Examples¶
Source code for the examples described in this section is available in the examples folder of the NVSHMEM package.
Attribute-Based Initialization Example¶
The following code shows an MPI version of the simple shift program that was explained in The NVSHMEM Programming Model. It shows the use of the NVSHMEM attribute-based initialization API where the MPI communicator can be used to set up NVSHMEM.
#include <stdio.h>
#include "mpi.h"
#include "nvshmem.h"
#include "nvshmemx.h"
#define CUDA_CHECK(stmt) \
do { \
cudaError_t result = (stmt); \
if (cudaSuccess != result) { \
fprintf(stderr, "[%s:%d] CUDA failed with %s \n", \
__FILE__, __LINE__, cudaGetErrorString(result)); \
exit(-1); \
} \
} while (0)
__global__ void simple_shift(int *destination) {
int mype = nvshmem_my_pe();
int npes = nvshmem_n_pes();
int peer = (mype + 1) % npes;
nvshmem_int_p(destination, mype, peer);
}
int main (int argc, char *argv[]) {
int mype_node, msg;
cudaStream_t stream;
int rank, nranks;
MPI_Comm mpi_comm = MPI_COMM_WORLD;
nvshmemx_init_attr_t attr = NVSHMEMX_INIT_ATTR_INITIALIZER;
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nranks);
attr.mpi_comm = &mpi_comm;
nvshmemx_init_attr(NVSHMEMX_INIT_WITH_MPI_COMM, &attr);
mype_node = nvshmem_team_my_pe(NVSHMEMX_TEAM_NODE);
CUDA_CHECK(cudaSetDevice(mype_node));
CUDA_CHECK(cudaStreamCreate(&stream));
int *destination = (int *) nvshmem_malloc (sizeof(int));
simple_shift<<<1, 1, 0, stream>>>(destination);
nvshmemx_barrier_all_on_stream(stream);
CUDA_CHECK(cudaMemcpyAsync(&msg, destination, sizeof(int),
cudaMemcpyDeviceToHost, stream));
CUDA_CHECK(cudaStreamSynchronize(stream));
printf("%d: received message %d\n", nvshmem_my_pe(), msg);
nvshmem_free(destination);
nvshmem_finalize();
MPI_Finalize();
return 0;
}
The following code shows a Unique ID version of the simple shift program that was explained in The NVSHMEM Programming Model. It shows the use of the NVSHMEM attribute-based initializion API where the Unique ID arguments can be used to set up NVSHMEM.
#include <stdio.h>
#include "mpi.h"
#include "nvshmem.h"
#include "nvshmemx.h"
#define CUDA_CHECK(stmt) \
do { \
cudaError_t result = (stmt); \
if (cudaSuccess != result) { \
fprintf(stderr, "[%s:%d] CUDA failed with %s \n", \
__FILE__, __LINE__, cudaGetErrorString(result)); \
exit(-1); \
} \
} while (0)
__global__ void simple_shift(int *destination) {
int mype = nvshmem_my_pe();
int npes = nvshmem_n_pes();
int peer = (mype + 1) % npes;
nvshmem_int_p(destination, mype, peer);
}
int main (int argc, char *argv[]) {
int mype_node, msg;
cudaStream_t stream;
int rank, nranks;
nvshmemx_init_attr_t attr = NVSHMEMX_INIT_ATTR_INITIALIZER;
nvshmemx_uniqueid_t id = NVSHMEMX_UNIQUEID_INITIALIZER;
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nranks);
// PE 0 queries the unique ID
if (rank == 0) {
nvshmemx_get_uniqueid(&id);
}
// PE 0 broadcast the unique ID to all peers
MPI_Bcast(&id, sizeof(nvshmemx_uniqueid_t), MPI_UINT8_T, 0, MPI_COMM_WORLD);
nvshmemx_set_attr_uniqueid_args(rank, nranks, &id, &attr);
nvshmemx_init_attr(NVSHMEMX_INIT_WITH_UNIQUEID, &attr);
mype_node = nvshmem_team_my_pe(NVSHMEMX_TEAM_NODE);
CUDA_CHECK(cudaSetDevice(mype_node));
CUDA_CHECK(cudaStreamCreate(&stream));
int *destination = (int *) nvshmem_malloc (sizeof(int));
simple_shift<<<1, 1, 0, stream>>>(destination);
nvshmemx_barrier_all_on_stream(stream);
CUDA_CHECK(cudaMemcpyAsync(&msg, destination, sizeof(int),
cudaMemcpyDeviceToHost, stream));
CUDA_CHECK(cudaStreamSynchronize(stream));
printf("%d: received message %d\n", nvshmem_my_pe(), msg);
nvshmem_free(destination);
nvshmem_finalize();
MPI_Finalize();
return 0;
}
Collective Launch Example¶
The following code shows an example implementation of a single ring-based reduction where multiple iterations of the code, including computation, communication and synchronization are expressed as a single kernel.
This example also demonstrates the use of NVSHMEM collective launch, required when the NVSHMEM synchronization API is used from inside the CUDA kernel.
There is no MPI dependency for the example. NVSHMEM can be used to port existing MPI applications and develop new applications.
#include <stdio.h>
#include "nvshmem.h"
#include "nvshmemx.h"
#ifdef NVSHMEM_MPI_SUPPORT
#include "mpi.h"
#endif
#undef CUDA_CHECK
#define CUDA_CHECK(stmt) \
do { \
cudaError_t result = (stmt); \
if (cudaSuccess != result) { \
fprintf(stderr, "[%s:%d] cuda failed with %s \n", __FILE__, __LINE__, \
cudaGetErrorString(result)); \
exit(-1); \
} \
} while (0)
#define NVSHMEM_CHECK(stmt) \
do { \
int result = (stmt); \
if (NVSHMEMX_SUCCESS != result) { \
fprintf(stderr, "[%s:%d] nvshmem failed with error %d \n", __FILE__, __LINE__, \
result); \
exit(-1); \
} \
} while (0)
__global__ void reduce_ring(int *target, int mype, int npes) {
int peer = (mype + 1) % npes;
int lvalue = mype;
for (int i = 1; i < npes; i++) {
nvshmem_int_p(target, lvalue, peer);
nvshmem_barrier_all();
lvalue = *target + mype;
nvshmem_barrier_all();
}
}
int main(int c, char *v[]) {
int mype, npes, mype_node;
#ifdef NVSHMEM_MPI_SUPPORT
bool use_mpi = false;
char *value = getenv("NVSHMEMTEST_USE_MPI_LAUNCHER");
if (value) use_mpi = atoi(value);
#endif
#ifdef NVSHMEM_MPI_SUPPORT
if (use_mpi) {
MPI_Init(&c, &v);
int rank, nranks;
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nranks);
MPI_Comm mpi_comm = MPI_COMM_WORLD;
nvshmemx_init_attr_t attr;
attr.mpi_comm = &mpi_comm;
nvshmemx_init_attr(NVSHMEMX_INIT_WITH_MPI_COMM, &attr);
} else
nvshmem_init();
#else
nvshmem_init();
#endif
mype = nvshmem_my_pe();
npes = nvshmem_n_pes();
mype_node = nvshmem_team_my_pe(NVSHMEMX_TEAM_NODE);
// application picks the device each PE will use
CUDA_CHECK(cudaSetDevice(mype_node));
int *u = (int *)nvshmem_calloc(1, sizeof(int));
int *h = (int *)calloc(1, sizeof(int));
void *args[] = {&u, &mype, &npes};
dim3 dimBlock(1);
dim3 dimGrid(1);
NVSHMEM_CHECK(
nvshmemx_collective_launch((const void *)reduce_ring, dimGrid, dimBlock, args, 0, 0));
CUDA_CHECK(cudaDeviceSynchronize());
cudaMemcpy(h, u, sizeof(int), cudaMemcpyDeviceToHost);
printf("results on device [%d] is %d \n",mype, h[0]);
nvshmem_free(u);
free(h);
nvshmem_finalize();
#ifdef NVSHMEM_MPI_SUPPORT
if (use_mpi) MPI_Finalize();
#endif
return 0;
}
On-Stream Example¶
The following example shows how nvshmemx_*_on_stream functions can be used to
enqueue a SHMEM operation onto a CUDA stream for execution in stream order.
Specifically, the example shows the following:
- How a collective SHMEM reduction operation can be made to wait on a preceding kernel in the stream.
- How a kernel can be made to wait for a communication result from a previous collective SHMEM reduction operation.
The example shows one use case for relieving CPU control over GPU compute and communication.
#include <stdio.h>
#include "nvshmem.h"
#include "nvshmemx.h"
#ifdef NVSHMEM_MPI_SUPPORT
#include "mpi.h"
#endif
#define THRESHOLD 42
#define CORRECTION 7
#undef CUDA_CHECK
#define CUDA_CHECK(stmt) \
do { \
cudaError_t result = (stmt); \
if (cudaSuccess != result) { \
fprintf(stderr, "[%s:%d] cuda failed with %s \n", __FILE__, __LINE__, \
cudaGetErrorString(result)); \
exit(-1); \
} \
} while (0)
__global__ void accumulate(int *input, int *partial_sum) {
int index = threadIdx.x;
if (0 == index) *partial_sum = 0;
__syncthreads();
atomicAdd(partial_sum, input[index]);
}
__global__ void correct_accumulate(int *input, int *partial_sum, int *full_sum) {
int index = threadIdx.x;
if (*full_sum > THRESHOLD) {
input[index] = input[index] - CORRECTION;
}
if (0 == index) *partial_sum = 0;
__syncthreads();
atomicAdd(partial_sum, input[index]);
}
int main(int c, char *v[]) {
int mype, npes, mype_node;
int *input;
int *partial_sum;
int *full_sum;
int input_nelems = 512;
int to_all_nelems = 1;
cudaStream_t stream;
#ifdef NVSHMEM_MPI_SUPPORT
bool use_mpi = false;
char *value = getenv("NVSHMEMTEST_USE_MPI_LAUNCHER");
if (value) use_mpi = atoi(value);
#endif
#ifdef NVSHMEM_MPI_SUPPORT
if (use_mpi) {
MPI_Init(&c, &v);
int rank, nranks;
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nranks);
MPI_Comm mpi_comm = MPI_COMM_WORLD;
nvshmemx_init_attr_t attr;
attr.mpi_comm = &mpi_comm;
nvshmemx_init_attr(NVSHMEMX_INIT_WITH_MPI_COMM, &attr);
} else
nvshmem_init();
#else
nvshmem_init();
#endif
mype = nvshmem_my_pe();
npes = nvshmem_n_pes();
mype_node = nvshmem_team_my_pe(NVSHMEMX_TEAM_NODE);
CUDA_CHECK(cudaSetDevice(mype_node));
CUDA_CHECK(cudaStreamCreate(&stream));
input = (int *)nvshmem_malloc(sizeof(int) * input_nelems);
partial_sum = (int *)nvshmem_malloc(sizeof(int));
full_sum = (int *)nvshmem_malloc(sizeof(int));
accumulate<<<1, input_nelems, 0, stream>>>(input, partial_sum);
nvshmemx_int_sum_reduce_on_stream(NVSHMEM_TEAM_WORLD, full_sum, partial_sum, to_all_nelems, stream);
correct_accumulate<<<1, input_nelems, 0, stream>>>(input, partial_sum, full_sum);
CUDA_CHECK(cudaStreamSynchronize(stream));
printf("[%d of %d] run complete \n", mype, npes);
CUDA_CHECK(cudaStreamDestroy(stream));
nvshmem_free(input);
nvshmem_free(partial_sum);
nvshmem_free(full_sum);
nvshmem_finalize();
#ifdef NVSHMEM_MPI_SUPPORT
if (use_mpi) MPI_Finalize();
#endif
return 0;
}
Threadgroup Example¶
The example in this section shows how nvshmemx_collect32_block can be used to
leverage threads to accelerate a SHMEM collect operation when all threads in
the block depend on the result of a preceding communication operation. For this
instance, partial vector sums are computed across different PEs and have a
SHMEM collect operation to obtain the complete sum across PEs.
#include <stdio.h>
#include "nvshmem.h"
#include "nvshmemx.h"
#ifdef NVSHMEM_MPI_SUPPORT
#include "mpi.h"
#endif
#define NTHREADS 512
#undef CUDA_CHECK
#define CUDA_CHECK(stmt) \
do { \
cudaError_t result = (stmt); \
if (cudaSuccess != result) { \
fprintf(stderr, "[%s:%d] cuda failed with %s \n", __FILE__, __LINE__, \
cudaGetErrorString(result)); \
exit(-1); \
} \
} while (0)
__global__ void distributed_vector_sum(int *x, int *y, int *partial_sum, int *sum,
int use_threadgroup, int mype, int npes) {
int index = threadIdx.x;
int nelems = blockDim.x;
partial_sum[index] = x[index] + y[index];
if (use_threadgroup) {
/* all threads realize the entire fcollect operation */
nvshmemx_int_fcollect_block(NVSHMEM_TEAM_WORLD, sum, partial_sum, nelems);
} else {
/* thread 0 realizes the entire fcollect operation */
if (0 == index) {
nvshmem_int_fcollect(NVSHMEM_TEAM_WORLD, sum, partial_sum, nelems);
}
}
}
int main(int c, char *v[]) {
int mype, npes, mype_node;
int *x;
int *y;
int *partial_sum;
int *sum;
int use_threadgroup = 1;
int nthreads = NTHREADS;
#ifdef NVSHMEM_MPI_SUPPORT
bool use_mpi = false;
char *value = getenv("NVSHMEMTEST_USE_MPI_LAUNCHER");
if (value) use_mpi = atoi(value);
#endif
#ifdef NVSHMEM_MPI_SUPPORT
if (use_mpi) {
MPI_Init(&c, &v);
int rank, nranks;
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nranks);
MPI_Comm mpi_comm = MPI_COMM_WORLD;
nvshmemx_init_attr_t attr;
attr.mpi_comm = &mpi_comm;
nvshmemx_init_attr(NVSHMEMX_INIT_WITH_MPI_COMM, &attr);
} else
nvshmem_init();
#else
nvshmem_init();
#endif
npes = nvshmem_n_pes();
mype = nvshmem_my_pe();
mype_node = nvshmem_team_my_pe(NVSHMEMX_TEAM_NODE);
CUDA_CHECK(cudaSetDevice(mype_node));
x = (int *)nvshmem_malloc(sizeof(int) * nthreads);
y = (int *)nvshmem_malloc(sizeof(int) * nthreads);
partial_sum = (int *)nvshmem_malloc(sizeof(int) * nthreads);
sum = (int *)nvshmem_malloc(sizeof(int) * nthreads * npes);
void *args[] = {&x, &y, &partial_sum, &sum, &use_threadgroup, &mype, &npes};
dim3 dimBlock(nthreads);
dim3 dimGrid(1);
nvshmemx_collective_launch((const void *)distributed_vector_sum, dimGrid, dimBlock, args, 0, 0);
CUDA_CHECK(cudaDeviceSynchronize());
printf("[%d of %d] run complete \n", mype, npes);
nvshmem_free(x);
nvshmem_free(y);
nvshmem_free(partial_sum);
nvshmem_free(sum);
nvshmem_finalize();
#ifdef NVSHMEM_MPI_SUPPORT
if (use_mpi) MPI_Finalize();
#endif
return 0;
}
Put on Block Example¶
In the example below, every thread in block 0 calls nvshmemx_float_put_block.
Alternatively, every thread can call nvshmem_float_p, but nvshmem_float_p
has a disadvantage that when the destination GPU is connected via InfiniBand,
there is one RMA message for every single element, which can be detrimental to
performance.
The disadvantage with using nvshmem_float_put in this case is that when the
destination GPU is P2P-connected, a single thread will copy the entire data to
the destination GPU. While nvshmemx_float_put_block can leverage all the
threads in the block to copy the data in parallel to the destination GPU.
#include <stdio.h>
#include <assert.h>
#include "nvshmem.h"
#include "nvshmemx.h"
#undef CUDA_CHECK
#define CUDA_CHECK(stmt) \
do { \
cudaError_t result = (stmt); \
if (cudaSuccess != result) { \
fprintf(stderr, "[%s:%d] cuda failed with %s \n", __FILE__, __LINE__, \
cudaGetErrorString(result)); \
exit(-1); \
} \
} while (0)
#define THREADS_PER_BLOCK 1024
__global__ void set_and_shift_kernel(float *send_data, float *recv_data, int num_elems, int mype,
int npes) {
int thread_idx = blockIdx.x * blockDim.x + threadIdx.x;
/* set the corresponding element of send_data */
if (thread_idx < num_elems) send_data[thread_idx] = mype;
int peer = (mype + 1) % npes;
/* Every thread in block 0 calls nvshmemx_float_put_block. Alternatively,
every thread can call shmem_float_p, but shmem_float_p has a disadvantage
that when the destination GPU is connected via IB, there will be one rma
message for every single element which can be detrimental to performance.
And the disadvantage with shmem_float_put is that when the destination GPU is p2p
connected, it cannot leverage multiple threads to copy the data to the destination
GPU. */
int block_offset = blockIdx.x * blockDim.x;
nvshmemx_float_put_block(recv_data + block_offset, send_data + block_offset,
min(blockDim.x, num_elems - block_offset),
peer); /* All threads in a block call the API
with the same arguments */
}
int main(int c, char *v[]) {
int mype, npes, mype_node;
float *send_data, *recv_data;
int num_elems = 8192;
int num_blocks;
cudaStream_t stream;
nvshmem_init();
mype = nvshmem_my_pe();
npes = nvshmem_n_pes();
mype_node = nvshmem_team_my_pe(NVSHMEMX_TEAM_NODE);
// application picks the device each PE will use
CUDA_CHECK(cudaSetDevice(mype_node));
CUDA_CHECK(cudaStreamCreate(&stream));
send_data = (float *)nvshmem_malloc(sizeof(float) * num_elems);
recv_data = (float *)nvshmem_malloc(sizeof(float) * num_elems);
assert(send_data != NULL && recv_data != NULL);
assert(num_elems % THREADS_PER_BLOCK == 0); /* for simplicity */
num_blocks = num_elems / THREADS_PER_BLOCK;
set_and_shift_kernel<<<num_blocks, THREADS_PER_BLOCK, 0, stream>>>(send_data, recv_data, num_elems, mype, npes);
nvshmemx_barrier_all_on_stream(stream);
CUDA_CHECK(cudaStreamSynchronize(stream));
/* Do data validation */
float *host = new float[num_elems];
CUDA_CHECK(cudaMemcpy(host, recv_data, num_elems * sizeof(float), cudaMemcpyDefault));
int ref = (mype - 1 + npes) % npes;
bool success = true;
for (int i = 0; i < num_elems; ++i) {
if (host[i] != ref) {
printf("Error at %d of rank %d: %f\n", i, mype, host[i]);
success = false;
break;
}
}
if (success) {
printf("[%d of %d] run complete \n", mype, npes);
} else {
printf("[%d of %d] run failure \n", mype, npes);
}
nvshmem_free(send_data);
nvshmem_free(recv_data);
nvshmem_finalize();
return 0;
}
Ring Broadcast Example¶
In the example below, PE 0 broadcasts a message by sending it to PE 1, which
sends the message to PE 2 and so on. This example demonstrates several NVSHMEM
APIs, including the use of nvshmem_fence to order communication and
nvshmem_signal_wait_until and nvshmemx_signal_op for point-to-point
synchronization.
#include <stdio.h>
#include <stdint.h>
#include <cuda.h>
#include <nvshmem.h>
#include <nvshmemx.h>
__global__ void ring_bcast(int *data, size_t nelem, int root, uint64_t *psync) {
int mype = nvshmem_my_pe();
int npes = nvshmem_n_pes();
int peer = (mype + 1) % npes;
if (mype == root)
*psync = 1;
nvshmem_signal_wait_until(psync, NVSHMEM_CMP_NE, 0);
if (mype == npes-1) return;
nvshmem_int_put(data, data, nelem, peer);
nvshmem_fence();
nvshmemx_signal_op(psync, 1, NVSHMEM_SIGNAL_SET, peer);
*psync = 0;
}
int main(void) {
size_t data_len = 32;
cudaStream_t stream;
nvshmem_init();
int mype = nvshmem_my_pe();
int mype_node = nvshmem_team_my_pe(NVSHMEMX_TEAM_NODE);
cudaSetDevice(mype_node);
cudaStreamCreate(&stream);
int *data = (int *) nvshmem_malloc(sizeof(int) * data_len);
int *data_h = (int *) malloc(sizeof(int) * data_len);
uint64_t *psync = (uint64_t *) nvshmem_calloc(1, sizeof(uint64_t));
for (size_t i = 0; i < data_len; i++)
data_h[i] = mype + i;
cudaMemcpyAsync(data, data_h, sizeof(int) * data_len, cudaMemcpyHostToDevice, stream);
int root = 0;
dim3 gridDim(1), blockDim(1);
void *args[] = { &data, &data_len, &root, &psync };
nvshmemx_barrier_all_on_stream(stream);
nvshmemx_collective_launch((const void *)ring_bcast, gridDim, blockDim, args, 0, stream);
nvshmemx_barrier_all_on_stream(stream);
cudaMemcpyAsync(data_h, data, sizeof(int) * data_len, cudaMemcpyDeviceToHost, stream);
cudaStreamSynchronize(stream);
for (size_t i = 0; i < data_len; i++) {
if (data_h[i] != i)
printf("PE %d error, data[%zu] = %d expected data[%zu] = %d\n",
mype, i, data_h[i], i, (int) i);
}
nvshmem_free(data);
nvshmem_free(psync);
free(data_h);
nvshmem_finalize();
return 0;
}
Ring Allreduce Example¶
In the example below, PE0 receives a message chunk from its left neighbor, performs
a local reduction and sends the resulting chunk to its right neighbor (PE1), and so on.
Eventually, every PE (but last) broadcast its own chunk to right neighbor. This
examples demonstrates several NVSHMEM APIs, including the use of nvshmem_int_put_signal_nbi
and nvshmem_signal_wait_until for point-to-point communication & synchronization.
/*
* Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
*
* NVIDIA CORPORATION and its licensors retain all intellectual property
* and proprietary rights in and to this software, related documentation
* and any modifications thereto. Any use, reproduction, disclosure or
* distribution of this software and related documentation without an express
* license agreement from NVIDIA CORPORATION is strictly prohibited.
*
* See COPYRIGHT.txt for license information
*/
/* This example performs an allreduce operation using ring algorithm when
GPUs are connected via remote interconect like IB/RoCE/EFA, etc.
It does ring reduce followed by ring broadcast. We use single threaded put_signal API
as single thread is sufficient for remote transfers. The example is expected
to be performant only when GPUs are connected via remote interconnect. */
#include <stdio.h>
#include <stdint.h>
#include <cuda.h>
#include <nvshmem.h>
#include <nvshmemx.h>
#include <unistd.h>
#include <ctype.h>
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#undef CUDA_CHECK
#define CUDA_CHECK(stmt) \
do { \
cudaError_t result = (stmt); \
if (cudaSuccess != result) { \
fprintf(stderr, "[%s:%d] cuda failed with %s \n", __FILE__, __LINE__, \
cudaGetErrorString(result)); \
exit(-1); \
} \
} while (0)
/* atol() + optional scaled suffix recognition: 1K, 2M, 3G, 1T */
static inline int atol_scaled(const char *str, size_t *out) {
int scale, n;
double p = -1.0;
char f;
n = sscanf(str, "%lf%c", &p, &f);
if (n == 2) {
switch (f) {
case 'k':
case 'K':
scale = 10;
break;
case 'm':
case 'M':
scale = 20;
break;
case 'g':
case 'G':
scale = 30;
break;
case 't':
case 'T':
scale = 40;
break;
default:
return 1;
}
} else if (p < 0) {
return 1;
} else
scale = 0;
*out = (size_t)ceil(p * (1lu << scale));
return 0;
}
size_t min_size = 1024 * 1024 * 32;
size_t max_size = min_size * 16;
size_t num_blocks = 32;
size_t threads_per_block = 512;
size_t iters = 4;
size_t warmup_iters = 1;
size_t step_factor = 2;
size_t chunk_size = 262144;
// perform Allreduce using ring
__global__ void ring_reduce(int *dst, const int *src, size_t nreduce, uint64_t *signal,
size_t chunk_size) {
int mype = nvshmem_my_pe();
int npes = nvshmem_n_pes();
int peer = (mype + 1) % npes;
int thread_id = threadIdx.x;
int num_threads = blockDim.x;
int num_blocks = gridDim.x;
int block_idx = blockIdx.x;
size_t elems_per_block = nreduce / num_blocks;
// Change src, dst, nreduce, signal to what this block is going to process
// Each CTA will work independently
if (elems_per_block * (blockIdx.x + 1) > nreduce) return;
src = src + block_idx * elems_per_block;
dst = dst + block_idx * elems_per_block;
nreduce = elems_per_block;
signal = signal + block_idx;
size_t chunk_elems = chunk_size / sizeof(int);
size_t num_chunks = nreduce / chunk_elems;
// reduce phase
for (size_t chunk = 0; chunk < num_chunks; chunk++) {
if (mype != 0) {
if (thread_id == 0) nvshmem_signal_wait_until(signal, NVSHMEM_CMP_GE, chunk + 1);
__syncthreads();
for (size_t i = thread_id; i < chunk_elems; i += num_threads) {
dst[i] = dst[i] + src[i];
}
__syncthreads();
}
if (thread_id == 0)
nvshmem_int_put_signal_nbi(dst, (mype == 0) ? src : dst, chunk_elems, signal, 1,
NVSHMEM_SIGNAL_ADD, peer);
src = src + chunk_elems;
dst = dst + chunk_elems;
}
// Broadcast phase
dst = dst - num_chunks * chunk_elems;
if (thread_id == 0) {
for (size_t chunk = 0; chunk < num_chunks; chunk++) {
if (mype < npes - 1) { // Last pe already has the final result
nvshmem_signal_wait_until(signal, NVSHMEM_CMP_GE,
(mype == 0) ? chunk + 1 : num_chunks + chunk + 1);
}
if (mype < npes - 2)
nvshmem_int_put_signal_nbi(dst, dst, chunk_elems, signal, 1, NVSHMEM_SIGNAL_ADD,
peer);
dst = dst + chunk_elems;
}
*signal = 0; // reset for next iteration
}
}
int main(int argc, char **argv) {
int c;
while ((c = getopt(argc, argv, "b:e:f:n:w:c:t:m:")) != -1) {
switch (c) {
case 'b':
atol_scaled(optarg, &min_size);
break;
case 'e':
atol_scaled(optarg, &max_size);
break;
case 'f':
atol_scaled(optarg, &step_factor);
break;
case 'n':
atol_scaled(optarg, &iters);
break;
case 'w':
atol_scaled(optarg, &warmup_iters);
break;
case 'c':
atol_scaled(optarg, &num_blocks);
break;
case 't':
atol_scaled(optarg, &threads_per_block);
break;
case 'm':
atol_scaled(optarg, &chunk_size);
break;
case '?':
if (optopt == 'c')
fprintf(stderr, "Option -%c requires an argument.\n", optopt);
else if (isprint(optopt))
fprintf(stderr, "Unknown option `-%c'.\n", optopt);
else
fprintf(stderr, "Unknown option character `\\x%x'.\n", optopt);
return 1;
default:
abort();
}
}
size_t min_ints = min_size / sizeof(int);
assert(min_ints % num_blocks == 0);
nvshmem_init();
int mype = nvshmem_my_pe();
int npes = nvshmem_n_pes();
int mype_node = nvshmem_team_my_pe(NVSHMEMX_TEAM_NODE);
cudaStream_t stream;
cudaEvent_t start, stop;
CUDA_CHECK(cudaEventCreate(&start));
CUDA_CHECK(cudaEventCreate(&stop));
CUDA_CHECK(cudaSetDevice(mype_node));
CUDA_CHECK(cudaStreamCreate(&stream));
size_t max_ints = max_size / sizeof(int);
int *dst = (int *)nvshmem_malloc(max_size);
int *src = (int *)nvshmem_malloc(max_size);
int *data_h = (int *)malloc(max_size);
uint64_t *signal = (uint64_t *)nvshmem_calloc(num_blocks, sizeof(uint64_t));
dim3 gridDim(num_blocks), blockDim(threads_per_block);
for (size_t i = 0; i < max_ints; i++) data_h[i] = i;
CUDA_CHECK(cudaMemcpyAsync(src, data_h, max_size, cudaMemcpyHostToDevice, stream));
nvshmemx_barrier_all_on_stream(stream);
for (size_t size = min_size; size <= max_size; size *= step_factor) {
size_t num_ints = size / sizeof(int);
void *args[] = {&dst, &src, &num_ints, &signal, &chunk_size};
// do warmup
for (size_t i = 0; i < warmup_iters; i++) {
nvshmemx_collective_launch((const void *)ring_reduce, gridDim, blockDim, args, 0,
stream);
nvshmemx_barrier_all_on_stream(stream);
}
CUDA_CHECK(cudaStreamSynchronize(stream));
// main loop
CUDA_CHECK(cudaEventRecord(start, stream));
for (size_t i = 0; i < iters; i++) {
nvshmemx_collective_launch((const void *)ring_reduce, gridDim, blockDim, args, 0,
stream);
nvshmemx_barrier_all_on_stream(stream);
}
CUDA_CHECK(cudaEventRecord(stop, stream));
CUDA_CHECK(cudaStreamSynchronize(stream));
if (!mype) {
float ms;
CUDA_CHECK(cudaEventElapsedTime(&ms, start, stop));
printf("%zuB \t %fms\n", size, ms / iters);
}
// validate output
CUDA_CHECK(cudaMemcpy(data_h, dst, size, cudaMemcpyDeviceToHost));
for (size_t i = 0; i < num_ints; i++) {
if (data_h[i] != (int)i * npes)
printf("PE %d error, data[%zu] = %d expected data[%zu] = %d\n", mype, i, data_h[i],
i, (int)i * npes);
}
}
CUDA_CHECK(cudaEventDestroy(start));
CUDA_CHECK(cudaEventDestroy(stop));
nvshmem_free(dst);
nvshmem_free(src);
nvshmem_free(signal);
free(data_h);
nvshmem_finalize();
return 0;
}
User Buffer Registration Example¶
The example below shows user allocated memory buffer being registered with NVSHMEM symmetric heap
and used for communication. The example uses CUDA VMM APIs to allocate memory buffer
(createUserBuffer function) and registers it to the heap using nvshmemx_buffer_register_symmetric
routine. Once registered, the buffer can be used for communication like any memory allocated using
nvshmem_malloc and/or its friends.
#include <stdio.h>
#include <iostream>
#include <stdlib.h>
#include <unistd.h>
#include "nvshmem.h"
#include "nvshmemx.h"
#include "cuda_runtime.h"
#define GRANULARITY 536870912UL
#define COLL_NELEMS 4096
#undef CUDA_CHECK
#define CUDA_CHECK(stmt) \
do { \
cudaError_t result = (stmt); \
if (cudaSuccess != result) { \
fprintf(stderr, "[%s:%d] cuda failed with %s \n", __FILE__, __LINE__, \
cudaGetErrorString(result)); \
exit(-1); \
} \
assert(cudaSuccess == result); \
} while (0)
#undef CU_CHECK
#define CU_CHECK(stmt) \
do { \
CUresult result = (stmt); \
const char *str; \
if (CUDA_SUCCESS != result) { \
CUresult ret = cuGetErrorString(result, &str); \
if (ret == CUDA_ERROR_INVALID_VALUE) str = "Unknown error"; \
fprintf(stderr, "[%s:%d] cuda failed with %s \n", __FILE__, __LINE__, str); \
exit(-1); \
} \
assert(CUDA_SUCCESS == result); \
} while (0)
__global__ void init_data_kernel(float *source, size_t nelems) {
for (int i = 0; i < nelems; ++i) {
source[i] = (float)i;
}
}
void *createUserBuffer(size_t size, CUmemAllocationProp &prop) {
void *bufAddr = nullptr;
CUmemAccessDesc accessDescriptor;
accessDescriptor.location.id = prop.location.id;
accessDescriptor.location.type = prop.location.type;
accessDescriptor.flags = CU_MEM_ACCESS_FLAGS_PROT_READWRITE;
CUmemGenericAllocationHandle userAllocHandle;
CU_CHECK(cuMemCreate(&userAllocHandle, size, (const CUmemAllocationProp *)&prop, 0));
CU_CHECK(cuMemAddressReserve((CUdeviceptr *)&bufAddr, size, 0, (CUdeviceptr)NULL, 0));
CU_CHECK(cuMemMap((CUdeviceptr)bufAddr, size, 0, userAllocHandle, 0));
CU_CHECK(
cuMemSetAccess((CUdeviceptr)bufAddr, size, (const CUmemAccessDesc *)&accessDescriptor, 1));
return bufAddr;
}
void releaseUserBuf(void *ptr, size_t size) {
CUmemGenericAllocationHandle memHandle;
CU_CHECK(cuMemRetainAllocationHandle(&memHandle, ptr));
CU_CHECK(cuMemUnmap((CUdeviceptr)ptr, size));
CU_CHECK(cuMemAddressFree((CUdeviceptr)ptr, size));
CU_CHECK(cuMemRelease(memHandle));
}
int main(int argc, char **argv) {
nvshmem_init();
int status = 0;
int mype, npes;
int npes_node, mype_node;
const size_t size = GRANULARITY;
void *buffer;
void *mmaped_buffer;
CUmemAllocationProp prop = {};
int dev_id;
float *source, *dest, *dest_h;
size_t nelems;
cudaStream_t stream;
nvshmem_team_t team = NVSHMEM_TEAM_WORLD;
mype = nvshmem_my_pe();
mype_node = nvshmem_team_my_pe(NVSHMEMX_TEAM_NODE);
npes_node = nvshmem_team_n_pes(NVSHMEMX_TEAM_NODE);
npes = nvshmem_n_pes();
dev_id = mype_node % npes_node;
CUDA_CHECK(cudaSetDevice(dev_id));
if (!mype) printf("creating and mmapping buffer of size: %lu\n", size);
// Allocation of user buffer is local
prop.type = CU_MEM_ALLOCATION_TYPE_PINNED;
prop.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
prop.location.id = dev_id;
prop.allocFlags.gpuDirectRDMACapable = 1;
prop.requestedHandleTypes =
(CUmemAllocationHandleType)(CU_MEM_HANDLE_TYPE_POSIX_FILE_DESCRIPTOR);
buffer = createUserBuffer(size, prop);
if (!buffer) {
fprintf(stderr, "Failed to create user buffer \n");
status = 1;
goto out;
}
mmaped_buffer = (void *)nvshmemx_buffer_register_symmetric(buffer, size, 0);
if (!mmaped_buffer) {
fprintf(stderr, "shmem_mmap failed \n");
status = 1;
goto out;
}
CUDA_CHECK(cudaMemset(mmaped_buffer, 0, size));
// test heap usage to verify mmap correctness
CUDA_CHECK(cudaStreamCreateWithFlags(&stream, cudaStreamNonBlocking));
nelems = COLL_NELEMS;
nelems = nelems / 2; // split the buffer into source and dest
source = (float *)mmaped_buffer;
dest = (float *)(mmaped_buffer) + nelems;
dest_h = (float *)malloc(nelems * sizeof(float));
init_data_kernel<<<1, 1, 0, stream>>>((float *)source, nelems);
nvshmemx_barrier_on_stream(team, stream);
nvshmemx_float_sum_reduce_on_stream(team, (float *)dest, (const float *)source, nelems, stream);
cudaStreamSynchronize(stream);
CUDA_CHECK(cudaMemcpy(dest_h, dest, nelems * sizeof(float), cudaMemcpyDeviceToHost));
for (size_t i = 0; i < nelems; i++) {
if (dest_h[i] != (float)i * npes) {
printf("PE %d error, data[%zu] = %f expected data[%zu] = %f\n", mype, i, dest_h[i], i,
(float)i * npes);
status = -1;
}
}
if (!status) {
fprintf(stderr, "No errors found\n");
}
CUDA_CHECK(cudaDeviceSynchronize());
nvshmem_barrier_all();
// free all buffers
status = nvshmemx_buffer_unregister_symmetric(mmaped_buffer, size);
if (status) {
fprintf(stderr, "nvshmemx_buffer_unregister_symmetric failed\n");
}
free(dest_h);
nvshmem_finalize();
releaseUserBuf(buffer, size);
out:
return status;
}
GEMM + AllReduce Fused Kernel Example¶
This example uses the Tile-granular APIs to create a fused GEMM + AllReduce kernel. This
example builds on Blackwell fp16 GEMM
from CUTLASS and integrates AllReduce within the kernel. Below we show a code snippet from the
host code which sets up the input matrices and launches the fused kernel. The complete example
can be found at /examples/gemm_allreduce. As this example builds on top of CUTLASS example, it
is recommended to go over the CUTLASS example.
Specifically, this example extends the cutlass/examples/70_blackwell_gemm/70_blackwell_fp16_gemm.cu
to perform tensor allocation in NVSHMEM symmetric heap and perform AllReduce collective at tile-granularity
(please refer gemmAR_fusion_blackwell_fp16.cu). The TMA warp specialized GEMM kernel from CUTLASS is modified
to invoke AllReduce collective operation on each output tile (please refer to sm100_gemm_tma_warpspecialized_allreduce.hpp).
For integrating AllReduce, a new class CollectiveAllReduceMulticastWarpSpecialized (please refer to allreduce_nvls_warpspecilaized.hpp)
is added which performs the tile collective operations and contains the required inputs and helper functions. do_allreduce()
function is called after every tile of GEMM output is computed, this function takes in the coordinates of the computed tile along with
matrix data pointers to create nvshmemx::Tensors on which AllReduce is performed. tile_collective_wait() ensures completion of the
collectives.
A new allocator class nvshmemAllocation is also added (please refer nvshmemAlloc.hpp) which allows allocating tensors from within
NVSHMEM symmetric heap.
/////////////////////////////////////////////////////////////////////
/// GEMM kernel configurations
/////////////////////////////////////////////////////////////////////
// A matrix configuration
using ElementA = half_t; // Element type for A matrix operand
using LayoutA = cutlass::layout::RowMajor; // Layout type for A matrix operand
constexpr int AlignmentA = 128 / cutlass::sizeof_bits<ElementA>::value;
// B matrix configuration
using ElementB = half_t;
using LayoutB = cutlass::layout::ColumnMajor;
constexpr int AlignmentB = 128 / cutlass::sizeof_bits<ElementB>::value;
// C/D matrix configuration
using ElementC = float;
using LayoutC = cutlass::layout::ColumnMajor;
constexpr int AlignmentC = 128 / cutlass::sizeof_bits<ElementC>::value;
// Kernel functional config
using ElementAccumulator = float;
using ArchTag = cutlass::arch::Sm100;
using OperatorClass = cutlass::arch::OpClassTensorOp;
using MmaTileShape_MNK = Shape<_256, _128, _64>;
// Shape of the threadblocks in a cluster
using ClusterShape_MNK = Shape<_2, _2, _1>;
// Build the epilogue
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
ArchTag, OperatorClass, MmaTileShape_MNK, ClusterShape_MNK,
cutlass::epilogue::collective::EpilogueTileAuto, ElementAccumulator,
ElementAccumulator, ElementC, LayoutC, AlignmentC, ElementC,
LayoutC, AlignmentC,
cutlass::epilogue::collective::EpilogueScheduleAuto>::CollectiveOp;
// Build the mainloop
using CollectiveMainloop = typename cutlass::gemm::collective::CollectiveBuilder<
ArchTag, OperatorClass, ElementA, LayoutA, AlignmentA, ElementB,
LayoutB, AlignmentB, ElementAccumulator, MmaTileShape_MNK,
ClusterShape_MNK,
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(
sizeof(typename CollectiveEpilogue::SharedStorage))>,
cutlass::gemm::collective::KernelScheduleAuto>::CollectiveOp;
using CollectiveAllReduce = cutlass::comm::collective::CollectiveAllReduceMulticastWarpSpecialized<
ElementC, MmaTileShape_MNK, typename CollectiveEpilogue::StrideD>;
// Compose into a kernel
using GemmKernel = cutlass::gemm::kernel::Sm100GemmARUniversal<
Shape<int, int, int, int>, // Indicates ProblemShape
CollectiveMainloop, CollectiveEpilogue,
cutlass::gemm::PersistentScheduler, CollectiveAllReduce>;
using Gemm = cutlass::gemm::device::GemmUniversalAdapter<GemmKernel>;
// Reference device GEMM implementation type
using DeviceGemmReference =
cutlass::reference::device::Gemm<ElementA, LayoutA, ElementB, LayoutB,
ElementC, LayoutC,
ElementAccumulator, ElementAccumulator>;
using StrideA = typename Gemm::GemmKernel::StrideA;
using StrideB = typename Gemm::GemmKernel::StrideB;
using StrideC = typename Gemm::GemmKernel::StrideC;
using StrideD = typename Gemm::GemmKernel::StrideD;
//
// Data members
//
/// Initialization
StrideA stride_A;
StrideB stride_B;
StrideC stride_C;
StrideD stride_D;
uint64_t seed = 1;
cutlass::DeviceAllocation<typename Gemm::ElementA> block_A;
cutlass::DeviceAllocation<typename Gemm::ElementB> block_B;
cutlass::DeviceAllocation<typename Gemm::ElementC> block_C;
nvshmemAllocation<typename Gemm::EpilogueOutputOp::ElementOutput> block_D;
nvshmemAllocation<typename Gemm::EpilogueOutputOp::ElementOutput>
block_D_red;
nvshmemAllocation<typename Gemm::EpilogueOutputOp::ElementOutput> block_ref_D;
cutlass::DeviceAllocation<typename Gemm::EpilogueOutputOp::ElementOutput> block_ref_D_red;
__global__ void ref_reduce_kernel(ElementC *out, ElementC **ref_D_ptr,
ElementC *arrD_red,
ElementC *arrD, size_t npes, size_t nelem) {
int tid = threadIdx.x + blockIdx.x * blockDim.x;
volatile ElementC *output = out;
volatile ElementC *val_ptr;
for (int i = tid; i < nelem; i += gridDim.x * blockDim.x) {
val_ptr = ref_D_ptr[0] + i;
output[i] = *(val_ptr);
for (int n = 1; n < npes; ++n) {
val_ptr = ref_D_ptr[n] + i;
output[i] += *(val_ptr);
}
}
}
__global__ void compare_kernel(ElementC *expected_out, ElementC *actual_out,
ElementC **ref_D_ptr,
ElementC *arrD_red, ElementC *arrD, int mype,
size_t npes,
size_t nelem) {
int tid = threadIdx.x + blockIdx.x * blockDim.x;
for (int i = tid; i < nelem; i += gridDim.x * blockDim.x) {
if (actual_out[i] != expected_out[i]) {
printf("%d elem: %d, mismatch expected_out: %f, actual: %f
computed: %f : %f \n", mype, i, expected_out[i],
actual_out[i], *(ref_D_ptr[0] + i), *(ref_D_ptr[1] + i));
}
}
}
////// nvshmem variables //////
nvshmem_team_t *teams_dev, *teams;
int num_teams;
int mype, npes;
/// Execute a given example GEMM computation
template <typename Gemm>
int run(Options &options) {
initialize(options); // Initializes the inputs
// Instantiate CUTLASS kernel depending on templates
Gemm gemm;
// Create a structure of gemm kernel arguments
auto arguments = args_from_options(options);
auto grid = gemm.get_grid_shape(arguments);
dim3 blockShape = GemmKernel::get_block_shape();
int sm_count;
CUDA_CHECK(cudaDeviceGetAttribute(&sm_count,
cudaDevAttrMultiProcessorCount, 0));
int max_active_blocks = gemm.maximum_active_blocks();
printf("%d Grid dimension: (%d, %d, %d), block: (%d, %d, %d),
occupancy: %d\n", mype, grid.x, grid.y, grid.z,
blockShape.x, blockShape.y, blockShape.z, sm_count);
int max_concurrent_blocks = sm_count * max_active_blocks;
if (max_concurrent_blocks < (grid.x * grid.y * grid.z)) {
fprintf(stderr,
"Grid size exceeds maximum concurrent blocks. Using Tile-granular "
"APIs requires all thread blocks to be concurrent across PEs\n");
exit(1);
}
// create teams
// each block has 1 warpgroup acting as epilogue, so num_teams = #blocks
num_teams = grid.x * grid.y * grid.z;
teams = (nvshmem_team_t *)malloc(num_teams * sizeof(nvshmem_team_t));
for (int i = 0; i < num_teams; ++i) {
nvshmem_team_split_strided(NVSHMEM_TEAM_WORLD, 0, 1, npes,
nullptr, 0, &teams[i]);
}
CUDA_CHECK(cudaMalloc((void **)&teams_dev,
num_teams * sizeof(nvshmem_team_t)));
CUDA_CHECK(cudaMemcpy(teams_dev, teams,
num_teams * sizeof(nvshmem_team_t), cudaMemcpyHostToDevice));
// populate AR arguments
arguments.allReduceArgs = {block_D.get(), block_D_red.get(), stride_D,
nvshmem_my_pe(), nvshmem_n_pes(), teams_dev};
// Using the arguments, query for extra workspace required
// for matrix multiplication computation
size_t workspace_size = Gemm::get_workspace_size(arguments);
// Allocate workspace memory
cutlass::device_memory::allocation<uint8_t> workspace(workspace_size);
// Check if the problem size is supported or not
CUTLASS_CHECK(gemm.can_implement(arguments));
// Initialize CUTLASS kernel with arguments and workspace pointer
CUTLASS_CHECK(gemm.initialize(arguments, workspace.get()));
// Correctness / Warmup iteration
CUTLASS_CHECK(gemm.run());
// Check if output result and reference kernel are equal or not
CUDA_CHECK(cudaDeviceSynchronize());
nvshmem_barrier_all();
Result result;
result.passed = verify(options, mype, npes);
std::cout << " Disposition: " << (result.passed ? "Passed" : "Failed")
<< std::endl;
if (!result.passed) {
exit(-1);
}
// Run profiling loop
if (options.iterations > 0) {
GpuTimer timer;
timer.start();
for (int iter = 0; iter < options.iterations; ++iter) {
CUTLASS_CHECK(gemm.initialize(arguments, workspace.get()));
CUTLASS_CHECK(gemm.run());
}
CUDA_CHECK(cudaDeviceSynchronize());
timer.stop();
// Compute average runtime and GFLOPs.
float elapsed_ms = timer.elapsed_millis();
result.avg_runtime_ms = double(elapsed_ms) /
double(options.iterations);
result.gflops = options.gflops(result.avg_runtime_ms / 1000.0);
std::cout << " Problem Size: " << options.m << 'x'
<< options.n << 'x' << options.k << std::endl;
std::cout << " Avg runtime: " << result.avg_runtime_ms
<< " ms" << std::endl;
std::cout << " GFLOPS: " << result.gflops << std::endl;
}
return 0;
}
//////////////////////////////////////////////////////////////////
template <class ElementT_, class TileShape_, class StrideMNL_>
class CollectiveAllReduceMulticastWarpSpecialized {
public:
using ElementT = ElementT_;
using TileShape = TileShape_;
using StrideMNL = StrideMNL_;
struct Arguments {
ElementT* ptr_aux = nullptr; // start pointer of matrix
ElementT* out_ptr = nullptr; // start pointer of matrix
StrideMNL stride;
int rank;
int world_size;
nvshmem_team_t* teams = nullptr;
};
struct Params {
ElementT* ptr_aux = nullptr;
ElementT* out_ptr = nullptr;
StrideMNL stride;
int rank;
int world_size;
Layout<Shape<int, int>> tile_layout;
nvshmem_team_t* teams = nullptr;
};
template <class ProblemShape>
static constexpr Params
to_underlying_arguments(ProblemShape const& problem_shape,
Arguments const& args) {
// Append 1s until problem shape is rank-4
auto problem_shape_mnkl = append<4>(problem_shape, 1);
auto [M, N, K, L] = problem_shape_mnkl;
int m_tiles = ceil_div(M, size<0>(TileShape{}));
int n_tiles = ceil_div(N, size<1>(TileShape{}));
// number of tiles in each dimension
auto tile_layout = make_layout(make_shape(m_tiles, n_tiles));
return {
args.ptr_aux, args.out_ptr, args.stride, args.rank,
args.world_size, tile_layout, args.teams,
};
}
const Params* params_ptr;
CUTLASS_HOST_DEVICE
CollectiveAllReduceMulticastWarpSpecialized() {}
CUTLASS_HOST_DEVICE
CollectiveAllReduceMulticastWarpSpecialized(Params const& params):
params_ptr(¶ms) {}
template <class ProblemShapeMNKL, class TileCoordMNKL>
CUTLASS_DEVICE void do_allreduce(ProblemShapeMNKL const& problem_shape,
TileCoordMNKL const& tile_coord) {
auto [M, N, K, L] = problem_shape;
auto [m, n, k, l] = tile_coord;
if (m >= size<0>(params_ptr->tile_layout.shape()) ||
n >= size<1>(params_ptr->tile_layout.shape())) {
// early exit if out of bound
return;
}
int tile_index = params_ptr->tile_layout(m, n);
int tiles_per_rank =
cute::ceil_div(cute::product(params_ptr->tile_layout.shape()),
params_ptr->world_size);
// only root PE will do reduction for this tile
// only needed if using two-shot algorithm
int root = tile_index / tiles_per_rank;
Tensor mAux = make_tensor(params_ptr->ptr_aux,
make_layout(make_shape(M, N, L),
params_ptr->stride)); // (M,N,L)
Tensor mAux_out = make_tensor(
params_ptr->out_ptr, make_layout(make_shape(M, N, L),
params_ptr->stride)); // (M,N,L)
Tensor gAux =
local_tile(mAux, take<0, 2>(TileShape{}), make_coord(m, n, l));
Tensor gAux_out =
local_tile(mAux_out, take<0, 2>(TileShape{}),
make_coord(m, n, l));
// predication tensor
Tensor coordAux = make_identity_tensor(shape(mAux));
Tensor pAux = local_tile(coordAux, take<0, 2>(TileShape{}),
make_coord(m, n, l));
auto boundary = nvshmemx::make_shape<int, int>(M, N);
auto start_coord = nvshmemx::make_shape<int, int>(
size<0>(pAux(0, 0)), size<1>(pAux(0, 0)));
// Call AR
auto tensor_shape = nvshmemx::make_shape(M, N);
auto tensor_stride = nvshmemx::make_stride(
size<0>(params_ptr->stride),
size<1>(params_ptr->stride));
nvshmemx::Tensor srcTensor = nvshmemx::Tensor(gAux.data(),
nvshmemx::make_layout(tensor_shape, tensor_stride));
nvshmemx::Tensor dstTensor = nvshmemx::Tensor(gAux_out.data(),
nvshmemx::make_layout(tensor_shape, tensor_stride));
int blkId = blockIdx.x + gridDim.x * blockIdx.y;
nvshmemx::tile_sum_allreduce_warpgroup<decltype(srcTensor),
decltype(dstTensor),
decltype(boundary),
nvshmemx::tile_coll_algo_t::NVLS_ONE_SHOT_PULL_NBI>(
params_ptr->teams[blkId], srcTensor, dstTensor, start_coord,
boundary, root, 0);
}
CUTLASS_DEVICE
void tile_collective_wait() {
int blkId = blockIdx.x + gridDim.x * blockIdx.y;
nvshmemx::tile_collective_wait_warpgroup<
nvshmemx::tile_coll_algo_t::NVLS_ONE_SHOT_PULL_NBI>(
params_ptr->teams[blkId], 0);
}
};
///////////////////////////////////////////////////////////////////
int main(int argc, char const **args) {
// initialize nvshmem
nvshmem_init();
mype = nvshmem_my_pe();
npes = nvshmem_n_pes();
CUDA_CHECK(cudaSetDevice(mype));
printf(" Executing PE: %d out of %d\n", mype, npes);
// CUTLASS must be compiled with CUDA 12.0 Toolkit to run this example
// and must have compute capability at least 100a.
if (__CUDACC_VER_MAJOR__ < 12 ||
(__CUDACC_VER_MAJOR__ == 12 && __CUDACC_VER_MINOR__ < 8)) {
std::cerr << "This example requires CUDA 12.8 or newer."
<< std::endl;
// Returning zero so this test passes on older Toolkits.
// Its actions are no-op.
return 0;
}
cudaDeviceProp props;
int current_device_id;
CUDA_CHECK(cudaGetDevice(¤t_device_id));
CUDA_CHECK(cudaGetDeviceProperties(&props, current_device_id));
cudaError_t error = cudaGetDeviceProperties(&props, 0);
if (props.major != 10 || props.minor != 0) { std::cerr
<< "This example requires a GPU with compute capability 100a)."
<< std::endl;
return 0;
}
//
// Parse options
//
Options options;
options.parse(argc, args);
if (options.help) {
options.print_usage(std::cout) << std::endl;
return 0;
}
//
// Evaluate CUTLASS kernels
//
#if defined(CUTLASS_ARCH_MMA_SM100_SUPPORTED)
run<Gemm>(options);
#endif // defined(CUTLASS_ARCH_MMA_SM100_SUPPORTED)
nvshmem_barrier_all();
for (int i = 0; i < num_teams; ++i) {
nvshmem_team_destroy(teams[i]);
}
nvshmem_barrier_all();
block_D.free();
block_D_red.free();
block_ref_D.free();
free(teams);
CUDA_CHECK(cudaFree(teams_dev));
nvshmem_finalize();
return 0;
}