more tests, enabled fusing quantize

Author: michal-miotk

src/plugins/intel_gpu/src/kernel_selector/cl_kernels/resample_pil_ref.cl

@@ -58,7 +58,7 @@ KERNEL (calculate_coefficients_gpu_ref)(__global float* coefficients, __global i
         ww += w;
     }
     for (x = 0; x < xmax; x++) {
-        if (ww != 0.0) {
+        if (ww != 0.0f) {
             k[x] /= ww;
         }
     }
@@ -72,10 +72,16 @@ KERNEL (calculate_coefficients_gpu_ref)(__global float* coefficients, __global i
 
 #elif RESAMPLE_PILLOW_STAGE == STAGE_RESAMPLE_HORIZONTAL
 
+#if ENABLE_VERTICAL_PASS
+#define RESAMPLE_HORIZONTAL_OUTPUT_TYPE INTERMEDIATE_BUF_TYPE
+#else
+#define RESAMPLE_HORIZONTAL_OUTPUT_TYPE OUTPUT_TYPE
+#endif
+
 KERNEL (resample_horizontal_gpu_ref)(  __global INPUT0_TYPE* input
                                      , __global float* coefficients
                                      , __global int* bounds
-                                     , __global OUTPUT_TYPE* output
+                                     , __global RESAMPLE_HORIZONTAL_OUTPUT_TYPE* output
 #if HAS_FUSED_OPS_DECLS
                                      , FUSED_OPS_DECLS
 #endif
@@ -97,7 +103,7 @@ KERNEL (resample_horizontal_gpu_ref)(  __global INPUT0_TYPE* input
 #endif
     int horizontal_min, horizontal_max;
     float* k;
-    OUTPUT_TYPE ss = 0.f;
+    ACCUMULATOR_TYPE ss = 0.f;
 #if BATCH_IS_HORIZONTAL_AXIS == 1
     horizontal_min = bounds[b * 2 + 0];
     horizontal_max = bounds[b * 2 + 1];
@@ -157,7 +163,13 @@ KERNEL (resample_horizontal_gpu_ref)(  __global INPUT0_TYPE* input
 
 #else // RESAMPLE_PILLOW_STAGE == STAGE_RESAMPLE_VERTICAL
 
-KERNEL (resample_vertical_gpu_ref)(  __global INPUT0_TYPE* input
+#if ENABLE_HORIZONTAL_PASS
+#define RESAMPLE_VERTICAL_INPUT_TYPE INTERMEDIATE_BUF_TYPE
+#else
+#define RESAMPLE_VERTICAL_INPUT_TYPE INPUT0_TYPE
+#endif
+
+KERNEL (resample_vertical_gpu_ref)(  __global RESAMPLE_VERTICAL_INPUT_TYPE* input
                                      , __global float* coefficients
                                      , __global int* bounds
                                      , __global OUTPUT_TYPE* output
@@ -194,7 +206,7 @@ KERNEL (resample_vertical_gpu_ref)(  __global INPUT0_TYPE* input
     vertical_max = bounds[x * 2 + 1];
 #endif
 
-    OUTPUT_TYPE ss = 0.f;
+    ACCUMULATOR_TYPE ss = 0.f;
     for (int vertical_dim = 0; vertical_dim < vertical_max; vertical_dim++) {
 #if ENABLE_HORIZONTAL_PASS
 
@@ -254,3 +266,9 @@ KERNEL (resample_vertical_gpu_ref)(  __global INPUT0_TYPE* input
 #endif
 
 #undef PILLOW_SUPPORT
+#ifdef RESAMPLE_HORIZONTAL_OUTPUT_TYPE
+#undef RESAMPLE_HORIZONTAL_OUTPUT_TYPE
+#endif
+#ifdef RESAMPLE_VERTICAL_INPUT_TYPE
+#undef RESAMPLE_VERTICAL_INPUT_TYPE
+#endif

src/plugins/intel_gpu/src/kernel_selector/kernels/resample/resample_kernel_pil_ref.cpp

@@ -122,7 +122,7 @@ DataTensor GetIntermediateBufferSize(const resample_params& params) {
     OPENVINO_ASSERT(channelIndex >= 0, "Invalid layout channel index");
 
     dims[channelIndex] = ybox_last - ybox_first;
-    DataTensor result{dims, output.GetDType(), layout};
+    DataTensor result{dims, Datatype::F32, layout};
     return result;
 }
 
@@ -134,6 +134,8 @@ ParamsKey ResampleKernelPilRef::GetSupportedKey() const {
     k.EnableInputDataType(Datatype::F32);
     k.EnableOutputDataType(Datatype::F16);
     k.EnableOutputDataType(Datatype::F32);
+    k.EnableOutputDataType(Datatype::UINT8);
+    k.EnableOutputDataType(Datatype::INT8);
     k.EnableDifferentTypes();
     k.EnableAllInputLayout();
     k.EnableAllOutputLayout();
@@ -281,6 +283,7 @@ JitConstants ResampleKernelPilRef::GetJitConstantsForKernel(KernelId id, const r
         MakeJitConstant("STAGE_CALC_VERTICAL_COEFFICIENTS", static_cast<int>(eCalcVerticalCoefficients)),
         MakeJitConstant("STAGE_RESAMPLE_VERTICAL", static_cast<int>(eResampleVertical)),
     });
+    jit_constants.Merge(MakeTypeJitConstants(GetAccumulatorType(params), "ACCUMULATOR"));
     switch (id) {
         case eCalcHorizontalCoefficients: {
             auto inputHorizontalSizeWithPadding = getInputHorizontalSize(params, true);

src/plugins/intel_gpu/tests/unit/fusions/resample_fusion_test.cpp

@@ -473,3 +473,54 @@ INSTANTIATE_TEST_SUITE_P(fusings_gpu, resample_bicubic_pillow_axes_scale_activat
         resample_axes_test_params{ CASE_RESAMPLE_BICUBIC_PILLOW_AXES_1, RESAMPLE_BICUBIC_PILLOW_AXES_SCALE_ACTIVATION_ELTWISE_CNT },
         resample_axes_test_params{ CASE_RESAMPLE_BICUBIC_PILLOW_AXES_2, RESAMPLE_BICUBIC_PILLOW_AXES_SCALE_ACTIVATION_ELTWISE_CNT },
 }));
+
+class resample_bicubic_pillow_axes_activation : public ResampleAxesPrimitiveFusingTest {};
+TEST_P(resample_bicubic_pillow_axes_activation, basic) {
+    auto p = GetParam();
+    auto sizes = get_sizes_for_axes(p);
+    create_topologies(
+        input_layout("input", get_input_layout(p)),
+        resample("resample_prim", input_info("input"), sizes, {}, p.axes, {}, {}, 0, -0.75f,
+                 p.type, resample::InterpolateOp::ShapeCalcMode::SIZES),
+        activation("activation", input_info("resample_prim"), activation_func::abs),
+        reorder("reorder_bfyx", input_info("activation"), p.default_format, data_types::f32)
+    );
+
+    tolerance = 1e-2f;
+    execute(p);
+}
+
+#define RESAMPLE_BICUBIC_PILLOW_AXES_ACTIVATION_CNT 2, 3
+INSTANTIATE_TEST_SUITE_P(fusings_gpu, resample_bicubic_pillow_axes_activation,
+    ::testing::ValuesIn(std::vector<resample_axes_test_params>{
+        resample_axes_test_params{ CASE_RESAMPLE_BICUBIC_PILLOW_AXES_1, RESAMPLE_BICUBIC_PILLOW_AXES_ACTIVATION_CNT },
+        resample_axes_test_params{ CASE_RESAMPLE_BICUBIC_PILLOW_AXES_2, RESAMPLE_BICUBIC_PILLOW_AXES_ACTIVATION_CNT },
+}));
+
+class resample_bicubic_pillow_axes_quantize : public ResampleAxesPrimitiveFusingTest {};
+TEST_P(resample_bicubic_pillow_axes_quantize, basic) {
+    auto p = GetParam();
+    auto sizes = get_sizes_for_axes(p);
+    create_topologies(
+        input_layout("input", get_input_layout(p)),
+        data("in_lo", get_mem(get_per_channel_layout(p), min_random, 0)),
+        data("in_hi", get_mem(get_per_channel_layout(p), 1, max_random)),
+        data("out_lo", get_mem(get_single_element_layout(p), -127)),
+        data("out_hi", get_mem(get_single_element_layout(p), 127)),
+        resample("resample_prim", input_info("input"), sizes, {}, p.axes, {}, {}, 0, -0.75f,
+                 p.type, resample::InterpolateOp::ShapeCalcMode::SIZES),
+        quantize("quantize", input_info("resample_prim"), input_info("in_lo"), input_info("in_hi"),
+                 input_info("out_lo"), input_info("out_hi"), 255, data_types::i8),
+        reorder("reorder_bfyx", input_info("quantize"), p.default_format, data_types::f32)
+    );
+
+    tolerance = 1;
+    execute(p);
+}
+
+#define RESAMPLE_BICUBIC_PILLOW_AXES_QUANTIZE_CNT 2, 3
+INSTANTIATE_TEST_SUITE_P(fusings_gpu, resample_bicubic_pillow_axes_quantize,
+    ::testing::ValuesIn(std::vector<resample_axes_test_params>{
+        resample_axes_test_params{ CASE_RESAMPLE_BICUBIC_PILLOW_AXES_1, RESAMPLE_BICUBIC_PILLOW_AXES_QUANTIZE_CNT },
+        resample_axes_test_params{ CASE_RESAMPLE_BICUBIC_PILLOW_AXES_2, RESAMPLE_BICUBIC_PILLOW_AXES_QUANTIZE_CNT },
+}));