ref: a7f8bd451b82f52fb9bb5d9e7926f4d09cc66831
parent: f7e767d8ee584aabc2074e425186b219d564cce3
author: Luca Barbato <lu_zero@gentoo.org>
date: Wed May 10 21:07:09 EDT 2017
ppc: Add comp_avg_pred_vsx Change-Id: I59788cd98231e707239c2ad95ae54f67cfe24e10
--- a/test/comp_avg_pred_test.cc
+++ b/test/comp_avg_pred_test.cc
@@ -156,4 +156,8 @@
INSTANTIATE_TEST_CASE_P(SSE2, AvgPredTest,
::testing::Values(&vpx_comp_avg_pred_sse2));
#endif // HAVE_SSE2
+#if HAVE_VSX
+INSTANTIATE_TEST_CASE_P(VSX, AvgPredTest,
+ ::testing::Values(&vpx_comp_avg_pred_vsx));
+#endif // HAVE_VSX
} // namespace
--- /dev/null
+++ b/vpx_dsp/ppc/variance_vsx.c
@@ -1,0 +1,61 @@
+/*
+ * Copyright (c) 2017 The WebM project authors. All Rights Reserved.
+ *
+ * Use of this source code is governed by a BSD-style license
+ * that can be found in the LICENSE file in the root of the source
+ * tree. An additional intellectual property rights grant can be found
+ * 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 "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/ppc/types_vsx.h"
+
+void vpx_comp_avg_pred_vsx(uint8_t *comp_pred, const uint8_t *pred, int width,
+ int height, const uint8_t *ref, int ref_stride) {+ int i, j;
+ /* comp_pred and pred must be 16 byte aligned. */
+ assert(((intptr_t)comp_pred & 0xf) == 0);
+ assert(((intptr_t)pred & 0xf) == 0);
+ if (width >= 16) {+ for (i = 0; i < height; ++i) {+ for (j = 0; j < width; j += 16) {+ const uint8x16_t v = vec_avg(vec_vsx_ld(j, pred), vec_vsx_ld(j, ref));
+ vec_vsx_st(v, j, comp_pred);
+ }
+ comp_pred += width;
+ pred += width;
+ ref += ref_stride;
+ }
+ } else if (width == 8) {+ // Process 2 lines at time
+ for (i = 0; i < height / 2; ++i) {+ const uint8x16_t r0 = vec_vsx_ld(0, ref);
+ const uint8x16_t r1 = vec_vsx_ld(0, ref + ref_stride);
+ const uint8x16_t r = xxpermdi(r0, r1, 0);
+ const uint8x16_t v = vec_avg(vec_vsx_ld(0, pred), r);
+ vec_vsx_st(v, 0, comp_pred);
+ comp_pred += 16; // width * 2;
+ pred += 16; // width * 2;
+ ref += ref_stride * 2;
+ }
+ } else {+ assert(width == 4);
+ // process 4 lines at time
+ for (i = 0; i < height / 4; ++i) {+ const uint32x4_t r0 = (uint32x4_t)vec_vsx_ld(0, ref);
+ const uint32x4_t r1 = (uint32x4_t)vec_vsx_ld(0, ref + ref_stride);
+ const uint32x4_t r2 = (uint32x4_t)vec_vsx_ld(0, ref + ref_stride * 2);
+ const uint32x4_t r3 = (uint32x4_t)vec_vsx_ld(0, ref + ref_stride * 3);
+ const uint8x16_t r =
+ (uint8x16_t)xxpermdi(vec_mergeh(r0, r1), vec_mergeh(r2, r3), 0);
+ const uint8x16_t v = vec_avg(vec_vsx_ld(0, pred), r);
+ vec_vsx_st(v, 0, comp_pred);
+ comp_pred += 16; // width * 4;
+ pred += 16; // width * 4;
+ ref += ref_stride * 4;
+ }
+ }
+}
--- a/vpx_dsp/vpx_dsp.mk
+++ b/vpx_dsp/vpx_dsp.mk
@@ -333,6 +333,7 @@
DSP_SRCS-$(HAVE_SSE2) += x86/variance_sse2.c # Contains SSE2 and SSSE3
DSP_SRCS-$(HAVE_AVX2) += x86/variance_avx2.c
DSP_SRCS-$(HAVE_AVX2) += x86/variance_impl_avx2.c
+DSP_SRCS-$(HAVE_VSX) += ppc/variance_vsx.c
ifeq ($(ARCH_X86_64),yes)
DSP_SRCS-$(HAVE_SSE2) += x86/ssim_opt_x86_64.asm
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -1177,7 +1177,7 @@
specialize qw/vpx_get4x4sse_cs neon msa/;
add_proto qw/void vpx_comp_avg_pred/, "uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride";
- specialize qw/vpx_comp_avg_pred sse2/;
+ specialize qw/vpx_comp_avg_pred sse2 vsx/;
#
# Subpixel Variance