#include "vpx/vpx_integer.h"
#include "vpx_dsp/variance.h"
+#include "vpx_dsp/arm/mem_neon.h"
static const uint8_t bilinear_filters[8][2] = {
{ 128, 0 }, { 112, 16 }, { 96, 32 }, { 80, 48 },
{ 64, 64 }, { 48, 80 }, { 32, 96 }, { 16, 112 },
};
+// Process a block exactly 4 wide and a multiple of 2 high.
+static void var_filter_block2d_bil_w4(const uint8_t *src_ptr,
+ uint8_t *output_ptr,
+ unsigned int src_pixels_per_line,
+ int pixel_step,
+ unsigned int output_height,
+ const uint8_t *filter) {
+ const uint8x8_t f0 = vmov_n_u8(filter[0]);
+ const uint8x8_t f1 = vmov_n_u8(filter[1]);
+ unsigned int i;
+ for (i = 0; i < output_height; i += 2) {
+ const uint8x8_t src_0 = load_unaligned_u8(src_ptr, src_pixels_per_line);
+ const uint8x8_t src_1 =
+ load_unaligned_u8(src_ptr + pixel_step, src_pixels_per_line);
+ const uint16x8_t a = vmull_u8(src_0, f0);
+ const uint16x8_t b = vmlal_u8(a, src_1, f1);
+ const uint8x8_t out = vrshrn_n_u16(b, FILTER_BITS);
+ store_unaligned_u8(output_ptr, 4, out);
+ // Next row...
+ src_ptr += 2 * src_pixels_per_line;
+ output_ptr += 8;
+ }
+}
+
// Process a block exactly 8 wide and any height.
static void var_filter_block2d_bil_w8(const uint8_t *src_ptr,
uint8_t *output_ptr,
}
}
-// TODO(johannkoenig): support 4xM block sizes.
-#define sub_pixel_varianceNxM(n, m) \
- uint32_t vpx_sub_pixel_variance##n##x##m##_neon( \
- const uint8_t *a, int a_stride, int xoffset, int yoffset, \
- const uint8_t *b, int b_stride, uint32_t *sse) { \
- DECLARE_ALIGNED(16, uint8_t, fdata3[n * (m + 1)]); \
- DECLARE_ALIGNED(16, uint8_t, temp2[n * m]); \
- \
- if (n == 8) { \
- var_filter_block2d_bil_w8(a, fdata3, a_stride, 1, (m + 1), \
- bilinear_filters[xoffset]); \
- var_filter_block2d_bil_w8(fdata3, temp2, n, n, m, \
- bilinear_filters[yoffset]); \
- } else { \
- var_filter_block2d_bil_w16(a, fdata3, a_stride, 1, (m + 1), n, \
- bilinear_filters[xoffset]); \
- var_filter_block2d_bil_w16(fdata3, temp2, n, n, m, n, \
- bilinear_filters[yoffset]); \
- } \
- return vpx_variance##n##x##m(temp2, n, b, b_stride, sse); \
+// 4xM filter writes an extra row to fdata because it processes two rows at a
+// time.
+#define sub_pixel_varianceNxM(n, m) \
+ uint32_t vpx_sub_pixel_variance##n##x##m##_neon( \
+ const uint8_t *a, int a_stride, int xoffset, int yoffset, \
+ const uint8_t *b, int b_stride, uint32_t *sse) { \
+ DECLARE_ALIGNED(16, uint8_t, fdata3[n * (m + (n == 4 ? 2 : 1))]); \
+ DECLARE_ALIGNED(16, uint8_t, temp2[n * m]); \
+ \
+ if (n == 4) { \
+ var_filter_block2d_bil_w4(a, fdata3, a_stride, 1, (m + 2), \
+ bilinear_filters[xoffset]); \
+ var_filter_block2d_bil_w4(fdata3, temp2, n, n, m, \
+ bilinear_filters[yoffset]); \
+ } else if (n == 8) { \
+ var_filter_block2d_bil_w8(a, fdata3, a_stride, 1, (m + 1), \
+ bilinear_filters[xoffset]); \
+ var_filter_block2d_bil_w8(fdata3, temp2, n, n, m, \
+ bilinear_filters[yoffset]); \
+ } else { \
+ var_filter_block2d_bil_w16(a, fdata3, a_stride, 1, (m + 1), n, \
+ bilinear_filters[xoffset]); \
+ var_filter_block2d_bil_w16(fdata3, temp2, n, n, m, n, \
+ bilinear_filters[yoffset]); \
+ } \
+ return vpx_variance##n##x##m(temp2, n, b, b_stride, sse); \
}
+sub_pixel_varianceNxM(4, 4);
+sub_pixel_varianceNxM(4, 8);
sub_pixel_varianceNxM(8, 4);
sub_pixel_varianceNxM(8, 8);
sub_pixel_varianceNxM(8, 16);