xref: /libCEED/rust/libceed-sys/c-src/backends/cuda-gen/ceed-cuda-gen-operator.c (revision e9c76bddc0f2a44f522e0176ed6b7e0c0aa1df73)

// Copyright (c) 2017-2024, Lawrence Livermore National Security, LLC and other CEED contributors.
// All Rights Reserved. See the top-level LICENSE and NOTICE files for details.
//
// SPDX-License-Identifier: BSD-2-Clause
//
// This file is part of CEED:  http://github.com/ceed

#include <ceed.h>
#include <ceed/backend.h>
#include <ceed/jit-source/cuda/cuda-types.h>
#include <cuda.h>
#include <cuda_runtime.h>
#include <stddef.h>
#include <string.h>

#include "../cuda/ceed-cuda-common.h"
#include "../cuda/ceed-cuda-compile.h"
#include "ceed-cuda-gen-operator-build.h"
#include "ceed-cuda-gen.h"

//------------------------------------------------------------------------------
// Destroy operator
//------------------------------------------------------------------------------
static int CeedOperatorDestroy_Cuda_gen(CeedOperator op) {
  Ceed                   ceed;
  CeedOperator_Cuda_gen *impl;

  CeedCallBackend(CeedOperatorGetCeed(op, &ceed));
  CeedCallBackend(CeedOperatorGetData(op, &impl));
  if (impl->points.num_per_elem) CeedCallCuda(ceed, cudaFree((void **)impl->points.num_per_elem));
  CeedCallBackend(CeedFree(&impl));
  CeedCallBackend(CeedDestroy(&ceed));
  return CEED_ERROR_SUCCESS;
}

static int Waste(int threads_per_sm, int warp_size, int threads_per_elem, int elems_per_block) {
  int useful_threads_per_block = threads_per_elem * elems_per_block;
  // round up to nearest multiple of warp_size
  int block_size    = CeedDivUpInt(useful_threads_per_block, warp_size) * warp_size;
  int blocks_per_sm = threads_per_sm / block_size;
  return threads_per_sm - useful_threads_per_block * blocks_per_sm;
}

// Choose the least wasteful block size constrained by blocks_per_sm of max_threads_per_block.
//
// The x and y part of block[] contains per-element sizes (specified on input) while the z part is number of elements.
//
// Problem setting: we'd like to make occupancy high with relatively few inactive threads. CUDA (cuOccupancyMaxPotentialBlockSize) can tell us how
// many threads can run.
//
// Note that full occupancy sometimes can't be achieved by one thread block.
// For example, an SM might support 1536 threads in total, but only 1024 within a single thread block.
// So cuOccupancyMaxPotentialBlockSize may suggest a block size of 768 so that two blocks can run, versus one block of 1024 will prevent a second
// block from running. The cuda-gen kernels are pretty heavy with lots of instruction-level parallelism (ILP) so we'll generally be okay with
// relatively low occupancy and smaller thread blocks, but we solve a reasonably general problem here. Empirically, we find that blocks bigger than
// about 256 have higher latency and worse load balancing when the number of elements is modest.
//
// cuda-gen can't choose block sizes arbitrarily; they need to be a multiple of the number of quadrature points (or number of basis functions).
// They also have a lot of __syncthreads(), which is another point against excessively large thread blocks.
// Suppose I have elements with 7x7x7 quadrature points.
// This will loop over the last dimension, so we have 7*7=49 threads per element.
// Suppose we have two elements = 2*49=98 useful threads.
// CUDA schedules in units of full warps (32 threads), so 128 CUDA hardware threads are effectively committed to that block.
// Now suppose cuOccupancyMaxPotentialBlockSize returned 352.
// We can schedule 2 blocks of size 98 (196 useful threads using 256 hardware threads), but not a third block (which would need a total of 384
// hardware threads).
//
// If instead, we had packed 3 elements, we'd have 3*49=147 useful threads occupying 160 slots, and could schedule two blocks.
// Alternatively, we could pack a single block of 7 elements (2*49=343 useful threads) into the 354 slots.
// The latter has the least "waste", but __syncthreads() over-synchronizes and it might not pay off relative to smaller blocks.
static int BlockGridCalculate(CeedInt num_elem, int blocks_per_sm, int max_threads_per_block, int max_threads_z, int warp_size, int block[3],
                              int *grid) {
  const int threads_per_sm   = blocks_per_sm * max_threads_per_block;
  const int threads_per_elem = block[0] * block[1];
  int       elems_per_block  = 1;
  int       waste            = Waste(threads_per_sm, warp_size, threads_per_elem, 1);

  for (int i = 2; i <= CeedIntMin(max_threads_per_block / threads_per_elem, num_elem); i++) {
    int i_waste = Waste(threads_per_sm, warp_size, threads_per_elem, i);

    // We want to minimize waste, but smaller kernels have lower latency and less __syncthreads() overhead so when a larger block size has the same
    // waste as a smaller one, go ahead and prefer the smaller block.
    if (i_waste < waste || (i_waste == waste && threads_per_elem * i <= 128)) {
      elems_per_block = i;
      waste           = i_waste;
    }
  }
  // In low-order elements, threads_per_elem may be sufficiently low to give an elems_per_block greater than allowable for the device, so we must
  // check before setting the z-dimension size of the block.
  block[2] = CeedIntMin(elems_per_block, max_threads_z);
  *grid    = CeedDivUpInt(num_elem, elems_per_block);
  return CEED_ERROR_SUCCESS;
}

// callback for cuOccupancyMaxPotentialBlockSize, providing the amount of dynamic shared memory required for a thread block of size threads.
static size_t dynamicSMemSize(int threads) { return threads * sizeof(CeedScalar); }

//------------------------------------------------------------------------------
// Apply and add to output
//------------------------------------------------------------------------------
static int CeedOperatorApplyAddCore_Cuda_gen(CeedOperator op, CUstream stream, const CeedScalar *input_arr, CeedScalar *output_arr, bool *is_run_good,
                                             CeedRequest *request) {
  bool                    is_at_points, is_tensor;
  Ceed                    ceed;
  Ceed_Cuda              *cuda_data;
  CeedInt                 num_elem, num_input_fields, num_output_fields;
  CeedEvalMode            eval_mode;
  CeedQFunctionField     *qf_input_fields, *qf_output_fields;
  CeedQFunction_Cuda_gen *qf_data;
  CeedQFunction           qf;
  CeedOperatorField      *op_input_fields, *op_output_fields;
  CeedOperator_Cuda_gen  *data;

  // Build the operator kernel
  CeedCallBackend(CeedOperatorBuildKernel_Cuda_gen(op, is_run_good));
  if (!(*is_run_good)) return CEED_ERROR_SUCCESS;

  CeedCallBackend(CeedOperatorGetCeed(op, &ceed));
  CeedCallBackend(CeedGetData(ceed, &cuda_data));
  CeedCallBackend(CeedOperatorGetData(op, &data));
  CeedCallBackend(CeedOperatorGetQFunction(op, &qf));
  CeedCallBackend(CeedQFunctionGetData(qf, &qf_data));
  CeedCallBackend(CeedOperatorGetNumElements(op, &num_elem));
  CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields));
  CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields));

  // Input vectors
  for (CeedInt i = 0; i < num_input_fields; i++) {
    CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode));
    if (eval_mode == CEED_EVAL_WEIGHT) {  // Skip
      data->fields.inputs[i] = NULL;
    } else {
      bool       is_active;
      CeedVector vec;

      // Get input vector
      CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec));
      is_active = vec == CEED_VECTOR_ACTIVE;
      if (is_active) data->fields.inputs[i] = input_arr;
      else CeedCallBackend(CeedVectorGetArrayRead(vec, CEED_MEM_DEVICE, &data->fields.inputs[i]));
      CeedCallBackend(CeedVectorDestroy(&vec));
    }
  }

  // Output vectors
  for (CeedInt i = 0; i < num_output_fields; i++) {
    CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode));
    if (eval_mode == CEED_EVAL_WEIGHT) {  // Skip
      data->fields.outputs[i] = NULL;
    } else {
      bool       is_active;
      CeedVector vec;

      // Get output vector
      CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[i], &vec));
      is_active = vec == CEED_VECTOR_ACTIVE;
      if (is_active) data->fields.outputs[i] = output_arr;
      else CeedCallBackend(CeedVectorGetArrayWrite(vec, CEED_MEM_DEVICE, &data->fields.outputs[i]));
      CeedCallBackend(CeedVectorDestroy(&vec));
    }
  }

  // Point coordinates, if needed
  CeedCallBackend(CeedOperatorIsAtPoints(op, &is_at_points));
  if (is_at_points) {
    // Coords
    CeedVector vec;

    CeedCallBackend(CeedOperatorAtPointsGetPoints(op, NULL, &vec));
    CeedCallBackend(CeedVectorGetArrayRead(vec, CEED_MEM_DEVICE, &data->points.coords));
    CeedCallBackend(CeedVectorDestroy(&vec));

    // Points per elem
    if (num_elem != data->points.num_elem) {
      CeedInt            *points_per_elem;
      const CeedInt       num_bytes   = num_elem * sizeof(CeedInt);
      CeedElemRestriction rstr_points = NULL;

      data->points.num_elem = num_elem;
      CeedCallBackend(CeedOperatorAtPointsGetPoints(op, &rstr_points, NULL));
      CeedCallBackend(CeedCalloc(num_elem, &points_per_elem));
      for (CeedInt e = 0; e < num_elem; e++) {
        CeedInt num_points_elem;

        CeedCallBackend(CeedElemRestrictionGetNumPointsInElement(rstr_points, e, &num_points_elem));
        points_per_elem[e] = num_points_elem;
      }
      if (data->points.num_per_elem) CeedCallCuda(ceed, cudaFree((void **)data->points.num_per_elem));
      CeedCallCuda(ceed, cudaMalloc((void **)&data->points.num_per_elem, num_bytes));
      CeedCallCuda(ceed, cudaMemcpy((void *)data->points.num_per_elem, points_per_elem, num_bytes, cudaMemcpyHostToDevice));
      CeedCallBackend(CeedElemRestrictionDestroy(&rstr_points));
      CeedCallBackend(CeedFree(&points_per_elem));
    }
  }

  // Get context data
  CeedCallBackend(CeedQFunctionGetInnerContextData(qf, CEED_MEM_DEVICE, &qf_data->d_c));

  // Apply operator
  void         *opargs[]  = {(void *)&num_elem, &qf_data->d_c, &data->indices, &data->fields, &data->B, &data->G, &data->W, &data->points};
  const CeedInt dim       = data->dim;
  const CeedInt Q_1d      = data->Q_1d;
  const CeedInt P_1d      = data->max_P_1d;
  const CeedInt thread_1d = CeedIntMax(Q_1d, P_1d);
  int           max_threads_per_block, min_grid_size, grid;

  CeedCallBackend(CeedOperatorHasTensorBases(op, &is_tensor));
  CeedCallCuda(ceed, cuOccupancyMaxPotentialBlockSize(&min_grid_size, &max_threads_per_block, data->op, dynamicSMemSize, 0, 0x10000));
  int block[3] = {thread_1d, ((!is_tensor || dim == 1) ? 1 : thread_1d), -1};

  if (is_tensor) {
    CeedCallBackend(BlockGridCalculate(num_elem, min_grid_size / cuda_data->device_prop.multiProcessorCount, max_threads_per_block,
                                       cuda_data->device_prop.maxThreadsDim[2], cuda_data->device_prop.warpSize, block, &grid));
  } else {
    CeedInt elems_per_block = CeedIntMin(cuda_data->device_prop.maxThreadsDim[2], CeedIntMax(512 / thread_1d, 1));

    grid     = num_elem / elems_per_block + (num_elem % elems_per_block > 0);
    block[2] = elems_per_block;
  }
  CeedInt shared_mem = block[0] * block[1] * block[2] * sizeof(CeedScalar);

  CeedCallBackend(CeedTryRunKernelDimShared_Cuda(ceed, data->op, stream, grid, block[0], block[1], block[2], shared_mem, is_run_good, opargs));

  // Restore input arrays
  for (CeedInt i = 0; i < num_input_fields; i++) {
    CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode));
    if (eval_mode == CEED_EVAL_WEIGHT) {  // Skip
    } else {
      bool       is_active;
      CeedVector vec;

      CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec));
      is_active = vec == CEED_VECTOR_ACTIVE;
      if (!is_active) CeedCallBackend(CeedVectorRestoreArrayRead(vec, &data->fields.inputs[i]));
      CeedCallBackend(CeedVectorDestroy(&vec));
    }
  }

  // Restore output arrays
  for (CeedInt i = 0; i < num_output_fields; i++) {
    CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode));
    if (eval_mode == CEED_EVAL_WEIGHT) {  // Skip
    } else {
      bool       is_active;
      CeedVector vec;

      CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[i], &vec));
      is_active = vec == CEED_VECTOR_ACTIVE;
      if (!is_active) CeedCallBackend(CeedVectorRestoreArray(vec, &data->fields.outputs[i]));
      CeedCallBackend(CeedVectorDestroy(&vec));
    }
  }

  // Restore point coordinates, if needed
  if (is_at_points) {
    CeedVector vec;

    CeedCallBackend(CeedOperatorAtPointsGetPoints(op, NULL, &vec));
    CeedCallBackend(CeedVectorRestoreArrayRead(vec, &data->points.coords));
    CeedCallBackend(CeedVectorDestroy(&vec));
  }

  // Restore context data
  CeedCallBackend(CeedQFunctionRestoreInnerContextData(qf, &qf_data->d_c));

  // Cleanup
  CeedCallBackend(CeedDestroy(&ceed));
  CeedCallBackend(CeedQFunctionDestroy(&qf));
  if (!(*is_run_good)) data->use_fallback = true;
  return CEED_ERROR_SUCCESS;
}

static int CeedOperatorApplyAdd_Cuda_gen(CeedOperator op, CeedVector input_vec, CeedVector output_vec, CeedRequest *request) {
  bool              is_run_good = false;
  const CeedScalar *input_arr   = NULL;
  CeedScalar       *output_arr  = NULL;

  // Try to run kernel
  if (input_vec != CEED_VECTOR_NONE) CeedCallBackend(CeedVectorGetArrayRead(input_vec, CEED_MEM_DEVICE, &input_arr));
  if (output_vec != CEED_VECTOR_NONE) CeedCallBackend(CeedVectorGetArray(output_vec, CEED_MEM_DEVICE, &output_arr));
  CeedCallBackend(CeedOperatorApplyAddCore_Cuda_gen(op, NULL, input_arr, output_arr, &is_run_good, request));
  if (input_vec != CEED_VECTOR_NONE) CeedCallBackend(CeedVectorRestoreArrayRead(input_vec, &input_arr));
  if (output_vec != CEED_VECTOR_NONE) CeedCallBackend(CeedVectorRestoreArray(output_vec, &output_arr));

  // Fallback on unsuccessful run
  if (!is_run_good) {
    CeedOperator op_fallback;

    CeedDebug256(CeedOperatorReturnCeed(op), CEED_DEBUG_COLOR_SUCCESS, "Falling back to /gpu/cuda/ref CeedOperator");
    CeedCallBackend(CeedOperatorGetFallback(op, &op_fallback));
    CeedCallBackend(CeedOperatorApplyAdd(op_fallback, input_vec, output_vec, request));
  }
  return CEED_ERROR_SUCCESS;
}

//------------------------------------------------------------------------------
// Create operator
//------------------------------------------------------------------------------
int CeedOperatorCreate_Cuda_gen(CeedOperator op) {
  Ceed                   ceed;
  CeedOperator_Cuda_gen *impl;

  CeedCallBackend(CeedOperatorGetCeed(op, &ceed));
  CeedCallBackend(CeedCalloc(1, &impl));
  CeedCallBackend(CeedOperatorSetData(op, impl));
  CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "ApplyAdd", CeedOperatorApplyAdd_Cuda_gen));
  CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "Destroy", CeedOperatorDestroy_Cuda_gen));
  CeedCallBackend(CeedDestroy(&ceed));
  return CEED_ERROR_SUCCESS;
}

//------------------------------------------------------------------------------