Skip to content

Commit

Permalink
32b dim_t fixes (pytorch#4227)
Browse files Browse the repository at this point in the history 
Summary:
Also add unit tests to the 32b CI since it seems a lot of 32b issues are not revealed only via build checking.
Pull Request resolved: pytorch#4227

Test Plan: With 32b and 64b, also OpenCL enabled.

Differential Revision: D20331898

Pulled By: jfix71

fbshipit-source-id: f7fd0cbcf31f1425a68d4b5c95e820f36e84ffad
  • Loading branch information
@pjaaskel @facebook-github-bot
pjaaskel authored and facebook-github-bot committed Mar 9, 2020
1 parent 0f8107a commit 20cd623
Show file tree
Hide file tree
Showing 7 changed files with 280 additions and 308 deletions.
4 changes: 3 additions & 1 deletion .circleci/test.sh
View file
Original file line number Diff line number Diff line change
Expand Up @@ -81,7 +81,9 @@ case ${CIRCLE_JOB} in
# No tests with shared libs; it's similar to DEBUG.
;;
32B_DIM_T)
# No tests with 32b dim_t; it's similar to DEBUG.
# A lot of 32b dim_t issues are not revealed at build time, thus
# run the unit test suite also.
run_unit_tests check
;;
RELEASE_WITH_EXPENSIVE_TESTS)
run_unit_tests check_expensive
Expand Down
38 changes: 19 additions & 19 deletions lib/Backends/CPU/libjit/libjit.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1127,10 +1127,10 @@ static bool checkIOU(const Box &sb, const Box &cb, float iouThreshold,
template <typename T>
static void
libjit_nms_generic(T *indices, T *numDetected, const float *boxTensor,
const dim_t *boxTensorDims, size_t boxTensorDimSize,
const dim_t *boxTensorDims, dim_t boxTensorDimSize,
const float *scoresTensor, const dim_t *scoresTensorDims,
size_t scoresTensorDimSize, const dim_t *resultTensorDims,
size_t resultTensorDimSize, unsigned centerPointBox,
dim_t scoresTensorDimSize, const dim_t *resultTensorDims,
dim_t resultTensorDimSize, unsigned centerPointBox,
unsigned maxOutputBoxesPerClass, float iouThreshold,
float scoreThreshold, bool isV4) {
int boxesBoxDim = boxTensorDimSize - 2;
Expand Down Expand Up @@ -1439,7 +1439,7 @@ int8_t libjit_element_cmp_eq_kernel_u(dim_t idx, const size_t *LHS,
return LHS[idx] == RHS[idx] ? 1 : 0;
}

int8_t libjit_element_cmp_eq_kernel_i32(size_t idx, const int32_t *LHS,
int8_t libjit_element_cmp_eq_kernel_i32(dim_t idx, const int32_t *LHS,
const int32_t *RHS) {
return LHS[idx] == RHS[idx] ? 1 : 0;
}
Expand Down Expand Up @@ -1754,7 +1754,7 @@ void libjit_gatherranges32_u(uint64_t *output, int32_t *lengths,

void libjit_gatherranges32_i32(int32_t *output, int32_t *lengths,
const int32_t *data, const int32_t *ranges,
size_t numExamples, size_t exampleSize) {
dim_t numExamples, dim_t exampleSize) {
libjit_gatherranges(output, lengths, data, ranges, numExamples, exampleSize);
}

Expand Down Expand Up @@ -1782,11 +1782,11 @@ void libjit_scatterdata_f_i32(float *data, const dim_t *dataDims,
}

void libjit_scatterdata_i8_u(int8_t *data, const dim_t *dataDims,
const size_t *indices, const int8_t *slices,
size_t numIndices, size_t indexSize,
size_t sliceSize, bool isCumulative,
float dataScale, int32_t dataOffset,
float sliceScale, int32_t sliceOffset) {
const int64_t *indices, const int8_t *slices,
dim_t numIndices, dim_t indexSize, dim_t sliceSize,
bool isCumulative, float dataScale,
int32_t dataOffset, float sliceScale,
int32_t sliceOffset) {
if (isCumulative) {
libjit_scatterdataaddquantized(data, dataDims, indices, slices, numIndices,
indexSize, sliceSize, dataScale, dataOffset,
Expand All @@ -1799,8 +1799,8 @@ void libjit_scatterdata_i8_u(int8_t *data, const dim_t *dataDims,

void libjit_scatterdata_i8_i32(int8_t *data, const dim_t *dataDims,
const int32_t *indices, const int8_t *slices,
size_t numIndices, size_t indexSize,
size_t sliceSize, bool isCumulative,
dim_t numIndices, dim_t indexSize,
dim_t sliceSize, bool isCumulative,
float dataScale, int32_t dataOffset,
float sliceScale, int32_t sliceOffset) {
if (isCumulative) {
Expand Down Expand Up @@ -2729,11 +2729,11 @@ libjit_quantization_profile(float *inputTensor, dim_t tensorSize,
}

__attribute__((noinline)) void
libjit_nms_u(size_t *indices, size_t *numDetected, const float *boxTensor,
const dim_t *boxTensorDims, size_t boxTensorDimSize,
libjit_nms_u(uint64_t *indices, uint64_t *numDetected, const float *boxTensor,
const dim_t *boxTensorDims, dim_t boxTensorDimSize,
const float *scoresTensor, const dim_t *scoresTensorDims,
size_t scoresTensorDimSize, const dim_t *resultTensorDims,
size_t resultTensorDimSize, unsigned centerPointBox,
dim_t scoresTensorDimSize, const dim_t *resultTensorDims,
dim_t resultTensorDimSize, unsigned centerPointBox,
unsigned maxOutputBoxesPerClass, float iouThreshold,
float scoreThreshold, bool isV4) {
libjit_nms_generic(indices, numDetected, boxTensor, boxTensorDims,
Expand All @@ -2745,10 +2745,10 @@ libjit_nms_u(size_t *indices, size_t *numDetected, const float *boxTensor,

__attribute__((noinline)) void
libjit_nms_i32(int32_t *indices, int32_t *numDetected, const float *boxTensor,
const dim_t *boxTensorDims, size_t boxTensorDimSize,
const dim_t *boxTensorDims, dim_t boxTensorDimSize,
const float *scoresTensor, const dim_t *scoresTensorDims,
size_t scoresTensorDimSize, const dim_t *resultTensorDims,
size_t resultTensorDimSize, unsigned centerPointBox,
dim_t scoresTensorDimSize, const dim_t *resultTensorDims,
dim_t resultTensorDimSize, unsigned centerPointBox,
unsigned maxOutputBoxesPerClass, float iouThreshold,
float scoreThreshold, bool isV4) {
libjit_nms_generic(indices, numDetected, boxTensor, boxTensorDims,
Expand Down
Loading

0 comments on commit 20cd623

Please sign in to comment.