Skip to content

GitLab

Commit 5a9bb616 authored by Shucai Xiao's avatar Shucai Xiao
Browse files

add std namespace for size_t

parents f8fa90bd 22500e6c
Hide whitespace changes
Inline Side-by-side
Showing with 76 additions and 65 deletions
src/targets/cpu/lowering.cpp
View file @ 5a9bb616
......@@ -533,7 +533,7 @@ struct cpu_softmax
{
argument result{output_shape};
auto batch_lens = output_shape.lens();
size_t n_dims = batch_lens[op.axis];
std::size_t n_dims = batch_lens[op.axis];
batch_lens[op.axis] = 1;
shape batch_shape{shape::int32_type, batch_lens};
......@@ -544,26 +544,26 @@ struct cpu_softmax
std::vector<value_type> batch_sum(batch_shape.elements(), value_type(0));
par_for(batch_shape.elements(), [&](auto i) {
auto idx = batch_shape.multi(i);
for(size_t j = 0; j < n_dims; ++j)
for(std::size_t j = 0; j < n_dims; ++j)
{
idx[op.axis] = j;
batch_max[i] = std::max(batch_max[i], input(idx.begin(), idx.end()));
}
for(size_t j = 0; j < n_dims; ++j)
for(std::size_t j = 0; j < n_dims; ++j)
{
idx[op.axis] = j;
size_t index = output_shape.index(idx);
output[index] = std::exp(input[index] - batch_max[i]);
idx[op.axis] = j;
std::size_t index = output_shape.index(idx);
output[index] = std::exp(input[index] - batch_max[i]);
}
for(size_t j = 0; j < n_dims; ++j)
for(std::size_t j = 0; j < n_dims; ++j)
{
idx[op.axis] = j;
batch_sum[i] += output(idx.begin(), idx.end());
}
for(size_t j = 0; j < n_dims; ++j)
for(std::size_t j = 0; j < n_dims; ++j)
{
idx[op.axis] = j;
output(idx.begin(), idx.end()) /= batch_sum[i];
......@@ -591,7 +591,7 @@ struct cpu_logsoftmax
{
argument result{output_shape};
auto batch_lens = output_shape.lens();
size_t n_dims = batch_lens[op.axis];
std::size_t n_dims = batch_lens[op.axis];
batch_lens[op.axis] = 1;
shape batch_shape{shape::int32_type, batch_lens};
......@@ -605,20 +605,20 @@ struct cpu_logsoftmax
par_for(batch_shape.elements(), [&](auto i) {
auto idx = batch_shape.multi(i);
for(size_t j = 0; j < n_dims; ++j)
for(std::size_t j = 0; j < n_dims; ++j)
{
idx[op.axis] = j;
batch_max[i] = std::max(batch_max[i], input(idx.begin(), idx.end()));
}
for(size_t j = 0; j < n_dims; ++j)
for(std::size_t j = 0; j < n_dims; ++j)
{
idx[op.axis] = j;
size_t index = output_shape.index(idx);
output[index] = input[index] - batch_max[i];
idx[op.axis] = j;
std::size_t index = output_shape.index(idx);
output[index] = input[index] - batch_max[i];
}
for(size_t j = 0; j < n_dims; ++j)
for(std::size_t j = 0; j < n_dims; ++j)
{
idx[op.axis] = j;
batch_sum[i] += std::exp(output(idx.begin(), idx.end()));
......@@ -626,7 +626,7 @@ struct cpu_logsoftmax
batch_sum[i] = std::log(batch_sum[i]);
for(size_t j = 0; j < n_dims; ++j)
for(std::size_t j = 0; j < n_dims; ++j)
{
idx[op.axis] = j;
output(idx.begin(), idx.end()) -= batch_sum[i];
......@@ -655,7 +655,7 @@ struct cpu_argmax
{
argument result{output_shape};
auto batch_lens = args.front().get_shape().lens();
size_t batch_item_num = batch_lens[op.axis];
std::size_t batch_item_num = batch_lens[op.axis];
batch_lens[op.axis] = 1;
shape batch_shape{shape::int32_type, batch_lens};
......@@ -665,7 +665,7 @@ struct cpu_argmax
auto data_idx = batch_shape.multi(i);
auto max_val = input[i];
int64_t max_index = 0;
for(size_t j = 1; j < batch_item_num; ++j)
for(std::size_t j = 1; j < batch_item_num; ++j)
{
data_idx[op.axis] = j;
if(max_val < input(data_idx.begin(), data_idx.end()))
......@@ -701,7 +701,7 @@ struct cpu_argmin
{
argument result{output_shape};
auto batch_lens = args.front().get_shape().lens();
size_t batch_item_num = batch_lens[op.axis];
std::size_t batch_item_num = batch_lens[op.axis];
batch_lens[op.axis] = 1;
shape batch_shape{shape::int32_type, batch_lens};
......@@ -711,7 +711,7 @@ struct cpu_argmin
auto data_idx = batch_shape.multi(i);
auto min_val = input[i];
int64_t min_index = 0;
for(size_t j = 1; j < batch_item_num; ++j)
for(std::size_t j = 1; j < batch_item_num; ++j)
{
data_idx[op.axis] = j;
if(min_val > input(data_idx.begin(), data_idx.end()))
......
src/targets/gpu/device/logsoftmax.cpp
View file @ 5a9bb616
......@@ -23,26 +23,27 @@ void logsoftmax(hipStream_t stream, const argument& result, const argument& arg,
hip_visit_all(result, arg, batch_shape)([&](auto output, auto input, auto batch) {
// use one block for items in one batch.
const size_t max_block_size = 1024;
size_t block_size = 1;
const std::size_t max_block_size = 1024;
std::size_t block_size = 1;
while(block_size < max_block_size and block_size < batch_item_num)
{
block_size *= 2;
}
launch(stream, batch_shape.elements() * block_size, block_size)([=](auto idx) __device__ {
size_t thr_idx = idx.local;
size_t blk_idx = idx.group;
using type = device_type<std::remove_cv_t<typename decltype(output)::value_type>>;
std::size_t thr_idx = idx.local;
std::size_t blk_idx = idx.group;
using type = device_type<std::remove_cv_t<typename decltype(output)::value_type>>;
MIGRAPHX_DEVICE_SHARED type lds_data[max_block_size + 1];
auto batch_idx = batch.multi(blk_idx);
auto data_idx = batch_idx;
// load data to lds and compute the batch max
size_t remaining_item_num = batch_item_num;
size_t round_item_num = (batch_item_num + block_size - 1) / block_size * block_size;
lds_data[max_block_size] = input[0];
for(size_t i = thr_idx; i < round_item_num; i += block_size)
std::size_t remaining_item_num = batch_item_num;
std::size_t round_item_num =
(batch_item_num + block_size - 1) / block_size * block_size;
lds_data[max_block_size] = input[0];
for(std::size_t i = thr_idx; i < round_item_num; i += block_size)
{
if(i < batch_item_num)
{
......@@ -62,7 +63,7 @@ void logsoftmax(hipStream_t stream, const argument& result, const argument& arg,
lds_data[max_block_size] = 0;
remaining_item_num = batch_item_num;
for(size_t i = thr_idx; i < round_item_num; i += block_size)
for(std::size_t i = thr_idx; i < round_item_num; i += block_size)
{
if(i < batch_item_num)
{
......@@ -81,7 +82,7 @@ void logsoftmax(hipStream_t stream, const argument& result, const argument& arg,
auto log_batch_sum = ::log(to_hip_type(lds_data[max_block_size])) + batch_max;
for(size_t i = thr_idx; i < batch_item_num; i += block_size)
for(std::size_t i = thr_idx; i < batch_item_num; i += block_size)
{
data_idx[axis] = i;
output[data_idx] = input[data_idx] - log_batch_sum;
......
src/targets/gpu/device/softmax.cpp
View file @ 5a9bb616
......@@ -15,34 +15,35 @@ namespace device {
void softmax(hipStream_t stream, const argument& result, const argument& arg, int axis)
{
auto lens = result.get_shape().lens();
auto batch_lens = lens;
size_t batch_item_num = lens[axis];
batch_lens[axis] = 1;
auto lens = result.get_shape().lens();
auto batch_lens = lens;
std::size_t batch_item_num = lens[axis];
batch_lens[axis] = 1;
migraphx::shape batch_shape{result.get_shape().type(), batch_lens};
hip_visit_all(result, arg, batch_shape)([&](auto output, auto input, auto batch) {
// use one block for items in one batch.
const size_t max_block_size = 1024;
size_t block_size = 1;
const std::size_t max_block_size = 1024;
std::size_t block_size = 1;
while(block_size < max_block_size and block_size < batch_item_num)
{
block_size *= 2;
}
launch(stream, batch_shape.elements() * block_size, block_size)([=](auto idx) __device__ {
size_t thr_idx = idx.local;
size_t blk_idx = idx.group;
using type = device_type<std::remove_cv_t<typename decltype(output)::value_type>>;
std::size_t thr_idx = idx.local;
std::size_t blk_idx = idx.group;
using type = device_type<std::remove_cv_t<typename decltype(output)::value_type>>;
MIGRAPHX_DEVICE_SHARED type lds_data[max_block_size + 1];
auto batch_idx = batch.multi(blk_idx);
auto data_idx = batch_idx;
// load data to lds and compute the batch max
size_t remaining_item_num = batch_item_num;
size_t round_item_num = (batch_item_num + block_size - 1) / block_size * block_size;
lds_data[max_block_size] = input[0];
for(size_t i = thr_idx; i < round_item_num; i += block_size)
std::size_t remaining_item_num = batch_item_num;
std::size_t round_item_num =
(batch_item_num + block_size - 1) / block_size * block_size;
lds_data[max_block_size] = input[0];
for(std::size_t i = thr_idx; i < round_item_num; i += block_size)
{
if(i < batch_item_num)
{
......@@ -63,7 +64,7 @@ void softmax(hipStream_t stream, const argument& result, const argument& arg, in
lds_data[max_block_size] = 0;
remaining_item_num = batch_item_num;
for(size_t i = thr_idx; i < round_item_num; i += block_size)
for(std::size_t i = thr_idx; i < round_item_num; i += block_size)
{
if(i < batch_item_num)
{
......@@ -81,7 +82,7 @@ void softmax(hipStream_t stream, const argument& result, const argument& arg, in
}
auto batch_sum = lds_data[max_block_size];
for(size_t i = thr_idx; i < batch_item_num; i += block_size)
for(std::size_t i = thr_idx; i < batch_item_num; i += block_size)
{
data_idx[axis] = i;
auto val = input[data_idx] - batch_max;
......
src/targets/gpu/include/migraphx/gpu/device/reduce_opers.hpp
View file @ 5a9bb616
......@@ -11,14 +11,17 @@ namespace gpu {
namespace device {
template <class T>
inline __device__ void
reduce_max(T* data_ptr, size_t block_size, size_t thr_idx, size_t item_num, size_t max_index)
inline __device__ void reduce_max(T* data_ptr,
std::size_t block_size,
std::size_t thr_idx,
std::size_t item_num,
std::size_t max_index)
{
while(true)
{
auto stride = (item_num + 1) / 2;
auto size = item_num / 2;
for(size_t i = thr_idx; i < size; i += block_size)
for(std::size_t i = thr_idx; i < size; i += block_size)
{
data_ptr[i] = ::max(to_hip_type(data_ptr[i]), to_hip_type(data_ptr[i + stride]));
}
......@@ -39,14 +42,17 @@ reduce_max(T* data_ptr, size_t block_size, size_t thr_idx, size_t item_num, size
}
template <class T>
inline __device__ void
reduce_min(T* data_ptr, size_t block_size, size_t thr_idx, size_t item_num, size_t min_index)
inline __device__ void reduce_min(T* data_ptr,
std::size_t block_size,
std::size_t thr_idx,
std::size_t item_num,
std::size_t min_index)
{
while(true)
{
auto stride = (item_num + 1) / 2;
auto size = item_num / 2;
for(size_t i = thr_idx; i < size; i += block_size)
for(std::size_t i = thr_idx; i < size; i += block_size)
{
data_ptr[i] = ::min(to_hip_type(data_ptr[i]), to_hip_type(data_ptr[i + stride]));
}
......@@ -69,16 +75,16 @@ reduce_min(T* data_ptr, size_t block_size, size_t thr_idx, size_t item_num, size
template <class T>
inline __device__ void reduce_argmax(T* data_ptr,
int64_t* index_ptr,
size_t block_size,
size_t thr_idx,
size_t item_num,
size_t max_index)
std::size_t block_size,
std::size_t thr_idx,
std::size_t item_num,
std::size_t max_index)
{
while(true)
{
auto stride = (item_num + 1) / 2;
auto size = item_num / 2;
for(size_t i = thr_idx; i < size; i += block_size)
for(std::size_t i = thr_idx; i < size; i += block_size)
{
if(data_ptr[i] < data_ptr[i + stride])
{
......@@ -108,16 +114,16 @@ inline __device__ void reduce_argmax(T* data_ptr,
template <class T>
inline __device__ void reduce_argmin(T* data_ptr,
int64_t* index_ptr,
size_t block_size,
size_t thr_idx,
size_t item_num,
size_t min_index)
std::size_t block_size,
std::size_t thr_idx,
std::size_t item_num,
std::size_t min_index)
{
while(true)
{
auto stride = (item_num + 1) / 2;
auto size = item_num / 2;
for(size_t i = thr_idx; i < size; i += block_size)
for(std::size_t i = thr_idx; i < size; i += block_size)
{
if(data_ptr[i] > data_ptr[i + stride])
{
......@@ -145,14 +151,17 @@ inline __device__ void reduce_argmin(T* data_ptr,
}
template <class T>
inline __device__ void
reduce_sum(T* data_ptr, size_t block_size, size_t thr_idx, size_t item_num, size_t sum_index)
inline __device__ void reduce_sum(T* data_ptr,
std::size_t block_size,
std::size_t thr_idx,
std::size_t item_num,
std::size_t sum_index)
{
while(true)
{
auto stride = (item_num + 1) / 2;
auto size = item_num / 2;
for(size_t i = thr_idx; i < size; i += block_size)
for(std::size_t i = thr_idx; i < size; i += block_size)
{
data_ptr[i] += data_ptr[i + stride];
}
......
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