C++ ATL_MulBySize函数代码示例

static void geinv
   (const enum CBLAS_ORDER Order, const int N, TYPE *A, const int lda)
{
   int *ipiv;
   TYPE *wrk;
   int lwrk;

   ipiv = malloc(sizeof(int)*N);
   ATL_assert(ipiv);
   #ifdef TimeF77
      lwrk = N * Mjoin(PATL,GetNB)();
      wrk = malloc(ATL_MulBySize(lwrk));
      if (Order == AtlasRowMajor) Mjoin(PATL,tstsqtran)(N, A, lda);
      ATL_assert(Mjoin(PATL,f77getrf)(AtlasColMajor, N, N, A, lda, ipiv) == 0);
      ATL_assert(Mjoin(PATL,f77getri)
         (AtlasColMajor, N, A, lda, ipiv, wrk, &lwrk) == 0);
      if (Order == AtlasRowMajor) Mjoin(PATL,tstsqtran)(N, A, lda);
      free(wrk);
   #elif defined(TimeC)
      ATL_assert(Mjoin(CLP,getrf)(Order, N, N, A, lda, ipiv) == 0);
      ATL_assert(Mjoin(CLP,getri)(Order, N, A, lda, ipiv) == 0);
  #else
      lwrk = N * Mjoin(PATL,GetNB)();
      wrk = malloc(ATL_MulBySize(lwrk));
      ATL_assert(Mjoin(PATL,getrf)(Order, N, N, A, lda, ipiv) == 0);
      ATL_assert(Mjoin(PATL,getri)(Order, N, A, lda, ipiv, wrk, &lwrk) == 0);
      free(wrk);
   #endif
   free(ipiv);
}

static TYPE *DupMat(enum ATLAS_ORDER Order, int M, int N, TYPE *A, int lda,
                    int ldc)
/*
 * returns a duplicate of the A matrix, with new leading dimension
 */
{
   int i, j, M2;
   const int ldc2 = (ldc SHIFT), lda2 = (lda SHIFT);
   TYPE *C;
   if (Order == CblasRowMajor)
   {
      i = M;
      M = N;
      N = i;
   }
   M2 = M SHIFT;
   ATL_assert(ldc >= M);
   C = malloc(ATL_MulBySize(ldc)*N);
   ATL_assert(C);
   #if defined(ATL_USEPTHREADS) && !defined(ATL_NONUMATOUCH)
      ATL_NumaTouchSpread(ATL_MulBySize(ldc)*N, C);
   #endif
   for (j=0; j != N; j++)
   {
      for (i=0; i != M2; i++) C[i] = A[i];
      C += ldc2;
      A += lda2;
   }
   return(C-N*ldc2);
}

void cblas_cher2(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
                 const int N, const void *alpha,
                 const void *X, const int incX,
                 const void *Y, const int incY, void *A, const int lda)
{
   int info = 2000;
   void *vx, *vy;
   float *x0, *y0;
   const float *x=X, *y=Y, *alp=alpha;
   const float one[2]={ATL_rone, ATL_rzero};

#ifndef NoCblasErrorChecks
   if (Order != CblasColMajor && Order != CblasRowMajor)
      info = cblas_errprn(1, info, "Order must be %d or %d, but is set to %d",
                          CblasRowMajor, CblasColMajor, Order);
   if (Uplo != CblasUpper && Uplo != CblasLower)
      info = cblas_errprn(2, info, "UPLO must be %d or %d, but is set to %d",
                          CblasUpper, CblasLower, Uplo);
   if (N < 0) info = cblas_errprn(3, info,
                        "N cannot be less than zero; is set to %d.", N);
   if (!incX) info = cblas_errprn(6, info,
                                  "incX cannot be zero; is set to %d.", incX);
   if (!incY) info = cblas_errprn(8, info,
                                  "incY cannot be zero; is set to %d.", incY);
   if (lda < N || lda < 1)
      info = cblas_errprn(10, info, "lda must be >= MAX(N,1): lda=%d N=%d",
                          lda, N);
   if (info != 2000)
   {
      cblas_xerbla(info, "cblas_cher2", "");
      return;
   }
#endif

   if (incX < 0) x += (1-N)*incX<<1;
   if (incY < 0) y += (1-N)*incY<<1;

   if (Order == CblasColMajor)
      ATL_cher2(Uplo, N, alpha, x, incX, y, incY, A, lda);
   else if (alp[0] != ATL_rzero || alp[1] != ATL_rzero)
   {
      vx = malloc(ATL_Cachelen + ATL_MulBySize(N));
      vy = malloc(ATL_Cachelen + ATL_MulBySize(N));
      ATL_assert(vx != NULL && vy != NULL);
      x0 = ATL_AlignPtr(vx);
      y0 = ATL_AlignPtr(vy);
      ATL_cmoveConj(N, alpha, y, incY, y0, 1);
      ATL_ccopyConj(N, x, incX, x0, 1);
      ATL_cher2(( (Uplo == CblasUpper) ? CblasLower : CblasUpper ),
                N, one, y0, 1, x0, 1, A, lda);
      free(vx);
      free(vy);
   }
   else ATL_cher2(( (Uplo == CblasUpper) ? CblasLower : CblasUpper ),
                  N, alpha, y, incY, x, incX, A, lda);
}

static TYPE uumtest(enum ATLAS_ORDER Order, enum ATLAS_UPLO Uplo,
                    int CacheSize, int N, int lda, double *tim)
{
   TYPE *A, *Ag, *LmLt;
   double t0, t1;
   TYPE normA, eps, resid;
   enum ATLAS_UPLO MyUplo = Uplo;

   if (Order == CblasRowMajor)
   {
      if (Uplo == CblasUpper) MyUplo = CblasLower;
      else MyUplo = CblasUpper;
   }
   eps = Mjoin(PATL,epsilon)();
   A = malloc(ATL_MulBySize(lda)*N + ATL_MulBySize(N)*N);
   if (A == NULL) return(-1);
   Ag = A + lda*(N SHIFT);
   t0 = ATL_flushcache(CacheSize);
   lltgen(MyUplo, N, A, lda, N*1029+lda);
   lltgen(MyUplo, N, Ag, N, N*1029+lda);
   normA = lltnrm1(MyUplo, N, A, lda);
   #ifdef DEBUG
      Mjoin(PATL,geprint)("A", N, N, A, lda);
      Mjoin(PATL,geprint)("Ag", N, N, Ag, N);
   #endif

   t0 = ATL_flushcache(-1);

   t0 = time00();
   test_lauum(Order, Uplo, N, A, lda);
   t1 = time00() - t0;
   *tim = t1;

   t0 = ATL_flushcache(0);

   ATL_checkpad(MyUplo, N, A, lda);
   if (Uplo == CblasUpper) LmLt = ATL_UmulUt(Order, N, Ag, N);
   else LmLt = ATL_LtmulL(Order, N, Ag, N);
   #ifdef DEBUG
      Mjoin(PATL,geprint)("A", N, N, A, lda);
      Mjoin(PATL,geprint)("Ag", N, N, LmLt, N);
   #endif
   lltdiff(MyUplo, N, A, lda, LmLt, N);
   #ifdef DEBUG
      Mjoin(PATL,geprint)("A-L*Lt", N, N, LmLt, N);
   #endif
   resid = lltnrm1(MyUplo, N, LmLt, N) / (normA * eps * N);
   if (resid > 10.0 || resid != resid)
      fprintf(stderr, "normA=%e, eps=%e, num=%e\n", normA, eps, resid);

   free(LmLt);
   free(A);

   return(resid);
}

TYPE *GetGE(int M, int N, int lda)
{
   TYPE *A;
   A = malloc(ATL_MulBySize(lda)*N);
   if (A)
   {
      #if defined(ATL_USEPTHREADS) && !defined(ATL_NONUMATOUCH)
         ATL_NumaTouchSpread(ATL_MulBySize(lda)*N, A);
      #endif
      Mjoin(PATL,gegen)(M, N, A, lda, M*N+lda);
   }
   return(A);
}

void Mjoin(Mjoin(Mjoin(PATL,syrk),UploNM),T)
   (const int N, const int K, const void *valpha, const void *A, const int lda,
    const void *vbeta, void *C, const int ldc)
{
   void *vc;
   TYPE *c;
   #ifdef TREAL
      const SCALAR alpha=*( (const SCALAR *)valpha );
      const SCALAR beta =*( (const SCALAR *)vbeta  );
      const SCALAR one=1.0, zero=0.0;
   #else
      #define alpha valpha
      const TYPE *beta=vbeta;
      const TYPE one[2]={1.0,0.0}, zero[2]={0.0,0.0};
   #endif

   if (K > SYRK_Xover)
   {
      vc = malloc(ATL_Cachelen+ATL_MulBySize(N)*N);
      ATL_assert(vc);
      c = ATL_AlignPtr(vc);
      CgemmTN(N, N, K, alpha, A, lda, A, lda, zero, c, N);
      if ( SCALAR_IS_ONE(beta) ) Mjoin(syr_put,_b1)(N, c, beta, C, ldc);
      else if ( SCALAR_IS_ZERO(beta) ) Mjoin(syr_put,_b0)(N, c, beta, C, ldc);
      #ifdef TCPLX
         else if ( SCALAR_IS_NONE(beta) )
            Mjoin(syr_put,_bn1)(N, c, beta, C, ldc);
         else if (beta[1] == *zero) Mjoin(syr_put,_bXi0)(N, c, beta, C, ldc);
      #endif
      else Mjoin(syr_put,_bX)(N, c, beta, C, ldc);
      free(vc);
   }
   else Mjoin(PATL,refsyrk)(Uplo_, AtlasTrans, N, K, alpha, A, lda,
                            beta, C, ldc);
}

void Mjoin(Mjoin(PATL,trmmL),ATLP)
   (const int M, const int N, const void *valpha, const void *A, const int lda,
    void *C, const int ldc)
{
   #ifdef TREAL
      const SCALAR alpha=*( (const SCALAR *)valpha );
      const SCALAR one=1.0, zero=0.0;
   #else
      const TYPE zero[2]={0.0,0.0};
      #define alpha valpha
   #endif
   void *va;
   TYPE *a;

   if (N > TRMM_Xover)
   {
      va = malloc(ATL_Cachelen + ATL_MulBySize(M)*M);
      ATL_assert(va);
      a = ATL_AlignPtr(va);
      #ifdef TREAL
         if ( SCALAR_IS_ONE(alpha) ) Mjoin(ATL_trcopy,_a1)(M, alpha, A, lda, a);
         else Mjoin(ATL_trcopy,_aX)(M, alpha, A, lda, a);
         CAgemmTN(M, N, M, one, a, M, C, ldc, zero, C, ldc);
      #else
         ATL_trcopy(M, A, lda, a);
         CAgemmTN(M, N, M, valpha, a, M, C, ldc, zero, C, ldc);
      #endif
      free(va);
   }
   else Mjoin(PATL,reftrmm)(AtlasLeft, Uplo_, Trans_, Unit_, M, N, alpha,
                            A, lda, C, ldc);
}

      int Mjoin(PATL,her2kLN)
   #endif
#endif
   (const int N, const int K, const void *valpha, const void *A, const int lda,
    const void *B, const int ldb, const void *vbeta, void *C, const int ldc)
{
   int i;
   void *vc=NULL;
   TYPE *c;
   const TYPE beta =*( (const TYPE *)vbeta  );
   const TYPE zero[2]={0.0, 0.0};

   i = ATL_MulBySize(N)*N;
   if (i <= ATL_MaxMalloc) vc = malloc(ATL_Cachelen+i);
   if (vc == NULL) return(1);
   c = ATL_AlignPtr(vc);
   #ifdef Transpose_
      ATL_ammm(AtlasConjTrans, AtlasNoTrans, N, N, K, valpha, A, lda, B, ldb,
   #else
      ATL_ammm(AtlasNoTrans, AtlasConjTrans, N, N, K, valpha, A, lda, B, ldb,
   #endif
               zero, c, N);
   if ( beta == 1.0 ) Mjoin(her2k_put,_b1)(N, c, vbeta, C, ldc);
   else if ( beta == 0.0 ) Mjoin(her2k_put,_b0)(N, c, vbeta, C, ldc);
   else Mjoin(her2k_put,_bXi0)(N, c, vbeta, C, ldc);
   free(vc);
   return(0);
}

void Mjoin(Mjoin(PATL,symmL),UploNM)
(const int M, const int N, const void *valpha, const void *A, const int lda,
 const void *B, const int ldb, const void *vbeta, void *C, const int ldc)
{
#ifdef TREAL
    const SCALAR alpha=*( (const SCALAR *)valpha );
    const SCALAR beta =*( (const SCALAR *)vbeta  );
    const SCALAR one=1.0;
#else
#define alpha valpha
#define beta vbeta
#endif
    TYPE *a;
    void *va;

    if (N > SYMM_Xover)
    {
        va = malloc(ATL_Cachelen + (ATL_MulBySize(M)*M));
        ATL_assert(va);
        a = ATL_AlignPtr(va);
#ifdef TREAL
        if ( SCALAR_IS_ONE(alpha) )
            Mjoin(Mjoin(Mjoin(PATL,sycopy),UploNM),_a1)(M, alpha, A, lda, a);
        else Mjoin(Mjoin(Mjoin(PATL,sycopy),UploNM),_aX)(M, alpha, A, lda, a);
        CgemmTN(M, N, M, one, a, M, B, ldb, beta, C, ldc);
#else
        Mjoin(Mjoin(PATL,sycopy),UploNM)(M, A, lda, a);
        CgemmTN(M, N, M, valpha, a, M, B, ldb, vbeta, C, ldc);
#endif
        free(va);
    }
    else Mjoin(PATL,refsymm)(AtlasLeft, Uplo_, M, N, alpha, A, lda, B, ldb,
                                 beta, C, ldc);
}

void Mjoin(Mjoin(PATL,symmR),UploNM)
   (const int M, const int N, const void *valpha, const void *A, const int lda,
    const void *B, const int ldb, const void *vbeta, void *C, const int ldc)
{
   #ifdef TREAL
      const SCALAR alpha=*( (const SCALAR *)valpha );
      const SCALAR beta =*( (const SCALAR *)vbeta  );
      const SCALAR one=1.0;
   #else
      #define alpha valpha
      #define beta  vbeta
   #endif
   void *va;
   TYPE *a;

   if (M > SYMM_Xover)
   {
      va = malloc(ATL_Cachelen + ATL_MulBySize(N)*N);
      ATL_assert(va);
      a = ATL_AlignPtr(va);
      #ifdef TREAL
         if ( SCALAR_IS_ONE(alpha) )
            Mjoin(Mjoin(Mjoin(PATL,sycopy),UploNM),_a1)(N, alpha, A, lda, a);
         else Mjoin(Mjoin(Mjoin(PATL,sycopy),UploNM),_aX)(N, alpha, A, lda, a);
         ATL_ammm(AtlasNoTrans, AtlasNoTrans, M, N, N, one, B, ldb, a, N, beta, C, ldc);
      #else
         Mjoin(Mjoin(PATL,sycopy),UploNM)(N, A, lda, a);
         ATL_ammm(AtlasNoTrans, AtlasNoTrans, M, N, N, valpha, B, ldb, a, N, vbeta, C, ldc);
      #endif
      free(va);
   }
   else Mjoin(PATL,refsymm)(AtlasRight, Uplo_, M, N, alpha, A, lda, B, ldb,
                            beta, C, ldc);
}

static TYPE geresid(enum CBLAS_ORDER Order, int N, TYPE *A, int lda,
                    TYPE *AI, int ldi)
/*
 * returns ||A - AI|| / (N * eps * ||A|| * ||AI||);
 * for row-major, we are not using 1-norm, since we are adding rows instead
 * of cols, but it should be an equally good norm, so don't worry about it.
 */
{
   TYPE numer, denom, eps;
   const int ldcp1 = (N+1)SHIFT;
   TYPE *C;
   int i;

   #ifdef TREAL
      TYPE one = ATL_rone, zero = ATL_rzero;
   #else
      TYPE one[2] = {ATL_rone, ATL_rzero}, zero[2] = {ATL_rzero, ATL_rzero};
   #endif

   eps = Mjoin(PATL,epsilon)();
   C = malloc(N*ATL_MulBySize(N));
   ATL_assert(C);
   cblas_gemm(Order, CblasNoTrans, CblasNoTrans, N, N, N, one, A, lda,
              AI, ldi, zero, C, N);                /* C now has A*inv(A) */
   for (i=0; i != N; i++) C[i*ldcp1] -= ATL_rone;  /* C now has A*inv(A)-I */
   numer = Mjoin(PATL,genrm1)(N, N, C, N);
   denom = Mjoin(PATL,genrm1)(N, N, A, lda) *
           Mjoin(PATL,genrm1)(N, N, AI, ldi) * N * eps;
   free(C);
   return(numer/denom);
}

static TYPE *ATL_LmulLt(const int N, const TYPE *L, const int ldl)
/*
 * A = L * L^H
 */
{
   const int incA = 1 SHIFT, incL = (ldl+1) SHIFT;
   TYPE *A;
   int i, j;
   #ifdef TCPLX
      int i1, i2;
      TYPE tmp;
   #endif

   A = malloc(N*ATL_MulBySize(N));
   ATL_assert(A);
   for (j=0; j < N; j++)
   {
      for (i=j; i < N; i++)
      {
      #ifdef TREAL
         A[i+j*N] = L[i+j*ldl] * L[j+j*ldl] +
                    Mjoin(PATL,dot)(j, L+i, ldl, L+j, ldl);
      #else
         tmp = L[(j+j*ldl)<<1];
         i1 = (i + j * N)<<1;
         i2 = (i + j * ldl)<<1;
         Mjoin(PATL,dotc_sub)(j, L+(j<<1), ldl, L+(i<<1), ldl, A+i1);
         A[i1] += L[i2] * tmp;
         if (i != j) A[i1+1] += tmp * L[i2+1];
      #endif
      }
   }
   return(A);
}

   int Mjoin(PATL,syr2kLT)
#endif
   (const int N, const int K, const void *valpha, const void *A, const int lda,
    const void *B, const int ldb, const void *vbeta, void *C, const int ldc)
{
   int i;
   void *vc=NULL;
   TYPE *c;
   #ifdef TREAL
      const SCALAR alpha=*( (const SCALAR *)valpha );
      const SCALAR beta =*( (const SCALAR *)vbeta  );
      const SCALAR one=1.0, zero=0.0;
   #else
      #define alpha valpha
      const TYPE *beta=vbeta;
      const TYPE one[2]={1.0,0.0}, zero[2]={0.0,0.0};
   #endif

   i = ATL_MulBySize(N)*N;
   if (i <= ATL_MaxMalloc) vc = malloc(ATL_Cachelen+i);
   if (vc == NULL) return(1);
   c = ATL_AlignPtr(vc);
   CgemmTN(N, N, K, alpha, A, lda, B, ldb, zero, c, N);
   if ( SCALAR_IS_ONE(beta) ) Mjoin(syr2k_put,_b1)(N, c, beta, C, ldc);
   else if ( SCALAR_IS_ZERO(beta) ) Mjoin(syr2k_put,_b0)(N, c, beta, C, ldc);
   #ifdef TCPLX
      else if (SCALAR_IS_NONE(beta)) Mjoin(syr2k_put,_bn1)(N, c, beta, C, ldc);
      else if (beta[1] == *zero) Mjoin(syr2k_put,_bXi0)(N, c, beta, C, ldc);
   #endif
   else Mjoin(syr2k_put,_bX)(N, c, beta, C, ldc);
   free(vc);
   return(0);
}

void Mjoin(Mjoin(Mjoin(PATL,herk),UploNM),N)
   (const int N, const int K, const void *valpha, const void *A, const int lda,
    const void *vbeta, void *C, const int ldc)
{
   void *vc;
   TYPE *c;
   TYPE alpha[2];
   const TYPE beta = *( (const TYPE *)vbeta  );
   const TYPE zero[2] = {0.0, 0.0};

   alpha[0] = *( (const TYPE *)valpha );
   if (K > HERK_Xover)
   {
      alpha[1] = 0.0;
      vc = malloc(ATL_Cachelen+ATL_MulBySize(N)*N);
      ATL_assert(vc);
      c = ATL_AlignPtr(vc);
      CgemmNC(N, N, K, alpha, A, lda, A, lda, zero, c, N);
      if ( beta == 1.0 ) Mjoin(her_put,_b1)(N, c, vbeta, C, ldc);
      else if ( beta == 0.0 ) Mjoin(her_put,_b0)(N, c, vbeta, C, ldc);
      else Mjoin(her_put,_bXi0)(N, c, vbeta, C, ldc);
      free(vc);
   }
   else Mjoin(PATL,refherk)(Uplo_, AtlasNoTrans, N, K, *alpha, A, lda,
                            beta, C, ldc);
}

static double RunTiming
   (enum CBLAS_ORDER Order, enum TEST_UPLO Uplo, int N, int lda,
    int CacheSize, int nreps)
{
   TYPE *A, *a;
   const int incA = N*lda;
   int i, k;
   double t0, t1=0.0;

   if (nreps < 1) nreps = 1;
   i = ATL_DivBySize(2*CacheSize) ATL_PTCACHEMUL;
   k = i = (i + N*N-1) / (N*N);
   if (nreps > i) k = i = nreps;
   a = A = malloc(i * ATL_MulBySize(incA));
   if (A)
   {
      if (Uplo == TestGE)
         for (i=0; i < k; i++) Mjoin(PATL,gegen)(N, N, A+i*incA, lda, N+lda);
      else for (i=0; i < k; i++) hegen(Order, Uplo, N, A+i*incA, lda);

      t0 = time00();
      for (i=nreps; i; i--, a += incA) test_inv(Order, Uplo, N, a, lda);
      t1 = time00() - t0;
      free(A);
   }
   else fprintf(stderr, "   WARNING: not enough mem to run timings!\n");
   return(t1/nreps);
}