CHW DWCONV with implicit padding
PiperOrigin-RevId: 310369233
diff --git a/src/f32-dwconv-spchw/3x3p1-scalar.c b/src/f32-dwconv-spchw/3x3p1-scalar.c
index d5d69e1..af60813 100644
--- a/src/f32-dwconv-spchw/3x3p1-scalar.c
+++ b/src/f32-dwconv-spchw/3x3p1-scalar.c
@@ -14,7 +14,9 @@
size_t n,
const float* input,
const float* weights,
+ const float* zero,
float* output,
+ uint32_t padding_top,
size_t input_tuple_stride,
size_t output_tuple_stride,
size_t input_width_stride,
@@ -22,18 +24,23 @@
const union xnn_f32_spchw_params params[restrict XNN_MIN_ELEMENTS(1)])
{
assert(n != 0);
+ assert(padding_top == 1);
- const size_t input_width_increment = input_width_stride - n * input_tuple_stride;
+ const size_t input_width_decrement = n * input_tuple_stride;
+ const size_t input_width_increment = input_width_stride - input_width_decrement;
const size_t output_width_increment = output_width_stride - (n - 1) * output_tuple_stride;
const float params_min = params->scalar.min;
const float params_max = params->scalar.max;
- // No vertical padding.
- const float* i0 = input;
- const float* i1 = (const float*) ((uintptr_t) i0 + input_width_stride);
+ const float* i0 = zero;
+ const float* i1 = input;
const float* i2 = (const float*) ((uintptr_t) i1 + input_width_stride);
+ if (m == 1) {
+ i2 = zero;
+ }
+
float* output0 = output;
const float vw0 = weights[0];
@@ -93,10 +100,13 @@
*output0 = voutput;
}
- i0 = (const float*) ((uintptr_t) i0 + input_width_increment);
+ i0 = (const float*) ((uintptr_t) i1 - input_width_decrement);
i1 = (const float*) ((uintptr_t) i1 + input_width_increment);
i2 = (const float*) ((uintptr_t) i2 + input_width_increment);
output0 = (float*) ((uintptr_t) output0 + output_width_increment);
m--;
+ if (m == 1) {
+ i2 = zero;
+ }
}
}