ref: 9b54e812f73e401762fbf3486f8bb83d98dce78a
parent: a45596cff7bce21db776dbae8df88844db46c0e8
author: Johann <[email protected]>
date: Wed May 11 09:26:19 EDT 2016
neon hadamard 8x8 Runs about 30% faster than the C BUG=webm:1021 Change-Id: I6809d6d84c3077ab619c53298296950e976bdaba
--- a/test/hadamard_test.cc
+++ b/test/hadamard_test.cc
@@ -123,4 +123,9 @@
INSTANTIATE_TEST_CASE_P(SSSE3, HadamardTest,
::testing::Values(&vpx_hadamard_8x8_ssse3));
#endif // HAVE_SSSE3 && CONFIG_USE_X86INC && ARCH_X86_64
+
+#if HAVE_NEON
+INSTANTIATE_TEST_CASE_P(NEON, HadamardTest,
+ ::testing::Values(&vpx_hadamard_8x8_neon));
+#endif // HAVE_NEON
} // namespace
--- /dev/null
+++ b/vpx_dsp/arm/hadamard_neon.c
@@ -1,0 +1,162 @@
+/*
+ * 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 <arm_neon.h>
+
+#include "./vpx_dsp_rtcd.h"
+
+static void hadamard8x8_one_pass(int16x8_t *a0, int16x8_t *a1,
+ int16x8_t *a2, int16x8_t *a3,
+ int16x8_t *a4, int16x8_t *a5,
+ int16x8_t *a6, int16x8_t *a7) {+ const int16x8_t b0 = vaddq_s16(*a0, *a1);
+ const int16x8_t b1 = vsubq_s16(*a0, *a1);
+ const int16x8_t b2 = vaddq_s16(*a2, *a3);
+ const int16x8_t b3 = vsubq_s16(*a2, *a3);
+ const int16x8_t b4 = vaddq_s16(*a4, *a5);
+ const int16x8_t b5 = vsubq_s16(*a4, *a5);
+ const int16x8_t b6 = vaddq_s16(*a6, *a7);
+ const int16x8_t b7 = vsubq_s16(*a6, *a7);
+
+ const int16x8_t c0 = vaddq_s16(b0, b2);
+ const int16x8_t c1 = vaddq_s16(b1, b3);
+ const int16x8_t c2 = vsubq_s16(b0, b2);
+ const int16x8_t c3 = vsubq_s16(b1, b3);
+ const int16x8_t c4 = vaddq_s16(b4, b6);
+ const int16x8_t c5 = vaddq_s16(b5, b7);
+ const int16x8_t c6 = vsubq_s16(b4, b6);
+ const int16x8_t c7 = vsubq_s16(b5, b7);
+
+ *a0 = vaddq_s16(c0, c4);
+ *a1 = vsubq_s16(c2, c6);
+ *a2 = vsubq_s16(c0, c4);
+ *a3 = vaddq_s16(c2, c6);
+ *a4 = vaddq_s16(c3, c7);
+ *a5 = vsubq_s16(c3, c7);
+ *a6 = vsubq_s16(c1, c5);
+ *a7 = vaddq_s16(c1, c5);
+}
+
+// TODO(johannkoenig): Make a transpose library and dedup with idct. Consider
+// reversing transpose order which may make it easier for the compiler to
+// reconcile the vtrn.64 moves.
+static void transpose8x8(int16x8_t *a0, int16x8_t *a1,
+ int16x8_t *a2, int16x8_t *a3,
+ int16x8_t *a4, int16x8_t *a5,
+ int16x8_t *a6, int16x8_t *a7) {+ // Swap 64 bit elements. Goes from:
+ // a0: 00 01 02 03 04 05 06 07
+ // a1: 08 09 10 11 12 13 14 15
+ // a2: 16 17 18 19 20 21 22 23
+ // a3: 24 25 26 27 28 29 30 31
+ // a4: 32 33 34 35 36 37 38 39
+ // a5: 40 41 42 43 44 45 46 47
+ // a6: 48 49 50 51 52 53 54 55
+ // a7: 56 57 58 59 60 61 62 63
+ // to:
+ // a04_lo: 00 01 02 03 32 33 34 35
+ // a15_lo: 08 09 10 11 40 41 42 43
+ // a26_lo: 16 17 18 19 48 49 50 51
+ // a37_lo: 24 25 26 27 56 57 58 59
+ // a04_hi: 04 05 06 07 36 37 38 39
+ // a15_hi: 12 13 14 15 44 45 46 47
+ // a26_hi: 20 21 22 23 52 53 54 55
+ // a37_hi: 28 29 30 31 60 61 62 63
+ const int16x8_t a04_lo = vcombine_s16(vget_low_s16(*a0), vget_low_s16(*a4));
+ const int16x8_t a15_lo = vcombine_s16(vget_low_s16(*a1), vget_low_s16(*a5));
+ const int16x8_t a26_lo = vcombine_s16(vget_low_s16(*a2), vget_low_s16(*a6));
+ const int16x8_t a37_lo = vcombine_s16(vget_low_s16(*a3), vget_low_s16(*a7));
+ const int16x8_t a04_hi = vcombine_s16(vget_high_s16(*a0), vget_high_s16(*a4));
+ const int16x8_t a15_hi = vcombine_s16(vget_high_s16(*a1), vget_high_s16(*a5));
+ const int16x8_t a26_hi = vcombine_s16(vget_high_s16(*a2), vget_high_s16(*a6));
+ const int16x8_t a37_hi = vcombine_s16(vget_high_s16(*a3), vget_high_s16(*a7));
+
+ // Swap 32 bit elements resulting in:
+ // a0246_lo:
+ // 00 01 16 17 32 33 48 49
+ // 02 03 18 19 34 35 50 51
+ // a1357_lo:
+ // 08 09 24 25 40 41 56 57
+ // 10 11 26 27 42 43 58 59
+ // a0246_hi:
+ // 04 05 20 21 36 37 52 53
+ // 06 07 22 23 38 39 54 55
+ // a1657_hi:
+ // 12 13 28 29 44 45 60 61
+ // 14 15 30 31 46 47 62 63
+ const int32x4x2_t a0246_lo = vtrnq_s32(vreinterpretq_s32_s16(a04_lo),
+ vreinterpretq_s32_s16(a26_lo));
+ const int32x4x2_t a1357_lo = vtrnq_s32(vreinterpretq_s32_s16(a15_lo),
+ vreinterpretq_s32_s16(a37_lo));
+ const int32x4x2_t a0246_hi = vtrnq_s32(vreinterpretq_s32_s16(a04_hi),
+ vreinterpretq_s32_s16(a26_hi));
+ const int32x4x2_t a1357_hi = vtrnq_s32(vreinterpretq_s32_s16(a15_hi),
+ vreinterpretq_s32_s16(a37_hi));
+
+ // Swap 16 bit elements resulting in:
+ // b0:
+ // 00 08 16 24 32 40 48 56
+ // 01 09 17 25 33 41 49 57
+ // b1:
+ // 02 10 18 26 34 42 50 58
+ // 03 11 19 27 35 43 51 59
+ // b2:
+ // 04 12 20 28 36 44 52 60
+ // 05 13 21 29 37 45 53 61
+ // b3:
+ // 06 14 22 30 38 46 54 62
+ // 07 15 23 31 39 47 55 63
+ const int16x8x2_t b0 = vtrnq_s16(vreinterpretq_s16_s32(a0246_lo.val[0]),
+ vreinterpretq_s16_s32(a1357_lo.val[0]));
+ const int16x8x2_t b1 = vtrnq_s16(vreinterpretq_s16_s32(a0246_lo.val[1]),
+ vreinterpretq_s16_s32(a1357_lo.val[1]));
+ const int16x8x2_t b2 = vtrnq_s16(vreinterpretq_s16_s32(a0246_hi.val[0]),
+ vreinterpretq_s16_s32(a1357_hi.val[0]));
+ const int16x8x2_t b3 = vtrnq_s16(vreinterpretq_s16_s32(a0246_hi.val[1]),
+ vreinterpretq_s16_s32(a1357_hi.val[1]));
+
+ *a0 = b0.val[0];
+ *a1 = b0.val[1];
+ *a2 = b1.val[0];
+ *a3 = b1.val[1];
+ *a4 = b2.val[0];
+ *a5 = b2.val[1];
+ *a6 = b3.val[0];
+ *a7 = b3.val[1];
+}
+
+void vpx_hadamard_8x8_neon(const int16_t *src_diff, int src_stride,
+ int16_t *coeff) {+ int16x8_t a0 = vld1q_s16(src_diff);
+ int16x8_t a1 = vld1q_s16(src_diff + src_stride);
+ int16x8_t a2 = vld1q_s16(src_diff + 2 * src_stride);
+ int16x8_t a3 = vld1q_s16(src_diff + 3 * src_stride);
+ int16x8_t a4 = vld1q_s16(src_diff + 4 * src_stride);
+ int16x8_t a5 = vld1q_s16(src_diff + 5 * src_stride);
+ int16x8_t a6 = vld1q_s16(src_diff + 6 * src_stride);
+ int16x8_t a7 = vld1q_s16(src_diff + 7 * src_stride);
+
+ hadamard8x8_one_pass(&a0, &a1, &a2, &a3, &a4, &a5, &a6, &a7);
+
+ transpose8x8(&a0, &a1, &a2, &a3, &a4, &a5, &a6, &a7);
+
+ hadamard8x8_one_pass(&a0, &a1, &a2, &a3, &a4, &a5, &a6, &a7);
+
+ // Skip the second transpose because it is not required.
+
+ vst1q_s16(coeff + 0, a0);
+ vst1q_s16(coeff + 8, a1);
+ vst1q_s16(coeff + 16, a2);
+ vst1q_s16(coeff + 24, a3);
+ vst1q_s16(coeff + 32, a4);
+ vst1q_s16(coeff + 40, a5);
+ vst1q_s16(coeff + 48, a6);
+ vst1q_s16(coeff + 56, a7);
+}
--- a/vpx_dsp/vpx_dsp.mk
+++ b/vpx_dsp/vpx_dsp.mk
@@ -265,6 +265,7 @@
DSP_SRCS-$(HAVE_SSE2) += x86/avg_intrin_sse2.c
DSP_SRCS-$(HAVE_NEON) += arm/avg_neon.c
DSP_SRCS-$(HAVE_MSA) += mips/avg_msa.c
+DSP_SRCS-$(HAVE_NEON) += arm/hadamard_neon.c
ifeq ($(ARCH_X86_64),yes)
ifeq ($(CONFIG_USE_X86INC),yes)
DSP_SRCS-$(HAVE_SSSE3) += x86/avg_ssse3_x86_64.asm
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -1017,7 +1017,7 @@
specialize qw/vpx_minmax_8x8 sse2 neon/;
add_proto qw/void vpx_hadamard_8x8/, "const int16_t *src_diff, int src_stride, int16_t *coeff";
- specialize qw/vpx_hadamard_8x8 sse2/, "$ssse3_x86_64_x86inc";
+ specialize qw/vpx_hadamard_8x8 sse2 neon/, "$ssse3_x86_64_x86inc";
add_proto qw/void vpx_hadamard_16x16/, "const int16_t *src_diff, int src_stride, int16_t *coeff";
specialize qw/vpx_hadamard_16x16 sse2/;