Skip to content

Topk fixes #239

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 6 commits into from
Oct 3, 2018
Merged

Topk fixes #239

Show file tree
Hide file tree
Changes from 4 commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
95afb10
Spec (https://docs.nvidia.com/cuda/pdf/ptx_isa_6.3.pdf) Sec 9.7.1.19 …
jithunnair-amd Oct 3, 2018
03e2734
Spec (https://docs.nvidia.com/cuda/pdf/ptx_isa_6.3.pdf) Sec 9.7.1.19 …
jithunnair-amd Oct 3, 2018
b4ae465
Correct logic for getLaneMaskLe. Previous logic would return 0x0 inst…
jithunnair-amd Oct 3, 2018
42a5ff1
Round up blockDim.x to prevent negative index for smem
jithunnair-amd Oct 3, 2018
1e39dac
Revert unintended change; back to __forceinline__
jithunnair-amd Oct 3, 2018
9103ba6
Revert change to TOPK_WARP_SIZE that crept in with previous commit
jithunnair-amd Oct 3, 2018
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
26 changes: 12 additions & 14 deletions aten/src/THC/THCAsmUtils.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,12 +11,11 @@ struct Bitfield<unsigned int> {
static __device__ __forceinline__
unsigned int getBitfield(unsigned int val, int pos, int len) {
#if defined(__HIP_PLATFORM_HCC__)
pos &= 0x1f;
len &= 0x1f;
pos &= 0xff;
len &= 0xff;

unsigned int m = (1u << len) - 1u;
m <<= pos;
return val & m;
return (val >> pos) & m;
#else
unsigned int ret;
asm("bfe.u32 %0, %1, %2, %3;" : "=r"(ret) : "r"(val), "r"(pos), "r"(len));
Expand All @@ -27,8 +26,8 @@ struct Bitfield<unsigned int> {
static __device__ __forceinline__
unsigned int setBitfield(unsigned int val, unsigned int toInsert, int pos, int len) {
#if defined(__HIP_PLATFORM_HCC__)
pos &= 0x1f;
len &= 0x1f;
pos &= 0xff;
len &= 0xff;

unsigned int m = (1u << len) - 1u;
toInsert &= m;
Expand All @@ -50,12 +49,11 @@ struct Bitfield<uint64_t> {
static __device__ __forceinline__
uint64_t getBitfield(uint64_t val, int pos, int len) {
#if defined(__HIP_PLATFORM_HCC__)
pos &= 0x1f;
len &= 0x1f;
pos &= 0xff;
len &= 0xff;

uint64_t m = (1u << len) - 1u;
m <<= pos;
return val & m;
return (val >> pos) & m;
#else
uint64_t ret;
asm("bfe.u64 %0, %1, %2, %3;" : "=l"(ret) : "l"(val), "r"(pos), "r"(len));
Expand All @@ -66,8 +64,8 @@ struct Bitfield<uint64_t> {
static __device__ __forceinline__
uint64_t setBitfield(uint64_t val, uint64_t toInsert, int pos, int len) {
#if defined(__HIP_PLATFORM_HCC__)
pos &= 0x1f;
len &= 0x1f;
pos &= 0xff;
len &= 0xff;

uint64_t m = (1u << len) - 1u;
toInsert &= m;
Expand Down Expand Up @@ -106,8 +104,8 @@ __device__ __forceinline__ unsigned getLaneMaskLt() {
}

#if defined (__HIP_PLATFORM_HCC__)
__device__ __forceinline__ unsigned long long int getLaneMaskLe() {
std::uint64_t m = (1ull << (getLaneId() + 1ull)) - 1ull;
__device__ inline unsigned long long int getLaneMaskLe() {
Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Given how __forceinline__ is now correctly defined, I'd like to keep it that way

std::uint64_t m = UINT64_MAX >> (sizeof(std::uint64_t) * CHAR_BIT - (getLaneId() + 1));
return m;
}
#else
Expand Down
2 changes: 1 addition & 1 deletion aten/src/THC/THCScanUtils.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -213,7 +213,7 @@ __device__ void exclusiveBinaryPrefixScan(T* smem, bool in, T* out, T* carry, Bi
*out -= (T) in;

// The outgoing carry for all threads is the last warp's sum
*carry = smem[(blockDim.x / SCAN_UTILS_WARP_SIZE) - 1];
*carry = smem[THCCeilDiv<int>(blockDim.x, SCAN_UTILS_WARP_SIZE) - 1];

if (KillWARDependency) {
__syncthreads();
Expand Down