IntelGPU backend: Fix Convert, Pad and Pool (#2837)

src/ngraph/runtime/intelgpu/intelgpu_backend.cpp

@@ -190,6 +190,11 @@ static void do_cldnn_unary(cldnn::topology& topology,
     topology.add(cldnn_unary);
 }
 
+static bool has_non_zero(const Shape& shape)
+{
+    return accumulate(shape.begin(), shape.end(), 0);
+}
+
 static void
     do_custom_unary(cldnn::topology& topology, const shared_ptr<Node>& op, const string& operation)
 {
@@ -802,7 +807,8 @@ shared_ptr<runtime::Executable>
 
             if ((op->get_input_shape(0).size() > 4) ||
                 (op->get_output_element_type(0) != element::f32) ||
-                !max_pool->get_padding_below().empty() || !max_pool->get_padding_above().empty())
+                has_non_zero(max_pool->get_padding_below()) ||
+                has_non_zero(max_pool->get_padding_above()))
             {
                 const shared_ptr<Node> def_val = max_pool->get_default_value();
                 const shared_ptr<op::Constant> def_const =
@@ -869,7 +875,8 @@ shared_ptr<runtime::Executable>
             if ((op->get_input_shape(0).size() > 4) ||
                 (op->get_output_element_type(0) != element::f32) ||
                 avg_pool->get_include_padding_in_avg_computation() ||
-                !avg_pool->get_padding_below().empty() || !avg_pool->get_padding_above().empty())
+                has_non_zero(avg_pool->get_padding_below()) ||
+                has_non_zero(avg_pool->get_padding_above()))
             {
                 const shared_ptr<Node> def_val = avg_pool->get_default_value();
                 const shared_ptr<op::Constant> def_const =
@@ -1194,13 +1201,29 @@ shared_ptr<runtime::Executable>
             break;
         }
         case OP_TYPEID::Ceiling:
+        {
+            if (!op->get_input_element_type(0).is_real())
+            {
+                do_equal_propagation(
+                    topology, op->get_input_tensor_name(0), op->get_output_tensor_name(0));
+            }
+            else
             {
                 do_custom_unary(topology, op, "ceil(input_var)");
+            }
             break;
         }
         case OP_TYPEID::Floor:
+        {
+            if (!op->get_input_element_type(0).is_real())
+            {
+                do_equal_propagation(
+                    topology, op->get_input_tensor_name(0), op->get_output_tensor_name(0));
+            }
+            else
             {
                 do_custom_unary(topology, op, "floor(input_var)");
+            }
             break;
         }
         case OP_TYPEID::Sign:

src/ngraph/runtime/intelgpu/intelgpu_op_custom_kernels.cpp

@@ -337,7 +337,12 @@ void runtime::intelgpu::do_pad_operation(cldnn::topology& topology,
     Shape pad_interior(pad_below.size(), 0);
 
     // The kernel name and parameters
-    gen_func_def(writer, entry_point_name, {2, "float"}, {input_shape, {1}}, "float", output_shape);
+    gen_func_def(writer,
+                 entry_point_name,
+                 {2, get_opencl_type_name(output_type)},
+                 {input_shape, {1}},
+                 get_opencl_type_name(output_type),
+                 output_shape);
 
     writer.block_begin();
     {
@@ -1677,6 +1682,31 @@ void runtime::intelgpu::do_one_hot_operation(cldnn::topology& topology,
     topology.add(op_one_hot);
 }
 
+static string emit_convert_bool(const string& input_type)
+{
+    CodeWriter writer;
+
+    writer << "bool convert_bool(const " << input_type << " input)";
+    writer.block_begin();
+    {
+        writer << "if (input)\n";
+        writer.block_begin();
+        {
+            writer << "return 1;\n";
+        }
+        writer.block_end();
+        writer << "else\n";
+        writer.block_begin();
+        {
+            writer << "return 0;\n";
+        }
+        writer.block_end();
+    }
+    writer.block_end();
+
+    return writer.get_code();
+}
+
 void runtime::intelgpu::do_convert_operation(cldnn::topology& topology,
                                              const string& input_name,
                                              const Shape& input_shape,
@@ -1692,6 +1722,11 @@ void runtime::intelgpu::do_convert_operation(cldnn::topology& topology,
     CodeWriter writer;
     vector<size_t> gws;
 
+    if (output_type == element::Type_t::boolean)
+    {
+        writer << emit_convert_bool(input_type_name);
+    }
+
     gen_func_def(
         writer, entry_point_name, {input_type_name}, {input_shape}, output_type_name, output_shape);
 
@@ -1764,8 +1799,12 @@ void runtime::intelgpu::do_sigmoid_backprop_operation(cldnn::topology& topology,
     CodeWriter writer;
     vector<size_t> gws;
 
-    gen_func_def(
-        writer, entry_point_name, {2, "float"}, {input_shape, delta_shape}, "float", output_shape);
+    gen_func_def(writer,
+                 entry_point_name,
+                 {2, get_opencl_type_name(output_type)},
+                 {input_shape, delta_shape},
+                 get_opencl_type_name(output_type),
+                 output_shape);
 
     writer.block_begin();
     {
@@ -1935,8 +1974,12 @@ void runtime::intelgpu::do_reshape_operation(cldnn::topology& topology,
 {
     const cldnn::layout layout = IntelGPULayout::create_cldnn_layout(output_type, output_shape);
     const string entry_point_name = "reshape_" + output_name;
-    const string& input_type_name = get_opencl_type_name(input_type);
-    const string& output_type_name = get_opencl_type_name(output_type);
+
+    // Workaround on openCL bool datatype. Need to be the same as CPU
+    const string& input_type_name =
+        (input_type == element::Type_t::boolean) ? "char" : get_opencl_type_name(input_type);
+    const string& output_type_name =
+        (output_type == element::Type_t::boolean) ? "char" : get_opencl_type_name(output_type);
     const size_t dst_shape_size = shape_size(output_shape);
     CodeWriter writer;
 

src/ngraph/runtime/intelgpu/unit_test.manifest

-all_2x2x3_eliminate_dims_0_1
 backwards_batch_norm_training
 backwards_dot_scalar_tensor
 backwards_dot_tensor_scalar
@@ -22,9 +21,6 @@ shape_of_5d
 shape_of_matrix
 shape_of_scalar
 shape_of_vector
-floor_int32
-convert_int32_bool
-convert_float32_bool
 prelu
 prelu_shared_slope
 prelu_negative_slope