ref: f3c97ed32ef6a9419df3b68de895af75b70d6166
parent: d204c4bf017dc8313fc315f5c4da4492acd7641f
author: Johann <johannkoenig@google.com>
date: Wed May 24 07:52:42 EDT 2017
subpel variance neon: reduce stack usage Unlike x86, arm does not impose additional alignment restrictions on vector loads. For incoming values to the first pass, it uses vld1_u32() which typically does impose a 4 byte alignment. However, as the first pass operates on user-supplied values we must prepare for unaligned values anyway (and have, see mem_neon.h). But for the local temporary values there is no stride and the load will use vld1_u8 which does not require 4 byte alignment. There are 3 temporary structures. In the C, one is uint16_t. The arm saturates between passes but still passes tests. If this becomes an issue new functions will be needed. Change-Id: I3c9d4701bfeb14b77c783d0164608e621bfecfb1
--- a/vpx_dsp/arm/subpel_variance_neon.c
+++ b/vpx_dsp/arm/subpel_variance_neon.c
@@ -12,7 +12,6 @@
#include "./vpx_dsp_rtcd.h"
#include "./vpx_config.h"
-#include "vpx_ports/mem.h"
#include "vpx/vpx_integer.h"
#include "vpx_dsp/variance.h"
@@ -40,8 +39,7 @@
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...
+ vst1_u8(output_ptr, out);
src_ptr += 2 * src_pixels_per_line;
output_ptr += 8;
}
@@ -63,8 +61,7 @@
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);
- vst1_u8(&output_ptr[0], out);
- // Next row...
+ vst1_u8(output_ptr, out);
src_ptr += src_pixels_per_line;
output_ptr += 8;
}
@@ -91,9 +88,8 @@
const uint16x8_t c = vmull_u8(vget_high_u8(src_0), f0);
const uint16x8_t d = vmlal_u8(c, vget_high_u8(src_1), f1);
const uint8x8_t out_hi = vrshrn_n_u16(d, FILTER_BITS);
- vst1q_u8(&output_ptr[j], vcombine_u8(out_lo, out_hi));
+ vst1q_u8(output_ptr + j, vcombine_u8(out_lo, out_hi));
}
- // Next row...
src_ptr += src_pixels_per_line;
output_ptr += output_width;
}
@@ -101,30 +97,30 @@
// 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); \
+#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) { \+ uint8_t temp0[n * (m + (n == 4 ? 2 : 1))]; \
+ uint8_t temp1[n * m]; \
+ \
+ if (n == 4) { \+ var_filter_block2d_bil_w4(a, temp0, a_stride, 1, (m + 2), \
+ bilinear_filters[xoffset]); \
+ var_filter_block2d_bil_w4(temp0, temp1, n, n, m, \
+ bilinear_filters[yoffset]); \
+ } else if (n == 8) { \+ var_filter_block2d_bil_w8(a, temp0, a_stride, 1, (m + 1), \
+ bilinear_filters[xoffset]); \
+ var_filter_block2d_bil_w8(temp0, temp1, n, n, m, \
+ bilinear_filters[yoffset]); \
+ } else { \+ var_filter_block2d_bil_w16(a, temp0, a_stride, 1, (m + 1), n, \
+ bilinear_filters[xoffset]); \
+ var_filter_block2d_bil_w16(temp0, temp1, n, n, m, n, \
+ bilinear_filters[yoffset]); \
+ } \
+ return vpx_variance##n##x##m(temp1, n, b, b_stride, sse); \
}
sub_pixel_varianceNxM(4, 4);
@@ -143,35 +139,34 @@
// 4xM filter writes an extra row to fdata because it processes two rows at a
// time.
-#define sub_pixel_avg_varianceNxM(n, m) \
- uint32_t vpx_sub_pixel_avg_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, \
- const uint8_t *second_pred) { \- DECLARE_ALIGNED(16, uint8_t, fdata3[n * (m + (n == 4 ? 2 : 1))]); \
- DECLARE_ALIGNED(16, uint8_t, temp2[n * m]); \
- DECLARE_ALIGNED(16, uint8_t, temp3[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]); \
- } \
- \
- vpx_comp_avg_pred(temp3, second_pred, n, m, temp2, n); \
- \
- return vpx_variance##n##x##m(temp3, n, b, b_stride, sse); \
+#define sub_pixel_avg_varianceNxM(n, m) \
+ uint32_t vpx_sub_pixel_avg_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, \
+ const uint8_t *second_pred) { \+ uint8_t temp0[n * (m + (n == 4 ? 2 : 1))]; \
+ uint8_t temp1[n * m]; \
+ \
+ if (n == 4) { \+ var_filter_block2d_bil_w4(a, temp0, a_stride, 1, (m + 2), \
+ bilinear_filters[xoffset]); \
+ var_filter_block2d_bil_w4(temp0, temp1, n, n, m, \
+ bilinear_filters[yoffset]); \
+ } else if (n == 8) { \+ var_filter_block2d_bil_w8(a, temp0, a_stride, 1, (m + 1), \
+ bilinear_filters[xoffset]); \
+ var_filter_block2d_bil_w8(temp0, temp1, n, n, m, \
+ bilinear_filters[yoffset]); \
+ } else { \+ var_filter_block2d_bil_w16(a, temp0, a_stride, 1, (m + 1), n, \
+ bilinear_filters[xoffset]); \
+ var_filter_block2d_bil_w16(temp0, temp1, n, n, m, n, \
+ bilinear_filters[yoffset]); \
+ } \
+ \
+ vpx_comp_avg_pred(temp0, second_pred, n, m, temp1, n); \
+ \
+ return vpx_variance##n##x##m(temp0, n, b, b_stride, sse); \
}
sub_pixel_avg_varianceNxM(4, 4);