Skip to content
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

Jump to bottom

Add support for int64_t indices and offsets in TBE inference [6/N] #3182

Open
wants to merge 6 commits into
base: main
Choose a base branch
from
+457 −209
Open

Add support for int64_t indices and offsets in TBE inference [6/N] #3182

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
ebbe476
Add support for int64_t indices in TBE inference [1/N]
Sep 26, 2024
fb08393
Add support for int64_t indices in TBE inference [2/N]
Sep 26, 2024
d67f9c0
Add support for int64_t indices and offsets in TBE inference [3/N]
Sep 26, 2024
278dbe5
Add support for int64_t indices and offsets in TBE inference [4/N]
Sep 26, 2024
681c82e
Add support for int64_t indices and offsets in TBE inference [5/N]
Sep 26, 2024
948a8bb
Add support for int64_t indices and offsets in TBE inference [6/N] (#…
q10 Sep 26, 2024
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
153 changes: 93 additions & 60 deletions fbgemm_gpu/codegen/inference/embedding_forward_quantized_cpu_template.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -41,6 +41,16 @@ inline uint32_t pruned_hash_function(uint32_t h) {
return h;
}

inline uint64_t pruned_hash_function(uint64_t k) {
// MurmorHash3 64-bit mixing function.
k ^= k >> 33;
k *= (0xff51afd7ed558ccd);
k ^= k >> 33;
k *= (0xc4ceb9fe1a85ec53);
k ^= k >> 33;
return k;
}

} // namespace

void pruned_hashmap_insert_{{ wdesc }}_cpu(
Expand Down Expand Up @@ -404,58 +414,72 @@ Tensor pruned_hashmap_lookup_{{ wdesc }}_cpu(
TENSOR_ON_CPU(offsets);
TENSOR_ON_CPU(hash_table);
TENSOR_ON_CPU(hash_table_offsets);
TENSORS_HAVE_SAME_SCALAR_TYPE(indices, offsets, hash_table);

int32_t T = hash_table_offsets.size(0) - 1;
int32_t B = (offsets.size(0) - 1) / T;
TORCH_CHECK(B > 0);

auto dense_indices = empty_like(indices);
const auto* indices_acc = indices.data_ptr<int32_t>();
auto* dense_indices_acc = dense_indices.data_ptr<int32_t>();

const auto* offsets_acc = offsets.data_ptr<int32_t>();
const auto hash_table_acc = hash_table.accessor<int32_t, 2>();
const auto hash_table_offsets_acc = hash_table_offsets.accessor<int64_t, 1>();
for (const auto t : c10::irange(T)) {
int64_t table_start = hash_table_offsets_acc[t];
int64_t table_end = hash_table_offsets_acc[t + 1];
int64_t capacity = table_end - table_start;
for (const auto b : c10::irange(B)) {
int32_t indices_start = offsets_acc[t * B + b];
int32_t indices_end = offsets_acc[t * B + b + 1];
int32_t L = indices_end - indices_start;
AT_DISPATCH_INDEX_TYPES(indices.scalar_type(), "pruned_hashmap_lookup_{{ wdesc }}_cpu", [&] {
using hash_t =
std::conditional_t<std::is_same_v<index_t, int64_t>, uint64_t, uint32_t>;

if (table_start == table_end) {
for (const auto l : c10::irange(L)) {
dense_indices_acc[indices_start + l] = indices_acc[indices_start + l];
}
} else {
for (const auto l : c10::irange(L)) {
int32_t idx = indices_acc[indices_start + l];
uint32_t slot = pruned_hash_function(static_cast<uint32_t>(idx)) % capacity;
while (true) {
int32_t slot_sparse_idx = hash_table_acc[table_start + static_cast<int64_t>(slot)][0];

// empty slot
if (slot_sparse_idx == -1) {
dense_indices_acc[indices_start + l] = -1;
break;
}
// already exists
if (slot_sparse_idx == idx) {
dense_indices_acc[indices_start + l] = hash_table_acc[table_start + static_cast<int64_t>(slot)][1];
break;
const auto* indices_acc = indices.data_ptr<index_t>();
auto* dense_indices_acc = dense_indices.data_ptr<index_t>();

const auto* offsets_acc = offsets.data_ptr<index_t>();
const auto hash_table_acc = hash_table.accessor<index_t, 2>();
const auto hash_table_offsets_acc = hash_table_offsets.accessor<int64_t, 1>();

for (const auto t : c10::irange(T)) {
const auto table_start = hash_table_offsets_acc[t];
const auto table_end = hash_table_offsets_acc[t + 1];
const auto capacity = table_end - table_start;

for (const auto b : c10::irange(B)) {
const auto indices_start = offsets_acc[t * B + b];
const auto indices_end = offsets_acc[t * B + b + 1];
const auto L = indices_end - indices_start;

if (table_start == table_end) {
for (const auto l : c10::irange(L)) {
dense_indices_acc[indices_start + l] = indices_acc[indices_start + l];
}

} else {
for (const auto l : c10::irange(L)) {
const auto idx = indices_acc[indices_start + l];
auto slot = pruned_hash_function(static_cast<hash_t>(idx)) % capacity;

while (true) {
const auto slot_sparse_idx = hash_table_acc[table_start + static_cast<int64_t>(slot)][0];

// empty slot
if (slot_sparse_idx == -1) {
dense_indices_acc[indices_start + l] = -1;
break;
}
// already exists
if (slot_sparse_idx == idx) {
dense_indices_acc[indices_start + l] = hash_table_acc[table_start + static_cast<int64_t>(slot)][1];
break;
}
// linear probe
slot = (slot + 1) % capacity;
}
// linear probe
slot = (slot + 1) % capacity;
}
}
}
}
}
});

return dense_indices;
}

{% if not weighted %}

Tensor pruned_array_lookup_cpu(
Tensor indices,
Tensor offsets,
Expand All @@ -465,37 +489,46 @@ Tensor pruned_array_lookup_cpu(
TENSOR_ON_CPU(offsets);
TENSOR_ON_CPU(index_remappings);
TENSOR_ON_CPU(index_remappings_offsets);
TENSORS_HAVE_SAME_SCALAR_TYPE(indices, offsets, index_remappings);

int32_t T = index_remappings_offsets.size(0) - 1;
int32_t B = (offsets.size(0) - 1) / T;
TORCH_CHECK(B > 0);

auto dense_indices = empty_like(indices);
const auto* indices_acc = indices.data_ptr<int32_t>();
auto* dense_indices_acc = dense_indices.data_ptr<int32_t>();
const auto* offsets_acc = offsets.data_ptr<int32_t>();

const auto index_remappings_acc = index_remappings.data_ptr<int32_t>();
const auto index_remappings_offsets_acc = index_remappings_offsets.data_ptr<int64_t>();
at::parallel_for(0, T, 1, [&](int64_t begin, int64_t end) {
for (const auto t : c10::irange(begin, end)) {
int64_t index_remappings_start = index_remappings_offsets_acc[t];
int64_t index_remappings_end = index_remappings_offsets_acc[t + 1];
int64_t capacity = index_remappings_end - index_remappings_start;
int32_t indices_start = offsets_acc[t * B];
int32_t indices_end = offsets_acc[(t + 1) * B];
if (capacity > 0) {
for (const auto i : c10::irange(indices_start,indices_end)) {
int32_t idx = indices_acc[i];
dense_indices_acc[i] = index_remappings_acc[index_remappings_start + idx];
}
} else {
std::memcpy(
dense_indices_acc + indices_start,
indices_acc + indices_start,
(indices_end - indices_start) * sizeof(int32_t));
}
}
AT_DISPATCH_INDEX_TYPES(indices.scalar_type(), "pruned_array_lookup_cpu", [&] {
const auto* indices_acc = indices.data_ptr<index_t>();
auto* dense_indices_acc = dense_indices.data_ptr<index_t>();
const auto* offsets_acc = offsets.data_ptr<index_t>();

const auto index_remappings_acc = index_remappings.data_ptr<index_t>();
const auto index_remappings_offsets_acc = index_remappings_offsets.data_ptr<int64_t>();

at::parallel_for(0, T, 1, [&](int64_t begin, int64_t end) {
for (const auto t : c10::irange(begin, end)) {
const auto index_remappings_start = index_remappings_offsets_acc[t];
const auto index_remappings_end = index_remappings_offsets_acc[t + 1];
const auto capacity = index_remappings_end - index_remappings_start;

const auto indices_start = offsets_acc[t * B];
const auto indices_end = offsets_acc[(t + 1) * B];

if (capacity > 0) {
for (const auto i : c10::irange(indices_start, indices_end)) {
auto idx = indices_acc[i];
dense_indices_acc[i] = index_remappings_acc[index_remappings_start + idx];
}
} else {
std::memcpy(
dense_indices_acc + indices_start,
indices_acc + indices_start,
(indices_end - indices_start) * sizeof(index_t));
}
}
});
});

return dense_indices;
}

Expand Down
39 changes: 24 additions & 15 deletions fbgemm_gpu/codegen/inference/embedding_forward_quantized_host_cpu.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -21,6 +21,7 @@
#include "fbgemm_gpu/embedding_common.h"
#include "fbgemm_gpu/utils/dispatch_macros.h"
#include "fbgemm_gpu/utils/ops_utils.h"
#include "fbgemm_gpu/utils/tensor_utils.h"

using Tensor = at::Tensor;
using namespace fbgemm_gpu;
Expand Down Expand Up @@ -374,29 +375,37 @@ class PrunedMapCPU : public torch::jit::CustomClassHolder {
}

Tensor lookup(Tensor indices, Tensor offsets) const {
TENSORS_HAVE_SAME_SCALAR_TYPE(indices, offsets);

int32_t T = maps_.size();
TORCH_CHECK(T > 0);
int32_t B = (offsets.size(0) - 1) / T;
TORCH_CHECK(B > 0);
TORCH_CHECK(maps_.size() == T);

auto dense_indices = empty_like(indices);
const auto* indices_acc = indices.data_ptr<int32_t>();
auto* dense_indices_acc = dense_indices.data_ptr<int32_t>();
const auto* offsets_acc = offsets.data_ptr<int32_t>();
for (const auto t : c10::irange(T)) {
auto& map = maps_[t];
for (const auto b : c10::irange(B)) {
int32_t indices_start = offsets_acc[t * B + b];
int32_t indices_end = offsets_acc[t * B + b + 1];
int32_t L = indices_end - indices_start;
for (const auto l : c10::irange(L)) {
int32_t slot_sparse_index = indices_acc[indices_start + l];
auto it = map.find(slot_sparse_index);
dense_indices_acc[indices_start + l] =
it != map.end() ? it->second : -1;

AT_DISPATCH_INDEX_TYPES(indices.scalar_type(), "PrunedMapCPU::lookup", [&] {
const auto* indices_acc = indices.data_ptr<index_t>();
auto* dense_indices_acc = dense_indices.data_ptr<index_t>();
const auto* offsets_acc = offsets.data_ptr<index_t>();

for (const auto t : c10::irange(T)) {
auto& map = maps_[t];
for (const auto b : c10::irange(B)) {
const auto indices_start = offsets_acc[t * B + b];
const auto indices_end = offsets_acc[t * B + b + 1];
const auto L = indices_end - indices_start;
for (const auto l : c10::irange(L)) {
const auto slot_sparse_index = indices_acc[indices_start + l];
const auto it = map.find(slot_sparse_index);
dense_indices_acc[indices_start + l] =
it != map.end() ? it->second : -1;
}
}
}
}
});

return dense_indices;
}

Expand Down
Loading
Loading