github
/
ColossalAI
mirror of https://github.com/hpcaitech/ColossalAI
1
0
Fork 0
Code Issues Projects Releases Wiki Activity

[NFC] polish colossalai/kernel/cuda_native/csrc/cpu_adam.cpp code style (#636)

pull/673/head
Sze-qq 2022-04-02 13:28:57 +08:00 committed by binmakeswell
parent 6fcb381801
commit 10591ecdf9
1 changed files with 372 additions and 375 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

193
colossalai/kernel/cuda_native/csrc/cpu_adam.cpp
View File

@ -20,29 +20,23 @@ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE
*/
#include "cpu_adam.h"
#include <math.h>
#include <omp.h>
#include <torch/extension.h>
#include <iostream>
#include <math.h>
#include <memory>
#include <omp.h>
#include <string.h>
#include <torch/extension.h>
#include <type_traits>
#include <unordered_map>
#include <string.h>
static std::unordered_map<int, std::shared_ptr<void>> s_optimizers;
// C++ interface
void Adam_Optimizer::Step_1(float* _params,
float* grads,
float* _exp_avg,
float* _exp_avg_sq,
size_t _param_size,
bool param_half_precision,
bool grad_half_precision,
float loss_scale)
{
void Adam_Optimizer::Step_1(float *_params, float *grads, float *_exp_avg,
float *_exp_avg_sq, size_t _param_size,
bool param_half_precision, bool grad_half_precision,
float loss_scale) {
size_t rounded_size = 0;
float betta1_minus1 = 1 - _betta1;
@ -82,12 +76,14 @@ void Adam_Optimizer::Step_1(float* _params,
AVX_Data weight_decay_4;
if (_weight_decay > 0)
weight_decay_4.data = (_adamw_mode ? SIMD_SET(w_decay) : SIMD_SET(_weight_decay));
weight_decay_4.data =
(_adamw_mode ? SIMD_SET(w_decay) : SIMD_SET(_weight_decay));
rounded_size = ROUND_DOWN(_param_size, SIMD_WIDTH);
for (size_t t = 0; t < rounded_size; t += TILE) {
size_t copy_size = TILE;
if ((t + TILE) > rounded_size) copy_size = rounded_size - t;
if ((t + TILE) > rounded_size)
copy_size = rounded_size - t;
size_t offset = copy_size + t;
#pragma omp parallel for
@ -120,16 +116,19 @@ void Adam_Optimizer::Step_1(float* _params,
grad_4.data = SIMD_FMA(param_4.data, weight_decay_4.data, grad_4.data);
}
momentum_4.data = SIMD_MUL(momentum_4.data, betta1_4.data);
momentum_4.data = SIMD_FMA(grad_4.data, betta1_minus1_4.data, momentum_4.data);
momentum_4.data =
SIMD_FMA(grad_4.data, betta1_minus1_4.data, momentum_4.data);
variance_4.data = SIMD_MUL(variance_4.data, betta2_4.data);
grad_4.data = SIMD_MUL(grad_4.data, grad_4.data);
variance_4.data = SIMD_FMA(grad_4.data, betta2_minus1_4.data, variance_4.data);
variance_4.data =
SIMD_FMA(grad_4.data, betta2_minus1_4.data, variance_4.data);
grad_4.data = SIMD_SQRT(variance_4.data);
grad_4.data = SIMD_FMA(grad_4.data, bias2_sqrt.data, eps_4.data);
grad_4.data = SIMD_DIV(momentum_4.data, grad_4.data);
if (_weight_decay > 0 && _adamw_mode) {
param_4.data = SIMD_FMA(param_4.data, weight_decay_4.data, param_4.data);
param_4.data =
SIMD_FMA(param_4.data, weight_decay_4.data, param_4.data);
}
param_4.data = SIMD_FMA(grad_4.data, step_size_4.data, param_4.data);
@ -146,17 +145,23 @@ void Adam_Optimizer::Step_1(float* _params,
if (_param_size > rounded_size) {
for (size_t t = rounded_size; t < _param_size; t += TILE) {
size_t copy_size = TILE;
if ((t + TILE) > _param_size) copy_size = _param_size - t;
if ((t + TILE) > _param_size)
copy_size = _param_size - t;
size_t offset = copy_size + t;
#pragma omp parallel for
for (size_t k = t; k < offset; k++) {
float grad = grad_half_precision ? (float)grads_cast_h[k] : grads[k];
if (loss_scale > 0) { grad /= loss_scale; }
float param = param_half_precision ? (float)params_cast_h[k] : _params[k];
if (loss_scale > 0) {
grad /= loss_scale;
}
float param =
param_half_precision ? (float)params_cast_h[k] : _params[k];
float momentum = _exp_avg[k];
float variance = _exp_avg_sq[k];
if (_weight_decay > 0 && !_adamw_mode) { grad = param * _weight_decay + grad; }
if (_weight_decay > 0 && !_adamw_mode) {
grad = param * _weight_decay + grad;
}
momentum = momentum * _betta1;
momentum = grad * betta1_minus1 + momentum;
@ -167,7 +172,9 @@ void Adam_Optimizer::Step_1(float* _params,
grad = sqrt(variance);
grad = grad * _bias_correction2 + _eps;
grad = momentum / grad;
if (_weight_decay > 0 && _adamw_mode) { param += w_decay * param; }
if (_weight_decay > 0 && _adamw_mode) {
param += w_decay * param;
}
param = grad * step_size + param;
if (param_half_precision)
@ -181,15 +188,10 @@ void Adam_Optimizer::Step_1(float* _params,
}
}
void Adam_Optimizer::Step_4(float* _params,
float* grads,
float* _exp_avg,
float* _exp_avg_sq,
size_t _param_size,
bool param_half_precision,
bool grad_half_precision,
float loss_scale)
{
void Adam_Optimizer::Step_4(float *_params, float *grads, float *_exp_avg,
float *_exp_avg_sq, size_t _param_size,
bool param_half_precision, bool grad_half_precision,
float loss_scale) {
size_t rounded_size = 0;
__half *params_cast_h = NULL;
@ -227,12 +229,14 @@ void Adam_Optimizer::Step_4(float* _params,
float w_decay = -1 * _alpha * _weight_decay;
AVX_Data weight_decay_4;
if (_weight_decay > 0)
weight_decay_4.data = (_adamw_mode ? SIMD_SET(w_decay) : SIMD_SET(_weight_decay));
weight_decay_4.data =
(_adamw_mode ? SIMD_SET(w_decay) : SIMD_SET(_weight_decay));
rounded_size = ROUND_DOWN(_param_size, SIMD_WIDTH * 4);
for (size_t t = 0; t < rounded_size; t += TILE) {
size_t copy_size = TILE;
if ((t + TILE) > rounded_size) copy_size = rounded_size - t;
if ((t + TILE) > rounded_size)
copy_size = rounded_size - t;
size_t offset = copy_size + t;
#pragma omp parallel for
@ -265,23 +269,29 @@ void Adam_Optimizer::Step_4(float* _params,
}
if (_weight_decay > 0 && !_adamw_mode) {
grad_4[j].data = SIMD_FMA(param_4[j].data, weight_decay_4.data, grad_4[j].data);
grad_4[j].data =
SIMD_FMA(param_4[j].data, weight_decay_4.data, grad_4[j].data);
}
momentum_4[j].data = SIMD_MUL(momentum_4[j].data, betta1_4.data);
momentum_4[j].data = SIMD_FMA(grad_4[j].data, betta1_minus1_4.data, momentum_4[j].data);
momentum_4[j].data =
SIMD_FMA(grad_4[j].data, betta1_minus1_4.data, momentum_4[j].data);
variance_4[j].data = SIMD_MUL(variance_4[j].data, betta2_4.data);
grad_4[j].data = SIMD_MUL(grad_4[j].data, grad_4[j].data);
variance_4[j].data = SIMD_FMA(grad_4[j].data, betta2_minus1_4.data, variance_4[j].data);
variance_4[j].data =
SIMD_FMA(grad_4[j].data, betta2_minus1_4.data, variance_4[j].data);
grad_4[j].data = SIMD_SQRT(variance_4[j].data);
grad_4[j].data = SIMD_FMA(grad_4[j].data, bias2_sqrt.data, eps_4.data);
grad_4[j].data = SIMD_DIV(momentum_4[j].data, grad_4[j].data);
if (_weight_decay > 0 && _adamw_mode) {
param_4[j].data = SIMD_FMA(param_4[j].data, weight_decay_4.data, param_4[j].data);
param_4[j].data =
SIMD_FMA(param_4[j].data, weight_decay_4.data, param_4[j].data);
}
param_4[j].data = SIMD_FMA(grad_4[j].data, step_size_4.data, param_4[j].data);
param_4[j].data =
SIMD_FMA(grad_4[j].data, step_size_4.data, param_4[j].data);
if (param_half_precision) {
SIMD_STORE_HALF((float*)(params_cast_h + i + SIMD_WIDTH * j), param_4[j].data);
SIMD_STORE_HALF((float *)(params_cast_h + i + SIMD_WIDTH * j),
param_4[j].data);
} else {
SIMD_STORE(_params + i + SIMD_WIDTH * j, param_4[j].data);
}
@ -292,27 +302,21 @@ void Adam_Optimizer::Step_4(float* _params,
}
#endif
if (_param_size > rounded_size)
Step_1((param_half_precision ? (float*)(params_cast_h + rounded_size) : _params + rounded_size),
(grad_half_precision ? (float*)(grads_cast_h + rounded_size) : grads + rounded_size),
(_exp_avg + rounded_size),
(_exp_avg_sq + rounded_size),
(_param_size - rounded_size),
param_half_precision,
grad_half_precision,
loss_scale);
Step_1((param_half_precision ? (float *)(params_cast_h + rounded_size)
: _params + rounded_size),
(grad_half_precision ? (float *)(grads_cast_h + rounded_size)
: grads + rounded_size),
(_exp_avg + rounded_size), (_exp_avg_sq + rounded_size),
(_param_size - rounded_size), param_half_precision,
grad_half_precision, loss_scale);
}
int create_adam_optimizer(int optimizer_id,
float alpha = 1e-3,
float betta1 = 0.9,
float betta2 = 0.999,
float eps = 1e-8,
float weight_decay = 0,
bool adamw_mode = true,
bool should_log = false)
{
auto opt =
std::make_shared<Adam_Optimizer>(alpha, betta1, betta2, eps, weight_decay, adamw_mode);
int create_adam_optimizer(int optimizer_id, float alpha = 1e-3,
float betta1 = 0.9, float betta2 = 0.999,
float eps = 1e-8, float weight_decay = 0,
bool adamw_mode = true, bool should_log = false) {
auto opt = std::make_shared<Adam_Optimizer>(alpha, betta1, betta2, eps,
weight_decay, adamw_mode);
s_optimizers[optimizer_id] = opt;
@ -329,28 +333,18 @@ int create_adam_optimizer(int optimizer_id,
#endif
#endif
printf("Adam Optimizer #%d is created with %s arithmetic capability.\n",
optimizer_id,
avx_type.c_str());
optimizer_id, avx_type.c_str());
printf("Config: alpha=%f, betas=(%f, %f), weight_decay=%f, adam_w=%d\n",
alpha,
betta1,
betta2,
weight_decay,
(int)adamw_mode);
alpha, betta1, betta2, weight_decay, (int)adamw_mode);
}
return 0;
}
void Adam_Optimizer::Step_8(float* _params,
float* grads,
float* _exp_avg,
float* _exp_avg_sq,
size_t _param_size,
bool param_half_precision,
bool grad_half_precision,
float loss_scale)
{
void Adam_Optimizer::Step_8(float *_params, float *grads, float *_exp_avg,
float *_exp_avg_sq, size_t _param_size,
bool param_half_precision, bool grad_half_precision,
float loss_scale) {
size_t rounded_size = 0;
__half *params_cast_h = NULL;
__half *grads_cast_h = NULL;
@ -386,12 +380,14 @@ void Adam_Optimizer::Step_8(float* _params,
float w_decay = -1 * _alpha * _weight_decay;
AVX_Data weight_decay_4;
if (_weight_decay > 0)
weight_decay_4.data = (_adamw_mode ? SIMD_SET(w_decay) : SIMD_SET(_weight_decay));
weight_decay_4.data =
(_adamw_mode ? SIMD_SET(w_decay) : SIMD_SET(_weight_decay));
rounded_size = ROUND_DOWN(_param_size, SIMD_WIDTH * 8);
for (size_t t = 0; t < rounded_size; t += TILE) {
size_t copy_size = TILE;
if ((t + TILE) > rounded_size) copy_size = rounded_size - t;
if ((t + TILE) > rounded_size)
copy_size = rounded_size - t;
size_t offset = copy_size + t;
#pragma omp parallel for
@ -424,23 +420,29 @@ void Adam_Optimizer::Step_8(float* _params,
}
if (_weight_decay > 0 && !_adamw_mode) {
grad_4[j].data = SIMD_FMA(param_4[j].data, weight_decay_4.data, grad_4[j].data);
grad_4[j].data =
SIMD_FMA(param_4[j].data, weight_decay_4.data, grad_4[j].data);
}
momentum_4[j].data = SIMD_MUL(momentum_4[j].data, betta1_4.data);
momentum_4[j].data = SIMD_FMA(grad_4[j].data, betta1_minus1_4.data, momentum_4[j].data);
momentum_4[j].data =
SIMD_FMA(grad_4[j].data, betta1_minus1_4.data, momentum_4[j].data);
variance_4[j].data = SIMD_MUL(variance_4[j].data, betta2_4.data);
grad_4[j].data = SIMD_MUL(grad_4[j].data, grad_4[j].data);
variance_4[j].data = SIMD_FMA(grad_4[j].data, betta2_minus1_4.data, variance_4[j].data);
variance_4[j].data =
SIMD_FMA(grad_4[j].data, betta2_minus1_4.data, variance_4[j].data);
grad_4[j].data = SIMD_SQRT(variance_4[j].data);
grad_4[j].data = SIMD_FMA(grad_4[j].data, bias2_sqrt.data, eps_4.data);
grad_4[j].data = SIMD_DIV(momentum_4[j].data, grad_4[j].data);
if (_weight_decay > 0 && _adamw_mode) {
param_4[j].data = SIMD_FMA(param_4[j].data, weight_decay_4.data, param_4[j].data);
param_4[j].data =
SIMD_FMA(param_4[j].data, weight_decay_4.data, param_4[j].data);
}
param_4[j].data = SIMD_FMA(grad_4[j].data, step_size_4.data, param_4[j].data);
param_4[j].data =
SIMD_FMA(grad_4[j].data, step_size_4.data, param_4[j].data);
if (param_half_precision) {
SIMD_STORE_HALF((float*)(params_cast_h + i + SIMD_WIDTH * j), param_4[j].data);
SIMD_STORE_HALF((float *)(params_cast_h + i + SIMD_WIDTH * j),
param_4[j].data);
} else {
SIMD_STORE(_params + i + SIMD_WIDTH * j, param_4[j].data);
}
@ -452,14 +454,13 @@ void Adam_Optimizer::Step_8(float* _params,
}
#endif
if (_param_size > rounded_size)
Step_4((param_half_precision ? (float*)(params_cast_h + rounded_size) : _params + rounded_size),
(grad_half_precision ? (float*)(grads_cast_h + rounded_size) : grads + rounded_size),
(_exp_avg + rounded_size),
(_exp_avg_sq + rounded_size),
(_param_size - rounded_size),
param_half_precision,
grad_half_precision,
loss_scale);
Step_4((param_half_precision ? (float *)(params_cast_h + rounded_size)
: _params + rounded_size),
(grad_half_precision ? (float *)(grads_cast_h + rounded_size)
: grads + rounded_size),
(_exp_avg + rounded_size), (_exp_avg_sq + rounded_size),
(_param_size - rounded_size), param_half_precision,
grad_half_precision, loss_scale);
}
int adam_step(int optimizer_id,
@ -501,16 +502,12 @@ int adam_step(int optimizer_id,
return 0;
}
int destroy_adam_optimizer(int optimizer_id)
{
int destroy_adam_optimizer(int optimizer_id) {
s_optimizers.erase(optimizer_id);
return 0;
}
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m)
{
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.def("adam_update", &adam_step, "CPU Adam update (C++)");
m.def("create_adam", &create_adam_optimizer, "CPU Adam (C++)");
m.def("destroy_adam", &destroy_adam_optimizer, "CPU Adam destroy (C++)");