Skip to content

GitLab

Commit c1bed0d9 authored by superhao1995's avatar superhao1995 Committed by binmakeswell
Browse files

[NFC] polish colossalai/kernel/cuda_native/csrc/multi_tensor_lamb.cu code stype (#628)

parent 0a96338b
Show whitespace changes
Inline Side-by-side
Showing with 304 additions and 377 deletions
colossalai/kernel/cuda_native/csrc/multi_tensor_lamb.cu
View file @ c1bed0d9
// modified from https://github.com/NVIDIA/apex/blob/master/csrc/multi_tensor_lamb.cu
// modified from
// https://github.com/NVIDIA/apex/blob/master/csrc/multi_tensor_lamb.cu
#include <ATen/ATen.h>
#include <ATen/AccumulateType.h>
#include <ATen/cuda/CUDAContext.h>
......@@ -8,57 +9,43 @@
#include <assert.h>
#include "type_shim.h"
#include "multi_tensor_apply.cuh"
#include "type_shim.h"
#define BLOCK_SIZE 512
#define ILP 4
template <typename T>
__device__ __forceinline__ bool is_aligned(T *p)
{
template <typename T> __device__ __forceinline__ bool is_aligned(T *p) {
return ((uint64_t)p) % (ILP * sizeof(T)) == 0;
}
template <typename T>
__device__ __forceinline__ void load_store(T *dst, T *src, int dst_offset, int src_offset)
{
typedef typename std::aligned_storage<ILP * sizeof(T), ILP * alignof(T)>::type LT;
__device__ __forceinline__ void load_store(T *dst, T *src, int dst_offset,
int src_offset) {
typedef
typename std::aligned_storage<ILP * sizeof(T), ILP * alignof(T)>::type LT;
((LT *)dst)[dst_offset] = ((LT *)src)[src_offset];
}
typedef enum
{
typedef enum {
MOMENT_MODE_0 = 0, // L2 regularization mode
MOMENT_MODE_1 = 1 // Decoupled weight decay mode
} adamMode_t;
std::tuple<at::Tensor, at::Tensor> multi_tensor_l2norm_cuda(
int chunk_size,
at::Tensor noop_flag,
std::tuple<at::Tensor, at::Tensor>
multi_tensor_l2norm_cuda(int chunk_size, at::Tensor noop_flag,
std::vector<std::vector<at::Tensor>> tensor_lists,
at::optional<bool> per_tensor_python);
using MATH_T = float;
template <typename T>
struct LAMBStage1Functor
{
__device__ __forceinline__ void operator()(
int chunk_size,
volatile int *noop_gmem,
TensorListMetadata<4> &tl,
const float beta1,
const float beta2,
const float beta3,
const float beta1_correction,
const float beta2_correction,
const float epsilon,
adamMode_t mode,
const float decay,
const float *global_grad_norm,
const float max_global_grad_norm)
{
template <typename T> struct LAMBStage1Functor {
__device__ __forceinline__ void
operator()(int chunk_size, volatile int *noop_gmem, TensorListMetadata<4> &tl,
const float beta1, const float beta2, const float beta3,
const float beta1_correction, const float beta2_correction,
const float epsilon, adamMode_t mode, const float decay,
const float *global_grad_norm, const float max_global_grad_norm) {
// I'd like this kernel to propagate infs/nans.
// if(*noop_gmem == 1)
// return;
......@@ -67,7 +54,10 @@ struct LAMBStage1Functor
int chunk_idx = tl.block_to_chunk[blockIdx.x];
int n = tl.sizes[tensor_loc];
float clipped_global_grad_norm = (*global_grad_norm) > max_global_grad_norm ? (*global_grad_norm) / max_global_grad_norm : 1.0f;
float clipped_global_grad_norm =
(*global_grad_norm) > max_global_grad_norm
? (*global_grad_norm) / max_global_grad_norm
: 1.0f;
T *g = (T *)tl.addresses[0][tensor_loc];
g += chunk_idx * chunk_size;
......@@ -88,19 +78,15 @@ struct LAMBStage1Functor
MATH_T r_m[ILP];
MATH_T r_v[ILP];
// to make things simple, we put aligned case in a different code path
if (n % ILP == 0 &&
chunk_size % ILP == 0 &&
is_aligned(g) &&
is_aligned(p) &&
is_aligned(m) &&
is_aligned(v))
{
if (n % ILP == 0 && chunk_size % ILP == 0 && is_aligned(g) &&
is_aligned(p) && is_aligned(m) && is_aligned(v)) {
T l_g[ILP];
T l_p[ILP];
T l_m[ILP];
T l_v[ILP];
for (int i_start = threadIdx.x; i_start * ILP < n && i_start * ILP < chunk_size; i_start += blockDim.x)
{
for (int i_start = threadIdx.x;
i_start * ILP < n && i_start * ILP < chunk_size;
i_start += blockDim.x) {
// load
load_store(l_g, g, 0, i_start);
if (decay != 0)
......@@ -109,25 +95,19 @@ struct LAMBStage1Functor
load_store(l_v, v, 0, i_start);
// unpack
#pragma unroll
for (int ii = 0; ii < ILP; ii++)
{
for (int ii = 0; ii < ILP; ii++) {
r_g[ii] = l_g[ii];
if (decay == 0)
{
if (decay == 0) {
r_p[ii] = MATH_T(0);
}
else
{
} else {
r_p[ii] = l_p[ii];
}
r_m[ii] = l_m[ii];
r_v[ii] = l_v[ii];
}
#pragma unroll
for (int ii = 0; ii < ILP; ii++)
{
if (mode == MOMENT_MODE_0)
{
for (int ii = 0; ii < ILP; ii++) {
if (mode == MOMENT_MODE_0) {
MATH_T scaled_grad = r_g[ii] / clipped_global_grad_norm;
// L2 on scaled grad
scaled_grad = scaled_grad + decay * r_p[ii];
......@@ -137,9 +117,7 @@ struct LAMBStage1Functor
MATH_T next_v_unbiased = r_v[ii] / beta2_correction;
MATH_T denom = sqrtf(next_v_unbiased) + epsilon;
r_p[ii] = next_m_unbiased / denom;
}
else
{
} else {
MATH_T scaled_grad = r_g[ii] / clipped_global_grad_norm;
r_m[ii] = r_m[ii] * beta1 + beta3 * scaled_grad;
r_v[ii] = r_v[ii] * beta2 + (1 - beta2) * scaled_grad * scaled_grad;
......@@ -150,8 +128,7 @@ struct LAMBStage1Functor
}
}
#pragma unroll
for (int ii = 0; ii < ILP; ii++)
{
for (int ii = 0; ii < ILP; ii++) {
l_p[ii] = r_p[ii];
l_m[ii] = r_m[ii];
l_v[ii] = r_v[ii];
......@@ -161,39 +138,28 @@ struct LAMBStage1Functor
load_store(m, l_m, i_start, 0);
load_store(v, l_v, i_start, 0);
}
}
else
{
} else {
// see note in multi_tensor_scale_kernel.cu
for (int i_start = 0;
i_start < n && i_start < chunk_size;
i_start += blockDim.x * ILP)
{
for (int i_start = 0; i_start < n && i_start < chunk_size;
i_start += blockDim.x * ILP) {
MATH_T r_g[ILP];
MATH_T r_p[ILP];
MATH_T r_m[ILP];
MATH_T r_v[ILP];
#pragma unroll
for (int ii = 0; ii < ILP; ii++)
{
for (int ii = 0; ii < ILP; ii++) {
int i = i_start + threadIdx.x + ii * blockDim.x;
if (i < n && i < chunk_size)
{
if (i < n && i < chunk_size) {
r_g[ii] = g[i];
// special ?optimization? for lamb stage 1
if (decay == 0)
{
if (decay == 0) {
r_p[ii] = MATH_T(0);
}
else
{
} else {
r_p[ii] = p[i];
}
r_m[ii] = m[i];
r_v[ii] = v[i];
}
else
{
} else {
r_g[ii] = MATH_T(0);
r_p[ii] = MATH_T(0);
r_m[ii] = MATH_T(0);
......@@ -201,10 +167,8 @@ struct LAMBStage1Functor
}
}
#pragma unroll
for (int ii = 0; ii < ILP; ii++)
{
if (mode == MOMENT_MODE_0)
{
for (int ii = 0; ii < ILP; ii++) {
if (mode == MOMENT_MODE_0) {
MATH_T scaled_grad = r_g[ii] / clipped_global_grad_norm;
// L2 on scaled grad
scaled_grad = scaled_grad + decay * r_p[ii];
......@@ -214,9 +178,7 @@ struct LAMBStage1Functor
MATH_T next_v_unbiased = r_v[ii] / beta2_correction;
MATH_T denom = sqrtf(next_v_unbiased) + epsilon;
r_p[ii] = next_m_unbiased / denom;
}
else
{
} else {
MATH_T scaled_grad = r_g[ii] / clipped_global_grad_norm;
r_m[ii] = r_m[ii] * beta1 + beta3 * scaled_grad;
r_v[ii] = r_v[ii] * beta2 + (1 - beta2) * scaled_grad * scaled_grad;
......@@ -227,11 +189,9 @@ struct LAMBStage1Functor
}
}
#pragma unroll
for (int ii = 0; ii < ILP; ii++)
{
for (int ii = 0; ii < ILP; ii++) {
int i = i_start + threadIdx.x + ii * blockDim.x;
if (i < n && i < chunk_size)
{
if (i < n && i < chunk_size) {
g[i] = r_p[ii];
m[i] = r_m[ii];
v[i] = r_v[ii];
......@@ -244,19 +204,12 @@ struct LAMBStage1Functor
// Step 2 reads in 'update' value and per-tensor param_norm and update_norm.
// It computes new parameter value.
template <typename T>
struct LAMBStage2Functor
{
__device__ __forceinline__ void operator()(
int chunk_size,
volatile int *noop_gmem,
TensorListMetadata<2> &tl,
template <typename T> struct LAMBStage2Functor {
__device__ __forceinline__ void
operator()(int chunk_size, volatile int *noop_gmem, TensorListMetadata<2> &tl,
const float *per_tensor_param_norm,
const float *per_tensor_update_norm,
const float learning_rate,
const float decay,
bool use_nvlamb)
{
const float *per_tensor_update_norm, const float learning_rate,
const float decay, bool use_nvlamb) {
// I'd like this kernel to propagate infs/nans.
// if(*noop_gmem == 1)
// return;
......@@ -269,11 +222,12 @@ struct LAMBStage2Functor
MATH_T ratio = learning_rate;
// nvlamb: apply adaptive learning rate to all parameters
// otherwise, only apply to those with non-zero weight decay
if (use_nvlamb || (decay != 0.0))
{
if (use_nvlamb || (decay != 0.0)) {
float param_norm = per_tensor_param_norm[tensor_num];
float update_norm = per_tensor_update_norm[tensor_num];
ratio = (update_norm != 0.0f && param_norm != 0.0f) ? learning_rate * (param_norm / update_norm) : learning_rate;
ratio = (update_norm != 0.0f && param_norm != 0.0f)
? learning_rate * (param_norm / update_norm)
: learning_rate;
}
T *update = (T *)tl.addresses[0][tensor_loc];
......@@ -285,55 +239,44 @@ struct LAMBStage2Functor
n -= chunk_idx * chunk_size;
// to make things simple, we put aligned case in a different code path
if (n % ILP == 0 &&
chunk_size % ILP == 0 &&
is_aligned(p) &&
is_aligned(update))
{
if (n % ILP == 0 && chunk_size % ILP == 0 && is_aligned(p) &&
is_aligned(update)) {
T r_p[ILP];
T r_update[ILP];
for (int i_start = threadIdx.x; i_start * ILP < n && i_start * ILP < chunk_size; i_start += blockDim.x)
{
for (int i_start = threadIdx.x;
i_start * ILP < n && i_start * ILP < chunk_size;
i_start += blockDim.x) {
// load
load_store(r_p, p, 0, i_start);
load_store(r_update, update, 0, i_start);
#pragma unroll
for (int ii = 0; ii < ILP; ii++)
{
r_p[ii] = static_cast<MATH_T>(r_p[ii]) - (ratio * static_cast<MATH_T>(r_update[ii]));
for (int ii = 0; ii < ILP; ii++) {
r_p[ii] = static_cast<MATH_T>(r_p[ii]) -
(ratio * static_cast<MATH_T>(r_update[ii]));
}
load_store(p, r_p, i_start, 0);
}
}
else
{
for (int i_start = 0;
i_start < n && i_start < chunk_size;
i_start += blockDim.x * ILP)
{
} else {
for (int i_start = 0; i_start < n && i_start < chunk_size;
i_start += blockDim.x * ILP) {
MATH_T r_p[ILP];
MATH_T r_update[ILP];
#pragma unroll
for (int ii = 0; ii < ILP; ii++)
{
for (int ii = 0; ii < ILP; ii++) {
int i = i_start + threadIdx.x + ii * blockDim.x;
if (i < n && i < chunk_size)
{
if (i < n && i < chunk_size) {
r_p[ii] = p[i];
r_update[ii] = update[i];
}
}
#pragma unroll
for (int ii = 0; ii < ILP; ii++)
{
for (int ii = 0; ii < ILP; ii++) {
r_p[ii] = r_p[ii] - (ratio * r_update[ii]);
}
#pragma unroll
for (int ii = 0; ii < ILP; ii++)
{
for (int ii = 0; ii < ILP; ii++) {
int i = i_start + threadIdx.x + ii * blockDim.x;
if (i < n && i < chunk_size)
{
if (i < n && i < chunk_size) {
p[i] = r_p[ii];
}
}
......@@ -342,33 +285,25 @@ struct LAMBStage2Functor
}
};
void multi_tensor_lamb_cuda(
int chunk_size,
at::Tensor noop_flag,
void multi_tensor_lamb_cuda(int chunk_size, at::Tensor noop_flag,
std::vector<std::vector<at::Tensor>> tensor_lists,
const float lr,
const float beta1,
const float beta2,
const float epsilon,
const int step,
const int bias_correction,
const float weight_decay,
const int grad_averaging,
const int mode,
at::Tensor global_grad_norm,
const float lr, const float beta1,
const float beta2, const float epsilon,
const int step, const int bias_correction,
const float weight_decay, const int grad_averaging,
const int mode, at::Tensor global_grad_norm,
const float max_grad_norm,
at::optional<bool> use_nvlamb_python)
{
at::optional<bool> use_nvlamb_python) {
using namespace at;
// Master weight and 32bit momentum(potentially changing) is not handled by this
// So we assume every tensor are all in the same type
// Master weight and 32bit momentum(potentially changing) is not handled by
// this So we assume every tensor are all in the same type
bool use_nvlamb = use_nvlamb_python.has_value() ? use_nvlamb_python.value() : false;
bool use_nvlamb =
use_nvlamb_python.has_value() ? use_nvlamb_python.value() : false;
// Handle bias correction mode
float bias_correction1 = 1.0f, bias_correction2 = 1.0f;
if (bias_correction == 1)
{
if (bias_correction == 1) {
bias_correction1 = 1 - std::pow(beta1, step);
bias_correction2 = 1 - std::pow(beta2, step);
}
......@@ -378,50 +313,42 @@ void multi_tensor_lamb_cuda(
if (grad_averaging == 1)
beta3 = 1 - beta1;
std::vector<std::vector<at::Tensor>> grad_list(tensor_lists.begin(), tensor_lists.begin() + 1);
std::vector<std::vector<at::Tensor>> param_list(tensor_lists.begin() + 1, tensor_lists.begin() + 2);
std::vector<std::vector<at::Tensor>> grad_list(tensor_lists.begin(),
tensor_lists.begin() + 1);
std::vector<std::vector<at::Tensor>> param_list(tensor_lists.begin() + 1,
tensor_lists.begin() + 2);
// Compute per tensor param norm
auto param_norm_tuple = multi_tensor_l2norm_cuda(chunk_size, noop_flag, param_list, true);
auto param_norm_tuple =
multi_tensor_l2norm_cuda(chunk_size, noop_flag, param_list, true);
// We now in-place modify grad to store update before compute its norm
// Generally this is not a issue since people modify grad in step() method all the time
// We can also grab list of empty tensor to avoid this, but I'd like to save space/cpu code
DISPATCH_FLOAT_AND_HALF(tensor_lists[0][0].scalar_type(), 0, "lamb_stage_1",
multi_tensor_apply<4>(
BLOCK_SIZE,
chunk_size,
noop_flag,
tensor_lists,
LAMBStage1Functor<scalar_t_0>(),
beta1,
beta2,
// Generally this is not a issue since people modify grad in step() method all
// the time We can also grab list of empty tensor to avoid this, but I'd like
// to save space/cpu code
DISPATCH_FLOAT_AND_HALF(
tensor_lists[0][0].scalar_type(), 0, "lamb_stage_1",
multi_tensor_apply<4>(BLOCK_SIZE, chunk_size, noop_flag, tensor_lists,
LAMBStage1Functor<scalar_t_0>(), beta1, beta2,
beta3, // 1-beta1 or 1 depends on averaging mode
bias_correction1,
bias_correction2,
epsilon,
(adamMode_t)mode,
weight_decay,
global_grad_norm.DATA_PTR<float>(),
max_grad_norm);)
bias_correction1, bias_correction2, epsilon,
(adamMode_t)mode, weight_decay,
global_grad_norm.DATA_PTR<float>(), max_grad_norm);)
// Compute update norms
auto update_norm_tuple = multi_tensor_l2norm_cuda(chunk_size, noop_flag, grad_list, true);
auto update_norm_tuple =
multi_tensor_l2norm_cuda(chunk_size, noop_flag, grad_list, true);
std::vector<std::vector<at::Tensor>> grad_param_list(tensor_lists.begin(), tensor_lists.begin() + 2);
std::vector<std::vector<at::Tensor>> grad_param_list(
tensor_lists.begin(), tensor_lists.begin() + 2);
DISPATCH_FLOAT_AND_HALF(tensor_lists[0][0].scalar_type(), 0, "lamb_stage_2",
multi_tensor_apply<2>(
BLOCK_SIZE,
chunk_size,
noop_flag,
grad_param_list,
DISPATCH_FLOAT_AND_HALF(
tensor_lists[0][0].scalar_type(), 0, "lamb_stage_2",
multi_tensor_apply<2>(BLOCK_SIZE, chunk_size, noop_flag, grad_param_list,
LAMBStage2Functor<scalar_t_0>(),
std::get<1>(param_norm_tuple).DATA_PTR<float>(),
std::get<1>(update_norm_tuple).DATA_PTR<float>(),
lr,
weight_decay,
use_nvlamb);)
lr, weight_decay, use_nvlamb);)
AT_CUDA_CHECK(cudaGetLastError());
}
\ No newline at end of file
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment