:
             for arg in node.args:
                 if isinstance(arg, bool):
@@ -446,10 +442,41 @@ def try_find_keepdim_arg(node: torch.fx.Node) -> bool:
 
         return True
 
+    def pick_io_storage_for_reduce(node: torch.fx.Node):
+        inputs_storage = utils.ANY_TEXTURE
+        outputs_storage = utils.ANY_TEXTURE
+
+        input_tensor = node.args[0]
+        ndim = input_tensor.meta["val"].ndim
+        dim_list = node.args[1]
+        if isinstance(dim_list, list) and len(dim_list) == 2:
+            reduce_dim1_whcn = utils.nchw_dim_to_whcn_dim(dim_list[0], ndim)
+            reduce_dim2_whcn = utils.nchw_dim_to_whcn_dim(dim_list[1], ndim)
+
+            possible_packed_dims = {0, 1, 2}
+            possible_packed_dims.discard(reduce_dim1_whcn)
+            possible_packed_dims.discard(reduce_dim2_whcn)
+
+            packed_dim = possible_packed_dims.pop()
+            assert packed_dim in [0, 1, 2]
+
+            if packed_dim == 0:
+                inputs_storage = utils.WIDTH_PACKED_TEXTURE
+                outputs_storage = utils.WIDTH_PACKED_TEXTURE
+            elif packed_dim == 1:
+                inputs_storage = utils.HEIGHT_PACKED_TEXTURE
+                outputs_storage = utils.HEIGHT_PACKED_TEXTURE
+            else:
+                inputs_storage = utils.CHANNELS_PACKED_TEXTURE
+                outputs_storage = utils.CHANNELS_PACKED_TEXTURE
+
+        return inputs_storage, outputs_storage
+
     return OpFeatures(
         inputs_storage=utils.ANY_TEXTURE,
         supports_resize=True,
         are_node_inputs_supported_fn=check_reduce_node,
+        pick_io_storage_fn=pick_io_storage_for_reduce,
     )
 
 
@@ -716,6 +743,7 @@ def register_ported_ops_with_prepacking():
     return OpFeatures(
         inputs_storage=utils.CHANNELS_PACKED_TEXTURE,
         supports_prepacking=True,
+        supports_resize=True,
     )
 
 
@@ -746,6 +774,7 @@ def register_ported_ops_with_prepacking_all_dims():
     return OpFeatures(
         inputs_storage=utils.ANY_TEXTURE,
         supports_prepacking=True,
+        supports_resize=True,
     )
 
 

backends/vulkan/partitioner/vulkan_partitioner.py

@@ -36,7 +36,7 @@
     Partitioner,
     PartitionResult,
 )
-from executorch.exir.backend.utils import tag_constant_data
+from executorch.exir.backend.utils import tag_constant_data, tag_mutated_buffer
 from executorch.exir.dialects._ops import ops as exir_ops
 
 from torch.export.exported_program import ExportedProgram
@@ -254,9 +254,10 @@ def _is_node_supported(self, node: torch.fx.Node) -> bool:  # noqa: C901
             self.log_skip(node, "permute node of non compatible linear node")
             return False
 
-        is_in_local_scalar_dense_chain, dst_node_is_compatible = (
-            self.is_in_local_scalar_dense_chain(node)
-        )
+        (
+            is_in_local_scalar_dense_chain,
+            dst_node_is_compatible,
+        ) = self.is_in_local_scalar_dense_chain(node)
         if is_in_local_scalar_dense_chain and dst_node_is_compatible:
             return True
         elif is_in_local_scalar_dense_chain:
@@ -419,6 +420,7 @@ def partition(self, exported_program: ExportedProgram) -> PartitionResult:
             logger.info(f"Found {pl} Vulkan subgraphs to be partitioned.")
 
         tag_constant_data(exported_program)
+        tag_mutated_buffer(exported_program)
 
         return PartitionResult(
             tagged_exported_program=exported_program, partition_tags=partition_tags

backends/vulkan/patterns/quantized_linear.py

@@ -92,9 +92,11 @@ def __init__(self, mm_node: torch.fx.Node) -> None:
             return
 
         # Identify input node
-        self.fp_input_node, self.quantize_input_node, dq_node = (
-            utils.maybe_skip_q_dq_arg_chain(self.anchor_node.args[0])
-        )
+        (
+            self.fp_input_node,
+            self.quantize_input_node,
+            dq_node,
+        ) = utils.maybe_skip_q_dq_arg_chain(self.anchor_node.args[0])
         assert self.fp_input_node is not None
         self.all_nodes.append(self.fp_input_node)
 
@@ -386,7 +388,7 @@ def make_linear_dq8ca_q4gsw_op(
         weight_sums_node = create_constant_placeholder(
             exp_program=ep,
             graph=graph_module.graph,
-            kind=InputKind.CONSTANT_TENSOR,
+            kind=InputKind.PARAMETER,
             name=sums_name,
             data=sum_per_quant_group,
         )
@@ -429,7 +431,7 @@ def make_linear_q8ta_q8csw_custom_op(
         weight_sums_node = create_constant_placeholder(
             exp_program=ep,
             graph=graph_module.graph,
-            kind=InputKind.CONSTANT_TENSOR,
+            kind=InputKind.PARAMETER,
             name=sums_name,
             data=sum_per_output_channel,
         )

backends/vulkan/runtime/graph/ops/glsl/full.yaml

@@ -14,5 +14,6 @@ full:
     DTYPE:
       - VALUE: half
       - VALUE: float
+      - VALUE: int32
   shader_variants:
     - NAME: full

backends/vulkan/runtime/graph/ops/glsl/softmax.glsl

@@ -42,7 +42,8 @@ layout(constant_id = 5) const int group_dim = 1;
 // work group will write into its assigned element in the shared array.
 #define MAX_NTHREADS 16
 
-shared vec4 shared_vecs[MAX_NTHREADS];
+shared vec4 shared_max[MAX_NTHREADS];
+shared vec4 shared_sum[MAX_NTHREADS];
 
 #include "indexing_utils.h"
 
@@ -102,13 +103,13 @@ void softmax_nonpacked_dim(const ivec2 tid, ivec3 scan_pos) {
        i += NWORKERS, scan_pos[reduce_dim] += NWORKERS) {
     max_elements = max(max_elements, load_texel(tin, scan_pos));
   }
-  shared_vecs[smi] = max_elements;
+  shared_max[smi] = max_elements;
   barrier();
   // Iterate over the partial maximums to obtain the overall maximum
   group_i = tid.y * NWORKERS;
-  max_elements = shared_vecs[group_i++];
+  max_elements = shared_max[group_i++];
   for (int i = 1; i < NWORKERS; ++i, group_i++) {
-    max_elements = max(max_elements, shared_vecs[group_i]);
+    max_elements = max(max_elements, shared_max[group_i]);
   }
 
   scan_pos[reduce_dim] = tid.x;
@@ -118,13 +119,13 @@ void softmax_nonpacked_dim(const ivec2 tid, ivec3 scan_pos) {
        i += NWORKERS, scan_pos[reduce_dim] += NWORKERS) {
     denominators += exp(load_texel(tin, scan_pos) - max_elements);
   }
-  shared_vecs[smi] = denominators;
+  shared_sum[smi] = denominators;
   barrier();
   // Iterate over the partial sums to obtain the overall sum
   group_i = tid.y * NWORKERS;
-  denominators = shared_vecs[group_i++];
+  denominators = shared_sum[group_i++];
   for (int i = 1; i < NWORKERS; ++i, group_i++) {
-    denominators += shared_vecs[group_i];
+    denominators += shared_sum[group_i];
   }
 
   // Determine if there are any padding elements in the final texel of the
@@ -184,13 +185,13 @@ void softmax_packed_dim(const ivec2 tid, ivec3 scan_pos) {
       max_elements.x = max(intex[i], max_elements.x);
     }
   }
-  shared_vecs[smi] = max_elements;
+  shared_max[smi] = max_elements;
   barrier();
   // Iterate over the partial maximums to obtain the overall maximum
   group_i = tid.y * NWORKERS;
-  max_elements = shared_vecs[group_i++];
+  max_elements = shared_max[group_i++];
   for (int i = 1; i < NWORKERS; ++i, group_i++) {
-    max_elements = max(max_elements, shared_vecs[group_i]);
+    max_elements = max(max_elements, shared_max[group_i]);
   }
   // Each element of the texel is itself a partial maximum; iterate over the
   // texel to find the actual maximum
@@ -214,13 +215,13 @@ void softmax_packed_dim(const ivec2 tid, ivec3 scan_pos) {
       denominators.x += exp(intex[i] - max_element);
     }
   }
-  shared_vecs[smi] = denominators;
+  shared_sum[smi] = denominators;
   barrier();
   // Iterate over the partial sums to obtain the overall sum
   group_i = tid.y * NWORKERS;
-  denominators = shared_vecs[group_i++];
+  denominators = shared_sum[group_i++];
   for (int i = 1; i < NWORKERS; ++i, group_i++) {
-    denominators += shared_vecs[group_i];
+    denominators += shared_sum[group_i];
   }
   // Reduce over the accumulated texel to find the overall sum
   float denominator = 0;