devtools/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2026-04-10 11:34:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

[SYCL] This PR adds the minimum modifications to the Eigen unsupported module required to run it on devices supporting SYCL.

Browse Source 
* Abstracting the pointer type so that both SYCL memory and pointer can be captured.
* Converting SYCL virtual pointer to SYCL device memory in Eigen evaluator class.
* Binding SYCL placeholder accessor to command group handler by using bind method in Eigen evaluator node.
* Adding SYCL macro for controlling loop unrolling.
* Modifying the TensorDeviceSycl.h and SYCL executor method to adopt the above changes.
This commit is contained in:
Mehdi Goli
2019-06-28 10:08:23 +01:00
parent 16a56b2ddd
commit 7d08fa805a
47 changed files with 1983 additions and 951 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h
View File

@@ -37,7 +37,7 @@ namespace {
{
#ifdef EIGEN_GPU_COMPILE_PHASE
return __clz(val);
#elif defined(__SYCL_DEVICE_ONLY__)
#elif defined(SYCL_DEVICE_ONLY)
return cl::sycl::clz(val);
#elif EIGEN_COMP_MSVC
unsigned long index;
@@ -55,8 +55,8 @@ namespace {
{
#ifdef EIGEN_GPU_COMPILE_PHASE
return __clzll(val);
#elif defined(__SYCL_DEVICE_ONLY__)
return cl::sycl::clz(val);
#elif defined(SYCL_DEVICE_ONLY)
return static_cast<int>(cl::sycl::clz(val));
#elif EIGEN_COMP_MSVC && EIGEN_ARCH_x86_64
unsigned long index;
_BitScanReverse64(&index, val);
@@ -92,7 +92,7 @@ namespace {
EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE uint32_t muluh(const uint32_t a, const T b) {
#if defined(EIGEN_GPU_COMPILE_PHASE)
return __umulhi(a, b);
#elif defined(__SYCL_DEVICE_ONLY__)
#elif defined(SYCL_DEVICE_ONLY)
return cl::sycl::mul_hi(a, static_cast<uint32_t>(b));
#else
return (static_cast<uint64_t>(a) * b) >> 32;
@@ -103,7 +103,7 @@ namespace {
EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE uint64_t muluh(const uint64_t a, const T b) {
#if defined(EIGEN_GPU_COMPILE_PHASE)
return __umul64hi(a, b);
#elif defined(__SYCL_DEVICE_ONLY__)
#elif defined(SYCL_DEVICE_ONLY)
return cl::sycl::mul_hi(a, static_cast<uint64_t>(b));
#elif defined(__SIZEOF_INT128__)
__uint128_t v = static_cast<__uint128_t>(a) * static_cast<__uint128_t>(b);
@@ -124,7 +124,7 @@ namespace {
template <typename T>
struct DividerHelper<64, T> {
static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE uint64_t computeMultiplier(const int log_div, const T divider) {
#if defined(__SIZEOF_INT128__) && !defined(EIGEN_GPU_COMPILE_PHASE) && !defined(__SYCL_DEVICE_ONLY__)
#if defined(__SIZEOF_INT128__) && !defined(EIGEN_GPU_COMPILE_PHASE) && !defined(SYCL_DEVICE_ONLY)
return static_cast<uint64_t>((static_cast<__uint128_t>(1) << (64+log_div)) / static_cast<__uint128_t>(divider) - (static_cast<__uint128_t>(1) << 64) + 1);
#else
const uint64_t shift = 1ULL << log_div;
@@ -205,8 +205,8 @@ class TensorIntDivisor<int32_t, true> {
EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE int divide(const int32_t n) const {
#ifdef EIGEN_GPU_COMPILE_PHASE
return (__umulhi(magic, n) >> shift);
#elif defined(__SYCL_DEVICE_ONLY__)
return (cl::sycl::mul_hi(static_cast<uint64_t>(magic), static_cast<uint64_t>(n)) >> shift);
#elif defined(SYCL_DEVICE_ONLY)
return (cl::sycl::mul_hi(magic, static_cast<uint32_t>(n)) >> shift);
#else
uint64_t v = static_cast<uint64_t>(magic) * static_cast<uint64_t>(n);
return (static_cast<uint32_t>(v >> 32) >> shift);