Skip to content

Migrate TBE UVM cache kernels to FBGEMM_LAUNCH_KERNEL #4817

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Migrate TBE UVM cache kernels to FBGEMM_LAUNCH_KERNEL #4817

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion fbgemm_gpu/src/split_embeddings_cache/lfu_cache_populate_byte.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -94,7 +94,7 @@ __launch_bounds__(kCacheMaxThreads) void lfu_cache_insert_byte_kernel(
const uint32_t sorted_slot = slots[0];
const int64_t sorted_lfu_cost = costs[0];

for (int32_t l = 0; l < min(SL, kWarpSize); ++l) {
for (auto l = 0; l < min(SL, kWarpSize); ++l) {
const int32_t insert_slot = shfl_sync(sorted_slot, l);
const int64_t insert_current_lfu_cost = shfl_sync(sorted_lfu_cost, l);
const index_t insert_idx = cache_set_sorted_indices[n + l];
Expand Down
116 changes: 50 additions & 66 deletions fbgemm_gpu/src/split_embeddings_cache/lru_cache_populate_byte.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -115,28 +115,24 @@ Tensor direct_mapped_lru_cache_find_uncached_cuda(
linear_cache_indices.scalar_type(),
"direct_mapped_lru_cache_find_uncached_cuda",
[&] {
#ifdef FBGEMM_GPU_MEMCHECK
const char* func_name = "direct_mapped_lru_cache_find_uncached_kernel";
#endif
// Find uncached indices
direct_mapped_lru_cache_find_uncached_kernel<<<
FBGEMM_LAUNCH_KERNEL(
(direct_mapped_lru_cache_find_uncached_kernel<index_t>),
std::min(
div_round_up(N, kMaxThreads),
get_max_thread_blocks_for_cache_kernels_()),
kMaxThreads,
0,
at::cuda::getCurrentCUDAStream()>>>(
MAKE_PTA_WITH_NAME(func_name, linear_cache_indices, index_t, 1, 32),
MAKE_PTA_WITH_NAME(func_name, cache_sets, int32_t, 1, 32),
at::cuda::getCurrentCUDAStream(),
PTA_B(linear_cache_indices, index_t, 1, 32),
PTA_B(cache_sets, int32_t, 1, 32),
max_indices,
MAKE_PTA_WITH_NAME(func_name, lxu_cache_state, int64_t, 2, 32),
PTA_B(lxu_cache_state, int64_t, 2, 32),
time_stamp,
MAKE_PTA_WITH_NAME(func_name, lru_state, int64_t, 2, 32),
PTA_B(lru_state, int64_t, 2, 32),
gather_cache_stats,
MAKE_PTA_WITH_NAME(func_name, uvm_cache_stats, int32_t, 1, 32),
MAKE_PTA_WITH_NAME(
func_name, lxu_cache_miss_timestamp, int64_t, 2, 32));
C10_CUDA_KERNEL_LAUNCH_CHECK();
PTA_B(uvm_cache_stats, int32_t, 1, 32),
PTA_B(lxu_cache_miss_timestamp, int64_t, 2, 32));
});

return cache_sets;
Expand Down Expand Up @@ -172,7 +168,7 @@ __global__ __launch_bounds__(kMaxThreads) void lru_cache_insert_byte_kernel(
const int64_t row_alignment) {
const int32_t C = lxu_cache_state.size(0);
int64_t n_conflict_misses = 0;
for (int32_t n = blockIdx.x * blockDim.y + threadIdx.y; n < *N_unique;
for (auto n = blockIdx.x * blockDim.y + threadIdx.y; n < *N_unique;
n += gridDim.x * blockDim.y) {
// check if this warp is responsible for this whole segment.
const bool segment_start =
Expand All @@ -197,16 +193,16 @@ __global__ __launch_bounds__(kMaxThreads) void lru_cache_insert_byte_kernel(

// now, we need to insert the (unique!) values in indices[n:n + SL] into
// our slots.
const int32_t slot = threadIdx.x;
const auto slot = threadIdx.x;
const int64_t slot_time = lru_state[cache_set][slot];
int64_t costs[1] = {slot_time};
int32_t slots[1] = {slot};
uint32_t slots[1] = {slot};

BitonicSort<int64_t, int32_t, 1, Comparator<int64_t>>::sort(costs, slots);
const int32_t sorted_slot = slots[0];
const int64_t sorted_lru_cost = costs[0];
BitonicSort<int64_t, uint32_t, 1, Comparator<int64_t>>::sort(costs, slots);
const auto sorted_slot = slots[0];
const auto sorted_lru_cost = costs[0];

for (int32_t l = 0; l < min(SL, kWarpSize); ++l) {
for (auto l = 0; l < min(SL, kWarpSize); ++l) {
const int32_t insert_slot = shfl_sync(sorted_slot, l);
const int64_t insert_current_lru_cost = shfl_sync(sorted_lru_cost, l);
if (insert_current_lru_cost == time_stamp) {
Expand All @@ -232,7 +228,7 @@ __global__ __launch_bounds__(kMaxThreads) void lru_cache_insert_byte_kernel(
&weights[weights_offset_insert + idx_insert * D_insert_bytes + 0]);
auto cache_row = reinterpret_cast<uint4*>(
&lxu_cache_weights[cache_set * kWarpSize + insert_slot][0]);
for (int32_t d = threadIdx.x; d * sizeof(uint4) < D_insert_bytes;
for (auto d = threadIdx.x; d * sizeof(uint4) < D_insert_bytes;
d += blockDim.x) {
cache_row[d] = row[d];
}
Expand Down Expand Up @@ -285,7 +281,7 @@ __launch_bounds__(kMaxThreads) void direct_mapped_lru_cache_insert_byte_kernel(

// one warp for each set (multiple times)
// (no divergence for each control branch)
for (int32_t pos = blockIdx.x * blockDim.y + threadIdx.y; pos < N;
for (auto pos = blockIdx.x * blockDim.y + threadIdx.y; pos < N;
pos += gridDim.x * blockDim.y) {
auto cache_set = cache_sets[pos];

Expand Down Expand Up @@ -346,7 +342,7 @@ __launch_bounds__(kMaxThreads) void direct_mapped_lru_cache_insert_byte_kernel(
auto row = reinterpret_cast<const uint4*>(
&weights[weights_offset_insert + idx_insert * D_insert_bytes + 0]);
auto cache_row = reinterpret_cast<uint4*>(&lxu_cache_weights[cache_set][0]);
for (int32_t d = threadIdx.x; d * sizeof(uint4) < D_insert_bytes;
for (auto d = threadIdx.x; d * sizeof(uint4) < D_insert_bytes;
d += blockDim.x) {
cache_row[d] = row[d];
}
Expand Down Expand Up @@ -398,36 +394,30 @@ void lru_cache_insert_byte_cuda(
cache_set_sorted_unique_indices.scalar_type(),
"lru_cache_insert_byte_cuda",
[&] {
#ifdef FBGEMM_GPU_MEMCHECK
const char* func_name = "lru_cache_insert_byte_kernel";
#endif
lru_cache_insert_byte_kernel<<<
FBGEMM_LAUNCH_KERNEL(
(lru_cache_insert_byte_kernel<index_t>),
std::min(
div_round_up(N, kMaxThreads / kWarpSize),
get_max_thread_blocks_for_cache_kernels_()),
dim3(kWarpSize, kMaxThreads / kWarpSize),
0,
at::cuda::getCurrentCUDAStream()>>>(
MAKE_PTA_WITH_NAME(func_name, weights, uint8_t, 1, 64),
MAKE_PTA_WITH_NAME(
func_name, cache_hash_size_cumsum, int64_t, 1, 32),
MAKE_PTA_WITH_NAME(
func_name, cache_index_table_map, int32_t, 1, 64),
MAKE_PTA_WITH_NAME(func_name, weights_offsets, int64_t, 1, 32),
MAKE_PTA_WITH_NAME(func_name, weights_tys, uint8_t, 1, 32),
MAKE_PTA_WITH_NAME(func_name, D_offsets, int32_t, 1, 32),
MAKE_PTA_WITH_NAME(func_name, sorted_cache_sets, int32_t, 1, 32),
MAKE_PTA_WITH_NAME(
func_name, cache_set_sorted_unique_indices, index_t, 1, 32),
at::cuda::getCurrentCUDAStream(),
PTA_B(weights, uint8_t, 1, 64),
PTA_B(cache_hash_size_cumsum, int64_t, 1, 32),
PTA_B(cache_index_table_map, int32_t, 1, 64),
PTA_B(weights_offsets, int64_t, 1, 32),
PTA_B(weights_tys, uint8_t, 1, 32),
PTA_B(D_offsets, int32_t, 1, 32),
PTA_B(sorted_cache_sets, int32_t, 1, 32),
PTA_B(cache_set_sorted_unique_indices, index_t, 1, 32),
unique_indices_length.data_ptr<int32_t>(),
MAKE_PTA_WITH_NAME(func_name, lxu_cache_state, int64_t, 2, 32),
MAKE_PTA_WITH_NAME(func_name, lxu_cache_weights, uint8_t, 2, 64),
PTA_B(lxu_cache_state, int64_t, 2, 32),
PTA_B(lxu_cache_weights, uint8_t, 2, 64),
time_stamp,
MAKE_PTA_WITH_NAME(func_name, lru_state, int64_t, 2, 32),
PTA_B(lru_state, int64_t, 2, 32),
gather_cache_stats,
MAKE_PTA_WITH_NAME(func_name, uvm_cache_stats, int32_t, 1, 32),
PTA_B(uvm_cache_stats, int32_t, 1, 32),
row_alignment);
C10_CUDA_KERNEL_LAUNCH_CHECK();
});
}

Expand Down Expand Up @@ -469,36 +459,30 @@ void direct_mapped_lru_cache_insert_byte_cuda(
linear_cache_indices.scalar_type(),
"direct_mapped_lru_cache_insert_byte_cuda",
[&] {
#ifdef FBGEMM_GPU_MEMCHECK
const char* func_name = "direct_mapped_lru_cache_insert_byte_kernel";
#endif
direct_mapped_lru_cache_insert_byte_kernel<<<
FBGEMM_LAUNCH_KERNEL(
(direct_mapped_lru_cache_insert_byte_kernel<index_t>),
std::min(
div_round_up(N, kMaxThreads / kWarpSize),
get_max_thread_blocks_for_cache_kernels_()),
dim3(kWarpSize, kMaxThreads / kWarpSize),
0,
at::cuda::getCurrentCUDAStream()>>>(
MAKE_PTA_WITH_NAME(func_name, weights, uint8_t, 1, 64),
MAKE_PTA_WITH_NAME(
func_name, cache_hash_size_cumsum, int64_t, 1, 32),
MAKE_PTA_WITH_NAME(
func_name, cache_index_table_map, int32_t, 1, 64),
MAKE_PTA_WITH_NAME(func_name, weights_offsets, int64_t, 1, 32),
MAKE_PTA_WITH_NAME(func_name, weights_tys, uint8_t, 1, 32),
MAKE_PTA_WITH_NAME(func_name, D_offsets, int32_t, 1, 32),
MAKE_PTA_WITH_NAME(func_name, lxu_cache_state, int64_t, 2, 32),
MAKE_PTA_WITH_NAME(func_name, lxu_cache_weights, uint8_t, 2, 64),
at::cuda::getCurrentCUDAStream(),
PTA_B(weights, uint8_t, 1, 64),
PTA_B(cache_hash_size_cumsum, int64_t, 1, 32),
PTA_B(cache_index_table_map, int32_t, 1, 64),
PTA_B(weights_offsets, int64_t, 1, 32),
PTA_B(weights_tys, uint8_t, 1, 32),
PTA_B(D_offsets, int32_t, 1, 32),
PTA_B(lxu_cache_state, int64_t, 2, 32),
PTA_B(lxu_cache_weights, uint8_t, 2, 64),
time_stamp,
MAKE_PTA_WITH_NAME(func_name, lru_state, int64_t, 2, 32),
MAKE_PTA_WITH_NAME(func_name, linear_cache_indices, index_t, 1, 32),
MAKE_PTA_WITH_NAME(
func_name, lxu_cache_miss_timestamp, int64_t, 2, 32),
MAKE_PTA_WITH_NAME(func_name, cache_sets, int32_t, 1, 32),
PTA_B(lru_state, int64_t, 2, 32),
PTA_B(linear_cache_indices, index_t, 1, 32),
PTA_B(lxu_cache_miss_timestamp, int64_t, 2, 32),
PTA_B(cache_sets, int32_t, 1, 32),
gather_cache_stats,
MAKE_PTA_WITH_NAME(func_name, uvm_cache_stats, int32_t, 1, 32),
PTA_B(uvm_cache_stats, int32_t, 1, 32),
row_alignment);
C10_CUDA_KERNEL_LAUNCH_CHECK();
});
}

Expand Down
Loading