Arm backend: Unsqueeze rank 0 tensor at vgf runtime

backends/arm/runtime/VGFSetup.cpp

@@ -24,6 +24,13 @@ namespace vgf {
 /* static function to map format to byte count */
 static uint32_t get_format_size(VkFormat format);
 
+// SPV_ARM_tensor does not support rank-0 representations according to the spec.
+// Use an unsqueezed dimension when the resource table contains an empty
+// shape. Tensors are output as rank 0 when copied back from the vgf backend.
+namespace {
+constexpr int64_t kScalarSentinelDimension = 1;
+}
+
 // Debug function to inspect memory properties
 static string memory_flags_to_string(VkMemoryPropertyFlags flags) {
   if (flags == 0)
@@ -264,7 +271,11 @@ static void debug_print_resources(
             the_shape.size(),
             the_stride.size());
         for (int j = 0; j < the_shape.size(); j++) {
-          ET_LOG(Info, "      %d: dim %ld", j, the_shape[j]);
+          ET_LOG(
+              Info,
+              "      %d: dim %lld",
+              j,
+              static_cast<long long>(the_shape[j]));
         }
         // Allocate a tensor with bound memory
         break;
@@ -387,6 +398,7 @@ bool VgfRepr::process_vgf(const char* vgf_data, ArrayRef<CompileSpec> specs) {
     // Get tensor shape and strides
     auto shape = resource_decoder->getTensorShape(i);
     auto stride = resource_decoder->getTensorStride(i);
+    const auto shape_size = shape.size();
 
     switch (resource_decoder->getCategory(i)) {
       case vgflib::ResourceCategory::INPUT:
@@ -409,9 +421,9 @@ bool VgfRepr::process_vgf(const char* vgf_data, ArrayRef<CompileSpec> specs) {
         result = allocate_tensor(
             vk_physical,
             vk_device,
-            vgflib::ToVkFormat(resource_decoder->getVkFormat(i)),
-            static_cast<uint32_t>(shape.size()),
-            shape.begin(),
+            resource_format,
+            shape_size == 0 ? 1 : static_cast<uint32_t>(shape_size),
+            shape_size == 0 ? &kScalarSentinelDimension : shape.begin(),
             static_cast<uint32_t>(stride.size()),
             stride.begin(),
             &tensor_description,
@@ -422,8 +434,7 @@ bool VgfRepr::process_vgf(const char* vgf_data, ArrayRef<CompileSpec> specs) {
           ET_LOG(Error, "Failed to allocate tensor for VGF resource %d", i);
           return false;
         }
-        size_t e_size = get_format_size(
-            vgflib::ToVkFormat(resource_decoder->getVkFormat(i)));
+        size_t e_size = get_format_size(resource_format);
         if (0 == e_size) {
           ET_LOG(Error, "failed to get element size of VkFormat");
           return false;
@@ -449,9 +460,11 @@ bool VgfRepr::process_vgf(const char* vgf_data, ArrayRef<CompileSpec> specs) {
             .sType = VK_STRUCTURE_TYPE_TENSOR_DESCRIPTION_ARM,
             .pNext = nullptr,
             .tiling = VK_TENSOR_TILING_LINEAR_ARM,
-            .format = vgflib::ToVkFormat(resource_decoder->getVkFormat(i)),
-            .dimensionCount = static_cast<uint32_t>(shape.size()),
-            .pDimensions = shape.begin(),
+            .format = resource_format,
+            .dimensionCount =
+                shape_size == 0 ? 1 : static_cast<uint32_t>(shape_size),
+            .pDimensions =
+                shape_size == 0 ? &kScalarSentinelDimension : shape.begin(),
             // Note: stride_data of 0's causes size==0, null means stride==size
             .pStrides = (0 == stride.size() ? nullptr : stride.begin()),
             .usage = VK_TENSOR_USAGE_DATA_GRAPH_BIT_ARM,

backends/arm/test/ops/test_addmm.py

@@ -167,7 +167,6 @@ def test_addmm_u85_INT(test_data: Tuple):
 
 @common.parametrize("test_data", test_data_suite)
 @common.SkipIfNoModelConverter
-@pytest.mark.xfail(reason="MLETORCH-1410: Tensor dimension count not supported: 0")
 def test_addmm_vgf_FP(test_data: input_t1):
     pipeline = VgfPipeline[input_t1](
         Addmm(),
@@ -181,7 +180,6 @@ def test_addmm_vgf_FP(test_data: input_t1):
 
 @common.parametrize("test_data", test_data_suite)
 @common.SkipIfNoModelConverter
-@pytest.mark.xfail(reason="MLETORCH-1410: Tensor dimension count not supported: 0")
 def test_addmm_vgf_INT(test_data: input_t1):
     pipeline = VgfPipeline[input_t1](
         Addmm(),

backends/arm/test/ops/test_amax.py

@@ -139,7 +139,6 @@ def test_max_dim_tosa_FP_not_delegated():
 
 @common.parametrize("test_data", Amax.test_data)
 @common.SkipIfNoModelConverter
-@pytest.mark.xfail(reason="MLETORCH-1410: Tensor dimension count not supported: 0")
 def test_amax_vgf_FP(test_data: Amax.input_t):
     data, dim, keep_dims = test_data()
     module = Amax(dim, keep_dims)
@@ -154,7 +153,6 @@ def test_amax_vgf_FP(test_data: Amax.input_t):
 
 @common.parametrize("test_data", Amax.test_data)
 @common.SkipIfNoModelConverter
-@pytest.mark.xfail(reason="MLETORCH-1410: Tensor dimension count not supported: 0")
 def test_amax_vgf_INT(test_data: Amax.input_t):
     data, dim, keep_dims = test_data()
     module = Amax(dim, keep_dims)
@@ -169,7 +167,6 @@ def test_amax_vgf_INT(test_data: Amax.input_t):
 
 @common.parametrize("test_data", Max.test_data)
 @common.SkipIfNoModelConverter
-@pytest.mark.xfail(reason="MLETORCH-1410: Tensor dimension count not supported: 0")
 def test_max_dim_vgf_FP_to_amax(test_data: Max.input_t):
     data, dim = test_data()
     pipeline = VgfPipeline[Max.input_t](
@@ -183,7 +180,6 @@ def test_max_dim_vgf_FP_to_amax(test_data: Max.input_t):
 
 @common.parametrize("test_data", Max.test_data)
 @common.SkipIfNoModelConverter
-@pytest.mark.xfail(reason="MLETORCH-1410: Tensor dimension count not supported: 0")
 def test_max_dim_vgf_INT_to_amax(test_data: Max.input_t):
     data, dim = test_data()
     pipeline = VgfPipeline[Max.input_t](

backends/arm/test/ops/test_amin.py

@@ -155,7 +155,6 @@ def test_min_dim_tosa_FP_not_delegated():
 
 @common.parametrize("test_data", Amin.test_data)
 @common.SkipIfNoModelConverter
-@pytest.mark.xfail(reason="MLETORCH-1410: Tensor dimension count not supported: 0")
 def test_amin_vgf_FP(test_data: Amin.input_t):
     data, dim, keep_dims = test_data()
     pipeline = VgfPipeline[Amin.input_t](
@@ -166,7 +165,6 @@ def test_amin_vgf_FP(test_data: Amin.input_t):
 
 @common.parametrize("test_data", Amin.test_data)
 @common.SkipIfNoModelConverter
-@pytest.mark.xfail(reason="MLETORCH-1410: Tensor dimension count not supported: 0")
 def test_amin_vgf_INT(test_data: Amin.input_t):
     data, dim, keep_dims = test_data()
     pipeline = VgfPipeline[Amin.input_t](
@@ -180,7 +178,6 @@ def test_amin_vgf_INT(test_data: Amin.input_t):
 
 @common.parametrize("test_data", Min.test_data)
 @common.SkipIfNoModelConverter
-@pytest.mark.xfail(reason="MLETORCH-1410: Tensor dimension count not supported: 0")
 def test_min_dim_vgf_FP_to_amin(test_data: Min.input_t):
     data, dim = test_data()
     pipeline = VgfPipeline[Min.input_t](
@@ -194,7 +191,6 @@ def test_min_dim_vgf_FP_to_amin(test_data: Min.input_t):
 
 @common.parametrize("test_data", Min.test_data)
 @common.SkipIfNoModelConverter
-@pytest.mark.xfail(reason="MLETORCH-1410: Tensor dimension count not supported: 0")
 def test_min_dim_vgf_INT_to_amin(test_data: Min.input_t):
     data, dim = test_data()
     pipeline = VgfPipeline[Min.input_t](

backends/arm/test/ops/test_any.py

@@ -6,7 +6,6 @@
 
 from typing import List, Tuple
 
-import pytest
 import torch
 from executorch.backends.arm.test import common
 from executorch.backends.arm.test.tester.test_pipeline import (
@@ -189,7 +188,6 @@ def test_any_u85_INT(test_data: input_t1):
 
 @common.parametrize("test_data", test_data)
 @common.SkipIfNoModelConverter
-@pytest.mark.xfail(reason="MLETORCH-1410: Tensor dimension count not supported: 0")
 def test_any_vgf_FP(test_data: input_t1):
     op, data_fn = test_data()
     pipeline = VgfPipeline[input_t1](
@@ -204,7 +202,6 @@ def test_any_vgf_FP(test_data: input_t1):
 
 @common.parametrize("test_data", test_data)
 @common.SkipIfNoModelConverter
-@pytest.mark.xfail(reason="MLETORCH-1410: Tensor dimension count not supported: 0")
 def test_any_vgf_INT(test_data: input_t1):
     op, data_fn = test_data()
     pipeline = VgfPipeline[input_t1](

backends/arm/test/ops/test_mean_dim.py

@@ -4,7 +4,6 @@
 #
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
-import pytest
 import torch
 from executorch.backends.arm.test import common
 from executorch.backends.arm.test.tester.test_pipeline import (
@@ -84,7 +83,6 @@ def test_adaptive_avg_pool2d_u85_INT(test_data):
 
 @common.parametrize("test_data", AdaptiveAveragePool2d.test_data_suite)
 @common.SkipIfNoModelConverter
-@pytest.mark.xfail(reason="MLETORCH-1410: Tensor dimension count not supported: 0")
 def test_adaptive_avg_pool2d_vgf_FP(test_data):
     pipeline = VgfPipeline[input_t](
         AdaptiveAveragePool2d(),
@@ -98,7 +96,6 @@ def test_adaptive_avg_pool2d_vgf_FP(test_data):
 
 @common.parametrize("test_data", AdaptiveAveragePool2d.test_data_suite)
 @common.SkipIfNoModelConverter
-@pytest.mark.xfail(reason="MLETORCH-1410: Tensor dimension count not supported: 0")
 def test_adaptive_avg_pool2d_vgf_INT(test_data):
     pipeline = VgfPipeline[input_t](
         AdaptiveAveragePool2d(),
@@ -331,7 +328,6 @@ def test_mean_dim_u85_INT(test_data):
 
 @common.parametrize("test_data", MeanDim.test_data_suite)
 @common.SkipIfNoModelConverter
-@pytest.mark.xfail(reason="MLETORCH-1410: Tensor dimension count not supported: 0")
 def test_mean_dim_vgf_FP(test_data):
     test_data_val, dim, keep_dim = test_data()
     pipeline = VgfPipeline[input_t](
@@ -346,7 +342,6 @@ def test_mean_dim_vgf_FP(test_data):
 
 @common.parametrize("test_data", MeanDim.test_data_suite)
 @common.SkipIfNoModelConverter
-@pytest.mark.xfail(reason="MLETORCH-1410: Tensor dimension count not supported: 0")
 def test_mean_dim_vgf_INT(test_data):
     test_data_val, dim, keep_dim = test_data()
     pipeline = VgfPipeline[input_t](

backends/arm/test/ops/test_scalar_tensor.py

@@ -2,7 +2,6 @@
 #
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
-import pytest
 import torch
 from executorch.backends.arm.test import common
 
@@ -102,7 +101,6 @@ def test_scalar_tensor_u85_INT(test_data):
 
 @common.parametrize("test_data", float_test_data_suite)
 @common.SkipIfNoModelConverter
-@pytest.mark.xfail(reason="MLETORCH-1410: Tensor dimension count not supported: 0")
 def test_scalar_tensor_vgf_FP(test_data):
     scalar, dtype, data = test_data()
     pipeline = VgfPipeline(
@@ -116,7 +114,6 @@ def test_scalar_tensor_vgf_FP(test_data):
 
 @common.parametrize("test_data", int_test_data_suite)
 @common.SkipIfNoModelConverter
-@pytest.mark.xfail(reason="MLETORCH-1410: Tensor dimension count not supported: 0")
 def test_scalar_tensor_vgf_INT(test_data):
     scalar, dtype, data = test_data()
     pipeline = VgfPipeline(

backends/arm/test/ops/test_select.py

@@ -7,7 +7,6 @@
 
 from typing import Tuple
 
-import pytest
 import torch
 
 from executorch.backends.arm.test import common
@@ -170,7 +169,6 @@ def test_select_int_u85_INT(test_data: Tuple):
 
 @common.parametrize("test_data", test_data_suite)
 @common.SkipIfNoModelConverter
-@pytest.mark.xfail(reason="MLETORCH-1410: Tensor dimension count not supported: 0")
 def test_select_int_vgf_FP_copy(test_data: Tuple):
     pipeline = VgfPipeline[input_t1](
         SelectCopy(), test_data(), aten_op_copy, [], tosa_version="TOSA-1.0+FP"
@@ -180,7 +178,6 @@ def test_select_int_vgf_FP_copy(test_data: Tuple):
 
 @common.parametrize("test_data", test_data_suite)
 @common.SkipIfNoModelConverter
-@pytest.mark.xfail(reason="MLETORCH-1410: Tensor dimension count not supported: 0")
 def test_select_int_vgf_FP(test_data: Tuple):
     pipeline = VgfPipeline[input_t1](
         SelectInt(), test_data(), aten_op_int, [], tosa_version="TOSA-1.0+FP"
@@ -190,7 +187,6 @@ def test_select_int_vgf_FP(test_data: Tuple):
 
 @common.parametrize("test_data", test_data_suite)
 @common.SkipIfNoModelConverter
-@pytest.mark.xfail(reason="MLETORCH-1410: Tensor dimension count not supported: 0")
 def test_select_int_vgf_INT_copy(test_data: Tuple):
     pipeline = VgfPipeline[input_t1](
         SelectCopy(),
@@ -204,7 +200,6 @@ def test_select_int_vgf_INT_copy(test_data: Tuple):
 
 @common.parametrize("test_data", test_data_suite)
 @common.SkipIfNoModelConverter
-@pytest.mark.xfail(reason="MLETORCH-1410: Tensor dimension count not supported: 0")
 def test_select_int_vgf_INT(test_data: Tuple):
     pipeline = VgfPipeline[input_t1](
         SelectInt(),

backends/arm/test/ops/test_var.py

@@ -6,7 +6,6 @@
 
 from typing import Tuple
 
-import pytest
 import torch
 
 from executorch.backends.arm.test import common
@@ -214,7 +213,6 @@ def test_var_dim_u85_INT_no_dim(test_data: Tuple):
 
 @common.parametrize("test_data", Var.test_parameters)
 @common.SkipIfNoModelConverter
-@pytest.mark.xfail(reason="MLETORCH-1410: Tensor dimension count not supported: 0")
 def test_var_dim_vgf_FP_no_dim(test_data: Tuple):
     data, keepdim, correction = test_data()
     pipeline = VgfPipeline[input_t1](
@@ -225,7 +223,6 @@ def test_var_dim_vgf_FP_no_dim(test_data: Tuple):
 
 @common.parametrize("test_data", Var.test_parameters)
 @common.SkipIfNoModelConverter
-@pytest.mark.xfail(reason="MLETORCH-1410: Tensor dimension count not supported: 0")
 def test_var_dim_vgf_INT_no_dim(test_data: Tuple):
     data, keepdim, correction = test_data()
     pipeline = VgfPipeline[input_t1](
@@ -296,7 +293,6 @@ def test_var_dim_u85_INT(test_data: Tuple):
 
 @common.parametrize("test_data", VarDim.test_parameters)
 @common.SkipIfNoModelConverter
-@pytest.mark.xfail(reason="MLETORCH-1410: Tensor dimension count not supported: 0")
 def test_var_dim_vgf_FP(test_data: Tuple):
     data, dim, keepdim, unbiased = test_data()
     pipeline = VgfPipeline[input_t1](
@@ -307,7 +303,6 @@ def test_var_dim_vgf_FP(test_data: Tuple):
 
 @common.parametrize("test_data", VarDim.test_parameters)
 @common.SkipIfNoModelConverter
-@pytest.mark.xfail(reason="MLETORCH-1410: Tensor dimension count not supported: 0")
 def test_var_dim_vgf_INT(test_data: Tuple):
     data, dim, keepdim, unbiased = test_data()
     pipeline = VgfPipeline[input_t1](
@@ -377,7 +372,6 @@ def test_var_dim_u85_INT_correction(test_data: Tuple):
 
 @common.parametrize("test_data", VarCorrection.test_parameters)
 @common.SkipIfNoModelConverter
-@pytest.mark.xfail(reason="MLETORCH-1410: Tensor dimension count not supported: 0")
 def test_var_dim_vgf_FP_correction(test_data: Tuple):
     data, dim, keepdim, corr = test_data()
     pipeline = VgfPipeline[input_t1](
@@ -388,7 +382,6 @@ def test_var_dim_vgf_FP_correction(test_data: Tuple):
 
 @common.parametrize("test_data", VarCorrection.test_parameters)
 @common.SkipIfNoModelConverter
-@pytest.mark.xfail(reason="MLETORCH-1410: Tensor dimension count not supported: 0")
 def test_var_dim_vgf_INT_correction(test_data: Tuple):
     data, dim, keepdim, corr = test_data()
     pipeline = VgfPipeline[input_t1](