ref: 89fbf3de501b5d7fd90047192521eae3198705cd
parent: 9f9f87c7f685eb743c8fd45b98d2f3c60ed93717
author: Alex Converse <aconverse@google.com>
date: Mon May 5 10:10:41 EDT 2014
Add an MMX fwht4x4 7% faster encoding a desktop lossless at RT speed 4. Change-Id: I41627f5b737752616b6512bb91a36ec45995bf64
--- a/test/fdct4x4_test.cc
+++ b/test/fdct4x4_test.cc
@@ -353,6 +353,13 @@
make_tuple(&vp9_fht4x4_c, &vp9_iht4x4_16_add_neon, 3)));
#endif
+#if HAVE_MMX
+INSTANTIATE_TEST_CASE_P(
+ MMX, Trans4x4WHT,
+ ::testing::Values(
+ make_tuple(&vp9_fwht4x4_mmx, &vp9_iwht4x4_16_add_c, 0)));
+#endif
+
#if HAVE_SSE2
INSTANTIATE_TEST_CASE_P(
SSE2, Trans4x4DCT,
--- a/vp9/common/vp9_rtcd_defs.pl
+++ b/vp9/common/vp9_rtcd_defs.pl
@@ -701,7 +701,7 @@
specialize qw/vp9_fht16x16 sse2 avx2/;
add_proto qw/void vp9_fwht4x4/, "const int16_t *input, int16_t *output, int stride";
-specialize qw/vp9_fwht4x4/;
+specialize qw/vp9_fwht4x4 mmx/;
add_proto qw/void vp9_fdct4x4/, "const int16_t *input, int16_t *output, int stride";
specialize qw/vp9_fdct4x4 sse2 avx2/;
--- /dev/null
+++ b/vp9/encoder/x86/vp9_dct_mmx.c
@@ -1,0 +1,67 @@
+/*
+ * Copyright (c) 2014 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 <mmintrin.h>
+#include <stdint.h>
+
+#include "./vpx_config.h"
+
+static void INLINE transpose_4x4_mmx(__m64* a, __m64* b, __m64* c, __m64* d) {+ __m64 w, x, y, z;
+ w = _mm_unpacklo_pi16(*a, *b);
+ x = _mm_unpackhi_pi16(*a, *b);
+ y = _mm_unpacklo_pi16(*c, *d);
+ z = _mm_unpackhi_pi16(*c, *d);
+ *a = _mm_unpacklo_pi32(w, y);
+ *b = _mm_unpackhi_pi32(w, y);
+ *c = _mm_unpacklo_pi32(x, z);
+ *d = _mm_unpackhi_pi32(x, z);
+}
+
+static void INLINE fwht_4x4_cols(__m64* out0,
+ __m64* out1,
+ __m64* out2,
+ __m64* out3,
+ __m64 a1,
+ __m64 b1,
+ __m64 c1,
+ __m64 d1) {+ __m64 e1;
+
+ a1 = _mm_add_pi16(a1, b1);
+ d1 = _mm_sub_pi16(d1, c1);
+ e1 = _mm_sub_pi16(a1, d1);
+ e1 = _mm_srai_pi16(e1, 1);
+ b1 = _mm_sub_pi16(e1, b1);
+ c1 = _mm_sub_pi16(e1, c1);
+ a1 = _mm_sub_pi16(a1, c1);
+ d1 = _mm_add_pi16(d1, b1);
+ *out0 = a1;
+ *out1 = c1;
+ *out2 = d1;
+ *out3 = b1;
+}
+
+void vp9_fwht4x4_mmx(const int16_t* input, int16_t* output, int stride) {+ __m64 a1 = *(const __m64*)input;
+ __m64 b1 = *(const __m64*)(input + stride);
+ __m64 c1 = *(const __m64*)(input + 2 * stride);
+ __m64 d1 = *(const __m64*)(input + 3 * stride);
+
+ fwht_4x4_cols(&a1, &b1, &c1, &d1, a1, b1, c1, d1);
+ transpose_4x4_mmx(&a1, &b1, &c1, &d1);
+ fwht_4x4_cols(&a1, &b1, &c1, &d1, a1, b1, c1, d1);
+ transpose_4x4_mmx(&a1, &b1, &c1, &d1);
+
+ *(__m64*)output = _mm_slli_pi16(a1, 2);
+ *(__m64*)(output + 4) = _mm_slli_pi16(b1, 2);
+ *(__m64*)(output + 8) = _mm_slli_pi16(c1, 2);
+ *(__m64*)(output + 12) = _mm_slli_pi16(d1, 2);
+}
--- a/vp9/vp9cx.mk
+++ b/vp9/vp9cx.mk
@@ -92,6 +92,7 @@
VP9_CX_SRCS-$(HAVE_MMX) += encoder/x86/vp9_variance_mmx.c
VP9_CX_SRCS-$(HAVE_MMX) += encoder/x86/vp9_variance_impl_mmx.asm
VP9_CX_SRCS-$(HAVE_MMX) += encoder/x86/vp9_sad_mmx.asm
+VP9_CX_SRCS-$(HAVE_MMX) += encoder/x86/vp9_dct_mmx.c
VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_variance_impl_sse2.asm
VP9_CX_SRCS-$(HAVE_AVX2) += encoder/x86/vp9_variance_impl_intrin_avx2.c
VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_sad4d_sse2.asm