xref: /libCEED/rust/libceed-sys/c-src/backends/magma/ceed-magma-gemm-selector.cpp (revision f80f4a748154eed4bc661c135f695b92b1bc45b9)

#include <stdio.h>
#include <sys/time.h>

#include <array>
#include <limits>
#include <vector>

#include "ceed-magma.h"
#include "tuning/indices.h"
#ifdef CEED_MAGMA_USE_HIP
#include "tuning/mi100.h"
#include "tuning/mi250x.h"
#include "tuning/mi250x_grad_rtc.h"
#include "tuning/mi250x_interp_rtc.h"
#else
#include "tuning/a100.h"
#include "tuning/a100_grad_rtc.h"
#include "tuning/a100_interp_rtc.h"
#include "tuning/v100.h"
#endif

////////////////////////////////////////////////////////////////////////////////
static void *gemm_selector_get_data(int gpu_arch, char precision, char trans_A) {
// a default
#ifdef CEED_MAGMA_USE_HIP
  void *data = (void *)&sgemm_nn_mi250x;
#else
  void *data = (void *)&sgemm_nn_a100;
#endif

#ifdef CEED_MAGMA_USE_HIP
  if (gpu_arch >= 910) {
    // gfx90a or newer
    data = (precision == 's') ? ((trans_A == 'n') ? (void *)&sgemm_nn_mi250x : (void *)&sgemm_tn_mi250x)
                              : ((trans_A == 'n') ? (void *)&dgemm_nn_mi250x : (void *)&dgemm_tn_mi250x);
  } else {
    // gfx908 or older
    data = (precision == 's') ? ((trans_A == 'n') ? (void *)&sgemm_nn_mi100 : (void *)&sgemm_tn_mi100)
                              : ((trans_A == 'n') ? (void *)&dgemm_nn_mi100 : (void *)&dgemm_tn_mi100);
  }
#else
  if (gpu_arch >= 800) {
    // sm80 or newer
    data = (precision == 's') ? ((trans_A == 'n') ? (void *)&sgemm_nn_a100 : (void *)&sgemm_tn_a100)
                              : ((trans_A == 'n') ? (void *)&dgemm_nn_a100 : (void *)&dgemm_tn_a100);
  } else {
    // sm70 or older
    data = (precision == 's') ? ((trans_A == 'n') ? (void *)&sgemm_nn_v100 : (void *)&sgemm_tn_v100)
                              : ((trans_A == 'n') ? (void *)&dgemm_nn_v100 : (void *)&dgemm_tn_v100);
  }
#endif

  return data;
}

////////////////////////////////////////////////////////////////////////////////
void gemm_selector(int gpu_arch, char precision, char trans_A, int m, int n, int k, int *n_batch, int *use_magma) {
  // defaults
  *n_batch                                           = n;
  *use_magma                                         = 0;
  std::vector<std::array<int, RECORD_LENGTH> > *data = NULL;
  data = (std::vector<std::array<int, RECORD_LENGTH> > *)gemm_selector_get_data(gpu_arch, precision, trans_A);

  int    ir   = -1;
  double norm = std::numeric_limits<double>::max();
  for (size_t i = 0; i < data->size(); i++) {
    int im = (*data)[i][M_INDEX];
    int in = (*data)[i][N_INDEX];
    int ik = (*data)[i][K_INDEX];

    double mdiff = (double)(im - m);
    double ndiff = (double)(in - n);
    double kdiff = (double)(ik - k);

    double nrm = sqrt(mdiff * mdiff + ndiff * ndiff + kdiff * kdiff);

    if (nrm < norm) {
      norm = nrm;
      ir   = i;
    }

    if (nrm == 0) {
      // the input (m, n, k) exactly matches a record in `data`
      // no need to search further
      break;
    }
  }

  if (ir >= 0) {
    *use_magma = (*data)[ir][USE_MAGMA_INDEX];
    // if the closest match indicates that n = n_batch,
    // that means calling the regular non-batch gemm.
    // So n_batch is set to n instead of the 'n_batch'
    // entry of the matching record
    int n_       = (*data)[ir][N_INDEX];
    int n_batch_ = (*data)[ir][N_BATCH_INDEX];
    *n_batch     = (n_ == n_batch_) ? n : n_batch_;
  }
}

////////////////////////////////////////////////////////////////////////////////
static void *nontensor_rtc_get_data(int gpu_arch, char precision, CeedEvalMode e_mode, CeedTransposeMode t_mode) {
// a default
#ifdef CEED_MAGMA_USE_HIP
  void *data = (void *)&dinterp_n_mi250x;
#else
  void *data = (void *)&dinterp_n_a100;
#endif

#ifdef CEED_MAGMA_USE_HIP
  if (e_mode == CEED_EVAL_INTERP) {
    data = (t_mode == CEED_TRANSPOSE) ? (void *)&dinterp_t_mi250x : (void *)&dinterp_n_mi250x;
  } else if (e_mode == CEED_EVAL_GRAD) {
    data = (t_mode == CEED_TRANSPOSE) ? (void *)&dgrad_t_mi250x : (void *)&dgrad_n_mi250x;
  }
#else
  if (e_mode == CEED_EVAL_INTERP) {
    data = (t_mode == CEED_TRANSPOSE) ? (void *)&dinterp_t_a100 : (void *)&dinterp_n_a100;
  } else if (e_mode == CEED_EVAL_GRAD) {
    data = (t_mode == CEED_TRANSPOSE) ? (void *)&dgrad_t_a100 : (void *)&dgrad_n_a100;
  }
#endif

  return data;
}

////////////////////////////////////////////////////////////////////////////////
CeedInt nontensor_rtc_get_nb(int gpu_arch, char precision, CeedEvalMode e_mode, CeedTransposeMode t_mode, int P_, int N, int Q_) {
  CeedInt P  = (t_mode == CEED_TRANSPOSE) ? P_ : Q_;
  CeedInt Q  = (t_mode == CEED_TRANSPOSE) ? Q_ : P_;
  CeedInt NB = 1;

  std::vector<std::array<int, RECORD_LENGTH_RTC> > *data = NULL;
  data = (std::vector<std::array<int, RECORD_LENGTH_RTC> > *)nontensor_rtc_get_data(gpu_arch, precision, e_mode, t_mode);

  int    ir   = -1;
  double norm = std::numeric_limits<double>::max();
  for (size_t i = 0; i < data->size(); i++) {
    int ip = (*data)[i][M_INDEX_RTC];
    int in = (*data)[i][N_INDEX_RTC];
    int iq = (*data)[i][K_INDEX_RTC];

    double pdiff = (double)(ip - P);
    double ndiff = (double)(in - N);
    double qdiff = (double)(iq - Q);
    double nrm   = sqrt(pdiff * pdiff + ndiff * ndiff + qdiff * qdiff);

    if (nrm < norm) {
      norm = nrm;
      ir   = i;
    }

    if (nrm == 0) {
      // the input (m, n, k) exactly matches a record in `data`
      // no need to search further
      break;
    }
  }

  if (ir >= 0) {
    NB = (*data)[ir][NB_INDEX_RTC];
  }

  return NB;
}