Skip to content

GitLab

Commit 6ffd41ae authored by carlushuang's avatar carlushuang
Browse files

fix a bug when upsampling value

parent d8fef836
Show whitespace changes
Inline Side-by-side
Showing with 330 additions and 290 deletions
include/ck/tensor_operation/cpu/grid/gridwise_direct_conv_avx2.hpp
View file @ 6ffd41ae
......@@ -421,19 +421,25 @@ struct GridwiseDirectConvNHWCAvx2
}
return t_;
};
const ck::index_t num_works_nho = N * Ho;
const ck::index_t num_works_n = N;
const ck::index_t num_works_ho = Ho;
// const ck::index_t num_works_nho = N * Ho;
const ck::index_t num_works_wo = math::integer_divide_ceil(Wo, m_per_thread);
const ck::index_t num_works_k = math::integer_divide_ceil(K, n_per_thread);
auto distribute_num_threads_nho_wo_k = [&](ck::index_t& num_threads_nho_,
auto distribute_num_threads_n_ho_wo_k = [&](ck::index_t& num_threads_n_,
ck::index_t& num_threads_ho_,
ck::index_t& num_threads_wo_,
ck::index_t& num_threads_k_) {
// TODO: only consider multiply of 2 to divide threads
ck::index_t num_threads = total_threads;
num_threads_nho_ = devide_thread(num_threads, num_works_nho, 2);
num_threads = num_threads / num_threads_nho_;
num_threads_n_ = devide_thread(num_threads, num_works_n, 2);
num_threads = num_threads / num_threads_n_;
num_threads_ho_ = devide_thread(num_threads, num_works_ho, 2);
num_threads = num_threads / num_threads_ho_;
num_threads_wo_ = devide_thread(num_threads, num_works_wo, 2);
num_threads = num_threads / num_threads_wo_;
......@@ -442,14 +448,18 @@ struct GridwiseDirectConvNHWCAvx2
// num_threads = num_threads / num_threads_k_;
};
ck::index_t num_threads_nho;
ck::index_t num_threads_n;
ck::index_t num_threads_ho;
ck::index_t num_threads_wo;
ck::index_t num_threads_k;
distribute_num_threads_nho_wo_k(num_threads_nho, num_threads_wo, num_threads_k);
distribute_num_threads_n_ho_wo_k(
num_threads_n, num_threads_ho, num_threads_wo, num_threads_k);
const ck::index_t num_works_nho_per_thread =
math::integer_divide_ceil(num_works_nho, num_threads_nho);
const ck::index_t num_works_n_per_thread =
math::integer_divide_ceil(num_works_n, num_threads_n);
const ck::index_t num_works_ho_per_thread =
math::integer_divide_ceil(num_works_ho, num_threads_ho);
const ck::index_t num_works_wo_per_thread =
math::integer_divide_ceil(num_works_wo, num_threads_wo);
const ck::index_t num_works_k_per_thread =
......@@ -460,6 +470,14 @@ struct GridwiseDirectConvNHWCAvx2
// num_threads_nho, num_threads_wo, num_threads_k, num_works_nho_per_thread,
// num_works_wo_per_thread, num_works_k_per_thread); fflush(stdout);
if((X - 1) * Dx + 1 <= Px || (Y - 1) * Dy + 1 <= Py)
{
// padding zero case, outpout will have zero due to upsampling
// TODO: This is ugly and slow
ck::cpu::avx2_util::memset32_avx2(&c_grid_buf.p_data_[0], 0, N * Ho * Wo * K);
// printf("___ clear\n");
}
if(dynamic_tunable.loop_over_spec ==
ck::tensor_operation::cpu::device::ConvolutionForwardBlockLoopOverSpecialization_t::
LoopOver_MNK)
......@@ -495,10 +513,14 @@ struct GridwiseDirectConvNHWCAvx2
UseCLocalBuffer ? c_block_mem.mMemSize / sizeof(FloatC)
: c_grid_desc.GetElementSpaceSize());
const ck::index_t tid = omp_get_thread_num();
const ck::index_t tid_k = tid % num_threads_k;
const ck::index_t tid_wo = (tid / num_threads_k) % num_threads_wo;
const ck::index_t tid_nho = tid / (num_threads_k * num_threads_wo);
ck::index_t tid = omp_get_thread_num();
const ck::index_t tid_n = tid % num_threads_n;
tid /= num_threads_n;
const ck::index_t tid_ho = tid % num_threads_ho;
tid /= num_threads_ho;
const ck::index_t tid_wo = tid % num_threads_wo;
tid /= num_threads_wo;
const ck::index_t tid_k = tid;
ck::cpu::ThreadwiseGemmParam param;
// param.Kr = k_per_block;
......@@ -510,35 +532,43 @@ struct GridwiseDirectConvNHWCAvx2
// ihi = iho * s_stride_h + iy * s_dilation_h - s_pad_h
// iwi = iwo * s_stride_w + ix * s_dilation_w - s_pad_w
ck::index_t i_nho = tid_nho * num_works_nho_per_thread;
ck::index_t i_ho = i_nho % Ho;
ck::index_t i_n = i_nho / Ho;
auto accumulate_n_ho = [&]() {
i_ho++;
if(i_ho >= Wo)
{
i_ho = 0;
i_n++;
}
};
for(; (i_nho < (tid_nho + 1) * num_works_nho_per_thread) && (i_nho < num_works_nho);
i_nho += 1, accumulate_n_ho())
// ck::index_t i_nho = tid_nho * num_works_nho_per_thread;
// ck::index_t i_ho = i_nho % Ho;
// ck::index_t i_n = i_nho / Ho;
// auto accumulate_n_ho = [&]() {
// i_ho++;
// if(i_ho >= Wo)
// {
// i_ho = 0;
// i_n++;
// }
// };
for(ck::index_t i_n = tid_n * num_works_n_per_thread;
(i_n < (tid_n + 1) * num_works_n_per_thread) && i_n < num_works_n;
i_n += 1)
{
for(ck::index_t i_ho = tid_ho * num_works_ho_per_thread;
(i_ho < (tid_ho + 1) * num_works_ho_per_thread) && i_ho < num_works_ho;
i_ho += 1)
{
// for input
ck::index_t i_hi_no_y = i_ho * Sy - Py;
for(ck::index_t i_wo = tid_wo * num_works_wo_per_thread * m_per_thread;
i_wo < (tid_wo + 1) * num_works_wo_per_thread * m_per_thread;
i_wo < (tid_wo + 1) * num_works_wo_per_thread * m_per_thread &&
i_wo < Wo;
i_wo += m_per_thread)
{
ck::index_t current_wo_size_no_dx = ck::math::min(Wo - i_wo, m_per_thread);
ck::index_t current_wo_size_no_dx =
ck::math::min(Wo - i_wo, m_per_thread);
ck::index_t i_wi_no_x = i_wo * Sx - Px;
// printf("-- i_nho:%d, i_wo:%d, num_works_nho:%d,
// num_threads_nho:%d(Hi:%d,nWi:%d)\n",
// i_nho, i_wo, num_works_nho, num_threads_nho, Hi, Wi);fflush(stdout);
// i_nho, i_wo, num_works_nho, num_threads_nho, Hi,
// Wi);fflush(stdout);
for(ck::index_t i_k = tid_k * num_works_k_per_thread * n_per_thread;
i_k < (tid_k + 1) * num_works_k_per_thread * n_per_thread;
......@@ -569,8 +599,8 @@ struct GridwiseDirectConvNHWCAvx2
ck::index_t i_wi = i_wi_no_x + i_dx;
ck::index_t current_wo_size = current_wo_size_no_dx;
ck::index_t pad_wo_size = 0; // when left pad, we may never have a
// chance to clear zero (like
ck::index_t pad_wo_size = 0; // when left pad, we may never have
// a chance to clear zero (like
// padding) we need to manually clear that
if(i_wi < 0)
......@@ -588,8 +618,8 @@ struct GridwiseDirectConvNHWCAvx2
current_i_wo += steps_wo_turn_possitive;
if(!accmulate_c)
pad_wo_size =
steps_wo_turn_possitive; // if already accumulating, no
// need to manually set
steps_wo_turn_possitive; // if already accumulating,
// no need to manually set
i_wi += steps_wo_turn_possitive *
Sx; // now i_wi will be a possitive number
}
......@@ -602,8 +632,8 @@ struct GridwiseDirectConvNHWCAvx2
{
// printf(" ->[r] i_wi:%d, r:%d(%d), ", i_wi, i_wi +
// ((current_wo_size - 1) * Sx), current_wo_size);
current_wo_size =
(Wi - 1 - i_wi) / Sx + 1; // NOTE: this be careful why here
current_wo_size = (Wi - 1 - i_wi) / Sx +
1; // NOTE: this be careful why here
// should be compute like this.
if(current_wo_size <= 0)
continue;
......@@ -621,10 +651,12 @@ struct GridwiseDirectConvNHWCAvx2
i_wo_pad++)
{
const intptr_t offset_c = GetCBlockStartOffset(
c_grid_desc, i_nho * Wo + i_wo_pad, i_k);
c_grid_desc,
(i_n * Ho + i_ho) * Wo + i_wo_pad,
i_k);
// printf("pad_wo_size:%d, current_k_block_size:%d, clear
// offset_c:%d\n",
// printf("pad_wo_size:%d, current_k_block_size:%d,
// clear offset_c:%d\n",
// pad_wo_size, current_k_size,
// offset_c);fflush(stdout);
ck::cpu::avx2_util::memset32_avx2(
......@@ -636,16 +668,16 @@ struct GridwiseDirectConvNHWCAvx2
const intptr_t offset_b =
GetBBlockStartOffset(b_grid_desc, i_yxc, i_k);
const intptr_t offset_c = GetCBlockStartOffset(
c_grid_desc, i_nho * Wo + current_i_wo, i_k);
c_grid_desc, (i_n * Ho + i_ho) * Wo + current_i_wo, i_k);
// printf("offset_a:%lu, offset_b:%lu, offset_c:%lu, i_n:%d,
// i_hi:%d, i_wi:%d, i_dx:%d, i_dy:%d, i_k:%d, i_ho:%d, i_wo:%d,
// current_wo_size:%d, current_k_size:%d, i_nho:%d, lda:%d, ldb:%d,
// ldc:%d, acc:%d\n",
// offset_a, offset_b, offset_c, i_n, i_hi, i_wi, i_dx, i_dy,
// i_k, i_ho, current_i_wo, current_wo_size, current_k_size,
// i_nho, param.lda / sizeof(FloatA), param.ldb /
// sizeof(FloatB), param.ldc / sizeof(FloatC),
// current_wo_size:%d, current_k_size:%d, i_nho:%d, lda:%d,
// ldb:%d, ldc:%d, acc:%d\n",
// offset_a, offset_b, offset_c, i_n, i_hi, i_wi, i_dx,
// i_dy, i_k, i_ho, current_i_wo, current_wo_size,
// current_k_size, i_nho, param.lda / sizeof(FloatA),
// param.ldb / sizeof(FloatB), param.ldc / sizeof(FloatC),
// param.accmulate_c); fflush(stdout);
param.p_a = &a_block_buf.p_data_[offset_a];
......@@ -660,6 +692,7 @@ struct GridwiseDirectConvNHWCAvx2
}
}
}
}
else if(dynamic_tunable.loop_over_spec ==
ck::tensor_operation::cpu::device::ConvolutionForwardBlockLoopOverSpecialization_t::
LoopOver_MKN)
......@@ -695,11 +728,6 @@ struct GridwiseDirectConvNHWCAvx2
UseCLocalBuffer ? c_block_mem.mMemSize / sizeof(FloatC)
: c_grid_desc.GetElementSpaceSize());
const ck::index_t tid = omp_get_thread_num();
const ck::index_t tid_k = tid % num_threads_k;
const ck::index_t tid_wo = (tid / num_threads_k) % num_threads_wo;
const ck::index_t tid_nho = tid / (num_threads_k * num_threads_wo);
ck::cpu::ThreadwiseGemmParam param;
// param.Kr = k_per_block;
param.lda = Sx * C * sizeof(FloatA);
......@@ -710,39 +738,42 @@ struct GridwiseDirectConvNHWCAvx2
// ihi = iho * s_stride_h + iy * s_dilation_h - s_pad_h
// iwi = iwo * s_stride_w + ix * s_dilation_w - s_pad_w
ck::index_t i_nho = tid_nho * num_works_nho_per_thread;
ck::index_t i_ho = i_nho % Ho;
ck::index_t i_n = i_nho / Ho;
auto accumulate_n_ho = [&]() {
i_ho++;
if(i_ho >= Wo)
{
i_ho = 0;
i_n++;
}
};
for(; (i_nho < (tid_nho + 1) * num_works_nho_per_thread) && (i_nho < num_works_nho);
i_nho += 1, accumulate_n_ho())
ck::index_t tid = omp_get_thread_num();
const ck::index_t tid_n = tid % num_threads_n;
tid /= num_threads_n;
const ck::index_t tid_ho = tid % num_threads_ho;
tid /= num_threads_ho;
const ck::index_t tid_wo = tid % num_threads_wo;
tid /= num_threads_wo;
const ck::index_t tid_k = tid;
for(ck::index_t i_n = tid_n * num_works_n_per_thread;
(i_n < (tid_n + 1) * num_works_n_per_thread) && i_n < num_works_n;
i_n += 1)
{
for(ck::index_t i_ho = tid_ho * num_works_ho_per_thread;
(i_ho < (tid_ho + 1) * num_works_ho_per_thread) && i_ho < num_works_ho;
i_ho += 1)
{
// for input
ck::index_t i_hi_no_y = i_ho * Sy - Py;
for(ck::index_t i_wo = tid_wo * num_works_wo_per_thread * m_per_thread;
i_wo < (tid_wo + 1) * num_works_wo_per_thread * m_per_thread;
i_wo < (tid_wo + 1) * num_works_wo_per_thread * m_per_thread &&
i_wo < Wo;
i_wo += m_per_thread)
{
ck::index_t current_wo_size_no_dx = ck::math::min(Wo - i_wo, m_per_thread);
ck::index_t current_wo_size_no_dx =
ck::math::min(Wo - i_wo, m_per_thread);
ck::index_t i_wi_no_x = i_wo * Sx - Px;
ck::index_t i_dx = 0;
ck::index_t i_dy = 0;
bool accmulate_c = false;
// printf("-- [%d] i_nho:%d, i_wo:%d, num_works_nho:%d,
// num_threads_nho:%d(Hi:%d, Wi:%d), current_wo_size_no_dx:%d,
// printf("-- [%d] i_n:%d, i_ho:%d, i_wo:%d, num_works_n:%d,
// num_threads_n:%d(Hi:%d, Wi:%d), current_wo_size_no_dx:%d,
// m_per_thread:%d\n",
// tid, i_nho, i_wo, num_works_nho, num_threads_nho, Hi, Wi,
// tid, i_n, i_ho, i_wo, num_works_n, num_threads_n, Hi, Wi,
// current_wo_size_no_dx, m_per_thread);fflush(stdout);
auto accumulate_dy_dx = [&]() {
......@@ -759,26 +790,25 @@ struct GridwiseDirectConvNHWCAvx2
{
ck::index_t current_i_wo = i_wo;
ck::index_t i_hi = i_hi_no_y + i_dy;
bool run_pad_only = false;
if(i_hi < 0 || i_hi >= Hi)
continue;
ck::index_t i_wi = i_wi_no_x + i_dx;
ck::index_t current_wo_size = current_wo_size_no_dx;
ck::index_t pad_wo_size =
0; // when left pad, we may never have a chance to clear zero (like
ck::index_t pad_wo_size = 0; // when left pad, we may never have a
// chance to clear zero (like
// padding) we need to manually clear that
/* left corner shift
* when i_wi is negative, need shift i_wo to right to make i_wi
* possitive sx px i_wi steps_wo_turn_possitive 1 0 0,
* 1, 2.... 0 2 0 0, 2, 4... 0 1 1 -1, 0, 1.... 1
* 2 1 -1, 1, 3.... 1
* 2 2 -2, 0, 2... 1
* 2 3 -3, -1, 1... 2
* 3 1 -1, 2, 5... 1
* 3 2 -2, 1, 4.... 1
* 3 3 -3, 0, 3 1
* 3 4 -4, -1, 2... 2
* possitive sx px i_wi steps_wo_turn_possitive
* 1 0
* 0, 1, 2.... 0 2 0 0, 2, 4... 0 1 1 -1,
* 0, 1.... 1 2 1 -1, 1, 3.... 1 2 2 -2, 0, 2... 1 2
* 3 -3, -1, 1... 2 3 1 -1, 2, 5... 1 3 2 -2,
* 1, 4.... 1 3 3 -3, 0, 3 1 3 4 -4,
* -1, 2... 2
*/
if(i_wi < 0)
{
......@@ -786,23 +816,27 @@ struct GridwiseDirectConvNHWCAvx2
-i_wi; // keep this a possitive number
ck::index_t steps_wo_turn_possitive =
(wi_to_zero_length + Sx - 1) /
Sx; // how many steps need to move wo, to let wi to be possitive
Sx; // how many steps need to move wo, to let wi to be
// possitive
current_wo_size -= steps_wo_turn_possitive;
// printf("--- current_wo_size:%d, i_wi:%d\n", current_wo_size,
// i_wi);
if(current_wo_size <= 0)
continue;
current_i_wo += steps_wo_turn_possitive;
if(!accmulate_c)
pad_wo_size =
steps_wo_turn_possitive; // if already accumulating, no need
// to manually set
steps_wo_turn_possitive; // if already accumulating, no
// need to manually set
i_wi += steps_wo_turn_possitive *
Sx; // now i_wi will be a possitive number
}
if(i_wi >= Wi)
{
continue;
}
// shrink right wi/wo
if((i_wi + ((current_wo_size - 1) * Sx)) >= Wi)
{
......@@ -823,44 +857,49 @@ struct GridwiseDirectConvNHWCAvx2
if(pad_wo_size != 0)
{
// manually clear zero. this may and only may need once along the
// gemm_k reduction
// manually clear zero. this may and only may need once along
// the gemm_k reduction
ck::index_t i_k = tid_k * num_works_k_per_thread * n_per_thread;
ck::index_t current_k_block_size =
ck::math::min(K - i_k, num_works_k_per_thread * n_per_thread);
ck::index_t current_k_block_size = ck::math::min(
K - i_k, num_works_k_per_thread * n_per_thread);
const intptr_t offset_c =
GetCBlockStartOffset(c_grid_desc, i_nho * Wo, i_k);
const intptr_t offset_c = GetCBlockStartOffset(
c_grid_desc, (i_n * Ho + i_ho) * Wo, i_k);
// printf("[%d] pad_wo_size:%d, current_k_block_size:%d,
// offset_c:%d\n",
// tid, pad_wo_size, current_k_block_size,
// offset_c);fflush(stdout);
ck::cpu::avx2_util::memset32_avx2(&c_block_buf.p_data_[offset_c],
// offset_c:%d, i_wo:%d\n",
// tid, pad_wo_size, current_k_block_size, offset_c,
// i_wo);fflush(stdout);
ck::cpu::avx2_util::memset32_avx2(
&c_block_buf.p_data_[offset_c],
0,
current_k_block_size *
pad_wo_size);
current_k_block_size * pad_wo_size);
}
if(run_pad_only)
continue;
for(ck::index_t i_k = tid_k * num_works_k_per_thread * n_per_thread;
i_k < (tid_k + 1) * num_works_k_per_thread * n_per_thread;
i_k += n_per_thread)
{
ck::index_t current_k_size = ck::math::min(K - i_k, n_per_thread);
ck::index_t current_k_size =
ck::math::min(K - i_k, n_per_thread);
const intptr_t offset_a = current_input_offset;
const intptr_t offset_b =
GetBBlockStartOffset(b_grid_desc, i_yxc, i_k);
const intptr_t offset_c = GetCBlockStartOffset(
c_grid_desc, i_nho * Wo + current_i_wo, i_k);
c_grid_desc, (i_n * Ho + i_ho) * Wo + current_i_wo, i_k);
// printf("[%d] offset_a:%lu, offset_b:%lu, offset_c:%lu, i_n:%d,
// i_hi:%d, i_wi:%d, i_dx:%d, i_dy:%d, i_k:%d, i_ho:%d, i_wo:%d,
// current_wo_size:%d, i_nho:%d, lda:%d, ldb:%d\n",
// printf("[%d] offset_a:%lu, offset_b:%lu, offset_c:%lu,
// i_n:%d, i_hi:%d, i_wi:%d, i_dx:%d, i_dy:%d, i_k:%d, i_ho:%d,
// i_wo:%d, current_wo_size:%d, i_n:%d, i_ho:%d, lda:%d,
// ldb:%d\n",
// tid, offset_a, offset_b, offset_c, i_n, i_hi, i_wi, i_dx,
// i_dy, i_k, i_ho, current_i_wo, current_wo_size, i_nho,
// param.lda / sizeof(FloatA), param.ldb / sizeof(FloatB));
// fflush(stdout);
// i_dy, i_k, i_ho, current_i_wo, current_wo_size, i_n,
// i_ho, param.lda / sizeof(FloatA), param.ldb /
// sizeof(FloatB)); fflush(stdout);
param.p_a = &a_block_buf.p_data_[offset_a];
param.p_b = &b_block_buf.p_data_[offset_b];
......@@ -875,6 +914,7 @@ struct GridwiseDirectConvNHWCAvx2
}
}
}
}
};
} // namespace cpu
......
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