fix specialization for small maxpooling kernels

Author: JaccovG

examples/example_cifar10_caffe/cifar10_model_chw.c

@@ -434,7 +434,7 @@ static void check_result(
 //========================================================================================
 #if (MODEL_BIT_DEPTH != MODEL_FX_8)
 static inline mli_status maxpool_chw(const mli_tensor *in, const mli_pool_cfg *cfg, mli_tensor *out) {
-    return mli_krn_maxpool_chw_fx16_k3x3(in, cfg, out);
+    return mli_krn_maxpool_chw_fx16_k3x3_krnpad(in, cfg, out);
 }
 
 static inline mli_status avepool_chw(const mli_tensor *in, const mli_pool_cfg *cfg, mli_tensor *out) {
@@ -455,7 +455,7 @@ static inline mli_status mli_krn_permute_fx(const mli_tensor *in, const mli_perm
 
 #else // MODEL_BIT_DEPTH == (MODEL_FX_8W16D || MODEL_FX_8W16D)
 static inline mli_status maxpool_chw(const mli_tensor *in, const mli_pool_cfg *cfg, mli_tensor *out) {
-    return mli_krn_maxpool_chw_fx8_k3x3(in, cfg, out);
+    return mli_krn_maxpool_chw_fx8_k3x3_krnpad(in, cfg, out);
 }
 
 static inline mli_status avepool_chw(const mli_tensor *in, const mli_pool_cfg *cfg, mli_tensor *out) {

include/api/mli_krn_maxpool_spec_api.h

@@ -101,6 +101,8 @@ mli_status mli_krn_maxpool_chw_fx16_k2x2(const mli_tensor * in, const mli_pool_c
 mli_status mli_krn_maxpool_chw_fx16_k2x2_ch1(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
 mli_status mli_krn_maxpool_chw_fx16_k3x3(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
 mli_status mli_krn_maxpool_chw_fx16_k3x3_ch1(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_maxpool_chw_fx16_k2x2_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_maxpool_chw_fx16_k3x3_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
 mli_status mli_krn_maxpool_chw_fx16_generic(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
 
 mli_status mli_krn_maxpool_chw_fx8_k2x2_str1_nopad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
@@ -180,6 +182,8 @@ mli_status mli_krn_maxpool_chw_fx8_k2x2(const mli_tensor * in, const mli_pool_cf
 mli_status mli_krn_maxpool_chw_fx8_k2x2_ch1(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
 mli_status mli_krn_maxpool_chw_fx8_k3x3(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
 mli_status mli_krn_maxpool_chw_fx8_k3x3_ch1(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_maxpool_chw_fx8_k2x2_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
+mli_status mli_krn_maxpool_chw_fx8_k3x3_krnpad(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
 mli_status mli_krn_maxpool_chw_fx8_generic(const mli_tensor * in, const mli_pool_cfg * cfg, mli_tensor * out);
 
 #ifdef __cplusplus

lib/gen/func.py

@@ -124,10 +124,15 @@ def print_padding_condition(self, split=False):
             else:
                 cond = "(1)"
         elif self.padding == "nopad":
-                cond  = "(padding_top == 0) && "
-                cond += "(padding_bot == 0) && "
-                cond += "(padding_left == 0) && "
-                cond += "(padding_right == 0)"
+            cond  = "(padding_top == 0) && "
+            cond += "(padding_bot == 0) && "
+            cond += "(padding_left == 0) && "
+            cond += "(padding_right == 0)"
+        elif self.padding == "krnpad" and (self.kernel_h > 0) and (self.kernel_w > 0):
+            cond  = "(padding_top <= " + str(int((self.kernel_h - 1) / 2)) + ") && "
+            cond += "(padding_bot <= " + str(int(self.kernel_h / 2)) + ") && "
+            cond += "(padding_left <= " + str(int((self.kernel_w -1) / 2)) + ") && "
+            cond += "(padding_right <= " + str(int(self.kernel_w / 2)) + ")"
         else:
             cond = "(1)"
         return cond

lib/gen/mli_krn_maxpool_gen.py

@@ -53,35 +53,41 @@
 channel_range = [0,1,3]
 f_list.extend([Func(fbase, k, k, ch, stride, stride, corefunc, "krnpad") for k in kernel_range for ch in channel_range])
 
-corefunc = "maxpool_chw_krnpad"
+corefunc = "maxpool_chw_pad"
 stride = 1
 kernel_range = range(4,11)
 channel_range = [0,1,3]
 f_list.extend([Func(fbase, k, k, ch, stride, stride, corefunc, "krnpad") for k in kernel_range for ch in channel_range])
 
 #stride = 1, 1xk and kx1 versions
-corefunc = "maxpool_chw_krnpad"
+corefunc = "maxpool_chw_pad"
 stride = 1
 kernel_range = range(2,4)
 channel_range = [0,1]
 f_list.extend([Func(fbase, 1, k, ch, stride, stride, corefunc, "krnpad") for k in kernel_range for ch in channel_range])
 f_list.extend([Func(fbase, k, 1, ch, stride, stride, corefunc, "krnpad") for k in kernel_range for ch in channel_range])
 
 #fix single dimension, others flex
-corefunc = "maxpool_chw_krnpad"
+corefunc = "maxpool_chw_pad"
 stride = 1
 f_list.extend([Func(fbase, 1, 0, 0, stride, stride, corefunc, "")]) #k_width == 1
 f_list.extend([Func(fbase, 0, 1, 0, stride, stride, corefunc, "")]) #k_heigth == 1
 f_list.extend([Func(fbase, 0, 0, 1, stride, stride, corefunc, "")]) #channels == 1
 
-corefunc = "maxpool_chw_krnpad_small"
+corefunc = "maxpool_chw_pad"
 stride = 0
 kernel_range = [2,3]
 channel_range = [0,1]
 f_list.extend([Func(fbase, k, k, ch, stride, stride, corefunc, "") for k in kernel_range for ch in channel_range])
 
+corefunc = "maxpool_chw_krnpad_small"
+stride = 0
+kernel_range = [2,3]
+channel_range = [0]
+f_list.extend([Func(fbase, k, k, ch, stride, stride, corefunc, "krnpad") for k in kernel_range for ch in channel_range])
+
 #at last add the generic function that can be used in the else branch in the wrapper.
-corefunc = "maxpool_chw_krnpad"
+corefunc = "maxpool_chw_pad"
 default_func = Func(fbase, 0, 0, 0, 0, 0, corefunc, generic=True)
 f_list.append(default_func)
 

lib/src/kernels/pooling/mli_krn_maxpool_chw.h

@@ -379,7 +379,7 @@ static inline void __attribute__((always_inline)) maxpool_chw_small(
 }
 
 template <typename io_T>
-static inline void __attribute__((always_inline)) maxpool_chw_krnpad(
+static inline void __attribute__((always_inline)) maxpool_chw_pad(
         const MLI_PTR(io_T) __restrict in_ftrs,
         MLI_OUT_PTR(io_T) __restrict out_ftrs,
         const int row_beg,