update avepool specializations

Author: JaccovG

include/api/mli_krn_avepool_spec_api.h

@@ -26,84 +26,64 @@ extern "C" {
 
 mli_status mli_krn_avepool_chw_fx16_k2x2_str1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
 mli_status mli_krn_avepool_chw_fx16_k4x4_str1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx16_k5x5_str1_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx16_k7x7_str1_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx16_k9x9_str1_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx16_k4x2_str1_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx16_k4x4_str1_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx16_k4x6_str1_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx16_k4x8_str1_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx16_k6x2_str1_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx16_k6x4_str1_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx16_k6x6_str1_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx16_k6x8_str1_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx16_k8x2_str1_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx16_k8x4_str1_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx16_k8x6_str1_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx16_k8x8_str1_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx16_k4x2_str1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx16_k6x2_str1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx16_k6x4_str1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx16_k6x6_str1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx16_k6x8_str1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx16_k8x2_str1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx16_k8x4_str1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx16_k8x6_str1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx16_k8x8_str1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx16_k3x3_str1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx16_k5x5_str1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx16_k7x7_str1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx16_k9x9_str1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx16_k2x2_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
 mli_status mli_krn_avepool_chw_fx16_k3x3_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx16_k4x4_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
 mli_status mli_krn_avepool_chw_fx16_k5x5_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx16_k6x6_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
 mli_status mli_krn_avepool_chw_fx16_k7x7_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx16_k8x8_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
 mli_status mli_krn_avepool_chw_fx16_k9x9_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx16_k10x10_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx16_k2x2_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx16_k4x4_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx16_k6x6_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx16_k8x8_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx16_k3x3_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx16_k5x5_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx16_k7x7_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx16_k9x9_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx16_k4x4_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx16_k6x6_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx16_k8x8_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx16_k1xn_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx16_k1x2_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx16_k1x3_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx16_knx1_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx16_k2x1_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx16_k3x1_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx16_k1xn_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx16_k1x2_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx16_k1x3_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx16_knx1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx16_k2x1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx16_k3x1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
 mli_status mli_krn_avepool_chw_fx16_generic(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
 
 mli_status mli_krn_avepool_chw_fx8_k2x2_str1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
 mli_status mli_krn_avepool_chw_fx8_k4x4_str1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx8_k5x5_str1_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx8_k7x7_str1_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx8_k9x9_str1_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx8_k4x2_str1_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx8_k4x4_str1_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx8_k4x6_str1_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx8_k4x8_str1_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx8_k6x2_str1_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx8_k6x4_str1_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx8_k6x6_str1_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx8_k6x8_str1_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx8_k8x2_str1_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx8_k8x4_str1_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx8_k8x6_str1_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx8_k8x8_str1_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx8_k4x2_str1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx8_k6x2_str1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx8_k6x4_str1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx8_k6x6_str1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx8_k6x8_str1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx8_k8x2_str1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx8_k8x4_str1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx8_k8x6_str1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx8_k8x8_str1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx8_k3x3_str1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx8_k5x5_str1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx8_k7x7_str1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx8_k9x9_str1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx8_k2x2_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
 mli_status mli_krn_avepool_chw_fx8_k3x3_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx8_k4x4_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
 mli_status mli_krn_avepool_chw_fx8_k5x5_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx8_k6x6_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
 mli_status mli_krn_avepool_chw_fx8_k7x7_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx8_k8x8_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
 mli_status mli_krn_avepool_chw_fx8_k9x9_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
-mli_status mli_krn_avepool_chw_fx8_k10x10_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx8_k2x2_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx8_k4x4_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx8_k6x6_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx8_k8x8_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx8_k3x3_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx8_k5x5_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx8_k7x7_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx8_k9x9_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx8_k4x4_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx8_k6x6_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx8_k8x8_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx8_k1xn_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx8_k1x2_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx8_k1x3_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx8_knx1_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx8_k2x1_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx8_k3x1_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx8_k1xn_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx8_k1x2_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx8_k1x3_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx8_knx1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx8_k2x1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_avepool_chw_fx8_k3x1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
 mli_status mli_krn_avepool_chw_fx8_generic(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
 
 #ifdef __cplusplus

lib/gen/codegen.py

@@ -111,13 +111,13 @@ def print_if_list(self, func_list, hierargy_list, default_func):
         for func in sorted_list:
             if func.generic:
                 continue #call to generic should be done at the end
-            string += sep + "if ("
+            string += sep + "if "
             cond = func.print_condition()
             if (len(cond) <= func.max_len_of_line):
                 string += cond
             else:
                 string += func.print_condition(split=True)
-            string += ") {\n"
+            string += " {\n"
             string += " "*8 + "return " + func.print_call()
             string += " "*4 + "}"
             sep = " else "

lib/gen/func.py

@@ -140,6 +140,7 @@ def print_padding_condition(self, split=False):
     def print_condition(self, split=False):
         indent = ""
         newline = ""
+        cond_count = 0
         if (split):
             indent = " "*12
             newline = "\n"
@@ -148,22 +149,32 @@ def print_condition(self, split=False):
         if (self.stride_w > 0):
             cond += sep + "(stride_w == " + str(self.stride_w) + ")"
             sep = " && "
+            cond_count += 1
         if (self.stride_h > 0):
             cond += sep + "(stride_h == " + str(self.stride_h) + ")"
             sep = " && "
+            cond_count += 1
         if (self.kernel_w > 0):
             cond += sep + newline + indent + "(kernel_w == " + str(self.kernel_w) + ")"
             sep = " && "
+            cond_count += 1
         if (self.kernel_h > 0):
             cond += sep + "(kernel_h == " + str(self.kernel_h) + ")"
             sep = " && "
+            cond_count += 1
         if (self.channels > 0):
             cond += sep + newline + indent + "(channels == " + str(self.channels) + ")"
             sep = " && "
+            cond_count += 1
         if (self.print_padding_condition() != "(1)"):
             #skip padding
             cond += sep + newline + indent + self.print_padding_condition(split=split)
-        return cond
+            cond_count += 1
+
+        if (cond_count > 1):
+            return "(" + cond + ")"
+        else:
+            return cond
 
     def get_types(self):
         if self.datatype == "fx16":

lib/gen/mli_krn_avepool_gen.py

@@ -58,48 +58,38 @@
 ch = 0
 f_list.extend([Func(fbase, k, k, ch, stride, stride, corefunc, "nopad")])
 
-#stride = 1, any kernel size, any channel size
 corefunc = "avepool_chw_krnpad"
-stride = 1
-kernel_range = range(5, 11, 2)
+stride = 0
+kernel_range = range(3, 11, 2)
 ch = 0
 f_list.extend([Func(fbase, k, k, ch, stride, stride, corefunc, "krnpad") for k in kernel_range])
 
 corefunc = "avepool_chw_krnpad_k4_Nx2_N_even"
-stride = 1
-width_range = range(4, 9, 2)
-height_range = range(2, 9, 2)
+stride = 0
+kernel_range = range(2, 9, 2)
 ch = 0
-f_list.extend([Func(fbase, w, h, ch, stride, stride, corefunc, "krnpad") for w in width_range for h in height_range])
+f_list.extend([Func(fbase, k, k, ch, stride, stride, corefunc, "krnpad") for k in kernel_range])
 
-corefunc = "avepool_chw_nopad_k4_Nx2_N_even"
-stride = 1
-w = 4
-h = 2
+corefunc = "avepool_chw_nopad"
+stride = 0
+kernel_range = range(3, 11, 2)
 ch = 0
-f_list.extend([Func(fbase, w, h, ch, stride, stride, corefunc, "nopad")])
+f_list.extend([Func(fbase, k, k, ch, stride, stride, corefunc, "nopad") for k in kernel_range])
 
-#stride = 1, any kernel size, any channel size
 corefunc = "avepool_chw_nopad_k4_Nx2_N_even"
-stride = 1
-width_range = range(6, 9, 2)
-height_range = range(2, 9, 2)
-ch = 0
-f_list.extend([Func(fbase, w, h, ch, stride, stride, corefunc, "nopad") for w in width_range for h in height_range])
-
-#stride = 1, any kernel size, any channel size
-corefunc = "avepool_chw_nopad"
-stride = 1
-kernel_range = range(3, 11, 2)
+stride = 0
+kernel_range = range(4, 9, 2)
 ch = 0
 f_list.extend([Func(fbase, k, k, ch, stride, stride, corefunc, "nopad") for k in kernel_range])
 
-#here construct the specializations for any stride, and multiple kernel sizes > 1
 corefunc = "avepool_chw_krnpad"
 stride = 0
-kernel_range = range(2, 11)
+kernel_range = [0, 2, 3]
 ch = 0
-f_list.extend([Func(fbase, k, k, ch, stride, stride, corefunc, "krnpad") for k in kernel_range])
+f_list.extend([Func(fbase, 1, k, ch, stride, stride, corefunc, "krnpad") for k in kernel_range])
+f_list.extend([Func(fbase, k, 1, ch, stride, stride, corefunc, "krnpad") for k in kernel_range])
+f_list.extend([Func(fbase, 1, k, ch, stride, stride, corefunc, "nopad") for k in kernel_range])
+f_list.extend([Func(fbase, k, 1, ch, stride, stride, corefunc, "nopad") for k in kernel_range])
 
 #at last add the generic function that can be used in the else branch in the wrapper.
 corefunc = "avepool_chw_krnpad"
@@ -117,7 +107,7 @@
 c.set_wrapper_variables({'padding_left' : "cfg->padding_left"})
 c.set_wrapper_variables({'padding_right' : "cfg->padding_right"})
 c.set_wrapper_hierarchy(['stride_w', 'stride_h', 'kernel_w', 'kernel_h', 'padding'])
-c.set_wrapper_if_tree(True)
+c.set_wrapper_if_tree(False)
 
 if "fx16" in sys.argv or no_args:
     f = open(output_file_fx16, "wb")