/ Drivers / CMSIS / DSP / Source / TransformFunctions / arm_dct4_f32.c
arm_dct4_f32.c
  1  /* ----------------------------------------------------------------------
  2   * Project:      CMSIS DSP Library
  3   * Title:        arm_dct4_f32.c
  4   * Description:  Processing function of DCT4 & IDCT4 F32
  5   *
  6   * $Date:        23 April 2021
  7   * $Revision:    V1.9.0
  8   *
  9   * Target Processor: Cortex-M and Cortex-A cores
 10   * -------------------------------------------------------------------- */
 11  /*
 12   * Copyright (C) 2010-2021 ARM Limited or its affiliates. All rights reserved.
 13   *
 14   * SPDX-License-Identifier: Apache-2.0
 15   *
 16   * Licensed under the Apache License, Version 2.0 (the License); you may
 17   * not use this file except in compliance with the License.
 18   * You may obtain a copy of the License at
 19   *
 20   * www.apache.org/licenses/LICENSE-2.0
 21   *
 22   * Unless required by applicable law or agreed to in writing, software
 23   * distributed under the License is distributed on an AS IS BASIS, WITHOUT
 24   * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 25   * See the License for the specific language governing permissions and
 26   * limitations under the License.
 27   */
 28  
 29  #include "dsp/transform_functions.h"
 30  
 31  /**
 32    @ingroup groupTransforms
 33   */
 34  
 35  /**
 36    @defgroup DCT4_IDCT4 DCT Type IV Functions
 37  
 38    Representation of signals by minimum number of values is important for storage and transmission.
 39    The possibility of large discontinuity between the beginning and end of a period of a signal
 40    in DFT can be avoided by extending the signal so that it is even-symmetric.
 41    Discrete Cosine Transform (DCT) is constructed such that its energy is heavily concentrated in the lower part of the
 42    spectrum and is very widely used in signal and image coding applications.
 43    The family of DCTs (DCT type- 1,2,3,4) is the outcome of different combinations of homogeneous boundary conditions.
 44    DCT has an excellent energy-packing capability, hence has many applications and in data compression in particular.
 45    
 46    DCT is essentially the Discrete Fourier Transform(DFT) of an even-extended real signal.
 47    Reordering of the input data makes the computation of DCT just a problem of
 48    computing the DFT of a real signal with a few additional operations.
 49    This approach provides regular, simple, and very efficient DCT algorithms for practical hardware and software implementations.
 50    
 51    DCT type-II can be implemented using Fast fourier transform (FFT) internally, as the transform is applied on real values, Real FFT can be used.
 52    DCT4 is implemented using DCT2 as their implementations are similar except with some added pre-processing and post-processing.
 53    DCT2 implementation can be described in the following steps:
 54    - Re-ordering input
 55    - Calculating Real FFT
 56    - Multiplication of weights and Real FFT output and getting real part from the product.
 57    
 58    This process is explained by the block diagram below:
 59    \image html DCT4.gif "Discrete Cosine Transform - type-IV"
 60   
 61    @par           Algorithm
 62                     The N-point type-IV DCT is defined as a real, linear transformation by the formula:
 63                     \image html DCT4Equation.gif
 64                     where <code>k = 0, 1, 2, ..., N-1</code>
 65    @par
 66                     Its inverse is defined as follows:
 67                     \image html IDCT4Equation.gif
 68                     where <code>n = 0, 1, 2, ..., N-1</code>
 69    @par
 70                     The DCT4 matrices become involutory (i.e. they are self-inverse) by multiplying with an overall scale factor of sqrt(2/N).
 71                     The symmetry of the transform matrix indicates that the fast algorithms for the forward
 72                     and inverse transform computation are identical.
 73                     Note that the implementation of Inverse DCT4 and DCT4 is same, hence same process function can be used for both.
 74   
 75    @par           Lengths supported by the transform:
 76                     As DCT4 internally uses Real FFT, it supports all the lengths 128, 512, 2048 and 8192.
 77                     The library provides separate functions for Q15, Q31, and floating-point data types.
 78  
 79    @par           Instance Structure
 80                     The instances for Real FFT and FFT, cosine values table and twiddle factor table are stored in an instance data structure.
 81                     A separate instance structure must be defined for each transform.
 82                     There are separate instance structure declarations for each of the 3 supported data types.
 83                   
 84    @par           Initialization Functions
 85                     There is also an associated initialization function for each data type.
 86                     The initialization function performs the following operations:
 87                     - Sets the values of the internal structure fields.
 88                     - Initializes Real FFT as its process function is used internally in DCT4, by calling \ref arm_rfft_init_f32().
 89    @par
 90                     Use of the initialization function is optional.
 91                     However, if the initialization function is used, then the instance structure cannot be placed into a const data section.
 92                     To place an instance structure into a const data section, the instance structure must be manually initialized.
 93                     Manually initialize the instance structure as follows:
 94    <pre>
 95        arm_dct4_instance_f32 S = {N, Nby2, normalize, pTwiddle, pCosFactor, pRfft, pCfft};
 96        arm_dct4_instance_q31 S = {N, Nby2, normalize, pTwiddle, pCosFactor, pRfft, pCfft};
 97        arm_dct4_instance_q15 S = {N, Nby2, normalize, pTwiddle, pCosFactor, pRfft, pCfft};
 98    </pre>
 99                     where \c N is the length of the DCT4; \c Nby2 is half of the length of the DCT4;
100                     \c normalize is normalizing factor used and is equal to <code>sqrt(2/N)</code>;
101                     \c pTwiddle points to the twiddle factor table;
102                     \c pCosFactor points to the cosFactor table;
103                     \c pRfft points to the real FFT instance;
104                     \c pCfft points to the complex FFT instance;
105                     The CFFT and RFFT structures also needs to be initialized, refer to arm_cfft_radix4_f32()
106                     and arm_rfft_f32() respectively for details regarding static initialization.
107   
108    @par           Fixed-Point Behavior
109                     Care must be taken when using the fixed-point versions of the DCT4 transform functions.
110                     In particular, the overflow and saturation behavior of the accumulator used in each function must be considered.
111                     Refer to the function specific documentation below for usage guidelines.
112   */
113  
114   /**
115    @addtogroup DCT4_IDCT4
116    @{
117   */
118  
119  /**
120    @brief         Processing function for the floating-point DCT4/IDCT4.
121    @param[in]     S             points to an instance of the floating-point DCT4/IDCT4 structure
122    @param[in]     pState        points to state buffer
123    @param[in,out] pInlineBuffer points to the in-place input and output buffer
124    @return        none
125   */
126  
127  void arm_dct4_f32(
128    const arm_dct4_instance_f32 * S,
129          float32_t * pState,
130          float32_t * pInlineBuffer)
131  {
132    const float32_t *weights = S->pTwiddle;              /* Pointer to the Weights table */
133    const float32_t *cosFact = S->pCosFactor;            /* Pointer to the cos factors table */
134          float32_t *pS1, *pS2, *pbuff;                  /* Temporary pointers for input buffer and pState buffer */
135          float32_t in;                                  /* Temporary variable */
136          uint32_t i;                                    /* Loop counter */
137  
138  
139    /* DCT4 computation involves DCT2 (which is calculated using RFFT)
140     * along with some pre-processing and post-processing.
141     * Computational procedure is explained as follows:
142     * (a) Pre-processing involves multiplying input with cos factor,
143     *     r(n) = 2 * u(n) * cos(pi*(2*n+1)/(4*n))
144     *              where,
145     *                 r(n) -- output of preprocessing
146     *                 u(n) -- input to preprocessing(actual Source buffer)
147     * (b) Calculation of DCT2 using FFT is divided into three steps:
148     *                  Step1: Re-ordering of even and odd elements of input.
149     *                  Step2: Calculating FFT of the re-ordered input.
150     *                  Step3: Taking the real part of the product of FFT output and weights.
151     * (c) Post-processing - DCT4 can be obtained from DCT2 output using the following equation:
152     *                   Y4(k) = Y2(k) - Y4(k-1) and Y4(-1) = Y4(0)
153     *                        where,
154     *                           Y4 -- DCT4 output,   Y2 -- DCT2 output
155     * (d) Multiplying the output with the normalizing factor sqrt(2/N).
156     */
157  
158    /*-------- Pre-processing ------------*/
159    /* Multiplying input with cos factor i.e. r(n) = 2 * x(n) * cos(pi*(2*n+1)/(4*n)) */
160    arm_scale_f32(pInlineBuffer, 2.0f, pInlineBuffer, S->N);
161    arm_mult_f32(pInlineBuffer, cosFact, pInlineBuffer, S->N);
162  
163    /* ----------------------------------------------------------------
164     * Step1: Re-ordering of even and odd elements as
165     *             pState[i] =  pInlineBuffer[2*i] and
166     *             pState[N-i-1] = pInlineBuffer[2*i+1] where i = 0 to N/2
167     ---------------------------------------------------------------------*/
168  
169    /* pS1 initialized to pState */
170    pS1 = pState;
171  
172    /* pS2 initialized to pState+N-1, so that it points to the end of the state buffer */
173    pS2 = pState + (S->N - 1U);
174  
175    /* pbuff initialized to input buffer */
176    pbuff = pInlineBuffer;
177  
178  
179  #if defined (ARM_MATH_LOOPUNROLL)
180  
181    /* Initializing the loop counter to N/2 >> 2 for loop unrolling by 4 */
182    i = S->Nby2 >> 2U;
183  
184    /* First part of the processing with loop unrolling.  Compute 4 outputs at a time.
185     ** a second loop below computes the remaining 1 to 3 samples. */
186    do
187    {
188      /* Re-ordering of even and odd elements */
189      /* pState[i] =  pInlineBuffer[2*i] */
190      *pS1++ = *pbuff++;
191      /* pState[N-i-1] = pInlineBuffer[2*i+1] */
192      *pS2-- = *pbuff++;
193  
194      *pS1++ = *pbuff++;
195      *pS2-- = *pbuff++;
196  
197      *pS1++ = *pbuff++;
198      *pS2-- = *pbuff++;
199  
200      *pS1++ = *pbuff++;
201      *pS2-- = *pbuff++;
202  
203      /* Decrement loop counter */
204      i--;
205    } while (i > 0U);
206  
207    /* pbuff initialized to input buffer */
208    pbuff = pInlineBuffer;
209  
210    /* pS1 initialized to pState */
211    pS1 = pState;
212  
213    /* Initializing the loop counter to N/4 instead of N for loop unrolling */
214    i = S->N >> 2U;
215  
216    /* Processing with loop unrolling 4 times as N is always multiple of 4.
217     * Compute 4 outputs at a time */
218    do
219    {
220      /* Writing the re-ordered output back to inplace input buffer */
221      *pbuff++ = *pS1++;
222      *pbuff++ = *pS1++;
223      *pbuff++ = *pS1++;
224      *pbuff++ = *pS1++;
225  
226      /* Decrement the loop counter */
227      i--;
228    } while (i > 0U);
229  
230  
231    /* ---------------------------------------------------------
232     *     Step2: Calculate RFFT for N-point input
233     * ---------------------------------------------------------- */
234    /* pInlineBuffer is real input of length N , pState is the complex output of length 2N */
235    arm_rfft_f32 (S->pRfft, pInlineBuffer, pState);
236  
237    /*----------------------------------------------------------------------
238     *  Step3: Multiply the FFT output with the weights.
239     *----------------------------------------------------------------------*/
240    arm_cmplx_mult_cmplx_f32 (pState, weights, pState, S->N);
241  
242    /* ----------- Post-processing ---------- */
243    /* DCT-IV can be obtained from DCT-II by the equation,
244     *       Y4(k) = Y2(k) - Y4(k-1) and Y4(-1) = Y4(0)
245     *       Hence, Y4(0) = Y2(0)/2  */
246    /* Getting only real part from the output and Converting to DCT-IV */
247  
248    /* Initializing the loop counter to N >> 2 for loop unrolling by 4 */
249    i = (S->N - 1U) >> 2U;
250  
251    /* pbuff initialized to input buffer. */
252    pbuff = pInlineBuffer;
253  
254    /* pS1 initialized to pState */
255    pS1 = pState;
256  
257    /* Calculating Y4(0) from Y2(0) using Y4(0) = Y2(0)/2 */
258    in = *pS1++ * (float32_t) 0.5;
259    /* input buffer acts as inplace, so output values are stored in the input itself. */
260    *pbuff++ = in;
261  
262    /* pState pointer is incremented twice as the real values are located alternatively in the array */
263    pS1++;
264  
265    /* First part of the processing with loop unrolling.  Compute 4 outputs at a time.
266     ** a second loop below computes the remaining 1 to 3 samples. */
267    do
268    {
269      /* Calculating Y4(1) to Y4(N-1) from Y2 using equation Y4(k) = Y2(k) - Y4(k-1) */
270      /* pState pointer (pS1) is incremented twice as the real values are located alternatively in the array */
271      in = *pS1++ - in;
272      *pbuff++ = in;
273      /* points to the next real value */
274      pS1++;
275  
276      in = *pS1++ - in;
277      *pbuff++ = in;
278      pS1++;
279  
280      in = *pS1++ - in;
281      *pbuff++ = in;
282      pS1++;
283  
284      in = *pS1++ - in;
285      *pbuff++ = in;
286      pS1++;
287  
288      /* Decrement the loop counter */
289      i--;
290    } while (i > 0U);
291  
292    /* If the blockSize is not a multiple of 4, compute any remaining output samples here.
293     ** No loop unrolling is used. */
294    i = (S->N - 1U) % 0x4U;
295  
296    while (i > 0U)
297    {
298      /* Calculating Y4(1) to Y4(N-1) from Y2 using equation Y4(k) = Y2(k) - Y4(k-1) */
299      /* pState pointer (pS1) is incremented twice as the real values are located alternatively in the array */
300      in = *pS1++ - in;
301      *pbuff++ = in;
302  
303      /* points to the next real value */
304      pS1++;
305  
306      /* Decrement the loop counter */
307      i--;
308    }
309  
310  
311    /*------------ Normalizing the output by multiplying with the normalizing factor ----------*/
312  
313    /* Initializing the loop counter to N/4 instead of N for loop unrolling */
314    i = S->N >> 2U;
315  
316    /* pbuff initialized to the pInlineBuffer(now contains the output values) */
317    pbuff = pInlineBuffer;
318  
319    /* Processing with loop unrolling 4 times as N is always multiple of 4.  Compute 4 outputs at a time */
320    do
321    {
322      /* Multiplying pInlineBuffer with the normalizing factor sqrt(2/N) */
323      in = *pbuff;
324      *pbuff++ = in * S->normalize;
325  
326      in = *pbuff;
327      *pbuff++ = in * S->normalize;
328  
329      in = *pbuff;
330      *pbuff++ = in * S->normalize;
331  
332      in = *pbuff;
333      *pbuff++ = in * S->normalize;
334  
335      /* Decrement the loop counter */
336      i--;
337    } while (i > 0U);
338  
339  
340  #else
341  
342    /* Initializing the loop counter to N/2 */
343    i = S->Nby2;
344  
345    do
346    {
347      /* Re-ordering of even and odd elements */
348      /* pState[i] =  pInlineBuffer[2*i] */
349      *pS1++ = *pbuff++;
350      /* pState[N-i-1] = pInlineBuffer[2*i+1] */
351      *pS2-- = *pbuff++;
352  
353      /* Decrement the loop counter */
354      i--;
355    } while (i > 0U);
356  
357    /* pbuff initialized to input buffer */
358    pbuff = pInlineBuffer;
359  
360    /* pS1 initialized to pState */
361    pS1 = pState;
362  
363    /* Initializing the loop counter */
364    i = S->N;
365  
366    do
367    {
368      /* Writing the re-ordered output back to inplace input buffer */
369      *pbuff++ = *pS1++;
370  
371      /* Decrement the loop counter */
372      i--;
373    } while (i > 0U);
374  
375  
376    /* ---------------------------------------------------------
377     *     Step2: Calculate RFFT for N-point input
378     * ---------------------------------------------------------- */
379    /* pInlineBuffer is real input of length N , pState is the complex output of length 2N */
380    arm_rfft_f32 (S->pRfft, pInlineBuffer, pState);
381  
382    /*----------------------------------------------------------------------
383     *  Step3: Multiply the FFT output with the weights.
384     *----------------------------------------------------------------------*/
385    arm_cmplx_mult_cmplx_f32 (pState, weights, pState, S->N);
386  
387    /* ----------- Post-processing ---------- */
388    /* DCT-IV can be obtained from DCT-II by the equation,
389     *       Y4(k) = Y2(k) - Y4(k-1) and Y4(-1) = Y4(0)
390     *       Hence, Y4(0) = Y2(0)/2  */
391    /* Getting only real part from the output and Converting to DCT-IV */
392  
393    /* pbuff initialized to input buffer. */
394    pbuff = pInlineBuffer;
395  
396    /* pS1 initialized to pState */
397    pS1 = pState;
398  
399    /* Calculating Y4(0) from Y2(0) using Y4(0) = Y2(0)/2 */
400    in = *pS1++ * (float32_t) 0.5;
401    /* input buffer acts as inplace, so output values are stored in the input itself. */
402    *pbuff++ = in;
403  
404    /* pState pointer is incremented twice as the real values are located alternatively in the array */
405    pS1++;
406  
407    /* Initializing the loop counter */
408    i = (S->N - 1U);
409  
410    do
411    {
412      /* Calculating Y4(1) to Y4(N-1) from Y2 using equation Y4(k) = Y2(k) - Y4(k-1) */
413      /* pState pointer (pS1) is incremented twice as the real values are located alternatively in the array */
414      in = *pS1++ - in;
415      *pbuff++ = in;
416  
417      /* points to the next real value */
418      pS1++;
419  
420      /* Decrement loop counter */
421      i--;
422    } while (i > 0U);
423  
424    /*------------ Normalizing the output by multiplying with the normalizing factor ----------*/
425  
426    /* Initializing loop counter */
427    i = S->N;
428  
429    /* pbuff initialized to the pInlineBuffer (now contains the output values) */
430    pbuff = pInlineBuffer;
431  
432    do
433    {
434      /* Multiplying pInlineBuffer with the normalizing factor sqrt(2/N) */
435      in = *pbuff;
436      *pbuff++ = in * S->normalize;
437  
438      /* Decrement loop counter */
439      i--;
440    } while (i > 0U);
441  
442  #endif /* #if defined (ARM_MATH_LOOPUNROLL) */
443  
444  }
445  
446  /**
447    @} end of DCT4_IDCT4 group
448   */