[GPU] softmax_kernel_items_class_optimized fix (#18178)

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

@@ -88,7 +88,6 @@ KERNEL(reverse_ref)(
         }
 #if INPUT0_DIMS >= 5
         else if (axis[i] == Z_INDEX) {
-            printf("outputZ %d", OUTPUT_SIZE_Z);
             z = OUTPUT_SIZE_Z - z - 1;
         }
 #if INPUT0_DIMS == 6

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

@@ -52,7 +52,7 @@ KERNEL (softmax_gpu_continuous_bfyx)(
     INPUT0_TYPE my_chunk[STACK_SIZE];
     INPUT0_TYPE my_maximum = -UNIT_VAL_MAX;
     INPUT0_TYPE my_sum = UNIT_VAL_ZERO;
-    
+
     __local INPUT0_TYPE lg_storage[SLM_SIZE];
 
     uint i=0;
@@ -61,9 +61,6 @@ KERNEL (softmax_gpu_continuous_bfyx)(
     {
         for (; i<items_num - (items_num % SUBGROUP_BLOCK_SIZE); i+=SUBGROUP_BLOCK_SIZE)
         {
-            INPUT0_TYPE tmp = input[data_set_offset + subgroup_offset + i * get_sub_group_size() + get_sub_group_local_id()];
-            my_maximum = max(my_maximum, tmp);
-            my_chunk[i] = tmp;
             BLOCK_TYPE vec_tmp = BLOCK_READ(input, data_set_offset + subgroup_offset + i * get_sub_group_size());
             for (int j = 0; j < SUBGROUP_BLOCK_SIZE; j++)
             {
@@ -87,7 +84,7 @@ KERNEL (softmax_gpu_continuous_bfyx)(
         my_chunk[items_num] = tmp;
     }
     my_maximum = sub_group_reduce_max(my_maximum);
-    
+
     if (get_sub_group_local_id() == 0)
         lg_storage[get_sub_group_id()] = my_maximum;
 
@@ -119,7 +116,7 @@ KERNEL (softmax_gpu_continuous_bfyx)(
         my_sum += tmp;
         my_chunk[items_num] = tmp;
     }
-    
+
     my_sum = sub_group_reduce_add(my_sum);
 
     if (get_sub_group_local_id() == 0)
@@ -136,7 +133,7 @@ KERNEL (softmax_gpu_continuous_bfyx)(
     barrier(CLK_LOCAL_MEM_FENCE);
 
     my_sum = lg_storage[0];
-    
+
 
     i=0;
 #if HAS_FUSED_OPS

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

@@ -126,7 +126,7 @@ KERNEL(softmax_items_class_optimized)(
     denominator = sub_group_reduce_add(denominator);
 
     // PART 3. Write out results
-    uint output_idx = out_depth_offset;// + simd_lane * OUTPUT_CLASS_PITCH;
+    uint output_idx = out_depth_offset;
     cls = 0;
 #if FULL_ITERATIONS_NUM >= SUB_GROUP_SIZE && IS_SUBGROUP_BLOCK_IO_ENABLED
     for (; cls < FULL_ITERATIONS_NUM - (FULL_ITERATIONS_NUM % BLOCK_SIZE); cls += BLOCK_SIZE)

src/plugins/intel_gpu/src/kernel_selector/kernels/softmax/softmax_kernel_items_class_optimized.cpp

@@ -80,9 +80,9 @@ JitConstants SoftmaxKerneItemsClassOptimized::GetJitConstants(const softmax_para
     auto jit = SoftmaxItemsClassKernelBase::GetJitConstants(params, dispatchData);
 
     // sub_group_block_write requires aligned memory,
-    // therefore it can be utilized if either memory is aligned by 16 bytes or there is only 1 dataset
+    // therefore it can be utilized if either memory is aligned by 16 bytes
     bool isSubGroupBlockIOEnabled = params.dim != SoftmaxDim::BATCH &&
-        (dispatchData.dataSetSize % (16 / params.inputs[0].ElementSize()) == 0 || dispatchData.dataSetsCount == 1);
+        (dispatchData.dataSetSize * params.outputs[0].ElementSize()) % 16 == 0;
 
     jit.AddConstants({
         MakeJitConstant("LEFTOVERS", dispatchData.leftovers),

src/plugins/intel_gpu/tests/functional/shared_tests_instances/single_layer_tests/softmax.cpp

@@ -71,6 +71,7 @@ const std::vector<ov::Shape> inputShapes4D = {
     {1, 100, 1, 1},
     {1, 3, 4, 3},
     {2, 3, 4, 5},
+    {1, 32, 1, 286},
 };
 
 const std::vector<int64_t> axis4D = {-4, -3, -2, -1, 0, 1, 2, 3};