1 /*
2 * Copyright (c) 2018, Alliance for Open Media. All rights reserved
3 *
4 * This source code is subject to the terms of the BSD 2 Clause License and
5 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
6 * was not distributed with this source code in the LICENSE file, you can
7 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
8 * Media Patent License 1.0 was not distributed with this source code in the
9 s * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
10 */
11
12 #include <xmmintrin.h>
13
14 #include "config/aom_dsp_rtcd.h"
15 #include "aom_dsp/aom_dsp_common.h"
16 #include "aom_dsp/fft_common.h"
17
transpose4x4(const float * A,float * B,const int lda,const int ldb)18 static INLINE void transpose4x4(const float *A, float *B, const int lda,
19 const int ldb) {
20 __m128 row1 = _mm_load_ps(&A[0 * lda]);
21 __m128 row2 = _mm_load_ps(&A[1 * lda]);
22 __m128 row3 = _mm_load_ps(&A[2 * lda]);
23 __m128 row4 = _mm_load_ps(&A[3 * lda]);
24 _MM_TRANSPOSE4_PS(row1, row2, row3, row4);
25 _mm_store_ps(&B[0 * ldb], row1);
26 _mm_store_ps(&B[1 * ldb], row2);
27 _mm_store_ps(&B[2 * ldb], row3);
28 _mm_store_ps(&B[3 * ldb], row4);
29 }
30
aom_transpose_float_sse2(const float * A,float * B,int n)31 void aom_transpose_float_sse2(const float *A, float *B, int n) {
32 for (int y = 0; y < n; y += 4) {
33 for (int x = 0; x < n; x += 4) {
34 transpose4x4(A + y * n + x, B + x * n + y, n, n);
35 }
36 }
37 }
38
aom_fft_unpack_2d_output_sse2(const float * packed,float * output,int n)39 void aom_fft_unpack_2d_output_sse2(const float *packed, float *output, int n) {
40 const int n2 = n / 2;
41 output[0] = packed[0];
42 output[1] = 0;
43 output[2 * (n2 * n)] = packed[n2 * n];
44 output[2 * (n2 * n) + 1] = 0;
45
46 output[2 * n2] = packed[n2];
47 output[2 * n2 + 1] = 0;
48 output[2 * (n2 * n + n2)] = packed[n2 * n + n2];
49 output[2 * (n2 * n + n2) + 1] = 0;
50
51 for (int c = 1; c < n2; ++c) {
52 output[2 * (0 * n + c)] = packed[c];
53 output[2 * (0 * n + c) + 1] = packed[c + n2];
54 output[2 * (n2 * n + c) + 0] = packed[n2 * n + c];
55 output[2 * (n2 * n + c) + 1] = packed[n2 * n + c + n2];
56 }
57 for (int r = 1; r < n2; ++r) {
58 output[2 * (r * n + 0)] = packed[r * n];
59 output[2 * (r * n + 0) + 1] = packed[(r + n2) * n];
60 output[2 * (r * n + n2) + 0] = packed[r * n + n2];
61 output[2 * (r * n + n2) + 1] = packed[(r + n2) * n + n2];
62
63 for (int c = 1; c < AOMMIN(n2, 4); ++c) {
64 output[2 * (r * n + c)] =
65 packed[r * n + c] - packed[(r + n2) * n + c + n2];
66 output[2 * (r * n + c) + 1] =
67 packed[(r + n2) * n + c] + packed[r * n + c + n2];
68 }
69
70 for (int c = 4; c < n2; c += 4) {
71 __m128 real1 = _mm_load_ps(packed + r * n + c);
72 __m128 real2 = _mm_load_ps(packed + (r + n2) * n + c + n2);
73 __m128 imag1 = _mm_load_ps(packed + (r + n2) * n + c);
74 __m128 imag2 = _mm_load_ps(packed + r * n + c + n2);
75 real1 = _mm_sub_ps(real1, real2);
76 imag1 = _mm_add_ps(imag1, imag2);
77 _mm_store_ps(output + 2 * (r * n + c), _mm_unpacklo_ps(real1, imag1));
78 _mm_store_ps(output + 2 * (r * n + c + 2), _mm_unpackhi_ps(real1, imag1));
79 }
80
81 int r2 = r + n2;
82 int r3 = n - r2;
83 output[2 * (r2 * n + 0)] = packed[r3 * n];
84 output[2 * (r2 * n + 0) + 1] = -packed[(r3 + n2) * n];
85 output[2 * (r2 * n + n2)] = packed[r3 * n + n2];
86 output[2 * (r2 * n + n2) + 1] = -packed[(r3 + n2) * n + n2];
87 for (int c = 1; c < AOMMIN(4, n2); ++c) {
88 output[2 * (r2 * n + c)] =
89 packed[r3 * n + c] + packed[(r3 + n2) * n + c + n2];
90 output[2 * (r2 * n + c) + 1] =
91 -packed[(r3 + n2) * n + c] + packed[r3 * n + c + n2];
92 }
93 for (int c = 4; c < n2; c += 4) {
94 __m128 real1 = _mm_load_ps(packed + r3 * n + c);
95 __m128 real2 = _mm_load_ps(packed + (r3 + n2) * n + c + n2);
96 __m128 imag1 = _mm_load_ps(packed + (r3 + n2) * n + c);
97 __m128 imag2 = _mm_load_ps(packed + r3 * n + c + n2);
98 real1 = _mm_add_ps(real1, real2);
99 imag1 = _mm_sub_ps(imag2, imag1);
100 _mm_store_ps(output + 2 * (r2 * n + c), _mm_unpacklo_ps(real1, imag1));
101 _mm_store_ps(output + 2 * (r2 * n + c + 2),
102 _mm_unpackhi_ps(real1, imag1));
103 }
104 }
105 }
106
107 // Generate definitions for 1d transforms using float and __mm128
108 GEN_FFT_4(static INLINE void, sse2, float, __m128, _mm_load_ps, _mm_store_ps,
109 _mm_set1_ps, _mm_add_ps, _mm_sub_ps);
110 GEN_FFT_8(static INLINE void, sse2, float, __m128, _mm_load_ps, _mm_store_ps,
111 _mm_set1_ps, _mm_add_ps, _mm_sub_ps, _mm_mul_ps);
112 GEN_FFT_16(static INLINE void, sse2, float, __m128, _mm_load_ps, _mm_store_ps,
113 _mm_set1_ps, _mm_add_ps, _mm_sub_ps, _mm_mul_ps);
114 GEN_FFT_32(static INLINE void, sse2, float, __m128, _mm_load_ps, _mm_store_ps,
115 _mm_set1_ps, _mm_add_ps, _mm_sub_ps, _mm_mul_ps);
116
aom_fft4x4_float_sse2(const float * input,float * temp,float * output)117 void aom_fft4x4_float_sse2(const float *input, float *temp, float *output) {
118 aom_fft_2d_gen(input, temp, output, 4, aom_fft1d_4_sse2,
119 aom_transpose_float_sse2, aom_fft_unpack_2d_output_sse2, 4);
120 }
121
aom_fft8x8_float_sse2(const float * input,float * temp,float * output)122 void aom_fft8x8_float_sse2(const float *input, float *temp, float *output) {
123 aom_fft_2d_gen(input, temp, output, 8, aom_fft1d_8_sse2,
124 aom_transpose_float_sse2, aom_fft_unpack_2d_output_sse2, 4);
125 }
126
aom_fft16x16_float_sse2(const float * input,float * temp,float * output)127 void aom_fft16x16_float_sse2(const float *input, float *temp, float *output) {
128 aom_fft_2d_gen(input, temp, output, 16, aom_fft1d_16_sse2,
129 aom_transpose_float_sse2, aom_fft_unpack_2d_output_sse2, 4);
130 }
131
aom_fft32x32_float_sse2(const float * input,float * temp,float * output)132 void aom_fft32x32_float_sse2(const float *input, float *temp, float *output) {
133 aom_fft_2d_gen(input, temp, output, 32, aom_fft1d_32_sse2,
134 aom_transpose_float_sse2, aom_fft_unpack_2d_output_sse2, 4);
135 }
136
137 // Generate definitions for 1d inverse transforms using float and mm128
138 GEN_IFFT_4(static INLINE void, sse2, float, __m128, _mm_load_ps, _mm_store_ps,
139 _mm_set1_ps, _mm_add_ps, _mm_sub_ps);
140 GEN_IFFT_8(static INLINE void, sse2, float, __m128, _mm_load_ps, _mm_store_ps,
141 _mm_set1_ps, _mm_add_ps, _mm_sub_ps, _mm_mul_ps);
142 GEN_IFFT_16(static INLINE void, sse2, float, __m128, _mm_load_ps, _mm_store_ps,
143 _mm_set1_ps, _mm_add_ps, _mm_sub_ps, _mm_mul_ps);
144 GEN_IFFT_32(static INLINE void, sse2, float, __m128, _mm_load_ps, _mm_store_ps,
145 _mm_set1_ps, _mm_add_ps, _mm_sub_ps, _mm_mul_ps);
146
aom_ifft4x4_float_sse2(const float * input,float * temp,float * output)147 void aom_ifft4x4_float_sse2(const float *input, float *temp, float *output) {
148 aom_ifft_2d_gen(input, temp, output, 4, aom_fft1d_4_float, aom_fft1d_4_sse2,
149 aom_ifft1d_4_sse2, aom_transpose_float_sse2, 4);
150 }
151
aom_ifft8x8_float_sse2(const float * input,float * temp,float * output)152 void aom_ifft8x8_float_sse2(const float *input, float *temp, float *output) {
153 aom_ifft_2d_gen(input, temp, output, 8, aom_fft1d_8_float, aom_fft1d_8_sse2,
154 aom_ifft1d_8_sse2, aom_transpose_float_sse2, 4);
155 }
156
aom_ifft16x16_float_sse2(const float * input,float * temp,float * output)157 void aom_ifft16x16_float_sse2(const float *input, float *temp, float *output) {
158 aom_ifft_2d_gen(input, temp, output, 16, aom_fft1d_16_float,
159 aom_fft1d_16_sse2, aom_ifft1d_16_sse2,
160 aom_transpose_float_sse2, 4);
161 }
162
aom_ifft32x32_float_sse2(const float * input,float * temp,float * output)163 void aom_ifft32x32_float_sse2(const float *input, float *temp, float *output) {
164 aom_ifft_2d_gen(input, temp, output, 32, aom_fft1d_32_float,
165 aom_fft1d_32_sse2, aom_ifft1d_32_sse2,
166 aom_transpose_float_sse2, 4);
167 }
168