1 files changed, 11 insertions, 13 deletions
diff --git a/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_native.cl b/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_native.cl
index e53ce3d1b2..bbe97b2781 100644
--- a/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_native.cl
+++ b/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_native.cl
@@ -170,7 +170,7 @@ __kernel void gemm_mm_native_post_act_eltwise_op_act(IMAGE_DECLARATION(lhs),
 #endif // defined(DUMMY_WORK_ITEMS)
 
     // Compute LHS matrix address
-    uint lhs_offset = lhs_offset_first_element_in_bytes + y * M0 * (uint)lhs_stride_y;
+    uint lhs_offset = lhs_offset_first_element_in_bytes + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y;
 
     // Compute RHS matrix address
     uint rhs_offset = rhs_offset_first_element_in_bytes + x * N0 * sizeof(DATA_TYPE);
@@ -187,7 +187,7 @@ __kernel void gemm_mm_native_post_act_eltwise_op_act(IMAGE_DECLARATION(lhs),
 
 #if defined(REINTERPRET_INPUT_AS_3D)
     // The plane (zlhs) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
-    CALCULATE_Z_OFFSET(M0, uint, zlhs, y * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);
+    CALCULATE_Z_OFFSET(M0, uint, zlhs, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);
 
     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply lhs_stride_z by DEPTH_GEMM3D
@@ -294,17 +294,13 @@ __kernel void gemm_mm_native_post_act_eltwise_op_act(IMAGE_DECLARATION(lhs),
         rhs_offset += rhs_stride_y;
     }
 
-    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * M0 * dst_stride_y);
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y);
 
     REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0);
 
-    // Boundary conditions: detect if current block is at the "bottom" or "right" boundary
-    const bool cond_y = ((y + 1) * M0 >= M);
-    const bool cond_x = ((x + 1) * N0 >= N);
-
 #if defined(REINTERPRET_OUTPUT_AS_3D)
     // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
-    CALCULATE_Z_OFFSET(M0, uint, zout, y * M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
+    CALCULATE_Z_OFFSET(M0, uint, zout, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
 
     // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
     // multiply dst_stride_z by DEPTH_GEMM3D
@@ -327,7 +323,7 @@ __kernel void gemm_mm_native_post_act_eltwise_op_act(IMAGE_DECLARATION(lhs),
 #if defined(BROADCAST_BIAS)
     __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE));
 
-    LOAD_BLOCK_BOUNDARY_AWARE(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, 1, PARTIAL_STORE_N0, false, cond_x);
+    LOAD_BLOCK(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero);
 
 #ifndef UNIT_BETA
     SCALE_BLOCK(1, DATA_TYPE, bias, BETA);
@@ -337,10 +333,9 @@ __kernel void gemm_mm_native_post_act_eltwise_op_act(IMAGE_DECLARATION(lhs),
     ADD_BLOCK_BROADCAST(M0, c, bias0);
 
 #else // defined(BROADCAST_BIAS)
-    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)) + (get_global_id(1) * (uint)M0 * bias_stride_y) + get_global_id(
-                                    2) * bias_stride_z;
+    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z;
 
-    LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
+    LOAD_BLOCK(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero);
 
 #ifndef UNIT_BETA
     SCALE_BLOCK(M0, DATA_TYPE, bias, BETA);
@@ -352,10 +347,13 @@ __kernel void gemm_mm_native_post_act_eltwise_op_act(IMAGE_DECLARATION(lhs),
 #endif // defined(BROADCAST_BIAS)
 #endif // defined(BETA)
 
+    const bool cond_y = y == 0;
+    const bool cond_x = ((x + 1) * N0 >= N);
+
     // c = act(c)
     POST_OP1_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c);
     // c = c + eltwise_operand (mix-precision, broadcast, boundary aware)
-    POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
+    POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, 1, PARTIAL_STORE_N0, false, cond_x);
     // c = act(c)
     POST_OP3_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c);