Update on "[ET-VK][Ops] choose_qparams op shaders and impl"

Creating the choose_qparams per_tensor and per_token logic shaders and impl which are linked with the testing framework

Differential Revision: [D76436933](https://our.internmc.facebook.com/intern/diff/D76436933/)

[ghstack-poisoned]

Author: morelos

.ci/scripts/utils.sh

@@ -156,13 +156,14 @@ build_executorch_runner() {
 }
 
 cmake_install_executorch_lib() {
+  build_type="${1:-Release}"
   echo "Installing libexecutorch.a and libportable_kernels.a"
   clean_executorch_install_folders
   retry cmake -DCMAKE_INSTALL_PREFIX=cmake-out \
-          -DCMAKE_BUILD_TYPE=Release \
+          -DCMAKE_BUILD_TYPE=${build_type} \
           -DPYTHON_EXECUTABLE="$PYTHON_EXECUTABLE" \
           -Bcmake-out .
-  cmake --build cmake-out -j9 --target install --config Release
+  cmake --build cmake-out -j9 --target install --config ${build_type}
 }
 
 download_stories_model_artifacts() {

.github/workflows/lint.yml

@@ -46,7 +46,7 @@ jobs:
         fi
 
         # This has already been cached in the docker image
-        lintrunner init 2> /dev/null
+        lintrunner init
 
         RC=0
         # Run lintrunner on all files

backends/qualcomm/_passes/i64_to_i32.py

@@ -28,8 +28,10 @@ class I64toI32(ExportPass):
     I64_OPS = {
         exir_ops.edge.aten.argmin.default,
         exir_ops.edge.aten.arange.start_step,
+        exir_ops.edge.aten.cumsum.default,
         exir_ops.edge.aten.full.default,
         exir_ops.edge.aten.scalar_tensor.default,
+        exir_ops.edge.dim_order_ops._to_dim_order_copy.default,
     }
     # This dict is to ensure that the input of the OPs are int64 due to Pytorch restrictions.
     # For example, scatter op can only accept args[2], the index, as int64.

backends/qualcomm/_passes/lift_constant_scalar_operands.py

@@ -86,7 +86,8 @@ def _build_tensor_constant(
             dtype=(
                 node.args[0].meta["val"].dtype
                 if not is_float_tensor(node)
-                and not SCALAR_OPS.get(node.target).use_self_dtype
+                and (info := SCALAR_OPS.get(node.target))
+                and not info.use_self_dtype
                 else node.meta["val"].dtype
             ),
             device=node.meta["val"].device,

backends/qualcomm/_passes/replace_inf_values.py

@@ -30,6 +30,12 @@ def call(self, graph_module: torch.fx.GraphModule):
                     arg_list[index] = torch.finfo(torch.float32).min
                 elif arg == float("inf"):
                     arg_list[index] = torch.finfo(torch.float32).max
+
+            if node.target == torch.ops.aten.masked_fill.Scalar:
+                if arg_list[2] == torch.finfo(torch.float32).max:
+                    arg_list[2] = 255
+                elif arg_list[2] == torch.finfo(torch.float32).min:
+                    arg_list[2] = -255
             node.args = tuple(arg_list)
 
         graph_module.recompile()

backends/qualcomm/builders/op_cum_sum.py

@@ -51,6 +51,8 @@ def define_node(
         dim = self.get_param(node, input_tensor)
 
         output_tensor = self.get_tensor(node, node)
+        if output_tensor.dtype == torch.int64:
+            output_tensor = output_tensor.to(torch.int32)
         output_tensor_wrapper = self.define_tensor(
             node,
             node,

backends/qualcomm/tests/models.py

@@ -1101,6 +1101,16 @@ def forward(self, x):
         return torch.mean(x, (-1, -2))
 
 
+class MaskedFill(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, attn_mask):
+        return attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(
+            attn_mask == 0, float(0.0)
+        )
+
+
 class Maximum(torch.nn.Module):
     def __init__(self):
         super().__init__()
@@ -1751,16 +1761,6 @@ def forward(self, x):
         )
 
 
-class MaskedFill(torch.nn.Module):
-    def __init__(self):
-        super().__init__()
-
-    def forward(self, attn_mask):
-        return attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(
-            attn_mask == 0, float(0.0)
-        )
-
-
 # Mimi Decoder has 0D tensor which QNN cannot handle.
 class ZeroDimTensor(torch.nn.Module):
     def __init__(self):

backends/qualcomm/tests/test_qnn_delegate.py

@@ -272,9 +272,24 @@ def test_qnn_backend_cos(self):
         self.lower_module_and_test_output(module, sample_input)
 
     def test_qnn_backend_cumsum(self):
-        module = CumSum()  # noqa: F405
-        sample_input = (torch.randn(4),)
-        self.lower_module_and_test_output(module, sample_input)
+        sample_input = ()
+        test_comb = [
+            {
+                QCOM_MODULE: [CumSum()],  # noqa: F405
+                QCOM_SAMPLE_INPUTS: [
+                    (torch.randn(4),),
+                    (torch.randint(0, 10, size=(4,)),),
+                ],
+            }
+        ]
+
+        index = 0
+        for comb in test_comb:
+            for module in comb[QCOM_MODULE]:
+                for sample_input in comb[QCOM_SAMPLE_INPUTS]:
+                    with self.subTest(i=index):
+                        self.lower_module_and_test_output(module, sample_input)
+                        index += 1
 
     def test_qnn_backend_einsum_outer_product(self):
         module = EinsumOuterProduct()  # noqa: F405
@@ -311,6 +326,12 @@ def test_qnn_backend_element_wise_add(self):
                 QCOM_MODULE: [AddConstantFloat()],  # noqa: F405
                 QCOM_SAMPLE_INPUTS: [(torch.randn(2, 5, 1, 3),)],
             },
+            {
+                QCOM_MODULE: [
+                    AddConstantLong(),  # noqa: F405
+                ],
+                QCOM_SAMPLE_INPUTS: [(torch.randint(0, 10, size=(2, 3)),)],
+            },
         ]
 
         index = 0
@@ -4526,6 +4547,40 @@ def test_retinanet(self):
             else:
                 self.assertGreaterEqual(msg["mAP"], 0.6)
 
+    def test_roberta(self):
+        if not self.required_envs([self.sentence_dataset]):
+            self.skipTest("missing required envs")
+        cmds = [
+            "python",
+            f"{self.executorch_root}/examples/qualcomm/oss_scripts/roberta.py",
+            "--dataset",
+            self.sentence_dataset,
+            "--artifact",
+            self.artifact_dir,
+            "--build_folder",
+            self.build_folder,
+            "--device",
+            self.device,
+            "--model",
+            self.model,
+            "--ip",
+            self.ip,
+            "--port",
+            str(self.port),
+        ]
+        if self.host:
+            cmds.extend(["--host", self.host])
+
+        p = subprocess.Popen(cmds, stdout=subprocess.DEVNULL)
+        with Listener((self.ip, self.port)) as listener:
+            conn = listener.accept()
+            p.communicate()
+            msg = json.loads(conn.recv())
+            if "Error" in msg:
+                self.fail(msg["Error"])
+            else:
+                self.assertGreaterEqual(msg["accuracy"], 0.5)
+
     def test_squeezenet(self):
         if not self.required_envs([self.image_dataset]):
             self.skipTest("missing required envs")
@@ -5344,6 +5399,11 @@ def setup_environment():
         help="Location for imagenet dataset",
         type=str,
     )
+    parser.add_argument(
+        "--sentence_dataset",
+        help="Location for sentence dataset",
+        type=str,
+    )
     parser.add_argument(
         "-p",
         "--pretrained_weight",
@@ -5402,6 +5462,7 @@ def setup_environment():
     TestQNN.executorch_root = args.executorch_root
     TestQNN.artifact_dir = args.artifact_dir
     TestQNN.image_dataset = args.image_dataset
+    TestQNN.sentence_dataset = args.sentence_dataset
     TestQNN.pretrained_weight = args.pretrained_weight
     TestQNN.model_name = args.model_name
     TestQNN.online_prepare = args.online_prepare

backends/qualcomm/tests/utils.py

@@ -183,6 +183,7 @@ class TestQNN(unittest.TestCase):
     executorch_root: str = ""
     artifact_dir: str = ""
     image_dataset: str = ""
+    sentence_dataset: str = ""
     pretrained_weight: str = ""
     enable_profile: bool = False
     op_package_dir: str = ""

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

@@ -12,16 +12,18 @@
 
 #define IN_T ${buffer_scalar_type(IN_DTYPE)}
 
+#define ${MODE}
+
 ${define_active_storage_type("buffer")}
 ${define_required_extensions(IN_DTYPE)}
 
 #extension GL_EXT_control_flow_attributes : require
 
 layout(std430) buffer;
 
-${layout_declare_tensor(B, "r", "t_in", IN_DTYPE, "buffer")}
 ${layout_declare_tensor(B, "w", "t_scale", "float", "buffer")}
 ${layout_declare_tensor(B, "w", "t_zero_point", "int", "buffer")}
+${layout_declare_tensor(B, "r", "t_in", IN_DTYPE, "buffer")}
 
 $if MODE == "per_tensor":
   layout(push_constant) uniform restrict Block {
@@ -53,11 +55,11 @@ layout(local_size_x_id = 0, local_size_y_id = 1, local_size_z_id = 2) in;
 shared float shared_min[NWORKERS];
 shared float shared_max[NWORKERS];
 
-void main() {
-$if MODE == "per_tensor":
+#ifdef per_tensor
+
+void choose_qparams_per_tensor() {
   uint global_id = gl_GlobalInvocationID.x;
   uint local_id = gl_LocalInvocationID.x;
-  uint group_id = gl_WorkGroupID.x;
   uint total_threads = gl_NumWorkGroups.x * gl_WorkGroupSize.x;
 
   uint total_elements = uint(t_in_sizes.x * t_in_sizes.y * t_in_sizes.z * t_in_sizes.w);
@@ -114,8 +116,11 @@ $if MODE == "per_tensor":
     t_scale[0] = scale_val;
     t_zero_point[0] = zero_point_val;
   }
+}
+
+#else
 
-$if MODE == "per_token":
+void choose_qparams_per_token() {
   uint global_id = gl_GlobalInvocationID.x;
   uint local_id = gl_LocalInvocationID.x;
   uint group_id = gl_WorkGroupID.x;
@@ -201,3 +206,9 @@ $if MODE == "per_token":
     barrier();
   }
 }
+
+#endif
+
+void main() {
+  choose_qparams_${MODE}();
+}

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

@@ -13,16 +13,18 @@
 #define IN_T ${buffer_scalar_type(IN_DTYPE)}
 #define FVEC4_T ${texel_load_type(IN_DTYPE, "texture3d")}
 
+#define ${MODE}
+
 ${define_active_storage_type("texture3d")}
 ${define_required_extensions(IN_DTYPE)}
 
 #extension GL_EXT_control_flow_attributes : require
 
 layout(std430) buffer;
 
-${layout_declare_tensor(B, "r", "t_in", IN_DTYPE, "texture3d")}
 ${layout_declare_tensor(B, "w", "t_scale", "float", "texture3d")}
 ${layout_declare_tensor(B, "w", "t_zero_point", "int", "texture3d")}
+${layout_declare_tensor(B, "r", "t_in", IN_DTYPE, "texture3d")}
 
 $if MODE == "per_tensor":
   layout(push_constant) uniform restrict Block {
@@ -51,8 +53,9 @@ layout(local_size_x_id = 0, local_size_y_id = 1, local_size_z_id = 2) in;
 shared float shared_min[NWORKERS];
 shared float shared_max[NWORKERS];
 
-void main() {
-$if MODE == "per_tensor":
+#ifdef per_tensor
+
+void choose_qparams_per_tensor() {
   uint global_id = gl_GlobalInvocationID.x;
   uint local_id = gl_LocalInvocationID.x;
   uint group_id = gl_WorkGroupID.x;
@@ -85,7 +88,7 @@ $if MODE == "per_tensor":
     // Calculate total tensor elements to determine padding
     int total_elements = t_in_limits.x * t_in_limits.y * t_in_limits.z * 4;
     int linear_tensor_idx = tensor_coord.x + tensor_coord.y * sizes.x +
-                           tensor_coord.z * sizes.x * sizes.y;
+                            tensor_coord.z * sizes.x * sizes.y;
     int remaining_elements = total_elements - (linear_tensor_idx);
     int valid_elements = min(4, remaining_elements);
 
@@ -168,8 +171,11 @@ $if MODE == "per_tensor":
     write_texel(t_scale, ivec3(0, 0, 0), vec4(scale_val, 0.0, 0.0, 0.0));
     write_texel(t_zero_point, ivec3(0, 0, 0), ivec4(zero_point_val, 0, 0, 0));
   }
+}
+
+#else
 
-$if MODE == "per_token":
+void choose_qparams_per_token() {
   // Each token is processed by multiple workgroups for parallel reduction
   uint local_id = gl_LocalInvocationID.x;
   uint group_id = gl_WorkGroupID.x;
@@ -219,7 +225,7 @@ $if MODE == "per_token":
       // Calculate total tensor elements to determine padding
       int total_elements = t_in_limits.x * t_in_limits.y * t_in_limits.z * 4;
       int linear_tensor_idx = tensor_coord.x + tensor_coord.y * sizes.x +
-                             tensor_coord.z * sizes.x * sizes.y;
+                              tensor_coord.z * sizes.x * sizes.y;
       int remaining_elements = total_elements - (linear_tensor_idx);
       int valid_elements = min(4, remaining_elements);
 
@@ -316,3 +322,9 @@ $if MODE == "per_token":
     barrier();
   }
 }
+
+#endif
+
+void main() {
+  choose_qparams_${MODE}();
+}