Files
concrete/compilers/concrete-compiler/compiler/lib/Runtime/GPUDFG.cpp

1655 lines
67 KiB
C++

// Part of the Concrete Compiler Project, under the BSD3 License with Zama
// Exceptions. See
// https://github.com/zama-ai/concrete/blob/main/LICENSE.txt
// for license information.
#include <atomic>
#include <cmath>
#include <cstdarg>
#include <err.h>
#include <hwloc.h>
#include <iostream>
#include <list>
#include <memory>
#include <mutex>
#include <numeric>
#include <queue>
#include <thread>
#include <utility>
#include <vector>
#include <concretelang/Runtime/stream_emulator_api.h>
#include <concretelang/Runtime/wrappers.h>
#ifdef CONCRETELANG_CUDA_SUPPORT
#include "bootstrap.h"
#include "device.h"
#include "keyswitch.h"
#include "linear_algebra.h"
using RuntimeContext = mlir::concretelang::RuntimeContext;
namespace mlir {
namespace concretelang {
namespace gpu_dfg {
namespace {
using MemRef2 = MemRefDescriptor<2>;
// When not using all accelerators on the machine, we distribute work
// by assigning the default accelerator for each SDFG to next_device
// round-robin.
static std::atomic<size_t> next_device = {0};
// Resources available (or set as requested by user through
// environment variables) on the machine. Defaults to using all
// available resources.
static size_t num_devices = 0; // Set SDFG_NUM_GPUS to configure
static size_t num_cores = 1; // Set SDFG_NUM_THREADS to configure
static size_t device_compute_factor = 16; // Set SDFG_DEVICE_TO_CORE_RATIO
// How much more memory than just input size is required on GPU to execute
static float gpu_memory_inflation_factor = 1.5;
// Get the byte size of a rank 2 MemRef
static inline size_t memref_get_data_size(MemRef2 &m) {
return m.sizes[0] * m.sizes[1] * sizeof(uint64_t);
}
// Copy contiguous rank 2 MemRef 'in' to 'out'
static inline void memref_copy_contiguous(MemRef2 &out, MemRef2 &in) {
assert(in.sizes[0] == out.sizes[0] && in.sizes[1] == out.sizes[1] &&
"memref_copy_contiguous sizes differ");
assert(in.strides[0] == out.strides[0] && in.strides[1] == out.strides[1] &&
"memref_copy_contiguous strides differ");
assert(in.strides[0] == in.sizes[1] && in.strides[1] == 1 &&
"memref_copy_contiguous strides not compatible with contiguous "
"storage.");
memcpy(out.aligned + out.offset, in.aligned + in.offset,
memref_get_data_size(in));
}
// Copy contiguous rank 2 MemRef to a newly allocated, returned MemRef
static inline MemRef2 memref_copy_alloc(MemRef2 &m) {
uint64_t *data = (uint64_t *)malloc(memref_get_data_size(m));
MemRef2 ret = {
data, data, 0, {m.sizes[0], m.sizes[1]}, {m.strides[0], m.strides[1]}};
memref_copy_contiguous(ret, m);
return ret;
}
// Parameter storage for KS and BS processes
struct Void {};
union Param {
Void _;
uint32_t val;
};
union Context {
Void _;
RuntimeContext *val;
};
// Tracking locations and state for dependences: location is either an
// integer >= 0 corresponding to the device/accelerator index on the
// machine or -1/-2 when it is only available on the host or is split
// across multiple locations.
static const int32_t host_location = -1;
static const int32_t split_location = -2;
// Similarly dependence chunks are either indexed (which does not
// always correlate to the device index on which they are located) or
// this dependence is split further.
static const int32_t single_chunk = -1;
static const int32_t split_chunks = -2;
struct Stream;
struct Dependence;
// Track buffer/scratchpad for the PBS to avoid re-allocating it on
// the device. Reuse where possible or reallocate if a larger buffer
// is required.
struct PBS_buffer {
PBS_buffer(void *stream, uint32_t gpu_idx, uint32_t glwe_dimension,
uint32_t polynomial_size, uint32_t input_lwe_ciphertext_count)
: max_pbs_buffer_samples(input_lwe_ciphertext_count),
glwe_dim(glwe_dimension), poly_size(polynomial_size),
gpu_stream(stream), gpu_index(gpu_idx) {
scratch_cuda_bootstrap_amortized_64(
gpu_stream, gpu_index, &pbs_buffer, glwe_dim, poly_size,
max_pbs_buffer_samples, cuda_get_max_shared_memory(gpu_index), true);
}
~PBS_buffer() {
cleanup_cuda_bootstrap_amortized(gpu_stream, gpu_index, &pbs_buffer);
}
int8_t *get_pbs_buffer(void *stream, uint32_t gpu_idx,
uint32_t glwe_dimension, uint32_t polynomial_size,
uint32_t input_lwe_ciphertext_count) {
assert(glwe_dimension <= glwe_dim);
assert(polynomial_size <= poly_size);
assert(input_lwe_ciphertext_count <= max_pbs_buffer_samples);
assert(stream == gpu_stream);
assert(gpu_idx == gpu_index);
return pbs_buffer;
}
uint32_t get_max_pbs_buffer_samples() { return max_pbs_buffer_samples; }
int8_t *pbs_buffer;
uint32_t max_pbs_buffer_samples;
uint32_t glwe_dim;
uint32_t poly_size;
void *gpu_stream;
uint32_t gpu_index;
};
// Keep track of the GPU/CUDA streams used for each accelerator and
// associated PBS buffer.
struct GPU_state {
uint32_t gpu_idx;
void *gpu_stream;
PBS_buffer *pbs_buffer;
GPU_state(uint32_t idx)
: gpu_idx(idx), gpu_stream(nullptr), pbs_buffer(nullptr) {}
~GPU_state() {
if (pbs_buffer != nullptr)
delete pbs_buffer;
if (gpu_stream != nullptr)
cuda_destroy_stream((cudaStream_t *)gpu_stream, gpu_idx);
}
inline int8_t *get_pbs_buffer(uint32_t glwe_dimension,
uint32_t polynomial_size,
uint32_t input_lwe_ciphertext_count) {
if (pbs_buffer != nullptr && (pbs_buffer->glwe_dim != glwe_dimension ||
pbs_buffer->poly_size != polynomial_size ||
pbs_buffer->get_max_pbs_buffer_samples() <
input_lwe_ciphertext_count)) {
delete pbs_buffer;
pbs_buffer = nullptr;
}
if (pbs_buffer == nullptr)
pbs_buffer = new PBS_buffer(get_gpu_stream(), gpu_idx, glwe_dimension,
polynomial_size, input_lwe_ciphertext_count);
return pbs_buffer->get_pbs_buffer(get_gpu_stream(), gpu_idx, glwe_dimension,
polynomial_size,
input_lwe_ciphertext_count);
}
inline void *get_gpu_stream() {
if (gpu_stream == nullptr)
gpu_stream = cuda_create_stream(gpu_idx);
return gpu_stream;
}
};
// Track resources required for the execution of a single DFG,
// including the GPU states of devices involved in its execution,
// streams and memory allocated which depends on execution progress on
// accelerators before it can be freed. As execution on accelerators
// is asynchronous, this must wait for the next synchronization point.
struct GPU_DFG {
std::vector<GPU_state> gpus;
uint32_t gpu_idx;
void *gpu_stream;
GPU_DFG(uint32_t idx) : gpu_idx(idx), pbs_buffer(nullptr) {
for (uint32_t i = 0; i < num_devices; ++i)
gpus.push_back(std::move(GPU_state(i)));
gpu_stream = gpus[idx].get_gpu_stream();
}
~GPU_DFG() {
free_streams();
free_stream_order_dependent_data();
}
inline void register_stream(Stream *s) { streams.push_back(s); }
inline void register_stream_order_dependent_allocation(void *p) {
std::lock_guard<std::mutex> guard(free_list_guard);
to_free_list.push_back(p);
}
inline void free_stream_order_dependent_data() {
std::lock_guard<std::mutex> guard(free_list_guard);
for (auto p : to_free_list)
free(p);
to_free_list.clear();
}
inline int8_t *get_pbs_buffer(uint32_t glwe_dimension,
uint32_t polynomial_size,
uint32_t input_lwe_ciphertext_count) {
if (pbs_buffer == nullptr) {
int8_t *ret = gpus[gpu_idx].get_pbs_buffer(
glwe_dimension, polynomial_size, input_lwe_ciphertext_count);
pbs_buffer = gpus[gpu_idx].pbs_buffer;
return ret;
}
return pbs_buffer->get_pbs_buffer(gpu_stream, gpu_idx, glwe_dimension,
polynomial_size,
input_lwe_ciphertext_count);
}
void free_streams() {
streams.sort();
streams.unique();
for (auto s : streams)
delete s;
}
inline void *get_gpu_stream(int32_t loc) {
if (loc < 0)
return nullptr;
return gpus[loc].get_gpu_stream();
}
private:
std::list<void *> to_free_list;
std::mutex free_list_guard;
std::list<Stream *> streams;
PBS_buffer *pbs_buffer;
};
struct Dependence;
static void sdfg_gpu_debug_print_mref(const char *c, MemRef2 m);
static MemRef2 sdfg_gpu_debug_dependence(Dependence *d, cudaStream_t *s);
static bool sdfg_gpu_debug_compare_memref(MemRef2 &a, MemRef2 &b,
char const *msg);
// Dependences track the location and state of each block of memory
// used as input/output to processes to allow either moving it on/off
// devices or determining when deallocation is possible.
struct Dependence {
int32_t location;
MemRef2 host_data;
void *device_data;
bool onHostReady;
bool hostAllocated;
int32_t chunk_id;
size_t stream_generation;
std::vector<Dependence *> chunks;
Dependence(int32_t l, MemRef2 hd, void *dd, bool ohr, bool alloc = false,
int32_t chunk_id = single_chunk, size_t gen = 0)
: location(l), host_data(hd), device_data(dd), onHostReady(ohr),
hostAllocated(alloc), chunk_id(chunk_id), stream_generation(gen) {}
Dependence(int32_t l, uint64_t val, void *dd, bool ohr, bool alloc = false,
int32_t chunk_id = single_chunk, size_t gen = 0)
: location(l), device_data(dd), onHostReady(ohr), hostAllocated(alloc),
chunk_id(chunk_id), stream_generation(gen) {
*host_data.aligned = val;
}
// Split a dependence into a number of chunks either to run on
// multiple GPUs or execute concurrently on the host.
void split_dependence(size_t num_chunks, size_t num_gpu_chunks,
size_t chunk_dim, bool constant) {
// If this dependence is already split, check that the split
// matches the new request
if (chunk_id == split_chunks) {
if (num_chunks + num_gpu_chunks != chunks.size())
warnx("WARNING: requesting to split dependence across different number "
"of chunks (%lu) than it already is split (%lu) which would "
"require remapping. This is not supported.",
num_chunks + num_gpu_chunks, chunks.size());
assert(num_chunks + num_gpu_chunks == chunks.size());
return;
}
if (!chunks.empty()) {
for (auto c : chunks)
delete c;
chunks.clear();
}
chunks.resize(num_chunks + num_gpu_chunks, nullptr);
assert(onHostReady && "Cannot split dependences located on a device.");
size_t num_samples = host_data.sizes[chunk_dim];
assert(num_samples > 0);
// If this is a constant (same data for each chunk), then copy a
// descriptor corresponding to the whole dependence for each
// chunk.
if (constant) {
for (size_t i = 0; i < num_chunks + num_gpu_chunks; ++i) {
MemRef2 m = host_data;
m.allocated = nullptr;
chunks[i] = new Dependence(host_location, m, nullptr, onHostReady,
false, i, stream_generation);
}
return;
}
size_t chunk_size =
num_samples / (num_chunks + num_gpu_chunks * device_compute_factor);
size_t gpu_chunk_size = chunk_size * device_compute_factor;
chunk_size = (num_samples - gpu_chunk_size * num_gpu_chunks) / num_chunks;
size_t chunk_remainder =
(num_samples - gpu_chunk_size * num_gpu_chunks) % num_chunks;
uint64_t offset = 0;
for (size_t i = 0; i < num_chunks; ++i) {
size_t chunk_size_ = (i < chunk_remainder) ? chunk_size + 1 : chunk_size;
MemRef2 m = host_data;
m.sizes[chunk_dim] = chunk_size_;
m.offset = offset + host_data.offset;
void *dd = (device_data == nullptr) ? device_data
: (uint64_t *)device_data + offset;
offset += chunk_size_ * host_data.strides[chunk_dim];
chunks[i] = new Dependence(location, m, dd, onHostReady, false, i,
stream_generation);
}
for (size_t i = num_chunks; i < num_chunks + num_gpu_chunks; ++i) {
MemRef2 m = host_data;
m.sizes[chunk_dim] = gpu_chunk_size;
m.offset = offset + host_data.offset;
void *dd = (device_data == nullptr) ? device_data
: (uint64_t *)device_data + offset;
offset += gpu_chunk_size * host_data.strides[chunk_dim];
chunks[i] = new Dependence(location, m, dd, onHostReady, false, i,
stream_generation);
}
chunk_id = split_chunks;
location = split_location;
}
void merge_dependence(GPU_DFG *dfg) {
assert(!chunks.empty() && "Cannot merge dependence with no chunks");
size_t data_size = 0;
size_t num_samples = 0;
for (auto c : chunks) {
data_size += memref_get_data_size(c->host_data);
num_samples += c->host_data.sizes[0];
}
uint64_t *data = (uint64_t *)malloc(data_size);
MemRef2 output = {data,
data,
0,
{num_samples, chunks.front()->host_data.sizes[1]},
{chunks.front()->host_data.sizes[1], 1}};
std::list<cudaStream_t *> custreams_used;
for (auto c : chunks) {
// Write out the piece in the final target dependence
size_t csize = memref_get_data_size(c->host_data);
if (c->onHostReady) {
memcpy(((char *)output.aligned) + output.offset, c->host_data.aligned,
csize);
} else {
assert(c->location > host_location);
cudaStream_t *s = (cudaStream_t *)dfg->get_gpu_stream(c->location);
cuda_memcpy_async_to_cpu(((char *)output.aligned) + output.offset,
c->device_data, csize, s, c->location);
custreams_used.push_back(s);
}
output.offset += csize;
}
output.offset = 0;
for (auto c : chunks)
c->free_data(dfg, true);
chunks.clear();
custreams_used.sort();
custreams_used.unique();
for (auto s : custreams_used)
cudaStreamSynchronize(*s);
location = host_location;
onHostReady = true;
assert(host_data.allocated == nullptr);
host_data = output;
assert(device_data == nullptr);
hostAllocated = true;
chunk_id = single_chunk;
}
void move_chunk_off_device(int32_t chunk_id, GPU_DFG *dfg) {
chunks[chunk_id]->copy(host_location, dfg);
cuda_drop_async(
chunks[chunk_id]->device_data,
(cudaStream_t *)dfg->get_gpu_stream(chunks[chunk_id]->location),
chunks[chunk_id]->location);
chunks[chunk_id]->location = host_location;
}
void free_chunk_host_data(int32_t chunk_id, GPU_DFG *dfg) {
assert(chunks[chunk_id]->location == host_location &&
chunks[chunk_id]->onHostReady && chunks[chunk_id]->hostAllocated);
free(chunks[chunk_id]->host_data.allocated);
chunks[chunk_id]->host_data.allocated = nullptr;
chunks[chunk_id]->hostAllocated = false;
chunks[chunk_id]->onHostReady = false;
}
void free_chunk_device_data(int32_t chunk_id, GPU_DFG *dfg) {
assert(chunks[chunk_id]->location > host_location &&
chunks[chunk_id]->device_data != nullptr);
cuda_drop_async(
chunks[chunk_id]->device_data,
(cudaStream_t *)dfg->get_gpu_stream(chunks[chunk_id]->location),
chunks[chunk_id]->location);
chunks[chunk_id]->device_data = nullptr;
}
inline void free_data(GPU_DFG *dfg, bool immediate = false) {
if (location >= 0 && device_data != nullptr) {
cuda_drop_async(device_data,
(cudaStream_t *)dfg->get_gpu_stream(location), location);
}
if (onHostReady && host_data.allocated != nullptr && hostAllocated) {
// As streams are not synchronized aside from the GET operation,
// we cannot free host-side data until after the synchronization
// point as it could still be used by an asynchronous operation.
if (immediate)
free(host_data.allocated);
else
dfg->register_stream_order_dependent_allocation(host_data.allocated);
}
for (auto c : chunks)
c->free_data(dfg, immediate);
chunks.clear();
delete (this);
}
inline void copy(int32_t loc, GPU_DFG *dfg) {
size_t data_size = memref_get_data_size(host_data);
if (loc == location)
return;
if (loc == host_location) {
if (onHostReady)
return;
if (host_data.allocated == nullptr) {
host_data.allocated = host_data.aligned = (uint64_t *)malloc(data_size);
hostAllocated = true;
}
cudaStream_t *s = (cudaStream_t *)dfg->get_gpu_stream(location);
cuda_memcpy_async_to_cpu(host_data.aligned, device_data, data_size, s,
location);
cudaStreamSynchronize(*s);
onHostReady = true;
} else {
assert(onHostReady &&
"Device-to-device data transfers not supported yet.");
cudaStream_t *s = (cudaStream_t *)dfg->get_gpu_stream(loc);
if (device_data != nullptr)
cuda_drop_async(device_data, s, location);
device_data = cuda_malloc_async(data_size, s, loc);
cuda_memcpy_async_to_gpu(
device_data, host_data.aligned + host_data.offset, data_size, s, loc);
location = loc;
}
}
};
// Set of input/output streams required to execute a process'
// activation, along with any parameters (for KS / BS) and the
// associated work-function.
struct Process {
std::vector<Stream *> input_streams;
std::vector<Stream *> output_streams;
GPU_DFG *dfg;
Param level;
Param base_log;
Param input_lwe_dim;
Param output_lwe_dim;
Param poly_size;
Param glwe_dim;
Param sk_index;
Param output_size;
Context ctx;
void (*fun)(Process *, int32_t, int32_t, uint64_t *);
char name[80];
bool batched_process;
};
void memref_keyswitch_lwe_u64_process(Process *p, int32_t loc, int32_t chunk_id,
uint64_t *out_ptr);
void memref_bootstrap_lwe_u64_process(Process *p, int32_t loc, int32_t chunk_id,
uint64_t *out_ptr);
void memref_add_lwe_ciphertexts_u64_process(Process *p, int32_t loc,
int32_t chunk_id,
uint64_t *out_ptr);
void memref_add_plaintext_lwe_ciphertext_u64_process(Process *p, int32_t loc,
int32_t chunk_id,
uint64_t *out_ptr);
void memref_mul_cleartext_lwe_ciphertext_u64_process(Process *p, int32_t loc,
int32_t chunk_id,
uint64_t *out_ptr);
void memref_negate_lwe_ciphertext_u64_process(Process *p, int32_t loc,
int32_t chunk_id,
uint64_t *out_ptr);
static inline void schedule_kernel(Process *p, int32_t loc, int32_t chunk_id,
uint64_t *out_ptr) {
p->fun(p, loc, chunk_id, out_ptr);
}
struct Stream {
Dependence *dep;
Dependence *saved_dependence;
stream_type type;
Process *producer;
std::vector<Process *> consumers;
GPU_DFG *dfg;
bool batched_stream;
bool const_stream; // When a batched op uses the same value on this
// stream for all ops
bool ct_stream;
bool pt_stream;
size_t generation;
const char *name;
Stream(stream_type t, const char *sname = nullptr)
: dep(nullptr), type(t), producer(nullptr), dfg(nullptr),
batched_stream(false), const_stream(false), ct_stream(false),
pt_stream(false), generation(0) {
if (sname == nullptr) {
static unsigned long stream_id = 0;
char *n = new char[16];
snprintf(n, 16, "stream%llu", stream_id++);
name = n;
} else {
name = sname;
}
}
~Stream() {
if (dep != nullptr)
dep->free_data(dfg);
if (producer != nullptr)
delete producer;
}
void put(Dependence *d, int32_t chunk_id = single_chunk) {
assert(chunk_id != split_chunks);
// If this is a chunk of a larger dependence, then insert it in its
// position.
if (chunk_id != single_chunk) {
// We must always have a dependence present and its split into
// chunks pre-determined here.
assert(dep != nullptr);
if (dep->chunks[chunk_id] != nullptr)
dep->chunks[chunk_id]->free_data(dfg, true);
assert(dep->chunks.size() > chunk_id);
dep->chunks[chunk_id] = d;
} else {
// If a dependence was already present, schedule deallocation.
if (dep != nullptr)
dep->free_data(dfg);
dep = d;
}
dep->stream_generation = generation;
}
// For a given dependence, traverse the DFG backwards to extract the lattice
// of kernels required to execute to produce this data
void extract_producing_graph(std::list<Process *> &queue) {
// If there's no producer process for this stream, it is fed by
// the control program - nothing to do
if (producer == nullptr) {
assert(dep != nullptr && "Data missing on control program stream.");
return;
}
// Recursively go up the DFG to check if new data is available
for (auto s : producer->input_streams)
s->extract_producing_graph(queue);
// Check if any of the inputs have changed - and if so recompute
// this value. Do not recompute if no changes.
size_t input_generation = 0;
for (auto s : producer->input_streams)
input_generation += s->generation;
if (generation < input_generation) {
queue.push_back(producer);
generation = input_generation;
}
}
void schedule_work(MemRef2 &out) {
std::list<Process *> queue;
extract_producing_graph(queue);
// Determine if this subgrah bootstraps
bool is_batched_subgraph = false;
size_t subgraph_bootstraps = 0;
for (auto p : queue) {
is_batched_subgraph |= p->batched_process;
subgraph_bootstraps +=
(p->fun == memref_bootstrap_lwe_u64_process) ? 1 : 0;
}
// If this subgraph is not batched, then use this DFG's allocated
// GPU to offload to. If this does not bootstrap, just execute on
// the host.
if (!is_batched_subgraph) {
for (auto p : queue) {
schedule_kernel(
p, (subgraph_bootstraps > 0) ? dfg->gpu_idx : host_location,
single_chunk, nullptr);
}
return;
}
// Identify all inputs to these processes that are not also
// outputs (i.e. data not produced within this subgraph) - and
// outputs of the subgraph.
std::list<Stream *> inputs;
std::list<Stream *> inputs_all;
std::list<Stream *> outputs;
std::list<Stream *> intermediate_values;
for (auto p : queue) {
inputs.insert(inputs.end(), p->input_streams.begin(),
p->input_streams.end());
outputs.insert(outputs.end(), p->output_streams.begin(),
p->output_streams.end());
}
inputs.sort();
inputs.unique();
inputs_all = inputs;
outputs.sort();
outputs.unique();
intermediate_values = outputs;
for (auto o : outputs)
inputs.remove(o);
for (auto i : inputs_all)
outputs.remove(i);
for (auto o : outputs)
intermediate_values.remove(o);
assert(!inputs.empty() && !outputs.empty());
// For now operations have a single output so this subgraph should
// also only have one output as only operations required to
// produce this value are taken.
assert(outputs.size() == 1);
// Decide on number of chunks to split -- TODO: refine this
size_t mem_per_sample = 0;
size_t const_mem_per_sample = 0;
size_t num_samples = 1;
size_t num_real_inputs = 0;
// Only the sizes of inputs is known ahead of execution
for (auto s : inputs) {
// Const streams data is required in whole for each computation,
// we treat this separately as it can be substantial.
if (s->const_stream) {
const_mem_per_sample += memref_get_data_size(s->dep->host_data);
continue;
}
// If this is a ciphertext
if (s->ct_stream) {
mem_per_sample += s->dep->host_data.sizes[1] * sizeof(uint64_t);
num_real_inputs++;
if (s->dep->host_data.sizes[0] > num_samples)
num_samples = s->dep->host_data.sizes[0];
if (!s->dep->chunks.empty())
num_samples = s->dep->chunks.size();
} else {
mem_per_sample += sizeof(uint64_t);
}
}
// Approximate the memory required for intermediate values and outputs
mem_per_sample += mem_per_sample *
(outputs.size() + intermediate_values.size()) /
(num_real_inputs ? num_real_inputs : 1);
size_t num_chunks = 1;
size_t num_gpu_chunks = 0;
int32_t num_devices_to_use = 0;
// If the subgraph does not have sufficient computational
// intensity (which we approximate by whether it bootstraps), then
// we assume (TODO: confirm with profiling) that it is not
// beneficial to offload to GPU.
if (subgraph_bootstraps) {
// Determine maximum GPU granulariry
size_t gpu_free_mem;
size_t gpu_total_mem;
// TODO: this could be improved
// Force deallocation with a synchronization point
for (size_t g = 0; g < num_devices; ++g)
cudaStreamSynchronize(*(cudaStream_t *)dfg->get_gpu_stream(g));
auto status = cudaMemGetInfo(&gpu_free_mem, &gpu_total_mem);
assert(status == cudaSuccess);
// TODO - for now assume each device on the system has roughly same
// available memory.
size_t available_mem = gpu_free_mem;
// Further assume (TODO) that kernel execution requires some
// magic factor more meory per sample to execute
size_t max_samples_per_chunk =
(available_mem - const_mem_per_sample) /
((mem_per_sample ? mem_per_sample : 1) * gpu_memory_inflation_factor);
if (num_samples < num_cores + device_compute_factor * num_devices) {
num_devices_to_use = 0;
num_chunks = std::min(num_cores, num_samples);
} else {
num_devices_to_use = num_devices;
size_t compute_resources =
num_cores + num_devices * device_compute_factor;
size_t gpu_chunk_size =
std::ceil((double)num_samples / compute_resources) *
device_compute_factor;
size_t scale_factor =
std::ceil((double)gpu_chunk_size / max_samples_per_chunk);
num_chunks = num_cores * scale_factor;
num_gpu_chunks = num_devices * scale_factor;
}
} else {
num_chunks = std::min(num_cores, num_samples);
}
for (auto i : inputs)
i->dep->split_dependence(num_chunks, num_gpu_chunks,
(i->ct_stream) ? 0 : 1, i->const_stream);
for (auto iv : intermediate_values) {
if (iv->need_new_gen()) {
iv->put(new Dependence(split_location,
{nullptr, nullptr, 0, {0, 0}, {0, 0}}, nullptr,
false, false, split_chunks));
iv->dep->chunks.resize(num_chunks + num_gpu_chunks, nullptr);
}
}
for (auto o : outputs) {
if (o->need_new_gen()) {
o->put(new Dependence(split_location,
{nullptr, nullptr, 0, {0, 0}, {0, 0}}, nullptr,
false, false, split_chunks));
o->dep->chunks.resize(num_chunks + num_gpu_chunks, nullptr);
}
}
// Execute graph
std::list<std::thread> workers;
std::list<std::thread> gpu_schedulers;
std::vector<std::list<size_t>> gpu_chunk_list;
gpu_chunk_list.resize(num_devices);
int32_t dev = 0;
for (size_t c = 0; c < num_chunks + num_gpu_chunks; ++c) {
if (!subgraph_bootstraps) {
workers.push_back(std::thread(
[&](std::list<Process *> queue, size_t c, int32_t host_location) {
for (auto p : queue)
schedule_kernel(p, host_location, c, nullptr);
for (auto iv : intermediate_values)
if (iv->consumers.size() == 1)
iv->dep->free_chunk_host_data(c, dfg);
},
queue, c, host_location));
} else {
if (c < num_chunks) {
workers.push_back(std::thread(
[&](std::list<Process *> queue, size_t c, int32_t host_location) {
for (auto p : queue) {
schedule_kernel(p, host_location, c, nullptr);
}
for (auto iv : intermediate_values)
if (iv->consumers.size() == 1)
iv->dep->free_chunk_host_data(c, dfg);
},
queue, c, host_location));
} else {
// Gather per-device chunk list
gpu_chunk_list[dev++ % num_devices].push_back(c);
}
}
}
for (dev = 0; dev < num_devices; ++dev) {
gpu_schedulers.push_back(std::thread(
[&](std::list<Process *> queue, int32_t dev) {
for (size_t c : gpu_chunk_list[dev]) {
size_t gpu_free_mem;
size_t gpu_total_mem;
auto status = cudaSetDevice(dev);
assert(status == cudaSuccess);
cudaMemGetInfo(&gpu_free_mem, &gpu_total_mem);
assert(status == cudaSuccess);
for (auto p : queue)
schedule_kernel(p, dev, c, nullptr);
for (auto iv : intermediate_values)
if (iv->consumers.size() > 1)
iv->dep->move_chunk_off_device(c, dfg);
else
iv->dep->free_chunk_device_data(c, dfg);
for (auto o : outputs)
o->dep->move_chunk_off_device(c, dfg);
cudaStreamSynchronize(*(cudaStream_t *)dfg->get_gpu_stream(dev));
}
},
queue, dev));
}
for (auto &w : workers)
w.join();
workers.clear();
for (auto &gs : gpu_schedulers)
gs.join();
gpu_schedulers.clear();
// Build output out of the separate chunks processed
for (auto o : outputs) {
assert(o->batched_stream && o->ct_stream &&
"Only operations with ciphertext output supported.");
o->dep->merge_dependence(dfg);
}
// We will assume that only one subgraph is being processed per
// DFG at a time, so we can safely free these here.
dfg->free_stream_order_dependent_data();
return;
}
Dependence *get_on_host(MemRef2 &out) {
schedule_work(out);
assert(dep != nullptr && "GET on empty stream not allowed.");
// If this was already copied to host, copy out
if (dep->onHostReady) {
memref_copy_contiguous(out, dep->host_data);
return dep;
} else if (dep->location == split_location) {
char *pos = (char *)(out.aligned + out.offset);
std::list<int32_t> devices_used;
for (auto c : dep->chunks) {
size_t data_size = memref_get_data_size(c->host_data);
cuda_memcpy_async_to_cpu(
pos, c->device_data, data_size,
(cudaStream_t *)dfg->get_gpu_stream(c->location), c->location);
pos += data_size;
devices_used.push_back(c->location);
}
// We should only synchronize devices that had data chunks
devices_used.sort();
devices_used.unique();
for (auto i : devices_used)
cudaStreamSynchronize(*(cudaStream_t *)dfg->get_gpu_stream(i));
} else {
size_t data_size = memref_get_data_size(dep->host_data);
cuda_memcpy_async_to_cpu(out.aligned + out.offset, dep->device_data,
data_size, (cudaStream_t *)dfg->gpu_stream,
dep->location);
cudaStreamSynchronize(*(cudaStream_t *)dfg->gpu_stream);
}
// After this synchronization point, all of the host-side
// allocated memory can be freed as we know all asynchronous
// operations have finished.
dfg->free_stream_order_dependent_data();
if (!dep->hostAllocated)
dep->host_data = memref_copy_alloc(out);
dep->onHostReady = true;
dep->hostAllocated = true;
return dep;
}
Dependence *get(int32_t location, int32_t chunk_id = single_chunk) {
assert(dep != nullptr && "Dependence could not be computed.");
assert(chunk_id != split_chunks);
if (chunk_id != single_chunk) {
Dependence *d = dep->chunks[chunk_id];
d->copy(location, dfg);
return d;
}
dep->copy(location, dfg);
return dep;
}
inline bool need_new_gen(int32_t chunk_id = single_chunk) {
if (dep == nullptr)
return true;
if (chunk_id == single_chunk && dep->stream_generation != generation)
return true;
if (chunk_id != single_chunk) {
assert(chunk_id < dep->chunks.size());
if (dep->chunks[chunk_id] == nullptr)
return true;
if (dep->chunks[chunk_id]->stream_generation != generation)
return true;
}
return false;
}
};
static inline mlir::concretelang::gpu_dfg::Process *
make_process_1_1(void *dfg, void *sin1, void *sout,
void (*fun)(Process *, int32_t, int32_t, uint64_t *)) {
mlir::concretelang::gpu_dfg::Process *p =
new mlir::concretelang::gpu_dfg::Process;
mlir::concretelang::gpu_dfg::Stream *s1 =
(mlir::concretelang::gpu_dfg::Stream *)sin1;
mlir::concretelang::gpu_dfg::Stream *so =
(mlir::concretelang::gpu_dfg::Stream *)sout;
p->input_streams.push_back(s1);
p->dfg = (GPU_DFG *)dfg;
p->fun = fun;
p->output_streams.push_back(so);
s1->consumers.push_back(p);
so->producer = p;
so->dfg = s1->dfg = (GPU_DFG *)dfg;
p->dfg->register_stream(s1);
p->dfg->register_stream(so);
p->batched_process = s1->batched_stream;
return p;
}
static inline mlir::concretelang::gpu_dfg::Process *
make_process_2_1(void *dfg, void *sin1, void *sin2, void *sout,
void (*fun)(Process *, int32_t, int32_t, uint64_t *)) {
mlir::concretelang::gpu_dfg::Process *p =
new mlir::concretelang::gpu_dfg::Process;
mlir::concretelang::gpu_dfg::Stream *s1 =
(mlir::concretelang::gpu_dfg::Stream *)sin1;
mlir::concretelang::gpu_dfg::Stream *s2 =
(mlir::concretelang::gpu_dfg::Stream *)sin2;
mlir::concretelang::gpu_dfg::Stream *so =
(mlir::concretelang::gpu_dfg::Stream *)sout;
p->input_streams.push_back(s1);
p->input_streams.push_back(s2);
p->dfg = (GPU_DFG *)dfg;
p->fun = fun;
p->output_streams.push_back(so);
s1->consumers.push_back(p);
s2->consumers.push_back(p);
so->producer = p;
so->dfg = s1->dfg = s2->dfg = (GPU_DFG *)dfg;
p->dfg->register_stream(s1);
p->dfg->register_stream(s2);
p->dfg->register_stream(so);
p->batched_process = s1->batched_stream;
return p;
}
[[maybe_unused]] static void sdfg_gpu_debug_print_mref(const char *c,
MemRef2 m) {
std::cout << c << " : " << m.allocated << ", " << m.aligned << ", "
<< m.offset << ", [" << m.sizes[0] << ", " << m.sizes[1] << "], ["
<< m.strides[0] << ", " << m.strides[1] << "]\n";
}
[[maybe_unused]] static MemRef2 sdfg_gpu_debug_dependence(Dependence *d,
cudaStream_t *s) {
if (d->onHostReady)
return d->host_data;
size_t data_size = memref_get_data_size(d->host_data);
uint64_t *data = (uint64_t *)malloc(data_size);
MemRef2 ret = {data,
data,
0,
{d->host_data.sizes[0], d->host_data.sizes[1]},
{d->host_data.strides[0], d->host_data.strides[1]}};
cuda_memcpy_async_to_cpu(data, d->device_data, data_size, s, d->location);
cudaStreamSynchronize(*s);
return ret;
}
[[maybe_unused]] static bool
sdfg_gpu_debug_compare_memref(MemRef2 &a, MemRef2 &b, char const *msg) {
if (a.sizes[0] != b.sizes[0] || a.sizes[1] != b.sizes[1] ||
a.strides[0] != b.strides[0] || a.strides[1] != b.strides[1])
return false;
size_t data_size = memref_get_data_size(a);
int num = 0;
for (size_t i = 0; i < data_size / sizeof(uint64_t); ++i)
if ((a.aligned + a.offset)[i] != (b.aligned + b.offset)[i]) {
std::cout << msg << " - memrefs differ at position " << i << " "
<< (a.aligned + a.offset)[i] << " " << (b.aligned + b.offset)[i]
<< "\n";
return false;
}
return true;
}
// Stream emulator processes
void memref_keyswitch_lwe_u64_process(Process *p, int32_t loc, int32_t chunk_id,
uint64_t *out_ptr) {
auto sched = [&](Dependence *d) {
uint64_t num_samples = d->host_data.sizes[0];
MemRef2 out = {
0, 0, 0, {num_samples, p->output_size.val}, {p->output_size.val, 1}};
size_t data_size = memref_get_data_size(out);
if (loc == host_location) {
// If it is not profitable to offload, schedule kernel on CPU
out.allocated = out.aligned =
(uint64_t *)((out_ptr != nullptr) ? out_ptr : malloc(data_size));
memref_batched_keyswitch_lwe_u64(
out.allocated, out.aligned, out.offset, out.sizes[0], out.sizes[1],
out.strides[0], out.strides[1], d->host_data.allocated,
d->host_data.aligned, d->host_data.offset, d->host_data.sizes[0],
d->host_data.sizes[1], d->host_data.strides[0],
d->host_data.strides[1], p->level.val, p->base_log.val,
p->input_lwe_dim.val, p->output_lwe_dim.val, p->sk_index.val,
p->ctx.val);
Dependence *dep =
new Dependence(loc, out, nullptr, true, true, d->chunk_id);
return dep;
} else {
// Schedule the keyswitch kernel on the GPU
cudaStream_t *s = (cudaStream_t *)p->dfg->get_gpu_stream(loc);
void *ct0_gpu = d->device_data;
void *out_gpu = cuda_malloc_async(data_size, s, loc);
void *ksk_gpu = p->ctx.val->get_ksk_gpu(
p->level.val, p->input_lwe_dim.val, p->output_lwe_dim.val, loc, s,
p->sk_index.val);
cuda_keyswitch_lwe_ciphertext_vector_64(
s, loc, out_gpu, ct0_gpu, ksk_gpu, p->input_lwe_dim.val,
p->output_lwe_dim.val, p->base_log.val, p->level.val, num_samples);
Dependence *dep =
new Dependence(loc, out, out_gpu, false, false, d->chunk_id);
return dep;
}
};
Dependence *idep = p->input_streams[0]->get(loc, chunk_id);
if (p->output_streams[0]->need_new_gen(chunk_id))
p->output_streams[0]->put(sched(idep), chunk_id);
}
void memref_bootstrap_lwe_u64_process(Process *p, int32_t loc, int32_t chunk_id,
uint64_t *out_ptr) {
assert(p->output_size.val == p->glwe_dim.val * p->poly_size.val + 1);
Dependence *idep1 = p->input_streams[1]->get(host_location, chunk_id);
MemRef2 &mtlu = idep1->host_data;
uint32_t num_lut_vectors = mtlu.sizes[0];
uint64_t glwe_ct_len =
p->poly_size.val * (p->glwe_dim.val + 1) * num_lut_vectors;
uint64_t glwe_ct_size = glwe_ct_len * sizeof(uint64_t);
uint64_t *glwe_ct = (uint64_t *)malloc(glwe_ct_size);
auto tlu = mtlu.aligned + mtlu.offset;
// Glwe trivial encryption
size_t pos = 0, postlu = 0;
for (size_t l = 0; l < num_lut_vectors; ++l) {
for (size_t i = 0; i < p->poly_size.val * p->glwe_dim.val; i++) {
glwe_ct[pos++] = 0;
}
for (size_t i = 0; i < p->poly_size.val; i++) {
glwe_ct[pos++] = tlu[postlu++];
}
}
auto sched = [&](Dependence *d0, Dependence *d1, uint64_t *glwe_ct,
std::vector<size_t> &lut_indexes, cudaStream_t *s,
int32_t loc) {
uint64_t num_samples = d0->host_data.sizes[0];
MemRef2 out = {
0, 0, 0, {num_samples, p->output_size.val}, {p->output_size.val, 1}};
size_t data_size = memref_get_data_size(out);
// Move test vector indexes to the GPU, the test vector indexes is set of 0
uint32_t lwe_idx = 0,
test_vector_idxes_size = num_samples * sizeof(uint64_t);
uint64_t *test_vector_idxes = (uint64_t *)malloc(test_vector_idxes_size);
if (lut_indexes.size() == 1) {
memset((void *)test_vector_idxes, lut_indexes[0], test_vector_idxes_size);
} else {
assert(lut_indexes.size() == num_samples);
for (size_t i = 0; i < num_samples; ++i)
test_vector_idxes[i] = lut_indexes[i];
}
if (loc == host_location) {
// If it is not profitable to offload, schedule kernel on CPU
out.allocated = out.aligned =
(uint64_t *)((out_ptr != nullptr) ? out_ptr : malloc(data_size));
if (lut_indexes.size() == 1)
memref_batched_bootstrap_lwe_u64(
out.allocated, out.aligned, out.offset, out.sizes[0], out.sizes[1],
out.strides[0], out.strides[1], d0->host_data.allocated,
d0->host_data.aligned, d0->host_data.offset, d0->host_data.sizes[0],
d0->host_data.sizes[1], d0->host_data.strides[0],
d0->host_data.strides[1], d1->host_data.allocated,
d1->host_data.aligned, d1->host_data.offset, d1->host_data.sizes[1],
d1->host_data.strides[1], p->input_lwe_dim.val, p->poly_size.val,
p->level.val, p->base_log.val, p->glwe_dim.val, p->sk_index.val,
p->ctx.val);
else
memref_batched_mapped_bootstrap_lwe_u64(
out.allocated, out.aligned, out.offset, out.sizes[0], out.sizes[1],
out.strides[0], out.strides[1], d0->host_data.allocated,
d0->host_data.aligned, d0->host_data.offset, d0->host_data.sizes[0],
d0->host_data.sizes[1], d0->host_data.strides[0],
d0->host_data.strides[1], d1->host_data.allocated,
d1->host_data.aligned, d1->host_data.offset, d1->host_data.sizes[0],
d1->host_data.sizes[1], d1->host_data.strides[0],
d1->host_data.strides[1], p->input_lwe_dim.val, p->poly_size.val,
p->level.val, p->base_log.val, p->glwe_dim.val, p->sk_index.val,
p->ctx.val);
Dependence *dep =
new Dependence(loc, out, nullptr, true, true, d0->chunk_id);
free(glwe_ct);
return dep;
} else {
// Schedule the bootstrap kernel on the GPU
void *glwe_ct_gpu = cuda_malloc_async(glwe_ct_size, s, loc);
cuda_memcpy_async_to_gpu(glwe_ct_gpu, glwe_ct, glwe_ct_size, s, loc);
void *test_vector_idxes_gpu =
cuda_malloc_async(test_vector_idxes_size, s, loc);
cuda_memcpy_async_to_gpu(test_vector_idxes_gpu, (void *)test_vector_idxes,
test_vector_idxes_size, s, loc);
int8_t *pbs_buffer = p->dfg->gpus[loc].get_pbs_buffer(
p->glwe_dim.val, p->poly_size.val, num_samples);
void *ct0_gpu = d0->device_data;
void *out_gpu = cuda_malloc_async(data_size, s, loc);
void *fbsk_gpu = p->ctx.val->get_bsk_gpu(
p->input_lwe_dim.val, p->poly_size.val, p->level.val, p->glwe_dim.val,
loc, s, p->sk_index.val);
cuda_bootstrap_amortized_lwe_ciphertext_vector_64(
s, loc, out_gpu, glwe_ct_gpu, test_vector_idxes_gpu, ct0_gpu,
fbsk_gpu, (int8_t *)pbs_buffer, p->input_lwe_dim.val, p->glwe_dim.val,
p->poly_size.val, p->base_log.val, p->level.val, num_samples,
lut_indexes.size(), lwe_idx, cuda_get_max_shared_memory(loc));
cuda_drop_async(test_vector_idxes_gpu, s, loc);
cuda_drop_async(glwe_ct_gpu, s, loc);
Dependence *dep =
new Dependence(loc, out, out_gpu, false, false, d0->chunk_id);
// As streams are not synchronized, we can only free this vector
// after a later synchronization point where we are guaranteed that
// this vector is no longer needed.
p->dfg->register_stream_order_dependent_allocation(test_vector_idxes);
p->dfg->register_stream_order_dependent_allocation(glwe_ct);
return dep;
}
};
// If this is a mapped TLU
// TODO: for now we do not provide more advanced ways of selecting
bool mapped = (p->input_streams[1]->dep->host_data.sizes[0] > 1);
std::vector<size_t> lut_indexes;
if (mapped) {
lut_indexes.resize(num_lut_vectors);
std::iota(lut_indexes.begin(), lut_indexes.end(), 0);
} else {
lut_indexes.push_back(0);
}
cudaStream_t *cstream = (cudaStream_t *)p->dfg->get_gpu_stream(loc);
Dependence *idep0 = p->input_streams[0]->get(loc, chunk_id);
if (p->output_streams[0]->need_new_gen(chunk_id))
p->output_streams[0]->put(
sched(idep0, idep1, glwe_ct, lut_indexes, cstream, loc), chunk_id);
}
void memref_add_lwe_ciphertexts_u64_process(Process *p, int32_t loc,
int32_t chunk_id,
uint64_t *out_ptr) {
auto sched = [&](Dependence *d0, Dependence *d1, cudaStream_t *s,
int32_t loc) {
assert(d0->host_data.sizes[0] == d1->host_data.sizes[0]);
assert(d0->host_data.sizes[1] == d1->host_data.sizes[1]);
assert(d0->chunk_id == d1->chunk_id);
uint64_t num_samples = d0->host_data.sizes[0];
MemRef2 out = {0,
0,
0,
{num_samples, d0->host_data.sizes[1]},
{d0->host_data.sizes[1], 1}};
size_t data_size = memref_get_data_size(out);
if (loc == host_location) {
// If it is not profitable to offload, schedule kernel on CPU
out.allocated = out.aligned =
(uint64_t *)((out_ptr != nullptr) ? out_ptr : malloc(data_size));
memref_batched_add_lwe_ciphertexts_u64(
out.allocated, out.aligned, out.offset, out.sizes[0], out.sizes[1],
out.strides[0], out.strides[1], d0->host_data.allocated,
d0->host_data.aligned, d0->host_data.offset, d0->host_data.sizes[0],
d0->host_data.sizes[1], d0->host_data.strides[0],
d0->host_data.strides[1], d1->host_data.allocated,
d1->host_data.aligned, d1->host_data.offset, d1->host_data.sizes[0],
d1->host_data.sizes[1], d1->host_data.strides[0],
d1->host_data.strides[1]);
Dependence *dep =
new Dependence(loc, out, nullptr, true, true, d0->chunk_id);
return dep;
} else {
// Schedule the kernel on the GPU
void *out_gpu = cuda_malloc_async(data_size, s, loc);
cuda_add_lwe_ciphertext_vector_64(
s, loc, out_gpu, d0->device_data, d1->device_data,
d0->host_data.sizes[1] - 1, num_samples);
Dependence *dep =
new Dependence(loc, out, out_gpu, false, false, d0->chunk_id);
return dep;
}
};
Dependence *idep0 = p->input_streams[0]->get(loc, chunk_id);
Dependence *idep1 = p->input_streams[1]->get(loc, chunk_id);
if (p->output_streams[0]->need_new_gen(chunk_id))
p->output_streams[0]->put(
sched(idep0, idep1, (cudaStream_t *)p->dfg->get_gpu_stream(loc), loc),
chunk_id);
}
void memref_add_plaintext_lwe_ciphertext_u64_process(Process *p, int32_t loc,
int32_t chunk_id,
uint64_t *out_ptr) {
auto sched = [&](Dependence *d0, Dependence *d1, cudaStream_t *s,
int32_t loc) {
assert(d0->host_data.sizes[0] == d1->host_data.sizes[1] ||
d1->host_data.sizes[1] == 1);
assert(d0->chunk_id == d1->chunk_id);
uint64_t num_samples = d0->host_data.sizes[0];
MemRef2 out = {0,
0,
0,
{num_samples, d0->host_data.sizes[1]},
{d0->host_data.sizes[1], 1}};
size_t data_size = memref_get_data_size(out);
if (loc == host_location) {
// If it is not profitable to offload, schedule kernel on CPU
out.allocated = out.aligned =
(uint64_t *)((out_ptr != nullptr) ? out_ptr : malloc(data_size));
if (d1->host_data.sizes[1] == 1) // Constant case - or single sample
memref_batched_add_plaintext_cst_lwe_ciphertext_u64(
out.allocated, out.aligned, out.offset, out.sizes[0], out.sizes[1],
out.strides[0], out.strides[1], d0->host_data.allocated,
d0->host_data.aligned, d0->host_data.offset, d0->host_data.sizes[0],
d0->host_data.sizes[1], d0->host_data.strides[0],
d0->host_data.strides[1],
*(d1->host_data.aligned + d1->host_data.offset));
else
memref_batched_add_plaintext_lwe_ciphertext_u64(
out.allocated, out.aligned, out.offset, out.sizes[0], out.sizes[1],
out.strides[0], out.strides[1], d0->host_data.allocated,
d0->host_data.aligned, d0->host_data.offset, d0->host_data.sizes[0],
d0->host_data.sizes[1], d0->host_data.strides[0],
d0->host_data.strides[1], d1->host_data.allocated,
d1->host_data.aligned, d1->host_data.offset, d1->host_data.sizes[1],
d1->host_data.strides[1]);
Dependence *dep =
new Dependence(loc, out, nullptr, true, true, d0->chunk_id);
return dep;
} else {
// Schedule the kernel on the GPU
void *out_gpu = cuda_malloc_async(data_size, s, loc);
cuda_add_lwe_ciphertext_vector_plaintext_vector_64(
s, loc, out_gpu, d0->device_data, d1->device_data,
d0->host_data.sizes[1] - 1, num_samples);
Dependence *dep =
new Dependence(loc, out, out_gpu, false, false, d0->chunk_id);
return dep;
}
};
Dependence *idep0 = p->input_streams[0]->get(loc, chunk_id);
Dependence *idep1 = p->input_streams[1]->get(loc, chunk_id);
if (p->output_streams[0]->need_new_gen(chunk_id))
p->output_streams[0]->put(
sched(idep0, idep1, (cudaStream_t *)p->dfg->get_gpu_stream(loc), loc),
chunk_id);
}
void memref_mul_cleartext_lwe_ciphertext_u64_process(Process *p, int32_t loc,
int32_t chunk_id,
uint64_t *out_ptr) {
auto sched = [&](Dependence *d0, Dependence *d1, cudaStream_t *s,
int32_t loc) {
assert(d0->host_data.sizes[0] == d1->host_data.sizes[1] ||
d1->host_data.sizes[1] == 1);
assert(d0->chunk_id == d1->chunk_id);
uint64_t num_samples = d0->host_data.sizes[0];
MemRef2 out = {0,
0,
0,
{num_samples, d0->host_data.sizes[1]},
{d0->host_data.sizes[1], 1}};
size_t data_size = memref_get_data_size(out);
if (loc == host_location) {
// If it is not profitable to offload, schedule kernel on CPU
out.allocated = out.aligned =
(uint64_t *)((out_ptr != nullptr) ? out_ptr : malloc(data_size));
if (d1->host_data.sizes[1] == 1) // Constant case
memref_batched_mul_cleartext_cst_lwe_ciphertext_u64(
out.allocated, out.aligned, out.offset, out.sizes[0], out.sizes[1],
out.strides[0], out.strides[1], d0->host_data.allocated,
d0->host_data.aligned, d0->host_data.offset, d0->host_data.sizes[0],
d0->host_data.sizes[1], d0->host_data.strides[0],
d0->host_data.strides[1],
*(d1->host_data.aligned + d1->host_data.offset));
else
memref_batched_mul_cleartext_lwe_ciphertext_u64(
out.allocated, out.aligned, out.offset, out.sizes[0], out.sizes[1],
out.strides[0], out.strides[1], d0->host_data.allocated,
d0->host_data.aligned, d0->host_data.offset, d0->host_data.sizes[0],
d0->host_data.sizes[1], d0->host_data.strides[0],
d0->host_data.strides[1], d1->host_data.allocated,
d1->host_data.aligned, d1->host_data.offset, d1->host_data.sizes[1],
d1->host_data.strides[1]);
Dependence *dep =
new Dependence(loc, out, nullptr, true, true, d0->chunk_id);
return dep;
} else {
// Schedule the keyswitch kernel on the GPU
void *out_gpu = cuda_malloc_async(data_size, s, loc);
cuda_mult_lwe_ciphertext_vector_cleartext_vector_64(
s, loc, out_gpu, d0->device_data, d1->device_data,
d0->host_data.sizes[1] - 1, num_samples);
Dependence *dep =
new Dependence(loc, out, out_gpu, false, false, d0->chunk_id);
return dep;
}
};
Dependence *idep0 = p->input_streams[0]->get(loc, chunk_id);
Dependence *idep1 = p->input_streams[1]->get(loc, chunk_id);
if (p->output_streams[0]->need_new_gen(chunk_id))
p->output_streams[0]->put(
sched(idep0, idep1, (cudaStream_t *)p->dfg->get_gpu_stream(loc), loc),
chunk_id);
}
void memref_negate_lwe_ciphertext_u64_process(Process *p, int32_t loc,
int32_t chunk_id,
uint64_t *out_ptr) {
auto sched = [&](Dependence *d0, cudaStream_t *s, int32_t loc) {
uint64_t num_samples = d0->host_data.sizes[0];
MemRef2 out = {0,
0,
0,
{num_samples, d0->host_data.sizes[1]},
{d0->host_data.sizes[1], 1}};
size_t data_size = memref_get_data_size(out);
if (loc == host_location) {
// If it is not profitable to offload, schedule kernel on CPU
out.allocated = out.aligned =
(uint64_t *)((out_ptr != nullptr) ? out_ptr : malloc(data_size));
memref_batched_negate_lwe_ciphertext_u64(
out.allocated, out.aligned, out.offset, out.sizes[0], out.sizes[1],
out.strides[0], out.strides[1], d0->host_data.allocated,
d0->host_data.aligned, d0->host_data.offset, d0->host_data.sizes[0],
d0->host_data.sizes[1], d0->host_data.strides[0],
d0->host_data.strides[1]);
Dependence *dep =
new Dependence(loc, out, nullptr, true, true, d0->chunk_id);
return dep;
} else {
// Schedule the kernel on the GPU
void *out_gpu = cuda_malloc_async(data_size, s, loc);
cuda_negate_lwe_ciphertext_vector_64(s, loc, out_gpu, d0->device_data,
d0->host_data.sizes[1] - 1,
num_samples);
Dependence *dep =
new Dependence(loc, out, out_gpu, false, false, d0->chunk_id);
return dep;
}
};
Dependence *idep0 = p->input_streams[0]->get(loc, chunk_id);
if (p->output_streams[0]->need_new_gen(chunk_id))
p->output_streams[0]->put(
sched(idep0, (cudaStream_t *)p->dfg->get_gpu_stream(loc), loc),
chunk_id);
}
} // namespace
} // namespace gpu_dfg
} // namespace concretelang
} // namespace mlir
using namespace mlir::concretelang::gpu_dfg;
// Code generation interface
void stream_emulator_make_memref_add_lwe_ciphertexts_u64_process(void *dfg,
void *sin1,
void *sin2,
void *sout) {
Process *p = make_process_2_1(dfg, sin1, sin2, sout,
memref_add_lwe_ciphertexts_u64_process);
static int count = 0;
sprintf(p->name, "add_lwe_ciphertexts_%d", count++);
}
void stream_emulator_make_memref_add_plaintext_lwe_ciphertext_u64_process(
void *dfg, void *sin1, void *sin2, void *sout) {
Process *p = make_process_2_1(
dfg, sin1, sin2, sout, memref_add_plaintext_lwe_ciphertext_u64_process);
static int count = 0;
sprintf(p->name, "add_plaintext_lwe_ciphertexts_%d", count++);
}
void stream_emulator_make_memref_mul_cleartext_lwe_ciphertext_u64_process(
void *dfg, void *sin1, void *sin2, void *sout) {
Process *p = make_process_2_1(
dfg, sin1, sin2, sout, memref_mul_cleartext_lwe_ciphertext_u64_process);
static int count = 0;
sprintf(p->name, "mul_cleartext_lwe_ciphertexts_%d", count++);
}
void stream_emulator_make_memref_negate_lwe_ciphertext_u64_process(void *dfg,
void *sin1,
void *sout) {
Process *p = make_process_1_1(dfg, sin1, sout,
memref_negate_lwe_ciphertext_u64_process);
static int count = 0;
sprintf(p->name, "negate_lwe_ciphertext_%d", count++);
}
void stream_emulator_make_memref_keyswitch_lwe_u64_process(
void *dfg, void *sin1, void *sout, uint32_t level, uint32_t base_log,
uint32_t input_lwe_dim, uint32_t output_lwe_dim, uint32_t output_size,
uint32_t ksk_index, void *context) {
Process *p =
make_process_1_1(dfg, sin1, sout, memref_keyswitch_lwe_u64_process);
p->level.val = level;
p->base_log.val = base_log;
p->input_lwe_dim.val = input_lwe_dim;
p->output_lwe_dim.val = output_lwe_dim;
p->sk_index.val = ksk_index;
p->output_size.val = output_size;
p->ctx.val = (RuntimeContext *)context;
static int count = 0;
sprintf(p->name, "keyswitch_%d", count++);
}
void stream_emulator_make_memref_bootstrap_lwe_u64_process(
void *dfg, void *sin1, void *sin2, void *sout, uint32_t input_lwe_dim,
uint32_t poly_size, uint32_t level, uint32_t base_log, uint32_t glwe_dim,
uint32_t output_size, uint32_t bsk_index, void *context) {
// The TLU does not need to be sent to GPU
((Stream *)sin2)->type = TS_STREAM_TYPE_X86_TO_X86_LSAP;
Process *p =
make_process_2_1(dfg, sin1, sin2, sout, memref_bootstrap_lwe_u64_process);
p->input_lwe_dim.val = input_lwe_dim;
p->poly_size.val = poly_size;
p->level.val = level;
p->base_log.val = base_log;
p->glwe_dim.val = glwe_dim;
p->sk_index.val = bsk_index;
p->output_size.val = output_size;
p->ctx.val = (RuntimeContext *)context;
static int count = 0;
sprintf(p->name, "bootstrap_%d", count++);
}
void stream_emulator_make_memref_batched_add_lwe_ciphertexts_u64_process(
void *dfg, void *sin1, void *sin2, void *sout) {
((Stream *)sin1)->batched_stream = true;
((Stream *)sin1)->ct_stream = true;
((Stream *)sin2)->batched_stream = true;
((Stream *)sin2)->ct_stream = true;
stream_emulator_make_memref_add_lwe_ciphertexts_u64_process(dfg, sin1, sin2,
sout);
((Stream *)sout)->batched_stream = true;
((Stream *)sout)->ct_stream = true;
}
void stream_emulator_make_memref_batched_add_plaintext_lwe_ciphertext_u64_process(
void *dfg, void *sin1, void *sin2, void *sout) {
((Stream *)sin1)->batched_stream = true;
((Stream *)sin1)->ct_stream = true;
((Stream *)sin2)->batched_stream = true;
((Stream *)sin2)->pt_stream = true;
stream_emulator_make_memref_add_plaintext_lwe_ciphertext_u64_process(
dfg, sin1, sin2, sout);
((Stream *)sout)->batched_stream = true;
((Stream *)sout)->ct_stream = true;
}
void stream_emulator_make_memref_batched_add_plaintext_cst_lwe_ciphertext_u64_process(
void *dfg, void *sin1, void *sin2, void *sout) {
((Stream *)sin1)->batched_stream = true;
((Stream *)sin1)->ct_stream = true;
((Stream *)sin2)->const_stream = true;
((Stream *)sin2)->pt_stream = true;
stream_emulator_make_memref_add_plaintext_lwe_ciphertext_u64_process(
dfg, sin1, sin2, sout);
((Stream *)sout)->batched_stream = true;
((Stream *)sout)->ct_stream = true;
}
void stream_emulator_make_memref_batched_mul_cleartext_lwe_ciphertext_u64_process(
void *dfg, void *sin1, void *sin2, void *sout) {
((Stream *)sin1)->batched_stream = true;
((Stream *)sin1)->ct_stream = true;
((Stream *)sin2)->batched_stream = true;
((Stream *)sin2)->pt_stream = true;
stream_emulator_make_memref_mul_cleartext_lwe_ciphertext_u64_process(
dfg, sin1, sin2, sout);
((Stream *)sout)->batched_stream = true;
((Stream *)sout)->ct_stream = true;
}
void stream_emulator_make_memref_batched_mul_cleartext_cst_lwe_ciphertext_u64_process(
void *dfg, void *sin1, void *sin2, void *sout) {
((Stream *)sin1)->batched_stream = true;
((Stream *)sin1)->ct_stream = true;
((Stream *)sin2)->const_stream = true;
((Stream *)sin2)->pt_stream = true;
stream_emulator_make_memref_mul_cleartext_lwe_ciphertext_u64_process(
dfg, sin1, sin2, sout);
((Stream *)sout)->batched_stream = true;
((Stream *)sout)->ct_stream = true;
}
void stream_emulator_make_memref_batched_negate_lwe_ciphertext_u64_process(
void *dfg, void *sin1, void *sout) {
((Stream *)sin1)->batched_stream = true;
((Stream *)sin1)->ct_stream = true;
stream_emulator_make_memref_negate_lwe_ciphertext_u64_process(dfg, sin1,
sout);
((Stream *)sout)->batched_stream = true;
((Stream *)sout)->ct_stream = true;
}
void stream_emulator_make_memref_batched_keyswitch_lwe_u64_process(
void *dfg, void *sin1, void *sout, uint32_t level, uint32_t base_log,
uint32_t input_lwe_dim, uint32_t output_lwe_dim, uint32_t output_size,
uint32_t ksk_index, void *context) {
((Stream *)sin1)->batched_stream = true;
((Stream *)sin1)->ct_stream = true;
stream_emulator_make_memref_keyswitch_lwe_u64_process(
dfg, sin1, sout, level, base_log, input_lwe_dim, output_lwe_dim,
output_size, ksk_index, context);
((Stream *)sout)->batched_stream = true;
((Stream *)sout)->ct_stream = true;
}
void stream_emulator_make_memref_batched_bootstrap_lwe_u64_process(
void *dfg, void *sin1, void *sin2, void *sout, uint32_t input_lwe_dim,
uint32_t poly_size, uint32_t level, uint32_t base_log, uint32_t glwe_dim,
uint32_t output_size, uint32_t bsk_index, void *context) {
((Stream *)sin1)->batched_stream = true;
((Stream *)sin1)->ct_stream = true;
((Stream *)sin2)->const_stream = true;
((Stream *)sin2)->ct_stream = true;
stream_emulator_make_memref_bootstrap_lwe_u64_process(
dfg, sin1, sin2, sout, input_lwe_dim, poly_size, level, base_log,
glwe_dim, output_size, bsk_index, context);
((Stream *)sout)->batched_stream = true;
((Stream *)sout)->ct_stream = true;
}
void stream_emulator_make_memref_batched_mapped_bootstrap_lwe_u64_process(
void *dfg, void *sin1, void *sin2, void *sout, uint32_t input_lwe_dim,
uint32_t poly_size, uint32_t level, uint32_t base_log, uint32_t glwe_dim,
uint32_t output_size, uint32_t bsk_index, void *context) {
((Stream *)sin1)->batched_stream = true;
((Stream *)sin1)->ct_stream = true;
((Stream *)sin2)->batched_stream = true;
((Stream *)sin2)->ct_stream = true;
stream_emulator_make_memref_bootstrap_lwe_u64_process(
dfg, sin1, sin2, sout, input_lwe_dim, poly_size, level, base_log,
glwe_dim, output_size, bsk_index, context);
((Stream *)sout)->batched_stream = true;
((Stream *)sout)->ct_stream = true;
}
void *stream_emulator_make_uint64_stream(const char *name, stream_type stype) {
return (void *)new Stream(stype, name);
}
void stream_emulator_put_uint64(void *stream, uint64_t e) {
Stream *s = (Stream *)stream;
uint64_t *data = (uint64_t *)malloc(sizeof(uint64_t));
*data = e;
MemRef2 m = {data, data, 0, {1, 1}, {1, 1}};
Dependence *dep = new Dependence(host_location, m, nullptr, true, true);
s->put(dep);
s->generation++;
}
uint64_t stream_emulator_get_uint64(void *stream) {
uint64_t res;
auto s = (Stream *)stream;
MemRef2 m = {&res, &res, 0, {1, 1}, {1, 1}};
s->get_on_host(m);
return res;
}
void *stream_emulator_make_memref_stream(const char *name, stream_type stype) {
return (void *)new Stream(stype, name);
}
void stream_emulator_put_memref(void *stream, uint64_t *allocated,
uint64_t *aligned, uint64_t offset,
uint64_t size, uint64_t stride) {
assert(stride == 1 && "Strided memrefs not supported");
Stream *s = (Stream *)stream;
MemRef2 m = {allocated, aligned, offset, {1, size}, {size, stride}};
Dependence *dep =
new Dependence(host_location, memref_copy_alloc(m), nullptr, true, true);
s->put(dep);
s->generation++;
}
void stream_emulator_get_memref(void *stream, uint64_t *out_allocated,
uint64_t *out_aligned, uint64_t out_offset,
uint64_t out_size, uint64_t out_stride) {
assert(out_stride == 1 && "Strided memrefs not supported");
MemRef2 mref = {out_allocated,
out_aligned,
out_offset,
{1, out_size},
{out_size, out_stride}};
auto s = (Stream *)stream;
s->get_on_host(mref);
}
void *stream_emulator_make_memref_batch_stream(const char *name,
stream_type stype) {
return (void *)new Stream(stype, name);
}
void stream_emulator_put_memref_batch(void *stream, uint64_t *allocated,
uint64_t *aligned, uint64_t offset,
uint64_t size0, uint64_t size1,
uint64_t stride0, uint64_t stride1) {
assert(stride1 == 1 && "Strided memrefs not supported");
Stream *s = (Stream *)stream;
MemRef2 m = {allocated, aligned, offset, {size0, size1}, {stride0, stride1}};
Dependence *dep =
new Dependence(host_location, memref_copy_alloc(m), nullptr, true, true);
s->put(dep);
s->generation++;
}
void stream_emulator_get_memref_batch(void *stream, uint64_t *out_allocated,
uint64_t *out_aligned,
uint64_t out_offset, uint64_t out_size0,
uint64_t out_size1, uint64_t out_stride0,
uint64_t out_stride1) {
assert(out_stride1 == 1 && "Strided memrefs not supported");
MemRef2 mref = {out_allocated,
out_aligned,
out_offset,
{out_size0, out_size1},
{out_stride0, out_stride1}};
auto s = (Stream *)stream;
s->get_on_host(mref);
}
void *stream_emulator_init() {
int num;
assert(cudaGetDeviceCount(&num) == cudaSuccess);
num_devices = num;
assert(num_devices > 0 && "No GPUs available on system.");
char *env = getenv("SDFG_NUM_GPUS");
if (env != nullptr) {
size_t requested_gpus = strtoul(env, NULL, 10);
if (requested_gpus == 0)
warnx("WARNING: no GPUs requested (%lu available) - "
"continuing with one (%lu) device.",
num_devices, requested_gpus = 1);
if (requested_gpus > num_devices)
warnx("WARNING: requested more GPUs (%lu) than available (%lu) - "
"continuing with available devices.",
requested_gpus, num_devices);
else
num_devices = requested_gpus;
}
// If the user has specified a ratio, use that. Otherwise we
// estimate the ratio based on the number of multiprocessors
// available per GPU.
env = getenv("SDFG_DEVICE_TO_CORE_RATIO");
if (env != nullptr) {
device_compute_factor = strtoul(env, NULL, 10);
} else {
cudaDeviceProp properties;
// For now we only querry one GPU, assuming all are the same.
assert(cudaGetDeviceProperties(&properties, 0) == cudaSuccess);
int smpc = properties.multiProcessorCount;
// Rough estimate - each SM has ballpark similar compute
// capability as a CPU core (so 2 HW threads generally).
device_compute_factor = smpc * 2;
}
hwloc_topology_t topology;
hwloc_topology_init(&topology);
hwloc_topology_set_all_types_filter(topology, HWLOC_TYPE_FILTER_KEEP_NONE);
hwloc_topology_set_type_filter(topology, HWLOC_OBJ_CORE,
HWLOC_TYPE_FILTER_KEEP_ALL);
hwloc_topology_load(topology);
num_cores = hwloc_get_nbobjs_by_type(topology, HWLOC_OBJ_PU) - num_devices;
env = getenv("SDFG_NUM_THREADS");
if (env != nullptr && strtoul(env, NULL, 10) != 0)
num_cores = strtoul(env, NULL, 10);
if (num_cores < 1)
num_cores = 1;
int device = next_device.fetch_add(1) % num_devices;
return new GPU_DFG(device);
}
void stream_emulator_run(void *dfg) {}
void stream_emulator_delete(void *dfg) { delete (GPU_DFG *)dfg; }
#endif