sigrid/ffmpeg

Changelog+1−1

@@ -5,7 +5,7 @@ releases are sorted from youngest to oldest. · 4 unmodified lines
5	5	version <next>:
6	6
7	7	- Lots of deprecated API cruft removed
8		−
	8	+- fft and imdct optimizations for AVX (Sandy Bridge) processors
9	9
10	10	version 0.7_beta1:
11	11

libavcodec/aac.h+5−5

@@ -223,9 +223,9 @@ typedef struct { · 222 unmodified lines
223	223	float sf[120]; ///< scalefactors
224	224	int sf_idx[128]; ///< scalefactor indices (used by encoder)
225	225	uint8_t zeroes[128]; ///< band is not coded (used by encoder)
226		− DECLARE_ALIGNED(16, float, coeffs)[1024]; ///< coefficients for IMDCT
227		− DECLARE_ALIGNED(16, float, saved)[1024]; ///< overlap
228		− DECLARE_ALIGNED(16, float, ret)[2048]; ///< PCM output
	226	+ DECLARE_ALIGNED(32, float, coeffs)[1024]; ///< coefficients for IMDCT
	227	+ DECLARE_ALIGNED(32, float, saved)[1024]; ///< overlap
	228	+ DECLARE_ALIGNED(32, float, ret)[2048]; ///< PCM output
229	229	DECLARE_ALIGNED(16, int16_t, ltp_state)[3072]; ///< time signal for LTP
230	230	PredictorState predictor_state[MAX_PREDICTORS];
231	231	} SingleChannelElement;
@@ -272,7 +272,7 @@ typedef struct { · 40 unmodified lines
272	272	* @defgroup temporary aligned temporary buffers (We do not want to have these on the stack.)
273	273	* @{
274	274	*/
275		− DECLARE_ALIGNED(16, float, buf_mdct)[1024];
	275	+ DECLARE_ALIGNED(32, float, buf_mdct)[1024];
276	276	/** @} */
277	277
278	278	/**
@@ -296,7 +296,7 @@ typedef struct { · 17 unmodified lines
296	296	int sf_offset; ///< offset into pow2sf_tab as appropriate for dsp.float_to_int16
297	297	/** @} */
298	298
299		− DECLARE_ALIGNED(16, float, temp)[128];
	299	+ DECLARE_ALIGNED(32, float, temp)[128];
300	300
301	301	enum OCStatus output_configured;
302	302	} AACContext;

libavcodec/aacenc.h+1−1

@@ -64,7 +64,7 @@ typedef struct AACEncContext { · 63 unmodified lines
64	64	int last_frame;
65	65	float lambda;
66	66	DECLARE_ALIGNED(16, int, qcoefs)[96]; ///< quantized coefficients
67		− DECLARE_ALIGNED(16, float, scoefs)[1024]; ///< scaled coefficients
	67	+ DECLARE_ALIGNED(32, float, scoefs)[1024]; ///< scaled coefficients
68	68	} AACEncContext;
69	69
70	70	#endif /* AVCODEC_AACENC_H */

libavcodec/ac3dec.h+5−5

@@ -200,11 +200,11 @@ typedef struct { · 199 unmodified lines
200	200
201	201	///@defgroup arrays aligned arrays
202	202	DECLARE_ALIGNED(16, int, fixed_coeffs)[AC3_MAX_CHANNELS][AC3_MAX_COEFS]; ///> fixed-point transform coefficients
203		− DECLARE_ALIGNED(16, float, transform_coeffs)[AC3_MAX_CHANNELS][AC3_MAX_COEFS]; ///< transform coefficients
204		− DECLARE_ALIGNED(16, float, delay)[AC3_MAX_CHANNELS][AC3_BLOCK_SIZE]; ///< delay - added to the next block
205		− DECLARE_ALIGNED(16, float, window)[AC3_BLOCK_SIZE]; ///< window coefficients
206		− DECLARE_ALIGNED(16, float, tmp_output)[AC3_BLOCK_SIZE]; ///< temporary storage for output before windowing
207		− DECLARE_ALIGNED(16, float, output)[AC3_MAX_CHANNELS][AC3_BLOCK_SIZE]; ///< output after imdct transform and windowing
	203	+ DECLARE_ALIGNED(32, float, transform_coeffs)[AC3_MAX_CHANNELS][AC3_MAX_COEFS]; ///< transform coefficients
	204	+ DECLARE_ALIGNED(32, float, delay)[AC3_MAX_CHANNELS][AC3_BLOCK_SIZE]; ///< delay - added to the next block
	205	+ DECLARE_ALIGNED(32, float, window)[AC3_BLOCK_SIZE]; ///< window coefficients
	206	+ DECLARE_ALIGNED(32, float, tmp_output)[AC3_BLOCK_SIZE]; ///< temporary storage for output before windowing
	207	+ DECLARE_ALIGNED(32, float, output)[AC3_MAX_CHANNELS][AC3_BLOCK_SIZE]; ///< output after imdct transform and windowing
208	208	///@}
209	209	} AC3DecodeContext;
210	210

libavcodec/ac3enc.c+1−1

@@ -201,7 +201,7 @@ typedef struct AC3EncodeContext { · 200 unmodified lines
201	201
202	202	uint8_t exp_strategy[AC3_MAX_CHANNELS][AC3_MAX_BLOCKS]; ///< exponent strategies
203	203
204		− DECLARE_ALIGNED(16, SampleType, windowed_samples)[AC3_WINDOW_SIZE];
	204	+ DECLARE_ALIGNED(32, SampleType, windowed_samples)[AC3_WINDOW_SIZE];
205	205	} AC3EncodeContext;
206	206
207	207	typedef struct AC3Mant {

libavcodec/atrac1.c+10−10

@@ -60,11 +60,11 @@ typedef struct { · 59 unmodified lines
60	60	int log2_block_count[AT1_QMF_BANDS]; ///< log2 number of blocks in a band
61	61	int num_bfus; ///< number of Block Floating Units
62	62	float* spectrum[2];
63		− DECLARE_ALIGNED(16, float, spec1)[AT1_SU_SAMPLES]; ///< mdct buffer
64		− DECLARE_ALIGNED(16, float, spec2)[AT1_SU_SAMPLES]; ///< mdct buffer
65		− DECLARE_ALIGNED(16, float, fst_qmf_delay)[46]; ///< delay line for the 1st stacked QMF filter
66		− DECLARE_ALIGNED(16, float, snd_qmf_delay)[46]; ///< delay line for the 2nd stacked QMF filter
67		− DECLARE_ALIGNED(16, float, last_qmf_delay)[256+23]; ///< delay line for the last stacked QMF filter
	63	+ DECLARE_ALIGNED(32, float, spec1)[AT1_SU_SAMPLES]; ///< mdct buffer
	64	+ DECLARE_ALIGNED(32, float, spec2)[AT1_SU_SAMPLES]; ///< mdct buffer
	65	+ DECLARE_ALIGNED(32, float, fst_qmf_delay)[46]; ///< delay line for the 1st stacked QMF filter
	66	+ DECLARE_ALIGNED(32, float, snd_qmf_delay)[46]; ///< delay line for the 2nd stacked QMF filter
	67	+ DECLARE_ALIGNED(32, float, last_qmf_delay)[256+23]; ///< delay line for the last stacked QMF filter
68	68	} AT1SUCtx;
69	69
70	70	/**
@@ -72,13 +72,13 @@ typedef struct { · 1 unmodified line
72	72	*/
73	73	typedef struct {
74	74	AT1SUCtx SUs[AT1_MAX_CHANNELS]; ///< channel sound unit
75		− DECLARE_ALIGNED(16, float, spec)[AT1_SU_SAMPLES]; ///< the mdct spectrum buffer
	75	+ DECLARE_ALIGNED(32, float, spec)[AT1_SU_SAMPLES]; ///< the mdct spectrum buffer
76	76
77		− DECLARE_ALIGNED(16, float, low)[256];
78		− DECLARE_ALIGNED(16, float, mid)[256];
79		− DECLARE_ALIGNED(16, float, high)[512];
	77	+ DECLARE_ALIGNED(32, float, low)[256];
	78	+ DECLARE_ALIGNED(32, float, mid)[256];
	79	+ DECLARE_ALIGNED(32, float, high)[512];
80	80	float* bands[3];
81		− DECLARE_ALIGNED(16, float, out_samples)[AT1_MAX_CHANNELS][AT1_SU_SAMPLES];
	81	+ DECLARE_ALIGNED(32, float, out_samples)[AT1_MAX_CHANNELS][AT1_SU_SAMPLES];
82	82	FFTContext mdct_ctx[3];
83	83	int channels;
84	84	DSPContext dsp;

libavcodec/atrac3.c+3−3

@@ -74,8 +74,8 @@ typedef struct { · 73 unmodified lines
74	74	int gcBlkSwitch;
75	75	gain_block gainBlock[2];
76	76
77		− DECLARE_ALIGNED(16, float, spectrum)[1024];
78		− DECLARE_ALIGNED(16, float, IMDCT_buf)[1024];
	77	+ DECLARE_ALIGNED(32, float, spectrum)[1024];
	78	+ DECLARE_ALIGNED(32, float, IMDCT_buf)[1024];
79	79
80	80	float delayBuf1[46]; ///<qmf delay buffers
81	81	float delayBuf2[46];
@@ -122,7 +122,7 @@ typedef struct { · 40 unmodified lines
122	122	FFTContext mdct_ctx;
123	123	} ATRAC3Context;
124	124
125		−static DECLARE_ALIGNED(16, float,mdct_window)[512];
	125	+static DECLARE_ALIGNED(32, float, mdct_window)[512];
126	126	static VLC spectral_coeff_tab[7];
127	127	static float gain_tab1[16];
128	128	static float gain_tab2[31];

libavcodec/binkaudio.c+1−1

@@ -55,7 +55,7 @@ typedef struct { · 54 unmodified lines
55	55	int num_bands;
56	56	unsigned int *bands;
57	57	float root;
58		− DECLARE_ALIGNED(16, FFTSample, coeffs)[BINK_BLOCK_MAX_SIZE];
	58	+ DECLARE_ALIGNED(32, FFTSample, coeffs)[BINK_BLOCK_MAX_SIZE];
59	59	DECLARE_ALIGNED(16, short, previous)[BINK_BLOCK_MAX_SIZE / 16]; ///< coeffs from previous audio block
60	60	float *coeffs_ptr[MAX_CHANNELS]; ///< pointers to the coeffs arrays for float_to_int16_interleave
61	61	union {

libavcodec/cook.c+1−1

@@ -153,7 +153,7 @@ typedef struct cook { · 152 unmodified lines
153	153	/* data buffers */
154	154
155	155	uint8_t* decoded_bytes_buffer;
156		− DECLARE_ALIGNED(16, float,mono_mdct_output)[2048];
	156	+ DECLARE_ALIGNED(32, float, mono_mdct_output)[2048];
157	157	float decode_buffer_1[1024];
158	158	float decode_buffer_2[1024];
159	159	float decode_buffer_0[1060]; /* static allocation for joint decode */

libavcodec/dca.c+5−5

@@ -321,16 +321,16 @@ typedef struct { · 320 unmodified lines
321	321
322	322	/* Subband samples history (for ADPCM) */
323	323	float subband_samples_hist[DCA_PRIM_CHANNELS_MAX][DCA_SUBBANDS][4];
324		− DECLARE_ALIGNED(16, float, subband_fir_hist)[DCA_PRIM_CHANNELS_MAX][512];
325		− DECLARE_ALIGNED(16, float, subband_fir_noidea)[DCA_PRIM_CHANNELS_MAX][32];
	324	+ DECLARE_ALIGNED(32, float, subband_fir_hist)[DCA_PRIM_CHANNELS_MAX][512];
	325	+ DECLARE_ALIGNED(32, float, subband_fir_noidea)[DCA_PRIM_CHANNELS_MAX][32];
326	326	int hist_index[DCA_PRIM_CHANNELS_MAX];
327		− DECLARE_ALIGNED(16, float, raXin)[32];
	327	+ DECLARE_ALIGNED(32, float, raXin)[32];
328	328
329	329	int output; ///< type of output
330	330	float scale_bias; ///< output scale
331	331
332		− DECLARE_ALIGNED(16, float, subband_samples)[DCA_BLOCKS_MAX][DCA_PRIM_CHANNELS_MAX][DCA_SUBBANDS][8];
333		− DECLARE_ALIGNED(16, float, samples)[(DCA_PRIM_CHANNELS_MAX+1)*256];
	332	+ DECLARE_ALIGNED(32, float, subband_samples)[DCA_BLOCKS_MAX][DCA_PRIM_CHANNELS_MAX][DCA_SUBBANDS][8];
	333	+ DECLARE_ALIGNED(32, float, samples)[(DCA_PRIM_CHANNELS_MAX+1)*256];
334	334	const float *samples_chanptr[DCA_PRIM_CHANNELS_MAX+1];
335	335
336	336	uint8_t dca_buffer[DCA_MAX_FRAME_SIZE + DCA_MAX_EXSS_HEADER_SIZE + DCA_BUFFER_PADDING_SIZE];

libavcodec/fft.c+48−5

@@ -93,6 +93,44 @@ av_cold void ff_init_ff_cos_tabs(int index) · 92 unmodified lines
93	93	#endif
94	94	}
95	95
	96	+static const int avx_tab[] = {
	97	+ 0, 4, 1, 5, 8, 12, 9, 13, 2, 6, 3, 7, 10, 14, 11, 15
	98	+};
	99	+
	100	+static int is_second_half_of_fft32(int i, int n)
	101	+{
	102	+ if (n <= 32)
	103	+ return i >= 16;
	104	+ else if (i < n/2)
	105	+ return is_second_half_of_fft32(i, n/2);
	106	+ else if (i < 3*n/4)
	107	+ return is_second_half_of_fft32(i - n/2, n/4);
	108	+ else
	109	+ return is_second_half_of_fft32(i - 3*n/4, n/4);
	110	+}
	111	+
	112	+static av_cold void fft_perm_avx(FFTContext *s)
	113	+{
	114	+ int i;
	115	+ int n = 1 << s->nbits;
	116	+
	117	+ for (i = 0; i < n; i += 16) {
	118	+ int k;
	119	+ if (is_second_half_of_fft32(i, n)) {
	120	+ for (k = 0; k < 16; k++)
	121	+ s->revtab[-split_radix_permutation(i + k, n, s->inverse) & (n - 1)] =
	122	+ i + avx_tab[k];
	123	+
	124	+ } else {
	125	+ for (k = 0; k < 16; k++) {
	126	+ int j = i + k;
	127	+ j = (j & ~7) \| ((j >> 1) & 3) \| ((j << 2) & 4);
	128	+ s->revtab[-split_radix_permutation(i + k, n, s->inverse) & (n - 1)] = j;
	129	+ }
	130	+ }
	131	+ }
	132	+}
	133	+
96	134	av_cold int ff_fft_init(FFTContext *s, int nbits, int inverse)
97	135	{
98	136	int i, j, n;
@@ -132,11 +170,16 @@ av_cold int ff_fft_init(FFTContext *s, int nbits, int inverse) · 33 unmodified lines
132	170	for(j=4; j<=nbits; j++) {
133	171	ff_init_ff_cos_tabs(j);
134	172	}
135		− for(i=0; i<n; i++) {
136		− int j = i;
137		− if (s->fft_permutation == FF_FFT_PERM_SWAP_LSBS)
138		− j = (j&~3) \| ((j>>1)&1) \| ((j<<1)&2);
139		− s->revtab[-split_radix_permutation(i, n, s->inverse) & (n-1)] = j;
	173	+
	174	+ if (s->fft_permutation == FF_FFT_PERM_AVX) {
	175	+ fft_perm_avx(s);
	176	+ } else {
	177	+ for(i=0; i<n; i++) {
	178	+ int j = i;
	179	+ if (s->fft_permutation == FF_FFT_PERM_SWAP_LSBS)
	180	+ j = (j&~3) \| ((j>>1)&1) \| ((j<<1)&2);
	181	+ s->revtab[-split_radix_permutation(i, n, s->inverse) & (n-1)] = j;
	182	+ }
140	183	}
141	184
142	185	return 0;

libavcodec/fft.h+2−1

@@ -85,6 +85,7 @@ struct FFTContext { · 84 unmodified lines
85	85	int fft_permutation;
86	86	#define FF_FFT_PERM_DEFAULT 0
87	87	#define FF_FFT_PERM_SWAP_LSBS 1
	88	+#define FF_FFT_PERM_AVX 2
88	89	int mdct_permutation;
89	90	#define FF_MDCT_PERM_NONE 0
90	91	#define FF_MDCT_PERM_INTERLEAVE 1
@@ -97,7 +98,7 @@ struct FFTContext { · 6 unmodified lines
97	98	#endif
98	99
99	100	#define COSTABLE(size) \
100		− COSTABLE_CONST DECLARE_ALIGNED(16, FFTSample, FFT_NAME(ff_cos_##size))[size/2]
	101	+ COSTABLE_CONST DECLARE_ALIGNED(32, FFTSample, FFT_NAME(ff_cos_##size))[size/2]
101	102
102	103	extern COSTABLE(16);
103	104	extern COSTABLE(32);

libavcodec/imc.c+1−1

@@ -88,7 +88,7 @@ typedef struct { · 87 unmodified lines
88	88
89	89	DSPContext dsp;
90	90	FFTContext fft;
91		− DECLARE_ALIGNED(16, FFTComplex, samples)[COEFFS/2];
	91	+ DECLARE_ALIGNED(32, FFTComplex, samples)[COEFFS/2];
92	92	float *out_samples;
93	93	} IMCContext;
94	94

libavcodec/nellymoserdec.c+2−2

@@ -47,7 +47,7 @@ · 46 unmodified lines
47	47
48	48	typedef struct NellyMoserDecodeContext {
49	49	AVCodecContext* avctx;
50		− DECLARE_ALIGNED(16, float,float_buf)[NELLY_SAMPLES];
	50	+ DECLARE_ALIGNED(32, float, float_buf)[NELLY_SAMPLES];
51	51	float state[128];
52	52	AVLFG random_state;
53	53	GetBitContext gb;
@@ -55,7 +55,7 @@ typedef struct NellyMoserDecodeContext { · 1 unmodified line
55	55	DSPContext dsp;
56	56	FFTContext imdct_ctx;
57	57	FmtConvertContext fmt_conv;
58		− DECLARE_ALIGNED(16, float,imdct_out)[NELLY_BUF_LEN * 2];
	58	+ DECLARE_ALIGNED(32, float, imdct_out)[NELLY_BUF_LEN * 2];
59	59	} NellyMoserDecodeContext;
60	60
61	61	static void overlap_and_window(NellyMoserDecodeContext s, float state, float audio, float a_in)

libavcodec/nellymoserenc.c+3−3

@@ -55,9 +55,9 @@ typedef struct NellyMoserEncodeContext { · 54 unmodified lines
55	55	int have_saved;
56	56	DSPContext dsp;
57	57	FFTContext mdct_ctx;
58		− DECLARE_ALIGNED(16, float, mdct_out)[NELLY_SAMPLES];
59		− DECLARE_ALIGNED(16, float, in_buff)[NELLY_SAMPLES];
60		− DECLARE_ALIGNED(16, float, buf)[2][3 * NELLY_BUF_LEN]; ///< sample buffer
	58	+ DECLARE_ALIGNED(32, float, mdct_out)[NELLY_SAMPLES];
	59	+ DECLARE_ALIGNED(32, float, in_buff)[NELLY_SAMPLES];
	60	+ DECLARE_ALIGNED(32, float, buf)[2][3 * NELLY_BUF_LEN]; ///< sample buffer
61	61	float (*opt )[NELLY_BANDS];
62	62	uint8_t (*path)[NELLY_BANDS];
63	63	} NellyMoserEncodeContext;

libavcodec/qdm2.c+1−1

@@ -120,7 +120,7 @@ typedef struct { · 119 unmodified lines
120	120	} FFTCoefficient;
121	121
122	122	typedef struct {
123		− DECLARE_ALIGNED(16, QDM2Complex, complex)[MPA_MAX_CHANNELS][256];
	123	+ DECLARE_ALIGNED(32, QDM2Complex, complex)[MPA_MAX_CHANNELS][256];
124	124	} QDM2FFT;
125	125
126	126	/**

libavcodec/wma.h+4−4

@@ -113,15 +113,15 @@ typedef struct WMACodecContext { · 112 unmodified lines
113	113	uint8_t ms_stereo; ///< true if mid/side stereo mode
114	114	uint8_t channel_coded[MAX_CHANNELS]; ///< true if channel is coded
115	115	int exponents_bsize[MAX_CHANNELS]; ///< log2 ratio frame/exp. length
116		− DECLARE_ALIGNED(16, float, exponents)[MAX_CHANNELS][BLOCK_MAX_SIZE];
	116	+ DECLARE_ALIGNED(32, float, exponents)[MAX_CHANNELS][BLOCK_MAX_SIZE];
117	117	float max_exponent[MAX_CHANNELS];
118	118	WMACoef coefs1[MAX_CHANNELS][BLOCK_MAX_SIZE];
119		− DECLARE_ALIGNED(16, float, coefs)[MAX_CHANNELS][BLOCK_MAX_SIZE];
120		− DECLARE_ALIGNED(16, FFTSample, output)[BLOCK_MAX_SIZE * 2];
	119	+ DECLARE_ALIGNED(32, float, coefs)[MAX_CHANNELS][BLOCK_MAX_SIZE];
	120	+ DECLARE_ALIGNED(32, FFTSample, output)[BLOCK_MAX_SIZE * 2];
121	121	FFTContext mdct_ctx[BLOCK_NB_SIZES];
122	122	float *windows[BLOCK_NB_SIZES];
123	123	/* output buffer for one frame and the last for IMDCT windowing */
124		− DECLARE_ALIGNED(16, float, frame_out)[MAX_CHANNELS][BLOCK_MAX_SIZE * 2];
	124	+ DECLARE_ALIGNED(32, float, frame_out)[MAX_CHANNELS][BLOCK_MAX_SIZE * 2];
125	125	/* last frame info */
126	126	uint8_t last_superframe[MAX_CODED_SUPERFRAME_SIZE + 4]; /* padding added */
127	127	int last_bitoffset;

libavcodec/wmaprodec.c+2−2

@@ -145,7 +145,7 @@ typedef struct { · 144 unmodified lines
145	145	uint8_t table_idx; ///< index in sf_offsets for the scale factor reference block
146	146	float* coeffs; ///< pointer to the subframe decode buffer
147	147	uint16_t num_vec_coeffs; ///< number of vector coded coefficients
148		− DECLARE_ALIGNED(16, float, out)[WMAPRO_BLOCK_MAX_SIZE + WMAPRO_BLOCK_MAX_SIZE / 2]; ///< output buffer
	148	+ DECLARE_ALIGNED(32, float, out)[WMAPRO_BLOCK_MAX_SIZE + WMAPRO_BLOCK_MAX_SIZE / 2]; ///< output buffer
149	149	} WMAProChannelCtx;
150	150
151	151	/**
@@ -170,7 +170,7 @@ typedef struct WMAProDecodeCtx { · 18 unmodified lines
170	170	FF_INPUT_BUFFER_PADDING_SIZE];///< compressed frame data
171	171	PutBitContext pb; ///< context for filling the frame_data buffer
172	172	FFTContext mdct_ctx[WMAPRO_BLOCK_SIZES]; ///< MDCT context per block size
173		− DECLARE_ALIGNED(16, float, tmp)[WMAPRO_BLOCK_MAX_SIZE]; ///< IMDCT output buffer
	173	+ DECLARE_ALIGNED(32, float, tmp)[WMAPRO_BLOCK_MAX_SIZE]; ///< IMDCT output buffer
174	174	float* windows[WMAPRO_BLOCK_SIZES]; ///< windows for the different block sizes
175	175
176	176	/* frame size dependent frame information (set during initialization) */

libavcodec/wmavoice.c+3−3

@@ -275,11 +275,11 @@ typedef struct { · 274 unmodified lines
275	275	///< by postfilter
276	276	float denoise_filter_cache[MAX_FRAMESIZE];
277	277	int denoise_filter_cache_size; ///< samples in #denoise_filter_cache
278		− DECLARE_ALIGNED(16, float, tilted_lpcs_pf)[0x80];
	278	+ DECLARE_ALIGNED(32, float, tilted_lpcs_pf)[0x80];
279	279	///< aligned buffer for LPC tilting
280		− DECLARE_ALIGNED(16, float, denoise_coeffs_pf)[0x80];
	280	+ DECLARE_ALIGNED(32, float, denoise_coeffs_pf)[0x80];
281	281	///< aligned buffer for denoise coefficients
282		− DECLARE_ALIGNED(16, float, synth_filter_out_buf)[0x80 + MAX_LSPS_ALIGN16];
	282	+ DECLARE_ALIGNED(32, float, synth_filter_out_buf)[0x80 + MAX_LSPS_ALIGN16];
283	283	///< aligned buffer for postfilter speech
284	284	///< synthesis
285	285	/**

libavcodec/x86/fft.c+8−1

@@ -25,7 +25,14 @@ av_cold void ff_fft_init_mmx(FFTContext *s) · 24 unmodified lines
25	25	{
26	26	#if HAVE_YASM
27	27	int has_vectors = av_get_cpu_flags();
28		− if (has_vectors & AV_CPU_FLAG_SSE && HAVE_SSE) {
	28	+ if (has_vectors & AV_CPU_FLAG_AVX && HAVE_AVX && s->nbits >= 5) {
	29	+ /* AVX for SB */
	30	+ s->imdct_calc = ff_imdct_calc_sse;
	31	+ s->imdct_half = ff_imdct_half_avx;
	32	+ s->fft_permute = ff_fft_permute_sse;
	33	+ s->fft_calc = ff_fft_calc_avx;
	34	+ s->fft_permutation = FF_FFT_PERM_AVX;
	35	+ } else if (has_vectors & AV_CPU_FLAG_SSE && HAVE_SSE) {
29	36	/* SSE for P3/P4/K8 */
30	37	s->imdct_calc = ff_imdct_calc_sse;
31	38	s->imdct_half = ff_imdct_half_sse;

libavcodec/x86/fft.h+2−0

@@ -22,6 +22,7 @@ · 21 unmodified lines
22	22	#include "libavcodec/fft.h"
23	23
24	24	void ff_fft_permute_sse(FFTContext s, FFTComplex z);
	25	+void ff_fft_calc_avx(FFTContext s, FFTComplex z);
25	26	void ff_fft_calc_sse(FFTContext s, FFTComplex z);
26	27	void ff_fft_calc_3dn(FFTContext s, FFTComplex z);
27	28	void ff_fft_calc_3dn2(FFTContext s, FFTComplex z);
@@ -32,6 +33,7 @@ void ff_imdct_calc_3dn2(FFTContext s, FFTSample output, const FFTSample *input · 4 unmodified lines
32	33	void ff_imdct_half_3dn2(FFTContext s, FFTSample output, const FFTSample *input);
33	34	void ff_imdct_calc_sse(FFTContext s, FFTSample output, const FFTSample *input);
34	35	void ff_imdct_half_sse(FFTContext s, FFTSample output, const FFTSample *input);
	36	+void ff_imdct_half_avx(FFTContext s, FFTSample output, const FFTSample *input);
35	37	void ff_dct32_float_sse(FFTSample out, const FFTSample in);
36	38
37	39	#endif

libavcodec/x86/fft_mmx.asm+334−150

@@ -1,6 +1,7 @@
1	1	;******************************************************************************
2	2	;* FFT transform with SSE/3DNow optimizations
3	3	;* Copyright (c) 2008 Loren Merritt
	4	+;* Copyright (c) 2011 Vitor Sessak
4	5	;*
5	6	;* This algorithm (though not any of the implementation details) is
6	7	;* based on libdjbfft by D. J. Bernstein.
@@ -49,9 +50,21 @@ endstruc · 42 unmodified lines
49	50	SECTION_RODATA
50	51
51	52	%define M_SQRT1_2 0.70710678118654752440
52		−ps_root2: times 4 dd M_SQRT1_2
53		−ps_root2mppm: dd -M_SQRT1_2, M_SQRT1_2, M_SQRT1_2, -M_SQRT1_2
54		−ps_p1p1m1p1: dd 0, 0, 1<<31, 0
	53	+%define M_COS_PI_1_8 0.923879532511287
	54	+%define M_COS_PI_3_8 0.38268343236509
	55	+
	56	+align 32
	57	+ps_cos16_1: dd 1.0, M_COS_PI_1_8, M_SQRT1_2, M_COS_PI_3_8, 1.0, M_COS_PI_1_8, M_SQRT1_2, M_COS_PI_3_8
	58	+ps_cos16_2: dd 0, M_COS_PI_3_8, M_SQRT1_2, M_COS_PI_1_8, 0, -M_COS_PI_3_8, -M_SQRT1_2, -M_COS_PI_1_8
	59	+
	60	+ps_root2: times 8 dd M_SQRT1_2
	61	+ps_root2mppm: dd -M_SQRT1_2, M_SQRT1_2, M_SQRT1_2, -M_SQRT1_2, -M_SQRT1_2, M_SQRT1_2, M_SQRT1_2, -M_SQRT1_2
	62	+ps_p1p1m1p1: dd 0, 0, 1<<31, 0, 0, 0, 1<<31, 0
	63	+
	64	+perm1: dd 0x00, 0x02, 0x03, 0x01, 0x03, 0x00, 0x02, 0x01
	65	+perm2: dd 0x00, 0x01, 0x02, 0x03, 0x01, 0x00, 0x02, 0x03
	66	+ps_p1p1m1p1root2: dd 1.0, 1.0, -1.0, 1.0, M_SQRT1_2, M_SQRT1_2, M_SQRT1_2, M_SQRT1_2
	67	+ps_m1m1p1m1p1m1m1m1: dd 1<<31, 1<<31, 0, 1<<31, 0, 1<<31, 1<<31, 1<<31
55	68	ps_m1p1: dd 1<<31, 0
56	69
57	70	%assign i 16
@@ -96,51 +109,80 @@ section .text align=16 · 38 unmodified lines
96	109	SWAP %3, %6
97	110	%endmacro
98	111
	112	+; in: %1 = {r0,i0,r2,i2,r4,i4,r6,i6}
	113	+; %2 = {r1,i1,r3,i3,r5,i5,r7,i7}
	114	+; %3, %4, %5 tmp
	115	+; out: %1 = {r0,r1,r2,r3,i0,i1,i2,i3}
	116	+; %2 = {r4,r5,r6,r7,i4,i5,i6,i7}
	117	+%macro T8_AVX 5
	118	+ vsubps %5, %1, %2 ; v = %1 - %2
	119	+ vaddps %3, %1, %2 ; w = %1 + %2
	120	+ vmulps %2, %5, [ps_p1p1m1p1root2] ; v *= vals1
	121	+ vpermilps %2, %2, [perm1]
	122	+ vblendps %1, %2, %3, 0x33 ; q = {w1,w2,v4,v2,w5,w6,v7,v6}
	123	+ vshufps %5, %3, %2, 0x4e ; r = {w3,w4,v1,v3,w7,w8,v8,v5}
	124	+ vsubps %4, %5, %1 ; s = r - q
	125	+ vaddps %1, %5, %1 ; u = r + q
	126	+ vpermilps %1, %1, [perm2] ; k = {u1,u2,u3,u4,u6,u5,u7,u8}
	127	+ vshufps %5, %4, %1, 0xbb
	128	+ vshufps %3, %4, %1, 0xee
	129	+ vperm2f128 %3, %3, %5, 0x13
	130	+ vxorps %4, %4, [ps_m1m1p1m1p1m1m1m1] ; s *= {1,1,-1,-1,1,-1,-1,-1}
	131	+ vshufps %2, %1, %4, 0xdd
	132	+ vshufps %1, %1, %4, 0x88
	133	+ vperm2f128 %4, %2, %1, 0x02 ; v = {k1,k3,s1,s3,k2,k4,s2,s4}
	134	+ vperm2f128 %1, %1, %2, 0x13 ; w = {k6,k8,s6,s8,k5,k7,s5,s7}
	135	+ vsubps %5, %1, %3
	136	+ vblendps %1, %5, %1, 0x55 ; w -= {0,s7,0,k7,0,s8,0,k8}
	137	+ vsubps %2, %4, %1 ; %2 = v - w
	138	+ vaddps %1, %4, %1 ; %1 = v + w
	139	+%endmacro
	140	+
	141	+; In SSE mode do one fft4 transforms
99	142	; in: %1={r0,i0,r2,i2} %2={r1,i1,r3,i3}
100	143	; out: %1={r0,r1,r2,r3} %2={i0,i1,i2,i3}
	144	+;
	145	+; In AVX mode do two fft4 transforms
	146	+; in: %1={r0,i0,r2,i2,r4,i4,r6,i6} %2={r1,i1,r3,i3,r5,i5,r7,i7}
	147	+; out: %1={r0,r1,r2,r3,r4,r5,r6,r7} %2={i0,i1,i2,i3,i4,i5,i6,i7}
101	148	%macro T4_SSE 3
102		− mova %3, %1
103		− addps %1, %2 ; {t1,t2,t6,t5}
104		− subps %3, %2 ; {t3,t4,-t8,t7}
105		− xorps %3, [ps_p1p1m1p1]
106		− mova %2, %1
107		− shufps %1, %3, 0x44 ; {t1,t2,t3,t4}
108		− shufps %2, %3, 0xbe ; {t6,t5,t7,t8}
109		− mova %3, %1
110		− addps %1, %2 ; {r0,i0,r1,i1}
111		− subps %3, %2 ; {r2,i2,r3,i3}
112		− mova %2, %1
113		− shufps %1, %3, 0x88 ; {r0,r1,r2,r3}
114		− shufps %2, %3, 0xdd ; {i0,i1,i2,i3}
	149	+ subps %3, %1, %2 ; {t3,t4,-t8,t7}
	150	+ addps %1, %1, %2 ; {t1,t2,t6,t5}
	151	+ xorps %3, %3, [ps_p1p1m1p1]
	152	+ shufps %2, %1, %3, 0xbe ; {t6,t5,t7,t8}
	153	+ shufps %1, %1, %3, 0x44 ; {t1,t2,t3,t4}
	154	+ subps %3, %1, %2 ; {r2,i2,r3,i3}
	155	+ addps %1, %1, %2 ; {r0,i0,r1,i1}
	156	+ shufps %2, %1, %3, 0xdd ; {i0,i1,i2,i3}
	157	+ shufps %1, %1, %3, 0x88 ; {r0,r1,r2,r3}
115	158	%endmacro
116	159
	160	+; In SSE mode do one FFT8
117	161	; in: %1={r0,r1,r2,r3} %2={i0,i1,i2,i3} %3={r4,i4,r6,i6} %4={r5,i5,r7,i7}
118	162	; out: %1={r0,r1,r2,r3} %2={i0,i1,i2,i3} %1={r4,r5,r6,r7} %2={i4,i5,i6,i7}
	163	+;
	164	+; In AVX mode do two FFT8
	165	+; in: %1={r0,i0,r2,i2,r8, i8, r10,i10} %2={r1,i1,r3,i3,r9, i9, r11,i11}
	166	+; %3={r4,i4,r6,i6,r12,i12,r14,i14} %4={r5,i5,r7,i7,r13,i13,r15,i15}
	167	+; out: %1={r0,r1,r2,r3,r8, r9, r10,r11} %2={i0,i1,i2,i3,i8, i9, i10,i11}
	168	+; %3={r4,r5,r6,r7,r12,r13,r14,r15} %4={i4,i5,i6,i7,i12,i13,i14,i15}
119	169	%macro T8_SSE 6
120		− mova %6, %3
121		− subps %3, %4 ; {r5,i5,r7,i7}
122		− addps %6, %4 ; {t1,t2,t3,t4}
123		− mova %4, %3
124		− shufps %4, %4, 0xb1 ; {i5,r5,i7,r7}
125		− mulps %3, [ps_root2mppm] ; {-r5,i5,r7,-i7}
126		− mulps %4, [ps_root2]
127		− addps %3, %4 ; {t8,t7,ta,t9}
128		− mova %4, %6
129		− shufps %6, %3, 0x36 ; {t3,t2,t9,t8}
130		− shufps %4, %3, 0x9c ; {t1,t4,t7,ta}
131		− mova %3, %6
132		− addps %6, %4 ; {t1,t2,t9,ta}
133		− subps %3, %4 ; {t6,t5,tc,tb}
134		− mova %4, %6
135		− shufps %6, %3, 0xd8 ; {t1,t9,t5,tb}
136		− shufps %4, %3, 0x8d ; {t2,ta,t6,tc}
137		− mova %3, %1
138		− mova %5, %2
139		− addps %1, %6 ; {r0,r1,r2,r3}
140		− addps %2, %4 ; {i0,i1,i2,i3}
141		− subps %3, %6 ; {r4,r5,r6,r7}
142		− subps %5, %4 ; {i4,i5,i6,i7}
143		− SWAP %4, %5
	170	+ addps %6, %3, %4 ; {t1,t2,t3,t4}
	171	+ subps %3, %3, %4 ; {r5,i5,r7,i7}
	172	+ shufps %4, %3, %3, 0xb1 ; {i5,r5,i7,r7}
	173	+ mulps %3, %3, [ps_root2mppm] ; {-r5,i5,r7,-i7}
	174	+ mulps %4, %4, [ps_root2]
	175	+ addps %3, %3, %4 ; {t8,t7,ta,t9}
	176	+ shufps %4, %6, %3, 0x9c ; {t1,t4,t7,ta}
	177	+ shufps %6, %6, %3, 0x36 ; {t3,t2,t9,t8}
	178	+ subps %3, %6, %4 ; {t6,t5,tc,tb}
	179	+ addps %6, %6, %4 ; {t1,t2,t9,ta}
	180	+ shufps %5, %6, %3, 0x8d ; {t2,ta,t6,tc}
	181	+ shufps %6, %6, %3, 0xd8 ; {t1,t9,t5,tb}
	182	+ subps %3, %1, %6 ; {r4,r5,r6,r7}
	183	+ addps %1, %1, %6 ; {r0,r1,r2,r3}
	184	+ subps %4, %2, %5 ; {i4,i5,i6,i7}
	185	+ addps %2, %2, %5 ; {i0,i1,i2,i3}
144	186	%endmacro
145	187
146	188	; scheduled for cpu-bound sizes
@@ -148,52 +190,44 @@ section .text align=16 · 1 unmodified line
148	190	IF%1 mova m4, Z(4)
149	191	IF%1 mova m5, Z(5)
150	192	mova m0, %2 ; wre
151		− mova m2, m4
152	193	mova m1, %3 ; wim
153		− mova m3, m5
154		− mulps m2, m0 ; r2*wre
	194	+ mulps m2, m4, m0 ; r2*wre
155	195	IF%1 mova m6, Z2(6)
156		− mulps m3, m1 ; i2*wim
	196	+ mulps m3, m5, m1 ; i2*wim
157	197	IF%1 mova m7, Z2(7)
158		− mulps m4, m1 ; r2*wim
159		− mulps m5, m0 ; i2*wre
160		− addps m2, m3 ; r2wre + i2wim
161		− mova m3, m1
162		− mulps m1, m6 ; r3*wim
163		− subps m5, m4 ; i2wre - r2wim
164		− mova m4, m0
165		− mulps m3, m7 ; i3*wim
166		− mulps m4, m6 ; r3*wre
167		− mulps m0, m7 ; i3*wre
168		− subps m4, m3 ; r3wre - i3wim
	198	+ mulps m4, m4, m1 ; r2*wim
	199	+ mulps m5, m5, m0 ; i2*wre
	200	+ addps m2, m2, m3 ; r2wre + i2wim
	201	+ mulps m3, m1, m7 ; i3*wim
	202	+ subps m5, m5, m4 ; i2wre - r2wim
	203	+ mulps m1, m1, m6 ; r3*wim
	204	+ mulps m4, m0, m6 ; r3*wre
	205	+ mulps m0, m0, m7 ; i3*wre
	206	+ subps m4, m4, m3 ; r3wre - i3wim
169	207	mova m3, Z(0)
170		− addps m0, m1 ; i3wre + r3wim
171		− mova m1, m4
172		− addps m4, m2 ; t5
173		− subps m1, m2 ; t3
174		− subps m3, m4 ; r2
175		− addps m4, Z(0) ; r0
	208	+ addps m0, m0, m1 ; i3wre + r3wim
	209	+ subps m1, m4, m2 ; t3
	210	+ addps m4, m4, m2 ; t5
	211	+ subps m3, m3, m4 ; r2
	212	+ addps m4, m4, Z(0) ; r0
176	213	mova m6, Z(2)
177	214	mova Z(4), m3
178	215	mova Z(0), m4
179		− mova m3, m5
180		− subps m5, m0 ; t4
181		− mova m4, m6
182		− subps m6, m5 ; r3
183		− addps m5, m4 ; r1
184		− mova Z2(6), m6
185		− mova Z(2), m5
	216	+ subps m3, m5, m0 ; t4
	217	+ subps m4, m6, m3 ; r3
	218	+ addps m3, m3, m6 ; r1
	219	+ mova Z2(6), m4
	220	+ mova Z(2), m3
186	221	mova m2, Z(3)
187		− addps m3, m0 ; t6
188		− subps m2, m1 ; i3
	222	+ addps m3, m5, m0 ; t6
	223	+ subps m2, m2, m1 ; i3
189	224	mova m7, Z(1)
190		− addps m1, Z(3) ; i1
	225	+ addps m1, m1, Z(3) ; i1
191	226	mova Z2(7), m2
192	227	mova Z(3), m1
193		− mova m4, m7
194		− subps m7, m3 ; i2
195		− addps m3, m4 ; i0
196		− mova Z(5), m7
	228	+ subps m4, m7, m3 ; i2
	229	+ addps m3, m3, m7 ; i0
	230	+ mova Z(5), m4
197	231	mova Z(1), m3
198	232	%endmacro
199	233
@@ -201,77 +235,55 @@ IF%1 mova m7, Z2(7) · 1 unmodified line
201	235	%macro PASS_BIG 1 ; (!interleave)
202	236	mova m4, Z(4) ; r2
203	237	mova m5, Z(5) ; i2
204		− mova m2, m4
205	238	mova m0, [wq] ; wre
206		− mova m3, m5
207	239	mova m1, [wq+o1q] ; wim
208		− mulps m2, m0 ; r2*wre
	240	+ mulps m2, m4, m0 ; r2*wre
209	241	mova m6, Z2(6) ; r3
210		− mulps m3, m1 ; i2*wim
	242	+ mulps m3, m5, m1 ; i2*wim
211	243	mova m7, Z2(7) ; i3
212		− mulps m4, m1 ; r2*wim
213		− mulps m5, m0 ; i2*wre
214		− addps m2, m3 ; r2wre + i2wim
215		− mova m3, m1
216		− mulps m1, m6 ; r3*wim
217		− subps m5, m4 ; i2wre - r2wim
218		− mova m4, m0
219		− mulps m3, m7 ; i3*wim
220		− mulps m4, m6 ; r3*wre
221		− mulps m0, m7 ; i3*wre
222		− subps m4, m3 ; r3wre - i3wim
	244	+ mulps m4, m4, m1 ; r2*wim
	245	+ mulps m5, m5, m0 ; i2*wre
	246	+ addps m2, m2, m3 ; r2wre + i2wim
	247	+ mulps m3, m1, m7 ; i3*wim
	248	+ mulps m1, m1, m6 ; r3*wim
	249	+ subps m5, m5, m4 ; i2wre - r2wim
	250	+ mulps m4, m0, m6 ; r3*wre
	251	+ mulps m0, m0, m7 ; i3*wre
	252	+ subps m4, m4, m3 ; r3wre - i3wim
223	253	mova m3, Z(0)
224		− addps m0, m1 ; i3wre + r3wim
225		− mova m1, m4
226		− addps m4, m2 ; t5
227		− subps m1, m2 ; t3
228		− subps m3, m4 ; r2
229		− addps m4, Z(0) ; r0
	254	+ addps m0, m0, m1 ; i3wre + r3wim
	255	+ subps m1, m4, m2 ; t3
	256	+ addps m4, m4, m2 ; t5
	257	+ subps m3, m3, m4 ; r2
	258	+ addps m4, m4, Z(0) ; r0
230	259	mova m6, Z(2)
231	260	mova Z(4), m3
232	261	mova Z(0), m4
233		− mova m3, m5
234		− subps m5, m0 ; t4
235		− mova m4, m6
236		− subps m6, m5 ; r3
237		− addps m5, m4 ; r1
238		−IF%1 mova Z2(6), m6
239		−IF%1 mova Z(2), m5
	262	+ subps m3, m5, m0 ; t4
	263	+ subps m4, m6, m3 ; r3
	264	+ addps m3, m3, m6 ; r1
	265	+IF%1 mova Z2(6), m4
	266	+IF%1 mova Z(2), m3
240	267	mova m2, Z(3)
241		− addps m3, m0 ; t6
242		− subps m2, m1 ; i3
	268	+ addps m5, m5, m0 ; t6
	269	+ subps m2, m2, m1 ; i3
243	270	mova m7, Z(1)
244		− addps m1, Z(3) ; i1
	271	+ addps m1, m1, Z(3) ; i1
245	272	IF%1 mova Z2(7), m2
246	273	IF%1 mova Z(3), m1
247		− mova m4, m7
248		− subps m7, m3 ; i2
249		− addps m3, m4 ; i0
250		−IF%1 mova Z(5), m7
251		−IF%1 mova Z(1), m3
	274	+ subps m6, m7, m5 ; i2
	275	+ addps m5, m5, m7 ; i0
	276	+IF%1 mova Z(5), m6
	277	+IF%1 mova Z(1), m5
252	278	%if %1==0
253		− mova m4, m5 ; r1
254		− mova m0, m6 ; r3
255		− unpcklps m5, m1
256		− unpckhps m4, m1
257		− unpcklps m6, m2
258		− unpckhps m0, m2
	279	+ INTERL m1, m3, m7, Z, 2
	280	+ INTERL m2, m4, m0, Z2, 6
	281	+
259	282	mova m1, Z(0)
260	283	mova m2, Z(4)
261		− mova Z(2), m5
262		− mova Z(3), m4
263		− mova Z2(6), m6
264		− mova Z2(7), m0
265		− mova m5, m1 ; r0
266		− mova m4, m2 ; r2
267		− unpcklps m1, m3
268		− unpckhps m5, m3
269		− unpcklps m2, m7
270		− unpckhps m4, m7
271		− mova Z(0), m1
272		− mova Z(1), m5
273		− mova Z(4), m2
274		− mova Z(5), m4
	284	+
	285	+ INTERL m5, m1, m3, Z, 0
	286	+ INTERL m6, m2, m7, Z, 4
275	287	%endif
276	288	%endmacro
277	289
@@ -281,13 +293,106 @@ IF%1 mova Z(1), m3 · 3 unmodified lines
281	293	punpckhdq %3, %2
282	294	%endmacro
283	295
284		−INIT_XMM
285		−%define mova movaps
286		−
287	296	%define Z(x) [r0+mmsize*x]
288	297	%define Z2(x) [r0+mmsize*x]
	298	+%define ZH(x) [r0+mmsize*x+mmsize/2]
	299	+
	300	+INIT_YMM
	301	+
	302	+align 16
	303	+fft8_avx:
	304	+ mova m0, Z(0)
	305	+ mova m1, Z(1)
	306	+ T8_AVX m0, m1, m2, m3, m4
	307	+ mova Z(0), m0
	308	+ mova Z(1), m1
	309	+ ret
	310	+
	311	+
	312	+align 16
	313	+fft16_avx:
	314	+ mova m2, Z(2)
	315	+ mova m3, Z(3)
	316	+ T4_SSE m2, m3, m7
	317	+
	318	+ mova m0, Z(0)
	319	+ mova m1, Z(1)
	320	+ T8_AVX m0, m1, m4, m5, m7
	321	+
	322	+ mova m4, [ps_cos16_1]
	323	+ mova m5, [ps_cos16_2]
	324	+ vmulps m6, m2, m4
	325	+ vmulps m7, m3, m5
	326	+ vaddps m7, m7, m6
	327	+ vmulps m2, m2, m5
	328	+ vmulps m3, m3, m4
	329	+ vsubps m3, m3, m2
	330	+ vblendps m2, m7, m3, 0xf0
	331	+ vperm2f128 m3, m7, m3, 0x21
	332	+ vaddps m4, m2, m3
	333	+ vsubps m2, m3, m2
	334	+ vperm2f128 m2, m2, m2, 0x01
	335	+ vsubps m3, m1, m2
	336	+ vaddps m1, m1, m2
	337	+ vsubps m5, m0, m4
	338	+ vaddps m0, m0, m4
	339	+ vextractf128 Z(0), m0, 0
	340	+ vextractf128 ZH(0), m1, 0
	341	+ vextractf128 Z(1), m0, 1
	342	+ vextractf128 ZH(1), m1, 1
	343	+ vextractf128 Z(2), m5, 0
	344	+ vextractf128 ZH(2), m3, 0
	345	+ vextractf128 Z(3), m5, 1
	346	+ vextractf128 ZH(3), m3, 1
	347	+ ret
	348	+
	349	+align 16
	350	+fft32_avx:
	351	+ call fft16_avx
	352	+
	353	+ mova m0, Z(4)
	354	+ mova m1, Z(5)
	355	+
	356	+ T4_SSE m0, m1, m4
	357	+
	358	+ mova m2, Z(6)
	359	+ mova m3, Z(7)
	360	+
	361	+ T8_SSE m0, m1, m2, m3, m4, m6
	362	+ ; m0={r0,r1,r2,r3,r8, r9, r10,r11} m1={i0,i1,i2,i3,i8, i9, i10,i11}
	363	+ ; m2={r4,r5,r6,r7,r12,r13,r14,r15} m3={i4,i5,i6,i7,i12,i13,i14,i15}
	364	+
	365	+ vperm2f128 m4, m0, m2, 0x20
	366	+ vperm2f128 m5, m1, m3, 0x20
	367	+ vperm2f128 m6, m0, m2, 0x31
	368	+ vperm2f128 m7, m1, m3, 0x31
	369	+
	370	+ PASS_SMALL 0, [cos_32], [cos_32+32]
	371	+
	372	+ ret
	373	+
	374	+fft32_interleave_avx:
	375	+ call fft32_avx
	376	+ mov r2d, 32
	377	+.deint_loop:
	378	+ mova m2, Z(0)
	379	+ mova m3, Z(1)
	380	+ vunpcklps m0, m2, m3
	381	+ vunpckhps m1, m2, m3
	382	+ vextractf128 Z(0), m0, 0
	383	+ vextractf128 ZH(0), m1, 0
	384	+ vextractf128 Z(1), m0, 1
	385	+ vextractf128 ZH(1), m1, 1
	386	+ add r0, mmsize*2
	387	+ sub r2d, mmsize/4
	388	+ jg .deint_loop
	389	+ ret
	390	+
	391	+INIT_XMM
	392	+%define movdqa movaps
289	393
290	394	align 16
	395	+fft4_avx:
291	396	fft4_sse:
292	397	mova m0, Z(0)
293	398	mova m1, Z(1)
@@ -406,6 +511,8 @@ FFT48_3DN _3dn · 112 unmodified lines
406	511
407	512	%define Z(x) [zq + o1q(x&6) + mmsize(x&1)]
408	513	%define Z2(x) [zq + o3q + mmsize*(x&1)]
	514	+%define ZH(x) [zq + o1q(x&6) + mmsize(x&1) + mmsize/2]
	515	+%define Z2H(x) [zq + o3q + mmsize*(x&1) + mmsize/2]
409	516
410	517	%macro DECL_PASS 2+ ; name, payload
411	518	align 16
@@ -423,8 +530,34 @@ DEFINE_ARGS z, w, n, o1, o3 · 11 unmodified lines
423	530	rep ret
424	531	%endmacro
425	532
	533	+INIT_YMM
	534	+
	535	+%macro INTERL_AVX 5
	536	+ vunpckhps %3, %2, %1
	537	+ vunpcklps %2, %2, %1
	538	+ vextractf128 %4(%5), %2, 0
	539	+ vextractf128 %4 %+ H(%5), %3, 0
	540	+ vextractf128 %4(%5 + 1), %2, 1
	541	+ vextractf128 %4 %+ H(%5 + 1), %3, 1
	542	+%endmacro
	543	+
	544	+%define INTERL INTERL_AVX
	545	+
	546	+DECL_PASS pass_avx, PASS_BIG 1
	547	+DECL_PASS pass_interleave_avx, PASS_BIG 0
	548	+
426	549	INIT_XMM
427		−%define mova movaps
	550	+
	551	+%macro INTERL_SSE 5
	552	+ mova %3, %2
	553	+ unpcklps %2, %1
	554	+ unpckhps %3, %1
	555	+ mova %4(%5), %2
	556	+ mova %4(%5+1), %3
	557	+%endmacro
	558	+
	559	+%define INTERL INTERL_SSE
	560	+
428	561	DECL_PASS pass_sse, PASS_BIG 1
429	562	DECL_PASS pass_interleave_sse, PASS_BIG 0
430	563
@@ -457,9 +590,12 @@ DECL_PASS pass_interleave_3dn, PASS_BIG 0 · 26 unmodified lines
457	590
458	591	%macro DECL_FFT 2-3 ; nbits, cpu, suffix
459	592	%xdefine list_of_fft fft4%2 SECTION_REL, fft8%2 SECTION_REL
460		−%if %1==5
	593	+%if %1>=5
461	594	%xdefine list_of_fft list_of_fft, fft16%2 SECTION_REL
462	595	%endif
	596	+%if %1>=6
	597	+%xdefine list_of_fft list_of_fft, fft32%3%2 SECTION_REL
	598	+%endif
463	599
464	600	%assign n 1<<%1
465	601	%rep 17-%1
@@ -492,9 +628,14 @@ section .text · 26 unmodified lines
492	628	; The others pass args in registers and don't spill anything.
493	629	cglobal fft_dispatch%3%2, 2,5,8, z, nbits
494	630	FFT_DISPATCH %3%2, nbits
	631	+%ifidn %2, _avx
	632	+ vzeroupper
	633	+%endif
495	634	RET
496	635	%endmacro ; DECL_FFT
497	636
	637	+DECL_FFT 6, _avx
	638	+DECL_FFT 6, _avx, _interleave
498	639	DECL_FFT 5, _sse
499	640	DECL_FFT 5, _sse, _interleave
500	641	DECL_FFT 4, _3dn
@@ -533,21 +674,53 @@ INIT_XMM · 32 unmodified lines
533	674	%endmacro
534	675
535	676	%macro CMUL 6 ;j, xmm0, xmm1, 3, 4, 5
536		− movaps xmm6, [%4+%1*2]
537		− movaps %2, [%4+%1*2+0x10]
538		− movaps %3, xmm6
539		− movaps xmm7, %2
540		− mulps xmm6, [%5+%1]
541		− mulps %2, [%6+%1]
542		− mulps %3, [%6+%1]
543		− mulps xmm7, [%5+%1]
544		− subps %2, xmm6
545		− addps %3, xmm7
	677	+ mulps m6, %3, [%5+%1]
	678	+ mulps m7, %2, [%5+%1]
	679	+ mulps %2, %2, [%6+%1]
	680	+ mulps %3, %3, [%6+%1]
	681	+ subps %2, %2, m6
	682	+ addps %3, %3, m7
	683	+%endmacro
	684	+
	685	+%macro POSROTATESHUF_AVX 5 ;j, k, z+n8, tcos+n8, tsin+n8
	686	+.post:
	687	+ vmovaps ymm1, [%3+%1*2]
	688	+ vmovaps ymm0, [%3+%1*2+0x20]
	689	+ vmovaps ymm3, [%3+%2*2]
	690	+ vmovaps ymm2, [%3+%2*2+0x20]
	691	+
	692	+ CMUL %1, ymm0, ymm1, %3, %4, %5
	693	+ CMUL %2, ymm2, ymm3, %3, %4, %5
	694	+ vshufps ymm1, ymm1, ymm1, 0x1b
	695	+ vshufps ymm3, ymm3, ymm3, 0x1b
	696	+ vperm2f128 ymm1, ymm1, ymm1, 0x01
	697	+ vperm2f128 ymm3, ymm3, ymm3, 0x01
	698	+ vunpcklps ymm6, ymm2, ymm1
	699	+ vunpckhps ymm4, ymm2, ymm1
	700	+ vunpcklps ymm7, ymm0, ymm3
	701	+ vunpckhps ymm5, ymm0, ymm3
	702	+
	703	+ vextractf128 [%3+%1*2], ymm7, 0
	704	+ vextractf128 [%3+%1*2+0x10], ymm5, 0
	705	+ vextractf128 [%3+%1*2+0x20], ymm7, 1
	706	+ vextractf128 [%3+%1*2+0x30], ymm5, 1
	707	+
	708	+ vextractf128 [%3+%2*2], ymm6, 0
	709	+ vextractf128 [%3+%2*2+0x10], ymm4, 0
	710	+ vextractf128 [%3+%2*2+0x20], ymm6, 1
	711	+ vextractf128 [%3+%2*2+0x30], ymm4, 1
	712	+ sub %2, 0x20
	713	+ add %1, 0x20
	714	+ jl .post
546	715	%endmacro
547	716
548	717	%macro POSROTATESHUF 5 ;j, k, z+n8, tcos+n8, tsin+n8
549	718	.post:
	719	+ movaps xmm1, [%3+%1*2]
	720	+ movaps xmm0, [%3+%1*2+0x10]
550	721	CMUL %1, xmm0, xmm1, %3, %4, %5
	722	+ movaps xmm5, [%3+%2*2]
	723	+ movaps xmm4, [%3+%2*2+0x10]
551	724	CMUL %2, xmm4, xmm5, %3, %4, %5
552	725	shufps xmm1, xmm1, 0x1b
553	726	shufps xmm5, xmm5, 0x1b
@@ -566,7 +739,8 @@ INIT_XMM · 12 unmodified lines
566	739	jl .post
567	740	%endmacro
568	741
569		−cglobal imdct_half_sse, 3,7,8; FFTContext s, FFTSample output, const FFTSample *input
	742	+%macro DECL_IMDCT 2
	743	+cglobal imdct_half%1, 3,7,8; FFTContext s, FFTSample output, const FFTSample *input
570	744	%ifdef ARCH_X86_64
571	745	%define rrevtab r10
572	746	%define rtcos r11
@@ -641,7 +815,7 @@ cglobal imdct_half_sse, 3,7,8; FFTContext s, FFTSample output, const FFTSample · 68 unmodified lines
641	815	mov r0, r1
642	816	mov r1d, [r5+FFTContext.nbits]
643	817
644		− FFT_DISPATCH _sse, r1
	818	+ FFT_DISPATCH %1, r1
645	819
646	820	mov r0d, [r5+FFTContext.mdctsize]
647	821	add r6, r0
@@ -653,14 +827,24 @@ cglobal imdct_half_sse, 3,7,8; FFTContext s, FFTSample output, const FFTSample · 5 unmodified lines
653	827	mov rtsin, [esp+4]
654	828	%endif
655	829	neg r0
656		− mov r1, -16
	830	+ mov r1, -mmsize
657	831	sub r1, r0
658		− POSROTATESHUF r0, r1, r6, rtcos, rtsin
	832	+ %2 r0, r1, r6, rtcos, rtsin
659	833	%ifdef ARCH_X86_64
660	834	pop r14
661	835	pop r13
662	836	pop r12
663	837	%else
664	838	add esp, 12
	839	+%endif
	840	+%ifidn avx_enabled, 1
	841	+ vzeroupper
665	842	%endif
666	843	RET
	844	+%endmacro
	845	+
	846	+DECL_IMDCT _sse, POSROTATESHUF
	847	+
	848	+INIT_YMM
	849	+
	850	+DECL_IMDCT _avx, POSROTATESHUF_AVX

libavcodec/x86/fft_sse.c+7−1

@@ -28,6 +28,12 @@ DECLARE_ASM_CONST(16, int, ff_m1m1m1m1)[4] = · 27 unmodified lines
28	28
29	29	void ff_fft_dispatch_sse(FFTComplex *z, int nbits);
30	30	void ff_fft_dispatch_interleave_sse(FFTComplex *z, int nbits);
	31	+void ff_fft_dispatch_interleave_avx(FFTComplex *z, int nbits);
	32	+
	33	+void ff_fft_calc_avx(FFTContext s, FFTComplex z)
	34	+{
	35	+ ff_fft_dispatch_interleave_avx(z, s->nbits);
	36	+}
31	37
32	38	void ff_fft_calc_sse(FFTContext s, FFTComplex z)
33	39	{
@@ -77,7 +83,7 @@ void ff_imdct_calc_sse(FFTContext s, FFTSample output, const FFTSample *input) · 43 unmodified lines
77	83	long n = s->mdct_size;
78	84	long n4 = n >> 2;
79	85
80		− ff_imdct_half_sse(s, output+n4, input);
	86	+ s->imdct_half(s, output + n4, input);
81	87
82	88	j = -n;
83	89	k = n-16;

Add AVX FFT implementation.

Keyboard shortcuts

General

Issues and pull requests

Issues, pull requests and Inbox

Repository code