sglang.0.4.8.post1/nvshmem_src/perftest/device/coll/bcast_latency.cu

337 lines
13 KiB
Plaintext

/*
* Copyright (c) 2019-2020, 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
*/
#define CUMODULE_NAME "bcast_latency.cubin"
#include "coll_test.h"
#define DATATYPE int64_t
#if defined __cplusplus || defined NVSHMEM_BITCODE_APPLICATION
extern "C" {
#endif
#define CALL_BCAST(TYPENAME, TYPE, TG_PRE, THREADGROUP, THREAD_COMP, ELEM_COMP) \
__global__ void test_##TYPENAME##_bcast_call_kern##THREADGROUP( \
nvshmem_team_t team, TYPE *dest, const TYPE *source, int nelems, int PE_root, int mype, \
int iter) { \
int i; \
\
if (!blockIdx.x && (threadIdx.x < THREAD_COMP) && (nelems < ELEM_COMP)) { \
for (i = 0; i < iter; i++) { \
nvshmem##TG_PRE##_##TYPENAME##_broadcast##THREADGROUP(team, dest, source, nelems, \
PE_root); \
} \
} \
} \
void test_##TYPENAME##_bcast_call_kern##THREADGROUP##_cubin( \
int num_blocks, int num_tpb, cudaStream_t stream, void **arglist) { \
CUfunction test_cubin; \
\
init_test_case_kernel( \
&test_cubin, NVSHMEMI_TEST_STRINGIFY(test_##TYPENAME##_bcast_call_kern##THREADGROUP)); \
CU_CHECK(cuLaunchCooperativeKernel(test_cubin, num_blocks, 1, 1, num_tpb, 1, 1, 0, stream, \
arglist)); \
}
#define CALL_BCAST_KERNEL(TYPENAME, THREADGROUP, BLOCKS, THREADS, ARG_LIST, STREAM) \
if (use_cubin) { \
test_##TYPENAME##_bcast_call_kern##THREADGROUP##_cubin(BLOCKS, THREADS, STREAM, ARG_LIST); \
} else { \
status = nvshmemx_collective_launch( \
(const void *)test_##TYPENAME##_bcast_call_kern##THREADGROUP, BLOCKS, THREADS, \
ARG_LIST, 0, STREAM); \
if (status != NVSHMEMX_SUCCESS) { \
fprintf(stderr, "shmemx_collective_launch failed %d \n", status); \
exit(-1); \
} \
}
CALL_BCAST(int32, int32_t, , , 1, 512);
CALL_BCAST(int64, int64_t, , , 1, 512);
CALL_BCAST(int32, int32_t, x, _warp, warpSize, 4096);
CALL_BCAST(int64, int64_t, x, _warp, warpSize, 4096);
CALL_BCAST(int32, int32_t, x, _block, INT_MAX, INT_MAX);
CALL_BCAST(int64, int64_t, x, _block, INT_MAX, INT_MAX);
#if defined __cplusplus || defined NVSHMEM_BITCODE_APPLICATION
}
#endif
int broadcast_calling_kernel(nvshmem_team_t team, void *dest, const void *source, int mype,
int PE_root, cudaStream_t stream, void **h_tables) {
int status = 0;
int nvshm_test_num_tpb = threads_per_block;
int num_blocks = 1;
size_t num_elems = 1;
size_t min_elems, max_elems;
int i;
int skip = warmup_iters;
int iter = iters;
uint64_t *h_size_array = (uint64_t *)h_tables[0];
double *h_thread_lat = (double *)h_tables[1];
double *h_warp_lat = (double *)h_tables[2];
double *h_block_lat = (double *)h_tables[3];
float milliseconds;
void *args_1[] = {&team, &dest, &source, &num_elems, &mype, &PE_root, &skip};
void *args_2[] = {&team, &dest, &source, &num_elems, &mype, &PE_root, &iter};
cudaEvent_t start, stop;
cudaEventCreate(&start);
cudaEventCreate(&stop);
float *ms_d = (float *)nvshmem_malloc(sizeof(float));
float *ms_sum_d = (float *)nvshmem_malloc(sizeof(float));
nvshmem_barrier_all();
min_elems = max(static_cast<size_t>(1), min_size / sizeof(int32_t));
max_elems = max(static_cast<size_t>(1), max_size / sizeof(int32_t));
i = 0;
for (num_elems = min_elems; num_elems < 512; num_elems *= step_factor) {
CALL_BCAST_KERNEL(int32, , num_blocks, nvshm_test_num_tpb, args_1, stream);
CUDA_CHECK(cudaStreamSynchronize(stream));
nvshmem_barrier_all();
cudaEventRecord(start, stream);
CALL_BCAST_KERNEL(int32, , num_blocks, nvshm_test_num_tpb, args_2, stream);
cudaEventRecord(stop, stream);
CUDA_CHECK(cudaStreamSynchronize(stream));
if (!mype) {
cudaEventElapsedTime(&milliseconds, start, stop);
h_thread_lat[i] = (milliseconds * 1000.0) / (float)iter;
}
i++;
nvshmem_barrier_all();
}
i = 0;
for (num_elems = min_elems; num_elems < 4096; num_elems *= step_factor) {
CALL_BCAST_KERNEL(int32, _warp, num_blocks, nvshm_test_num_tpb, args_1, stream);
CUDA_CHECK(cudaStreamSynchronize(stream));
nvshmem_barrier_all();
cudaEventRecord(start, stream);
CALL_BCAST_KERNEL(int32, _warp, num_blocks, nvshm_test_num_tpb, args_2, stream);
cudaEventRecord(stop, stream);
CUDA_CHECK(cudaStreamSynchronize(stream));
if (!mype) {
cudaEventElapsedTime(&milliseconds, start, stop);
h_warp_lat[i] = (milliseconds * 1000.0) / (float)iter;
}
i++;
nvshmem_barrier_all();
}
i = 0;
for (num_elems = min_elems; num_elems <= max_elems; num_elems *= step_factor) {
h_size_array[i] = num_elems * 4;
CALL_BCAST_KERNEL(int32, _block, num_blocks, nvshm_test_num_tpb, args_1, stream);
CUDA_CHECK(cudaStreamSynchronize(stream));
nvshmem_barrier_all();
cudaEventRecord(start, stream);
CALL_BCAST_KERNEL(int32, _block, num_blocks, nvshm_test_num_tpb, args_2, stream);
cudaEventRecord(stop, stream);
CUDA_CHECK(cudaStreamSynchronize(stream));
if (!mype) {
cudaEventElapsedTime(&milliseconds, start, stop);
h_block_lat[i] = (milliseconds * 1000.0) / (float)iter;
}
i++;
nvshmem_barrier_all();
}
if (!mype) {
print_table_v1("bcast_device", "32-bit-thread", "size (Bytes)", "latency", "us", '-',
h_size_array, h_thread_lat, i);
print_table_v1("bcast_device", "32-bit-warp", "size (Bytes)", "latency", "us", '-',
h_size_array, h_warp_lat, i);
print_table_v1("bcast_device", "32-bit-block", "size (Bytes)", "latency", "us", '-',
h_size_array, h_block_lat, i);
}
min_elems = max(static_cast<size_t>(1), min_size / sizeof(int64_t));
max_elems = max(static_cast<size_t>(1), max_size / sizeof(int64_t));
i = 0;
for (num_elems = min_elems; num_elems < 512; num_elems *= step_factor) {
CALL_BCAST_KERNEL(int64, , num_blocks, nvshm_test_num_tpb, args_1, stream);
CUDA_CHECK(cudaStreamSynchronize(stream));
nvshmem_barrier_all();
cudaEventRecord(start, stream);
CALL_BCAST_KERNEL(int64, , num_blocks, nvshm_test_num_tpb, args_2, stream);
cudaEventRecord(stop, stream);
CUDA_CHECK(cudaStreamSynchronize(stream));
if (!mype) {
cudaEventElapsedTime(&milliseconds, start, stop);
h_thread_lat[i] = (milliseconds * 1000.0) / (float)iter;
}
i++;
nvshmem_barrier_all();
}
i = 0;
for (num_elems = min_elems; num_elems < 4096; num_elems *= step_factor) {
CALL_BCAST_KERNEL(int64, _warp, num_blocks, nvshm_test_num_tpb, args_1, stream);
CUDA_CHECK(cudaStreamSynchronize(stream));
nvshmem_barrier_all();
cudaEventRecord(start, stream);
CALL_BCAST_KERNEL(int64, _warp, num_blocks, nvshm_test_num_tpb, args_2, stream);
cudaEventRecord(stop, stream);
CUDA_CHECK(cudaStreamSynchronize(stream));
if (!mype) {
cudaEventElapsedTime(&milliseconds, start, stop);
h_warp_lat[i] = (milliseconds * 1000.0) / (float)iter;
}
i++;
nvshmem_barrier_all();
}
i = 0;
for (num_elems = min_elems; num_elems <= max_elems; num_elems *= step_factor) {
h_size_array[i] = num_elems * 8;
CALL_BCAST_KERNEL(int64, _block, num_blocks, nvshm_test_num_tpb, args_1, stream);
CUDA_CHECK(cudaStreamSynchronize(stream));
nvshmem_barrier_all();
cudaEventRecord(start, stream);
CALL_BCAST_KERNEL(int64, _block, num_blocks, nvshm_test_num_tpb, args_2, stream);
cudaEventRecord(stop, stream);
CUDA_CHECK(cudaStreamSynchronize(stream));
cudaEventElapsedTime(&milliseconds, start, stop);
cudaMemcpy(ms_d, &milliseconds, sizeof(float), cudaMemcpyHostToDevice);
nvshmem_float_sum_reduce(NVSHMEM_TEAM_WORLD, ms_sum_d, ms_d, 1);
cudaMemcpy(&milliseconds, ms_sum_d, sizeof(float), cudaMemcpyDeviceToHost);
if (!mype) {
h_block_lat[i] =
(milliseconds * 1000.0) / ((float)iter * nvshmem_team_n_pes(NVSHMEM_TEAM_WORLD));
}
i++;
nvshmem_barrier_all();
}
if (!mype) {
print_table_v1("bcast_device", "64-bit-thread", "size (Bytes)", "latency", "us", '-',
h_size_array, h_thread_lat, i);
print_table_v1("bcast_device", "64-bit-warp", "size (Bytes)", "latency", "us", '-',
h_size_array, h_warp_lat, i);
print_table_v1("bcast_device", "64-bit-block", "size (Bytes)", "latency", "us", '-',
h_size_array, h_block_lat, i);
}
return status;
}
int main(int argc, char **argv) {
int status = 0;
int mype, array_size;
read_args(argc, argv);
size_t size = max_size * 2;
size_t alloc_size;
DATATYPE *buffer = NULL;
DATATYPE *h_buffer = NULL;
DATATYPE *d_source, *d_dest;
DATATYPE *h_source, *h_dest;
int root = 0;
char size_string[100];
cudaStream_t cstrm;
void **h_tables;
array_size = max_size_log;
DEBUG_PRINT("symmetric size %lu\n", size);
sprintf(size_string, "%lu", size);
status = setenv("NVSHMEM_SYMMETRIC_SIZE", size_string, 1);
if (status) {
fprintf(stderr, "setenv failed \n");
status = -1;
goto out;
}
init_wrapper(&argc, &argv);
alloc_tables(&h_tables, 4, array_size);
if (use_cubin) {
init_cumodule(CUMODULE_NAME);
}
mype = nvshmem_my_pe();
CUDA_CHECK(cudaStreamCreateWithFlags(&cstrm, cudaStreamNonBlocking));
DEBUG_PRINT("SHMEM: [%d of %d] hello shmem world! \n", mype,
nvshmem_team_n_pes(NVSHMEM_TEAM_WORLD));
alloc_size = max_size * 2;
CUDA_CHECK(cudaHostAlloc(&h_buffer, alloc_size, cudaHostAllocDefault));
h_source = (DATATYPE *)h_buffer;
h_dest = (DATATYPE *)&h_source[max_size / sizeof(DATATYPE)];
buffer = (DATATYPE *)nvshmem_malloc(alloc_size);
if (!buffer) {
fprintf(stderr, "nvshmem_malloc failed \n");
status = -1;
goto out;
}
d_source = (DATATYPE *)buffer;
d_dest = (DATATYPE *)&d_source[max_size / sizeof(DATATYPE)];
for (int i = 0; i < max_size / sizeof(DATATYPE); i++) {
h_source[i] = i;
}
CUDA_CHECK(cudaMemcpyAsync(d_source, h_source, max_size, cudaMemcpyHostToDevice, cstrm));
CUDA_CHECK(cudaMemcpyAsync(d_dest, h_dest, max_size, cudaMemcpyHostToDevice, cstrm));
broadcast_calling_kernel(NVSHMEM_TEAM_WORLD, d_dest, d_source, mype, root, cstrm, h_tables);
CUDA_CHECK(cudaMemcpyAsync(h_source, d_source, max_size, cudaMemcpyDeviceToHost, cstrm));
CUDA_CHECK(cudaMemcpyAsync(h_dest, d_dest, max_size, cudaMemcpyDeviceToHost, cstrm));
nvshmem_barrier_all();
CUDA_CHECK(cudaFreeHost(h_buffer));
nvshmem_free(buffer);
CUDA_CHECK(cudaStreamDestroy(cstrm));
free_tables(h_tables, 4);
finalize_wrapper();
out:
return 0;
}