Fork of mbed-dsp. CMSIS-DSP library of supporting NEON

Dependents:   mbed-os-example-cmsis_dsp_neon

Fork of mbed-dsp by mbed official

Information

Japanese version is available in lower part of this page.
このページの後半に日本語版が用意されています.

CMSIS-DSP of supporting NEON

What is this ?

A library for CMSIS-DSP of supporting NEON.
We supported the NEON to CMSIS-DSP Ver1.4.3(CMSIS V4.1) that ARM supplied, has achieved the processing speed improvement.
If you use the mbed-dsp library, you can use to replace this library.
CMSIS-DSP of supporting NEON is provied as a library.

Library Creation environment

CMSIS-DSP library of supporting NEON was created by the following environment.

  • Compiler
    ARMCC Version 5.03
  • Compile option switch[C Compiler]
   -DARM_MATH_MATRIX_CHECK -DARM_MATH_ROUNDING -O3 -Otime --cpu=Cortex-A9 --littleend --arm 
   --apcs=/interwork --no_unaligned_access --fpu=vfpv3_fp16 --fpmode=fast --apcs=/hardfp 
   --vectorize --asm
  • Compile option switch[Assembler]
   --cpreproc --cpu=Cortex-A9 --littleend --arm --apcs=/interwork --no_unaligned_access 
   --fpu=vfpv3_fp16 --fpmode=fast --apcs=/hardfp


Effects of NEON support

In the data which passes to each function, large size will be expected more effective than small size.
Also if the data is a multiple of 16, effect will be expected in every function in the CMSIS-DSP.


NEON対応CMSIS-DSP

概要

NEON対応したCMSIS-DSPのライブラリです。
ARM社提供のCMSIS-DSP Ver1.4.3(CMSIS V4.1)をターゲットにNEON対応を行ない、処理速度向上を実現しております。
mbed-dspライブラリを使用している場合は、本ライブラリに置き換えて使用することができます。
NEON対応したCMSIS-DSPはライブラリで提供します。

ライブラリ作成環境

NEON対応CMSIS-DSPライブラリは、以下の環境で作成しています。

  • コンパイラ
    ARMCC Version 5.03
  • コンパイルオプションスイッチ[C Compiler]
   -DARM_MATH_MATRIX_CHECK -DARM_MATH_ROUNDING -O3 -Otime --cpu=Cortex-A9 --littleend --arm 
   --apcs=/interwork --no_unaligned_access --fpu=vfpv3_fp16 --fpmode=fast --apcs=/hardfp 
   --vectorize --asm
  • コンパイルオプションスイッチ[Assembler]
   --cpreproc --cpu=Cortex-A9 --littleend --arm --apcs=/interwork --no_unaligned_access 
   --fpu=vfpv3_fp16 --fpmode=fast --apcs=/hardfp


NEON対応による効果について

CMSIS-DSP内の各関数へ渡すデータは、小さいサイズよりも大きいサイズの方が効果が見込めます。
また、16の倍数のデータであれば、CMSIS-DSP内のどの関数でも効果が見込めます。


Revision:
1:fdd22bb7aa52
Child:
2:da51fb522205
--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/cmsis_dsp/TransformFunctions/arm_dct4_f32.c	Wed Nov 28 12:30:09 2012 +0000
@@ -0,0 +1,453 @@
+/* ----------------------------------------------------------------------    
+* Copyright (C) 2010 ARM Limited. All rights reserved.    
+*    
+* $Date:        15. February 2012  
+* $Revision:     V1.1.0  
+*    
+* Project:         CMSIS DSP Library    
+* Title:        arm_dct4_f32.c    
+*    
+* Description:    Processing function of DCT4 & IDCT4 F32.    
+*    
+* Target Processor: Cortex-M4/Cortex-M3/Cortex-M0
+*  
+* Version 1.1.0 2012/02/15 
+*    Updated with more optimizations, bug fixes and minor API changes.  
+*   
+* Version 1.0.10 2011/7/15  
+*    Big Endian support added and Merged M0 and M3/M4 Source code.   
+*    
+* Version 1.0.3 2010/11/29   
+*    Re-organized the CMSIS folders and updated documentation.    
+*     
+* Version 1.0.2 2010/11/11    
+*    Documentation updated.     
+*    
+* Version 1.0.1 2010/10/05     
+*    Production release and review comments incorporated.    
+*    
+* Version 1.0.0 2010/09/20     
+*    Production release and review comments incorporated.    
+* -------------------------------------------------------------------- */
+
+#include "arm_math.h"
+
+/**    
+ * @ingroup groupTransforms    
+ */
+
+/**    
+ * @defgroup DCT4_IDCT4 DCT Type IV Functions    
+ * Representation of signals by minimum number of values is important for storage and transmission.    
+ * The possibility of large discontinuity between the beginning and end of a period of a signal    
+ * in DFT can be avoided by extending the signal so that it is even-symmetric.    
+ * Discrete Cosine Transform (DCT) is constructed such that its energy is heavily concentrated in the lower part of the    
+ * spectrum and is very widely used in signal and image coding applications.    
+ * The family of DCTs (DCT type- 1,2,3,4) is the outcome of different combinations of homogeneous boundary conditions.    
+ * DCT has an excellent energy-packing capability, hence has many applications and in data compression in particular.    
+ *    
+ * DCT is essentially the Discrete Fourier Transform(DFT) of an even-extended real signal.    
+ * Reordering of the input data makes the computation of DCT just a problem of    
+ * computing the DFT of a real signal with a few additional operations.    
+ * This approach provides regular, simple, and very efficient DCT algorithms for practical hardware and software implementations.    
+ *     
+ * 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.    
+ * DCT4 is implemented using DCT2 as their implementations are similar except with some added pre-processing and post-processing.    
+ * DCT2 implementation can be described in the following steps:    
+ * - Re-ordering input    
+ * - Calculating Real FFT    
+ * - Multiplication of weights and Real FFT output and getting real part from the product.    
+ *    
+ * This process is explained by the block diagram below:    
+ * \image html DCT4.gif "Discrete Cosine Transform - type-IV"    
+ *    
+ * \par Algorithm:    
+ * The N-point type-IV DCT is defined as a real, linear transformation by the formula:    
+ * \image html DCT4Equation.gif    
+ * where <code>k = 0,1,2,.....N-1</code>    
+ *\par    
+ * Its inverse is defined as follows:    
+ * \image html IDCT4Equation.gif    
+ * where <code>n = 0,1,2,.....N-1</code>    
+ *\par    
+ * The DCT4 matrices become involutory (i.e. they are self-inverse) by multiplying with an overall scale factor of sqrt(2/N).    
+ * The symmetry of the transform matrix indicates that the fast algorithms for the forward    
+ * and inverse transform computation are identical.    
+ * Note that the implementation of Inverse DCT4 and DCT4 is same, hence same process function can be used for both.    
+ *    
+ * \par Lengths supported by the transform:    
+ *  As DCT4 internally uses Real FFT, it supports all the lengths supported by arm_rfft_f32().    
+ * The library provides separate functions for Q15, Q31, and floating-point data types.    
+ * \par Instance Structure    
+ * The instances for Real FFT and FFT, cosine values table and twiddle factor table are stored in an instance data structure.    
+ * A separate instance structure must be defined for each transform.    
+ * There are separate instance structure declarations for each of the 3 supported data types.    
+ *    
+ * \par Initialization Functions    
+ * There is also an associated initialization function for each data type.    
+ * The initialization function performs the following operations:    
+ * - Sets the values of the internal structure fields.    
+ * - Initializes Real FFT as its process function is used internally in DCT4, by calling arm_rfft_init_f32().    
+ * \par    
+ * Use of the initialization function is optional.    
+ * However, if the initialization function is used, then the instance structure cannot be placed into a const data section.    
+ * To place an instance structure into a const data section, the instance structure must be manually initialized.    
+ * Manually initialize the instance structure as follows:    
+ * <pre>    
+ *arm_dct4_instance_f32 S = {N, Nby2, normalize, pTwiddle, pCosFactor, pRfft, pCfft};    
+ *arm_dct4_instance_q31 S = {N, Nby2, normalize, pTwiddle, pCosFactor, pRfft, pCfft};   
+ *arm_dct4_instance_q15 S = {N, Nby2, normalize, pTwiddle, pCosFactor, pRfft, pCfft};   
+ * </pre>   
+ * where \c N is the length of the DCT4; \c Nby2 is half of the length of the DCT4;   
+ * \c normalize is normalizing factor used and is equal to <code>sqrt(2/N)</code>;    
+ * \c pTwiddle points to the twiddle factor table;   
+ * \c pCosFactor points to the cosFactor table;   
+ * \c pRfft points to the real FFT instance;   
+ * \c pCfft points to the complex FFT instance;   
+ * The CFFT and RFFT structures also needs to be initialized, refer to arm_cfft_radix4_f32()   
+ * and arm_rfft_f32() respectively for details regarding static initialization.   
+ *   
+ * \par Fixed-Point Behavior    
+ * Care must be taken when using the fixed-point versions of the DCT4 transform functions.    
+ * In particular, the overflow and saturation behavior of the accumulator used in each function must be considered.    
+ * Refer to the function specific documentation below for usage guidelines.    
+ */
+
+ /**    
+ * @addtogroup DCT4_IDCT4    
+ * @{    
+ */
+
+/**    
+ * @brief Processing function for the floating-point DCT4/IDCT4.   
+ * @param[in]       *S             points to an instance of the floating-point DCT4/IDCT4 structure.   
+ * @param[in]       *pState        points to state buffer.   
+ * @param[in,out]   *pInlineBuffer points to the in-place input and output buffer.   
+ * @return none.   
+ */
+
+void arm_dct4_f32(
+  const arm_dct4_instance_f32 * S,
+  float32_t * pState,
+  float32_t * pInlineBuffer)
+{
+  uint32_t i;                                    /* Loop counter */
+  float32_t *weights = S->pTwiddle;              /* Pointer to the Weights table */
+  float32_t *cosFact = S->pCosFactor;            /* Pointer to the cos factors table */
+  float32_t *pS1, *pS2, *pbuff;                  /* Temporary pointers for input buffer and pState buffer */
+  float32_t in;                                  /* Temporary variable */
+
+
+  /* DCT4 computation involves DCT2 (which is calculated using RFFT)    
+   * along with some pre-processing and post-processing.    
+   * Computational procedure is explained as follows:    
+   * (a) Pre-processing involves multiplying input with cos factor,    
+   *     r(n) = 2 * u(n) * cos(pi*(2*n+1)/(4*n))    
+   *              where,    
+   *                 r(n) -- output of preprocessing    
+   *                 u(n) -- input to preprocessing(actual Source buffer)    
+   * (b) Calculation of DCT2 using FFT is divided into three steps:    
+   *                  Step1: Re-ordering of even and odd elements of input.    
+   *                  Step2: Calculating FFT of the re-ordered input.    
+   *                  Step3: Taking the real part of the product of FFT output and weights.    
+   * (c) Post-processing - DCT4 can be obtained from DCT2 output using the following equation:    
+   *                   Y4(k) = Y2(k) - Y4(k-1) and Y4(-1) = Y4(0)    
+   *                        where,    
+   *                           Y4 -- DCT4 output,   Y2 -- DCT2 output    
+   * (d) Multiplying the output with the normalizing factor sqrt(2/N).    
+   */
+
+        /*-------- Pre-processing ------------*/
+  /* Multiplying input with cos factor i.e. r(n) = 2 * x(n) * cos(pi*(2*n+1)/(4*n)) */
+  arm_scale_f32(pInlineBuffer, 2.0f, pInlineBuffer, S->N);
+  arm_mult_f32(pInlineBuffer, cosFact, pInlineBuffer, S->N);
+
+  /* ----------------------------------------------------------------    
+   * Step1: Re-ordering of even and odd elements as,    
+   *             pState[i] =  pInlineBuffer[2*i] and    
+   *             pState[N-i-1] = pInlineBuffer[2*i+1] where i = 0 to N/2    
+   ---------------------------------------------------------------------*/
+
+  /* pS1 initialized to pState */
+  pS1 = pState;
+
+  /* pS2 initialized to pState+N-1, so that it points to the end of the state buffer */
+  pS2 = pState + (S->N - 1u);
+
+  /* pbuff initialized to input buffer */
+  pbuff = pInlineBuffer;
+
+#ifndef ARM_MATH_CM0
+
+  /* Run the below code for Cortex-M4 and Cortex-M3 */
+
+  /* Initializing the loop counter to N/2 >> 2 for loop unrolling by 4 */
+  i = (uint32_t) S->Nby2 >> 2u;
+
+  /* First part of the processing with loop unrolling.  Compute 4 outputs at a time.    
+   ** a second loop below computes the remaining 1 to 3 samples. */
+  do
+  {
+    /* Re-ordering of even and odd elements */
+    /* pState[i] =  pInlineBuffer[2*i] */
+    *pS1++ = *pbuff++;
+    /* pState[N-i-1] = pInlineBuffer[2*i+1] */
+    *pS2-- = *pbuff++;
+
+    *pS1++ = *pbuff++;
+    *pS2-- = *pbuff++;
+
+    *pS1++ = *pbuff++;
+    *pS2-- = *pbuff++;
+
+    *pS1++ = *pbuff++;
+    *pS2-- = *pbuff++;
+
+    /* Decrement the loop counter */
+    i--;
+  } while(i > 0u);
+
+  /* pbuff initialized to input buffer */
+  pbuff = pInlineBuffer;
+
+  /* pS1 initialized to pState */
+  pS1 = pState;
+
+  /* Initializing the loop counter to N/4 instead of N for loop unrolling */
+  i = (uint32_t) S->N >> 2u;
+
+  /* Processing with loop unrolling 4 times as N is always multiple of 4.    
+   * Compute 4 outputs at a time */
+  do
+  {
+    /* Writing the re-ordered output back to inplace input buffer */
+    *pbuff++ = *pS1++;
+    *pbuff++ = *pS1++;
+    *pbuff++ = *pS1++;
+    *pbuff++ = *pS1++;
+
+    /* Decrement the loop counter */
+    i--;
+  } while(i > 0u);
+
+
+  /* ---------------------------------------------------------    
+   *     Step2: Calculate RFFT for N-point input    
+   * ---------------------------------------------------------- */
+  /* pInlineBuffer is real input of length N , pState is the complex output of length 2N */
+  arm_rfft_f32(S->pRfft, pInlineBuffer, pState);
+
+        /*----------------------------------------------------------------------    
+     *  Step3: Multiply the FFT output with the weights.    
+     *----------------------------------------------------------------------*/
+  arm_cmplx_mult_cmplx_f32(pState, weights, pState, S->N);
+
+  /* ----------- Post-processing ---------- */
+  /* DCT-IV can be obtained from DCT-II by the equation,    
+   *       Y4(k) = Y2(k) - Y4(k-1) and Y4(-1) = Y4(0)    
+   *       Hence, Y4(0) = Y2(0)/2  */
+  /* Getting only real part from the output and Converting to DCT-IV */
+
+  /* Initializing the loop counter to N >> 2 for loop unrolling by 4 */
+  i = ((uint32_t) S->N - 1u) >> 2u;
+
+  /* pbuff initialized to input buffer. */
+  pbuff = pInlineBuffer;
+
+  /* pS1 initialized to pState */
+  pS1 = pState;
+
+  /* Calculating Y4(0) from Y2(0) using Y4(0) = Y2(0)/2 */
+  in = *pS1++ * (float32_t) 0.5;
+  /* input buffer acts as inplace, so output values are stored in the input itself. */
+  *pbuff++ = in;
+
+  /* pState pointer is incremented twice as the real values are located alternatively in the array */
+  pS1++;
+
+  /* First part of the processing with loop unrolling.  Compute 4 outputs at a time.    
+   ** a second loop below computes the remaining 1 to 3 samples. */
+  do
+  {
+    /* Calculating Y4(1) to Y4(N-1) from Y2 using equation Y4(k) = Y2(k) - Y4(k-1) */
+    /* pState pointer (pS1) is incremented twice as the real values are located alternatively in the array */
+    in = *pS1++ - in;
+    *pbuff++ = in;
+    /* points to the next real value */
+    pS1++;
+
+    in = *pS1++ - in;
+    *pbuff++ = in;
+    pS1++;
+
+    in = *pS1++ - in;
+    *pbuff++ = in;
+    pS1++;
+
+    in = *pS1++ - in;
+    *pbuff++ = in;
+    pS1++;
+
+    /* Decrement the loop counter */
+    i--;
+  } while(i > 0u);
+
+  /* If the blockSize is not a multiple of 4, compute any remaining output samples here.    
+   ** No loop unrolling is used. */
+  i = ((uint32_t) S->N - 1u) % 0x4u;
+
+  while(i > 0u)
+  {
+    /* Calculating Y4(1) to Y4(N-1) from Y2 using equation Y4(k) = Y2(k) - Y4(k-1) */
+    /* pState pointer (pS1) is incremented twice as the real values are located alternatively in the array */
+    in = *pS1++ - in;
+    *pbuff++ = in;
+    /* points to the next real value */
+    pS1++;
+
+    /* Decrement the loop counter */
+    i--;
+  }
+
+
+        /*------------ Normalizing the output by multiplying with the normalizing factor ----------*/
+
+  /* Initializing the loop counter to N/4 instead of N for loop unrolling */
+  i = (uint32_t) S->N >> 2u;
+
+  /* pbuff initialized to the pInlineBuffer(now contains the output values) */
+  pbuff = pInlineBuffer;
+
+  /* Processing with loop unrolling 4 times as N is always multiple of 4.  Compute 4 outputs at a time */
+  do
+  {
+    /* Multiplying pInlineBuffer with the normalizing factor sqrt(2/N) */
+    in = *pbuff;
+    *pbuff++ = in * S->normalize;
+
+    in = *pbuff;
+    *pbuff++ = in * S->normalize;
+
+    in = *pbuff;
+    *pbuff++ = in * S->normalize;
+
+    in = *pbuff;
+    *pbuff++ = in * S->normalize;
+
+    /* Decrement the loop counter */
+    i--;
+  } while(i > 0u);
+
+
+#else
+
+  /* Run the below code for Cortex-M0 */
+
+  /* Initializing the loop counter to N/2 */
+  i = (uint32_t) S->Nby2;
+
+  do
+  {
+    /* Re-ordering of even and odd elements */
+    /* pState[i] =  pInlineBuffer[2*i] */
+    *pS1++ = *pbuff++;
+    /* pState[N-i-1] = pInlineBuffer[2*i+1] */
+    *pS2-- = *pbuff++;
+
+    /* Decrement the loop counter */
+    i--;
+  } while(i > 0u);
+
+  /* pbuff initialized to input buffer */
+  pbuff = pInlineBuffer;
+
+  /* pS1 initialized to pState */
+  pS1 = pState;
+
+  /* Initializing the loop counter */
+  i = (uint32_t) S->N;
+
+  do
+  {
+    /* Writing the re-ordered output back to inplace input buffer */
+    *pbuff++ = *pS1++;
+
+    /* Decrement the loop counter */
+    i--;
+  } while(i > 0u);
+
+
+  /* ---------------------------------------------------------    
+   *     Step2: Calculate RFFT for N-point input    
+   * ---------------------------------------------------------- */
+  /* pInlineBuffer is real input of length N , pState is the complex output of length 2N */
+  arm_rfft_f32(S->pRfft, pInlineBuffer, pState);
+
+        /*----------------------------------------------------------------------    
+     *  Step3: Multiply the FFT output with the weights.    
+     *----------------------------------------------------------------------*/
+  arm_cmplx_mult_cmplx_f32(pState, weights, pState, S->N);
+
+  /* ----------- Post-processing ---------- */
+  /* DCT-IV can be obtained from DCT-II by the equation,    
+   *       Y4(k) = Y2(k) - Y4(k-1) and Y4(-1) = Y4(0)    
+   *       Hence, Y4(0) = Y2(0)/2  */
+  /* Getting only real part from the output and Converting to DCT-IV */
+
+  /* pbuff initialized to input buffer. */
+  pbuff = pInlineBuffer;
+
+  /* pS1 initialized to pState */
+  pS1 = pState;
+
+  /* Calculating Y4(0) from Y2(0) using Y4(0) = Y2(0)/2 */
+  in = *pS1++ * (float32_t) 0.5;
+  /* input buffer acts as inplace, so output values are stored in the input itself. */
+  *pbuff++ = in;
+
+  /* pState pointer is incremented twice as the real values are located alternatively in the array */
+  pS1++;
+
+  /* Initializing the loop counter */
+  i = ((uint32_t) S->N - 1u);
+
+  do
+  {
+    /* Calculating Y4(1) to Y4(N-1) from Y2 using equation Y4(k) = Y2(k) - Y4(k-1) */
+    /* pState pointer (pS1) is incremented twice as the real values are located alternatively in the array */
+    in = *pS1++ - in;
+    *pbuff++ = in;
+    /* points to the next real value */
+    pS1++;
+
+
+    /* Decrement the loop counter */
+    i--;
+  } while(i > 0u);
+
+
+        /*------------ Normalizing the output by multiplying with the normalizing factor ----------*/
+
+  /* Initializing the loop counter */
+  i = (uint32_t) S->N;
+
+  /* pbuff initialized to the pInlineBuffer(now contains the output values) */
+  pbuff = pInlineBuffer;
+
+  do
+  {
+    /* Multiplying pInlineBuffer with the normalizing factor sqrt(2/N) */
+    in = *pbuff;
+    *pbuff++ = in * S->normalize;
+
+    /* Decrement the loop counter */
+    i--;
+  } while(i > 0u);
+
+#endif /* #ifndef ARM_MATH_CM0 */
+
+}
+
+/**    
+   * @} end of DCT4_IDCT4 group    
+   */