ref: 520055bd1a3b929020510f3c426a6d2bfad13566
parent: bd3c874cac4d655d888979f3831af644aec9fcb8
parent: 2f5840de3ec53cef99b30bd5eb1877f92f15a80f
author: Johann Koenig <johannkoenig@chromium.org>
date: Mon Apr 25 14:58:29 EDT 2016
Merge "vpx_minmax_8x8_neon and test"
--- /dev/null
+++ b/test/minmax_test.cc
@@ -1,0 +1,132 @@
+/*
+ * Copyright (c) 2016 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 <stdlib.h>
+#include <string.h>
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx/vpx_integer.h"
+
+#include "test/acm_random.h"
+#include "test/register_state_check.h"
+
+namespace {+
+using ::libvpx_test::ACMRandom;
+
+typedef void (*MinMaxFunc)(const uint8_t *a, int a_stride,
+ const uint8_t *b, int b_stride,
+ int *min, int *max);
+
+class MinMaxTest : public ::testing::TestWithParam<MinMaxFunc> {+ public:
+ virtual void SetUp() {+ mm_func_ = GetParam();
+ rnd_.Reset(ACMRandom::DeterministicSeed());
+ }
+
+ protected:
+ MinMaxFunc mm_func_;
+ ACMRandom rnd_;
+};
+
+void reference_minmax(const uint8_t *a, int a_stride,
+ const uint8_t *b, int b_stride,
+ int *min_ret, int *max_ret) {+ int min = 255;
+ int max = 0;
+ for (int i = 0; i < 8; i++) {+ for (int j = 0; j < 8; j++) {+ const int diff = abs(a[i * a_stride + j] - b[i * b_stride + j]);
+ if (min > diff) min = diff;
+ if (max < diff) max = diff;
+ }
+ }
+
+ *min_ret = min;
+ *max_ret = max;
+}
+
+TEST_P(MinMaxTest, MinValue) {+ for (int i = 0; i < 64; i++) {+ uint8_t a[64], b[64];
+ memset(a, 0, sizeof(a));
+ memset(b, 255, sizeof(b));
+ b[i] = i; // Set a minimum difference of i.
+
+ int min, max;
+ ASM_REGISTER_STATE_CHECK(mm_func_(a, 8, b, 8, &min, &max));
+ EXPECT_EQ(255, max);
+ EXPECT_EQ(i, min);
+ }
+}
+
+TEST_P(MinMaxTest, MaxValue) {+ for (int i = 0; i < 64; i++) {+ uint8_t a[64], b[64];
+ memset(a, 0, sizeof(a));
+ memset(b, 0, sizeof(b));
+ b[i] = i; // Set a maximum difference of i.
+
+ int min, max;
+ ASM_REGISTER_STATE_CHECK(mm_func_(a, 8, b, 8, &min, &max));
+ EXPECT_EQ(i, max);
+ EXPECT_EQ(0, min);
+ }
+}
+
+TEST_P(MinMaxTest, CompareReference) {+ uint8_t a[64], b[64];
+ for (int j = 0; j < 64; j++) {+ a[j] = rnd_.Rand8();
+ b[j] = rnd_.Rand8();
+ }
+
+ int min_ref, max_ref, min, max;
+ reference_minmax(a, 8, b, 8, &min_ref, &max_ref);
+ ASM_REGISTER_STATE_CHECK(mm_func_(a, 8, b, 8, &min, &max));
+ EXPECT_EQ(max_ref, max);
+ EXPECT_EQ(min_ref, min);
+}
+
+TEST_P(MinMaxTest, CompareReferenceAndVaryStride) {+ uint8_t a[8 * 64], b[8 * 64];
+ for (int i = 0; i < 8 * 64; i++) {+ a[i] = rnd_.Rand8();
+ b[i] = rnd_.Rand8();
+ }
+ for (int a_stride = 8; a_stride <= 64; a_stride += 8) {+ for (int b_stride = 8; b_stride <= 64; b_stride += 8) {+ int min_ref, max_ref, min, max;
+ reference_minmax(a, a_stride, b, b_stride, &min_ref, &max_ref);
+ ASM_REGISTER_STATE_CHECK(mm_func_(a, a_stride, b, b_stride, &min, &max));
+ EXPECT_EQ(max_ref, max) << "when a_stride = " << a_stride
+ << " and b_stride = " << b_stride;;
+ EXPECT_EQ(min_ref, min) << "when a_stride = " << a_stride
+ << " and b_stride = " << b_stride;;
+ }
+ }
+}
+
+INSTANTIATE_TEST_CASE_P(C, MinMaxTest, ::testing::Values(&vpx_minmax_8x8_c));
+
+#if HAVE_SSE2
+INSTANTIATE_TEST_CASE_P(SSE2, MinMaxTest,
+ ::testing::Values(&vpx_minmax_8x8_sse2));
+#endif
+
+#if HAVE_NEON
+INSTANTIATE_TEST_CASE_P(NEON, MinMaxTest,
+ ::testing::Values(&vpx_minmax_8x8_neon));
+#endif
+
+} // namespace
--- a/test/test.mk
+++ b/test/test.mk
@@ -144,6 +144,7 @@
LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += fdct4x4_test.cc
LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += fdct8x8_test.cc
LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += hadamard_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += minmax_test.cc
LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += variance_test.cc
LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += vp9_error_block_test.cc
LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += vp9_quantize_test.cc
--- a/vpx_dsp/arm/avg_neon.c
+++ b/vpx_dsp/arm/avg_neon.c
@@ -197,3 +197,60 @@
return s - ((t * t) >> shift_factor);
}
}
+
+void vpx_minmax_8x8_neon(const uint8_t *a, int a_stride,
+ const uint8_t *b, int b_stride,
+ int *min, int *max) {+ // Load and concatenate.
+ const uint8x16_t a01 = vcombine_u8(vld1_u8(a),
+ vld1_u8(a + a_stride));
+ const uint8x16_t a23 = vcombine_u8(vld1_u8(a + 2 * a_stride),
+ vld1_u8(a + 3 * a_stride));
+ const uint8x16_t a45 = vcombine_u8(vld1_u8(a + 4 * a_stride),
+ vld1_u8(a + 5 * a_stride));
+ const uint8x16_t a67 = vcombine_u8(vld1_u8(a + 6 * a_stride),
+ vld1_u8(a + 7 * a_stride));
+
+ const uint8x16_t b01 = vcombine_u8(vld1_u8(b),
+ vld1_u8(b + b_stride));
+ const uint8x16_t b23 = vcombine_u8(vld1_u8(b + 2 * b_stride),
+ vld1_u8(b + 3 * b_stride));
+ const uint8x16_t b45 = vcombine_u8(vld1_u8(b + 4 * b_stride),
+ vld1_u8(b + 5 * b_stride));
+ const uint8x16_t b67 = vcombine_u8(vld1_u8(b + 6 * b_stride),
+ vld1_u8(b + 7 * b_stride));
+
+ // Absolute difference.
+ const uint8x16_t ab01_diff = vabdq_u8(a01, b01);
+ const uint8x16_t ab23_diff = vabdq_u8(a23, b23);
+ const uint8x16_t ab45_diff = vabdq_u8(a45, b45);
+ const uint8x16_t ab67_diff = vabdq_u8(a67, b67);
+
+ // Max values between the Q vectors.
+ const uint8x16_t ab0123_max = vmaxq_u8(ab01_diff, ab23_diff);
+ const uint8x16_t ab4567_max = vmaxq_u8(ab45_diff, ab67_diff);
+ const uint8x16_t ab0123_min = vminq_u8(ab01_diff, ab23_diff);
+ const uint8x16_t ab4567_min = vminq_u8(ab45_diff, ab67_diff);
+
+ const uint8x16_t ab07_max = vmaxq_u8(ab0123_max, ab4567_max);
+ const uint8x16_t ab07_min = vminq_u8(ab0123_min, ab4567_min);
+
+ // Split to D and start doing pairwise.
+ uint8x8_t ab_max = vmax_u8(vget_high_u8(ab07_max), vget_low_u8(ab07_max));
+ uint8x8_t ab_min = vmin_u8(vget_high_u8(ab07_min), vget_low_u8(ab07_min));
+
+ // Enough runs of vpmax/min propogate the max/min values to every position.
+ ab_max = vpmax_u8(ab_max, ab_max);
+ ab_min = vpmin_u8(ab_min, ab_min);
+
+ ab_max = vpmax_u8(ab_max, ab_max);
+ ab_min = vpmin_u8(ab_min, ab_min);
+
+ ab_max = vpmax_u8(ab_max, ab_max);
+ ab_min = vpmin_u8(ab_min, ab_min);
+
+ *min = *max = 0; // Clear high bits
+ // Store directly to avoid costly neon->gpr transfer.
+ vst1_lane_u8((uint8_t *)max, ab_max, 0);
+ vst1_lane_u8((uint8_t *)min, ab_min, 0);
+}
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -1014,7 +1014,7 @@
specialize qw/vpx_avg_4x4 sse2 neon msa/;
add_proto qw/void vpx_minmax_8x8/, "const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max";
- specialize qw/vpx_minmax_8x8 sse2/;
+ specialize qw/vpx_minmax_8x8 sse2 neon/;
add_proto qw/void vpx_hadamard_8x8/, "const int16_t *src_diff, int src_stride, int16_t *coeff";
specialize qw/vpx_hadamard_8x8 sse2/, "$ssse3_x86_64_x86inc";