ref: e2ad89092d44afcfcc2c84076ebe4bf6bb941e97
parent: e6ca81ee6706e3c151cdc9b5b33ea695f35e062d
author: Luca Barbato <lu_zero@gentoo.org>
date: Mon May 1 13:02:03 EDT 2017
ppc: Add convolve8_vsx and convolve8_avg_vsx Change-Id: Ia5293d948003a7fff5a7cbad6e83d8a72717c857
--- a/test/convolve_test.cc
+++ b/test/convolve_test.cc
@@ -1226,7 +1226,7 @@
const ConvolveFunctions convolve8_vsx(
vpx_convolve_copy_vsx, vpx_convolve_avg_vsx, vpx_convolve8_horiz_vsx,
vpx_convolve8_avg_horiz_vsx, vpx_convolve8_vert_vsx,
- vpx_convolve8_avg_vert_vsx, vpx_convolve8_c, vpx_convolve8_avg_c,
+ vpx_convolve8_avg_vert_vsx, vpx_convolve8_vsx, vpx_convolve8_avg_vsx,
vpx_scaled_horiz_c, vpx_scaled_avg_horiz_c, vpx_scaled_vert_c,
vpx_scaled_avg_vert_c, vpx_scaled_2d_c, vpx_scaled_avg_2d_c, 0);
const ConvolveParam kArrayConvolve_vsx[] = { ALL_SIZES(convolve8_vsx) };
--- a/vpx_dsp/ppc/vpx_convolve_vsx.c
+++ b/vpx_dsp/ppc/vpx_convolve_vsx.c
@@ -7,6 +7,7 @@
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
+#include <assert.h>
#include <string.h>
#include "./vpx_dsp_rtcd.h"
#include "vpx_dsp/vpx_filter.h"
@@ -296,6 +297,38 @@
}
}
+static inline void convolve(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const InterpKernel *const x_filters, int x0_q4,
+ int x_step_q4, const InterpKernel *const y_filters,
+ int y0_q4, int y_step_q4, int w, int h) {
+ // Note: Fixed size intermediate buffer, temp, places limits on parameters.
+ // 2d filtering proceeds in 2 steps:
+ // (1) Interpolate horizontally into an intermediate buffer, temp.
+ // (2) Interpolate temp vertically to derive the sub-pixel result.
+ // Deriving the maximum number of rows in the temp buffer (135):
+ // --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative).
+ // --Largest block size is 64x64 pixels.
+ // --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the
+ // original frame (in 1/16th pixel units).
+ // --Must round-up because block may be located at sub-pixel position.
+ // --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails.
+ // --((64 - 1) * 32 + 15) >> 4 + 8 = 135.
+ DECLARE_ALIGNED(16, uint8_t, temp[64 * 135]);
+ const int intermediate_height =
+ (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS;
+
+ assert(w <= 64);
+ assert(h <= 64);
+ assert(y_step_q4 <= 32);
+ assert(x_step_q4 <= 32);
+
+ convolve_horiz(src - src_stride * (SUBPEL_TAPS / 2 - 1), src_stride, temp, 64,
+ x_filters, x0_q4, x_step_q4, w, intermediate_height);
+ convolve_vert(temp + 64 * (SUBPEL_TAPS / 2 - 1), 64, dst, dst_stride,
+ y_filters, y0_q4, y_step_q4, w, h);
+}
+
void vpx_convolve8_horiz_vsx(const uint8_t *src, ptrdiff_t src_stride,
uint8_t *dst, ptrdiff_t dst_stride,
const int16_t *filter_x, int x_step_q4,
@@ -354,4 +387,32 @@
convolve_avg_vert(src, src_stride, dst, dst_stride, filters_y, y0_q4,
y_step_q4, w, h);
+}
+
+void vpx_convolve8_vsx(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
+ ptrdiff_t dst_stride, const int16_t *filter_x,
+ int x_step_q4, const int16_t *filter_y, int y_step_q4,
+ int w, int h) {
+ const InterpKernel *const filters_x = get_filter_base(filter_x);
+ const int x0_q4 = get_filter_offset(filter_x, filters_x);
+ const InterpKernel *const filters_y = get_filter_base(filter_y);
+ const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+ convolve(src, src_stride, dst, dst_stride, filters_x, x0_q4, x_step_q4,
+ filters_y, y0_q4, y_step_q4, w, h);
+}
+
+void vpx_convolve8_avg_vsx(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x, int x_step_q4,
+ const int16_t *filter_y, int y_step_q4, int w,
+ int h) {
+ // Fixed size intermediate buffer places limits on parameters.
+ DECLARE_ALIGNED(16, uint8_t, temp[64 * 64]);
+ assert(w <= 64);
+ assert(h <= 64);
+
+ vpx_convolve8_vsx(src, src_stride, temp, 64, filter_x, x_step_q4, filter_y,
+ y_step_q4, w, h);
+ vpx_convolve_avg_vsx(temp, 64, dst, dst_stride, NULL, 0, NULL, 0, w, h);
}
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -338,7 +338,7 @@
specialize qw/vpx_convolve_avg neon dspr2 msa sse2 vsx/;
add_proto qw/void vpx_convolve8/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
-specialize qw/vpx_convolve8 sse2 ssse3 avx2 neon dspr2 msa/;
+specialize qw/vpx_convolve8 sse2 ssse3 avx2 neon dspr2 msa vsx/;
add_proto qw/void vpx_convolve8_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
specialize qw/vpx_convolve8_horiz sse2 ssse3 avx2 neon dspr2 msa vsx/;
@@ -347,7 +347,7 @@
specialize qw/vpx_convolve8_vert sse2 ssse3 avx2 neon dspr2 msa vsx/;
add_proto qw/void vpx_convolve8_avg/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
-specialize qw/vpx_convolve8_avg sse2 ssse3 neon dspr2 msa/;
+specialize qw/vpx_convolve8_avg sse2 ssse3 neon dspr2 msa vsx/;
add_proto qw/void vpx_convolve8_avg_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
specialize qw/vpx_convolve8_avg_horiz sse2 ssse3 neon dspr2 msa vsx/;