leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 25 Packages Projects Releases Wiki Activity

support setting load granularity for conv2d fprop

Browse Source
This commit is contained in:
mengchi.hmc
2021-04-22 15:20:57 +08:00
parent 7ec3a87f22
commit bb35a3ba6f
3 changed files with 86 additions and 265 deletions
Download Patch File Download Diff File Expand all files Collapse all files
include/cutlass/conv/threadblock/conv2d_fprop_activation_tile_access_iterator_optimized.h
+4 -5
View File
@@ -61,7 +61,7 @@ template <
typename Element_,
typename Layout_,
typename ThreadMap_,
bool Aligned = true
int AccessSize = ThreadMap_::kElementsPerAccess
>
class Conv2dFpropActivationTileAccessIteratorOptimized {
public:
@@ -75,7 +75,6 @@ public:
using Layout = Layout_;
using TensorCoord = typename Layout::TensorCoord;
using ThreadMap = ThreadMap_;
static int const AccessSize = Aligned ? ThreadMap::kElementsPerAccess : 1;
using AccessType = AlignedArray<Element, AccessSize>;
using TensorRef = cutlass::TensorRef<Element, Layout>;
using Index = typename Layout::Index;
@@ -216,7 +215,7 @@ public:
CUTLASS_PRAGMA_UNROLL
for (int v_idx = 0; v_idx < kAccessesPerVector; ++v_idx) {
clear_mask_(filter_c_ + v_idx >= problem_size_.C, v_idx);
clear_mask_(filter_c_ + v_idx * AccessSize >= problem_size_.C, v_idx);
}
set_iteration_index(0);
@@ -357,7 +356,7 @@ public:
CUTLASS_PRAGMA_UNROLL
for (int v_idx = 0; v_idx < kAccessesPerVector; ++v_idx) {
clear_mask_(filter_c_ + v_idx >= problem_size_.C, v_idx);
clear_mask_(filter_c_ + v_idx * AccessSize >= problem_size_.C, v_idx);
}
}
@@ -420,7 +419,7 @@ public:
static Status can_implement(Conv2dProblemSize const &problem_size) {
// check alignment constraint on iterator's contiguous dimension
if (Aligned && problem_size.C % (128/sizeof_bits<Element>::value)) {
if (problem_size.C % AccessSize) {
return Status::kErrorInvalidProblem;
}
include/cutlass/conv/threadblock/conv2d_fprop_filter_tile_access_iterator_optimized.h
+4 -5
View File
@@ -61,7 +61,7 @@ template <
typename Element_,
typename Layout_,
typename ThreadMap_,
bool Aligned = true
int AccessSize = ThreadMap_::kElementsPerAccess
>
class Conv2dFpropFilterTileAccessIteratorOptimized{
public:
@@ -74,7 +74,6 @@ public:
using Element = Element_;
using Layout = Layout_;
using ThreadMap = ThreadMap_;
static int const AccessSize = Aligned ? ThreadMap::kElementsPerAccess : 1;
using AccessType = AlignedArray<Element, AccessSize>;
using TensorRef = cutlass::TensorRef<Element, Layout>;
using TensorCoord = typename Layout::TensorCoord;
@@ -178,7 +177,7 @@ public:
}
for (int v_idx = 0; v_idx < kAccessesPerVector; ++v_idx) {
clear_mask_(filter_c_ + v_idx >= problem_size_.C, v_idx);
clear_mask_(filter_c_ + v_idx * AccessSize >= problem_size_.C, v_idx);
}
pointer_ += (
@@ -249,7 +248,7 @@ public:
}
for (int v_idx = 0; v_idx < kAccessesPerVector; ++v_idx) {
clear_mask_(filter_c_ + v_idx >= problem_size_.C, v_idx);
clear_mask_(filter_c_ + v_idx * AccessSize >= problem_size_.C, v_idx);
}
pointer_ += next;
@@ -301,7 +300,7 @@ public:
static Status can_implement(Conv2dProblemSize const &problem_size) {
// check alignment constraint on iterator's contiguous dimension
if (Aligned && problem_size.C % (128/sizeof_bits<Element>::value)) {
if (problem_size.C % AccessSize) {
return Status::kErrorInvalidProblem;
}