Added sub kernels to support different input data types and alpha values

xa_nnlib/algo/kernels/basic/xa_nn_mean_f32.c

@@ -27,12 +27,12 @@
 #include <string.h>
 
 // Function to check if two consecutive axes are contiguous
-WORD32 are_two_axes_contiguous(WORD32 a, WORD32 b) {
+static inline WORD32 are_two_axes_contiguous(WORD32 a, WORD32 b) {
     return (b == a + 1);
 }
 
 // Check if an axis is in the given axes list
-WORD32 is_axis_in_list(WORD32 axis, const WORD32 *axes, WORD32 num_axes) {
+static inline WORD32 is_axis_in_list(WORD32 axis, const WORD32 *axes, WORD32 num_axes) {
 	WORD32 i;
     for (i = 0; i < num_axes; i++) {
         if (axes[i] == axis)
@@ -44,7 +44,7 @@ WORD32 is_axis_in_list(WORD32 axis, const WORD32 *axes, WORD32 num_axes) {
 }
 
 // Sort axes in ascending order (Bubble Sort)
-void sort_axes(WORD32 *axes, WORD32 num_axes) {
+static inline void sort_axes(WORD32 *axes, WORD32 num_axes) {
 	WORD32 temp;
 	WORD32 i, j;
     for (i = 0; i < (num_axes - 1); i++) {
@@ -60,7 +60,7 @@ void sort_axes(WORD32 *axes, WORD32 num_axes) {
 
 // Merge contiguous axes
 // Merge contiguous dimensions other than axes
-void merge_axes_dims(const WORD32 *const input_shape, WORD32 num_dims, WORD32 *axes, WORD32 num_axes,
+static inline void merge_axes_dims(const WORD32 *const input_shape, WORD32 num_dims, WORD32 *axes, WORD32 num_axes,
                 WORD32 *new_input_shape, WORD32 *new_num_dims, WORD32 *new_axes, WORD32 *new_num_axes) {
     *new_num_dims = 0;
     *new_num_axes = 0;
@@ -223,7 +223,7 @@ WORD32 xa_nn_mean_f32_f32(FLOAT32 *__restrict__ p_out,
 	merge_axes_dims(p_inp_shape, num_inp_dims, new_axes, num_axis_dims, new_input_shape, &new_num_inp_dims, new_axes_data, &new_num_axis_dims);
 
 	WORD32 last_dim = 0;
-	
+
 	if(new_axes_data[new_num_axis_dims - CONST_ONE] == (new_num_inp_dims - CONST_ONE))
 	{
 		last_dim = CONST_ONE;
@@ -348,7 +348,7 @@ WORD32 xa_nn_mean_f32_f32(FLOAT32 *__restrict__ p_out,
 							/* Load input elements with stride "inner_stride" */
 							PDX_LAV_MXF32_XP(x1, align_src1, p_in_mxf32, rem_elm);
 
-							rem_sum = rem_sum + x1;//PDX_ADD_MXF32(rem_sum, x1);
+							rem_sum = PDX_ADD_MXF32(rem_sum, x1);
 						}
 
 						/* Store output */
@@ -684,7 +684,6 @@ WORD32 xa_nn_mean_f32_f32(FLOAT32 *__restrict__ p_out,
 						out = PDX_RADD_MXF32(sum);
 						xtfloat_storeip(out, p_z, 4);
 					}
-
 				}
 			}
 			else
@@ -772,7 +771,7 @@ WORD32 xa_nn_mean_f32_f32(FLOAT32 *__restrict__ p_out,
 						/* Store output */
 						PDX_SAV_MXF32_XP(rem_sum, align_dst, p_dst, rem_elm);
 					}
-					PDX_SAPOS_MXF32_FP(align_dst, p_dst);			
+					PDX_SAPOS_MXF32_FP(align_dst, p_dst);
 				}
 			}
 		}

xa_nnlib/build/common.mk

@@ -129,17 +129,14 @@ $(LIBOBJ): $(OBJ_LIBOBJS) $(OBJ_LIBOSOBJS)
 	$(QUIET) -$(RM) $(TEMPOBJ)
 endif 
 
-
 $(OBJ_LIBOBJS): $(OBJDIR)/%.o: %.c
 	@echo "Compiling $<"
 	$(QUIET) $(CC) -o $@ $(OPT_O3) $(CFLAGS) $(INCLUDES) -c $<
 
 $(OBJ_LIBOSOBJS): $(OBJDIR)/%.o: %.c
 	@echo "Compiling $<"
 	$(QUIET) $(CC) -o $@ $(OPT_OS) $(CFLAGS) $(INCLUDES) -c $<
-
-
-
+
 $(LIB): %.a: $(OBJDIR)/%.o
 	@echo "Creating Library $@"
 	$(QUIET) $(AR) rc $@ $^

xa_nnlib/build/ldscript_nnlib.txt

@@ -77,7 +77,18 @@ EXTERN(xa_nn_elm_less_scalar_f32xf32_bool)
 EXTERN(xa_nn_elm_less_broadcast_5D_f32xf32_bool)
 EXTERN(xa_nn_elm_where_f32xf32_f32)
 EXTERN(xa_nn_elm_where_broadcast_5D_f32xf32_f32)
-
+EXTERN(xa_nn_elm_sub_f32x32xf32_f32)
+EXTERN(xa_nn_elm_sub_scalar_f32x32xf32_f32)
+EXTERN(xa_nn_elm_sub_broadcast_5D_f32x32xf32_f32)
+EXTERN(xa_nn_elm_sub_f32x32x32_f32)
+EXTERN(xa_nn_elm_sub_scalar_f32x32x32_f32)
+EXTERN(xa_nn_elm_sub_broadcast_5D_f32x32x32_f32)
+EXTERN(xa_nn_elm_sub_32xf32xf32_f32)
+EXTERN(xa_nn_elm_sub_scalar_32xf32xf32_f32)
+EXTERN(xa_nn_elm_sub_broadcast_5D_32xf32xf32_f32)
+EXTERN(xa_nn_elm_sub_32xf32x32_f32)
+EXTERN(xa_nn_elm_sub_scalar_32xf32x32_f32)
+EXTERN(xa_nn_elm_sub_broadcast_5D_32xf32x32_f32)
 
 /* Normalization kernels */
 EXTERN(xa_nn_native_layer_norm_f32_f32)

xa_nnlib/build/makefile_nn_lib_fusion_g3

@@ -82,6 +82,8 @@ BASICOBJS =                          \
   xa_nn_elm_rsqrt_f32.o \
   xa_nn_elm_where_f32.o \
   xa_nn_elm_less_f32.o \
+  xa_nn_elm_sub_32xf32.o \
+  xa_nn_elm_sub_f32x32.o \
 
 
 NORMOBJS = \

xa_nnlib/build/symbols_nnlib.txt

@@ -77,3 +77,16 @@ xa_nn_elm_where_f32xf32_f32
 xa_nn_elm_where_broadcast_5D_f32xf32_f32
 xa_nn_sigmoid_f32_f32
 xa_nn_tanh_f32_f32
+xa_nn_elm_sub_f32x32xf32_f32
+xa_nn_elm_sub_scalar_f32x32xf32_f32
+xa_nn_elm_sub_broadcast_5D_f32x32xf32_f32
+xa_nn_elm_sub_f32x32x32_f32
+xa_nn_elm_sub_scalar_f32x32x32_f32
+xa_nn_elm_sub_broadcast_5D_f32x32x32_f32
+xa_nn_elm_sub_32xf32xf32_f32
+xa_nn_elm_sub_scalar_32xf32xf32_f32
+xa_nn_elm_sub_broadcast_5D_32xf32xf32_f32
+xa_nn_elm_sub_32xf32x32_f32
+xa_nn_elm_sub_scalar_32xf32x32_f32
+xa_nn_elm_sub_broadcast_5D_32xf32x32_f32
+

xa_nnlib/include/nnlib/xa_nnlib_kernels_api.h

@@ -475,6 +475,144 @@ WORD32 xa_nn_mean_f32_f32(FLOAT32 *__restrict__ p_out,
         const WORD32 *__restrict__ p_axis,
         WORD32 num_axis_dims);
 
+WORD32 xa_nn_elm_clamp_f32_f32(FLOAT32 *__restrict__ p_out,
+        const FLOAT32 *__restrict__ p_inp,
+        const FLOAT32 *__restrict__ p_min,
+        const FLOAT32 *__restrict__ p_max,
+        WORD32 num_elm);
+
+WORD32 xa_nn_elm_clamp_scalar_f32_f32(FLOAT32 *__restrict__ p_out,
+        const FLOAT32 *__restrict__ p_inp,
+        const FLOAT32 min,
+        const FLOAT32 max,
+        WORD32 num_elm);
+
+WORD32 xa_nn_elm_clamp_broadcast_5D_f32_f32(FLOAT32 *__restrict__ p_out,
+        const WORD32 *const p_out_shape,
+        const FLOAT32 *__restrict__ p_inp,
+        const WORD32 *const p_inp_shape,
+        const FLOAT32 *__restrict__ p_min,
+        const WORD32 *const p_min_shape,
+        const FLOAT32 *__restrict__ p_max,
+        const WORD32 *const p_max_shape,
+        WORD32 num_inp_dims);
+
+WORD32 xa_nn_elm_where_f32xf32_f32(FLOAT32 *p_out,
+        const FLOAT32 *p_inp1,
+        const FLOAT32 *p_inp2,
+        const UWORD8 *p_cond,
+        WORD32 num_elm);
+
+WORD32 xa_nn_tanh_f32_f32(FLOAT32 *p_out,
+        const FLOAT32 *p_inp,
+        WORD32 vec_length);
+
+WORD32 xa_nn_elm_where_broadcast_5D_f32xf32_f32(FLOAT32 *__restrict__ p_out,
+        const WORD32 *const p_out_shape,
+        const FLOAT32 *__restrict__ p_inp1,
+        const WORD32 *const p_inp1_shape,
+        const FLOAT32 *__restrict__ p_inp2,
+        const WORD32 *const p_inp2_shape,
+        const UWORD8 *p_cond,
+        const WORD32 *const p_cond_shape,
+        WORD32 num_inp_dims);
+
+WORD32 xa_nn_elm_less_f32xf32_bool(WORD8 *p_out,
+        const FLOAT32 *p_inp1,
+        const FLOAT32 *p_inp2,
+        WORD32 num_elm);
+
+WORD32 xa_nn_elm_less_scalar_f32xf32_bool(WORD8 *p_out,
+        const FLOAT32 *p_inp1,
+        const FLOAT32 inp2,
+        WORD32 num_elm);
+
+WORD32 xa_nn_elm_less_broadcast_5D_f32xf32_bool(WORD8 *__restrict__ p_out,
+        const WORD32 *const p_out_shape,
+        const FLOAT32 *__restrict__ p_inp1,
+        const WORD32 *const p_inp1_shape,
+        const FLOAT32 *__restrict__ p_inp2,
+        const WORD32 *const p_inp2_shape,
+        WORD32 num_inp_dims);
+
+WORD32 xa_nn_elm_sqrt_f32_f32(FLOAT32 *p_out,
+        const FLOAT32 *p_inp,
+        WORD32 num_elm);
+
+WORD32 xa_nn_elm_rsqrt_f32_f32(FLOAT32 *p_out,
+        const FLOAT32 *p_inp,
+        WORD32 num_elm);
+
+WORD32 xa_nn_sigmoid_f32_f32(FLOAT32 *p_out,
+        const FLOAT32 *p_inp,
+        WORD32 vec_length);
+
+WORD32 xa_nn_elm_clamp_8_8(WORD8 *p_out,
+        const WORD8 *p_inp,
+        const WORD8 *p_min,
+        const WORD8 *p_max,
+        WORD32 num_elm);
+
+WORD32 xa_nn_elm_clamp_scalar_8_8(WORD8 *p_out,
+        const WORD8 *p_inp,
+        const WORD8 min,
+        const WORD8 max,
+        WORD32 num_elm);
+
+WORD32 xa_nn_elm_clamp_broadcast_5D_8_8(WORD8 *__restrict__ p_out,
+        const WORD32 *const p_out_shape,
+        const WORD8 *__restrict__ p_inp,
+        const WORD32 *const p_inp_shape,
+        const WORD8 *__restrict__ p_min,
+        const WORD32 *const p_min_shape,
+        const WORD8 *__restrict__ p_max,
+        const WORD32 *const p_max_shape,
+        WORD32 num_inp_dims);
+
+WORD32 xa_nn_elm_clamp_8u_8u(UWORD8 *p_out,
+        const UWORD8 *p_inp,
+        const UWORD8 *p_min,
+        const UWORD8 *p_max,
+        WORD32 num_elm);
+
+WORD32 xa_nn_elm_clamp_scalar_8u_8u(UWORD8 *p_out,
+        const UWORD8 *p_inp,
+        const UWORD8 min,
+        const UWORD8 max,
+        WORD32 num_elm);
+
+WORD32 xa_nn_elm_clamp_broadcast_5D_8u_8u(UWORD8 *__restrict__ p_out,
+        const WORD32 *const p_out_shape,
+        const UWORD8 *__restrict__ p_inp,
+        const WORD32 *const p_inp_shape,
+        const UWORD8 *__restrict__ p_min,
+        const WORD32 *const p_min_shape,
+        const UWORD8 *__restrict__ p_max,
+        const WORD32 *const p_max_shape,
+        WORD32 num_inp_dims);
+
+WORD32 xa_nn_elm_clamp_16_16(WORD16 *p_out,
+        const WORD16 *p_inp,
+        const WORD16 *p_min,
+        const WORD16 *p_max,
+        WORD32 num_elm);
+
+WORD32 xa_nn_elm_clamp_scalar_16_16(WORD16 *p_out,
+        const WORD16 *p_inp,
+        const WORD16 min,
+        const WORD16 max,
+        WORD32 num_elm);
+
+WORD32 xa_nn_elm_clamp_broadcast_5D_16_16(WORD16 *__restrict__ p_out,
+        const WORD32 *const p_out_shape,
+        const WORD16 *__restrict__ p_inp,
+        const WORD32 *const p_inp_shape,
+        const WORD16 *__restrict__ p_min,
+        const WORD32 *const p_min_shape,
+        const WORD16 *__restrict__ p_max,
+        const WORD32 *const p_max_shape,
+        WORD32 num_inp_dims);
+
 #if defined(__cplusplus)
 }
 #endif