AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge branch 'master' into getsymboladdress

Browse Source 
[ROCm/hip commit: 8610128c3e]
This commit is contained in:
Michael Kuron
2018-11-20 12:03:22 +01:00
committed by GitHub
parent 2e8a94e90b 20fff19daa
commit c35dfb71d5
31 changed files with 1349 additions and 580 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/hip/include/hip/hcc_detail/functional_grid_launch.hpp
+32 -7
View File
@@ -33,6 +33,7 @@ THE SOFTWARE.
#include <cstddef>
#include <cstdint>
#include <cstring>
#include <functional>
#include <iostream>
#include <mutex>
@@ -56,7 +57,9 @@ template <
typename... Ts,
typename std::enable_if<n == sizeof...(Ts)>::type* = nullptr>
inline std::vector<std::uint8_t> make_kernarg(
std::vector<std::uint8_t> kernarg, const std::tuple<Ts...>&) {
const std::tuple<Ts...>&,
const std::vector<std::pair<std::size_t, std::size_t>>&,
std::vector<std::uint8_t> kernarg) {
return kernarg;
}
@@ -65,7 +68,9 @@ template <
typename... Ts,
typename std::enable_if<n != sizeof...(Ts)>::type* = nullptr>
inline std::vector<std::uint8_t> make_kernarg(
std::vector<std::uint8_t> kernarg, const std::tuple<Ts...>& formals) {
const std::tuple<Ts...>& formals,
const std::vector<std::pair<std::size_t, std::size_t>>& size_align,
std::vector<std::uint8_t> kernarg) {
using T = typename std::tuple_element<n, std::tuple<Ts...>>::type;
static_assert(
@@ -80,24 +85,44 @@ inline std::vector<std::uint8_t> make_kernarg(
#endif
kernarg.resize(round_up_to_next_multiple_nonnegative(
kernarg.size(), alignof(T)) + sizeof(T));
kernarg.size(), size_align[n].second) +
size_align[n].first);
new (kernarg.data() + kernarg.size() - sizeof(T)) T{std::get<n>(formals)};
std::memcpy(
kernarg.data() + kernarg.size() - size_align[n].first,
&std::get<n>(formals),
size_align[n].first);
return make_kernarg<n + 1>(std::move(kernarg), formals);
return make_kernarg<n + 1>(formals, size_align, std::move(kernarg));
}
template <typename... Formals, typename... Actuals>
inline std::vector<std::uint8_t> make_kernarg(
void (*)(Formals...), std::tuple<Actuals...> actuals) {
void (*kernel)(Formals...), std::tuple<Actuals...> actuals) {
static_assert(sizeof...(Formals) == sizeof...(Actuals),
"The count of formal arguments must match the count of actuals.");
if (sizeof...(Formals) == 0) return {};
const auto it = function_names().find(
reinterpret_cast<std::uintptr_t>(kernel));
if (it == function_names().cend()) {
throw std::runtime_error{"Undefined __global__ function."};
}
const auto it1 = kernargs().find(it->second);
if (it1 == kernargs().end()) {
throw std::runtime_error{
"Missing metadata for __global__ function: " + it->second};
}
std::tuple<Formals...> to_formals{std::move(actuals)};
std::vector<std::uint8_t> kernarg;
kernarg.reserve(sizeof(to_formals));
return make_kernarg<0>(std::move(kernarg), to_formals);
return make_kernarg<0>(to_formals, it1->second, std::move(kernarg));
}
void hipLaunchKernelGGLImpl(std::uintptr_t function_address, const dim3& numBlocks,
projects/hip/include/hip/hcc_detail/hip_memory.h
+6
View File
@@ -41,8 +41,14 @@ THE SOFTWARE.
#define __HIP_SIZE_OF_HEAP (__HIP_NUM_PAGES * __HIP_SIZE_OF_PAGE)
#if __HIP__ && __HIP_DEVICE_COMPILE__
__attribute__((weak)) __device__ char __hip_device_heap[__HIP_SIZE_OF_HEAP];
__attribute__((weak)) __device__
uint32_t __hip_device_page_flag[__HIP_NUM_PAGES];
#else
extern __device__ char __hip_device_heap[];
extern __device__ uint32_t __hip_device_page_flag[];
#endif
extern "C" inline __device__ void* __hip_malloc(size_t size) {
char* heap = (char*)__hip_device_heap;
projects/hip/include/hip/hcc_detail/math_functions.h
+38 -32
View File
@@ -514,38 +514,41 @@ float __exp10f(float x) { return __ocml_exp10_f32(x); }
__DEVICE__
inline
float __expf(float x) { return __ocml_exp_f32(x); }
#if defined OCML_BASIC_ROUNDED_OPERATIONS
__DEVICE__
inline
float __fadd_rd(float x, float y) { return __ocml_add_rtp_f32(x, y); }
float __fadd_rd(float x, float y) { return __ocml_add_rtn_f32(x, y); }
__DEVICE__
inline
float __fadd_rn(float x, float y) { return __ocml_add_rte_f32(x, y); }
__DEVICE__
inline
float __fadd_ru(float x, float y) { return __ocml_add_rtn_f32(x, y); }
float __fadd_ru(float x, float y) { return __ocml_add_rtp_f32(x, y); }
__DEVICE__
inline
float __fadd_rz(float x, float y) { return __ocml_add_rtz_f32(x, y); }
__DEVICE__
inline
float __fdiv_rd(float x, float y) { return x / y; }
float __fdiv_rd(float x, float y) { return __ocml_div_rtn_f32(x, y); }
__DEVICE__
inline
float __fdiv_rn(float x, float y) { return x / y; }
float __fdiv_rn(float x, float y) { return __ocml_div_rte_f32(x, y); }
__DEVICE__
inline
float __fdiv_ru(float x, float y) { return x / y; }
float __fdiv_ru(float x, float y) { return __ocml_div_rtp_f32(x, y); }
__DEVICE__
inline
float __fdiv_rz(float x, float y) { return x / y; }
float __fdiv_rz(float x, float y) { return __ocml_div_rtz_f32(x, y); }
#endif
__DEVICE__
inline
float __fdividef(float x, float y) { return x / y; }
#if defined OCML_BASIC_ROUNDED_OPERATIONS
__DEVICE__
inline
float __fmaf_rd(float x, float y, float z)
{
return __ocml_fma_rtp_f32(x, y, z);
return __ocml_fma_rtn_f32(x, y, z);
}
__DEVICE__
inline
@@ -557,7 +560,7 @@ __DEVICE__
inline
float __fmaf_ru(float x, float y, float z)
{
return __ocml_fma_rtn_f32(x, y, z);
return __ocml_fma_rtp_f32(x, y, z);
}
__DEVICE__
inline
@@ -567,13 +570,13 @@ float __fmaf_rz(float x, float y, float z)
}
__DEVICE__
inline
float __fmul_rd(float x, float y) { return __ocml_mul_rtp_f32(x, y); }
float __fmul_rd(float x, float y) { return __ocml_mul_rtn_f32(x, y); }
__DEVICE__
inline
float __fmul_rn(float x, float y) { return __ocml_mul_rte_f32(x, y); }
__DEVICE__
inline
float __fmul_ru(float x, float y) { return __ocml_mul_rtn_f32(x, y); }
float __fmul_ru(float x, float y) { return __ocml_mul_rtp_f32(x, y); }
__DEVICE__
inline
float __fmul_rz(float x, float y) { return __ocml_mul_rtz_f32(x, y); }
@@ -594,28 +597,29 @@ inline
float __frsqrt_rn(float x) { return __llvm_amdgcn_rsq_f32(x); }
__DEVICE__
inline
float __fsqrt_rd(float x) { return __ocml_sqrt_f32(x); }
float __fsqrt_rd(float x) { return __ocml_sqrt_rtn_f32(x); }
__DEVICE__
inline
float __fsqrt_rn(float x) { return __ocml_sqrt_f32(x); }
float __fsqrt_rn(float x) { return __ocml_sqrt_rte_f32(x); }
__DEVICE__
inline
float __fsqrt_ru(float x) { return __ocml_sqrt_f32(x); }
float __fsqrt_ru(float x) { return __ocml_sqrt_rtp_f32(x); }
__DEVICE__
inline
float __fsqrt_rz(float x) { return __ocml_sqrt_f32(x); }
float __fsqrt_rz(float x) { return __ocml_sqrt_rtz_f32(x); }
__DEVICE__
inline
float __fsub_rd(float x, float y) { return __ocml_sub_rtp_f32(x, y); }
float __fsub_rd(float x, float y) { return __ocml_sub_rtn_f32(x, y); }
__DEVICE__
inline
float __fsub_rn(float x, float y) { return __ocml_sub_rte_f32(x, y); }
__DEVICE__
inline
float __fsub_ru(float x, float y) { return __ocml_sub_rtn_f32(x, y); }
float __fsub_ru(float x, float y) { return __ocml_sub_rtp_f32(x, y); }
__DEVICE__
inline
float __fsub_rz(float x, float y) { return __ocml_sub_rtz_f32(x, y); }
#endif
__DEVICE__
inline
float __log10f(float x) { return __ocml_log10_f32(x); }
@@ -1034,39 +1038,40 @@ double yn(int n, double x)
}
// BEGIN INTRINSICS
#if defined OCML_BASIC_ROUNDED_OPERATIONS
__DEVICE__
inline
double __dadd_rd(double x, double y) { return __ocml_add_rtp_f64(x, y); }
double __dadd_rd(double x, double y) { return __ocml_add_rtn_f64(x, y); }
__DEVICE__
inline
double __dadd_rn(double x, double y) { return __ocml_add_rte_f64(x, y); }
__DEVICE__
inline
double __dadd_ru(double x, double y) { return __ocml_add_rtn_f64(x, y); }
double __dadd_ru(double x, double y) { return __ocml_add_rtp_f64(x, y); }
__DEVICE__
inline
double __dadd_rz(double x, double y) { return __ocml_add_rtz_f64(x, y); }
__DEVICE__
inline
double __ddiv_rd(double x, double y) { return x / y; }
double __ddiv_rd(double x, double y) { return __ocml_div_rtn_f64(x, y); }
__DEVICE__
inline
double __ddiv_rn(double x, double y) { return x / y; }
double __ddiv_rn(double x, double y) { return __ocml_div_rte_f64(x, y); }
__DEVICE__
inline
double __ddiv_ru(double x, double y) { return x / y; }
double __ddiv_ru(double x, double y) { return __ocml_div_rtp_f64(x, y); }
__DEVICE__
inline
double __ddiv_rz(double x, double y) { return x / y; }
double __ddiv_rz(double x, double y) { return __ocml_div_rtz_f64(x, y); }
__DEVICE__
inline
double __dmul_rd(double x, double y) { return __ocml_mul_rtp_f64(x, y); }
double __dmul_rd(double x, double y) { return __ocml_mul_rtn_f64(x, y); }
__DEVICE__
inline
double __dmul_rn(double x, double y) { return __ocml_mul_rte_f64(x, y); }
__DEVICE__
inline
double __dmul_ru(double x, double y) { return __ocml_mul_rtn_f64(x, y); }
double __dmul_ru(double x, double y) { return __ocml_mul_rtp_f64(x, y); }
__DEVICE__
inline
double __dmul_rz(double x, double y) { return __ocml_mul_rtz_f64(x, y); }
@@ -1084,25 +1089,25 @@ inline
double __drcp_rz(double x) { return __llvm_amdgcn_rcp_f64(x); }
__DEVICE__
inline
double __dsqrt_rd(double x) { return __ocml_sqrt_f64(x); }
double __dsqrt_rd(double x) { return __ocml_sqrt_rtn_f64(x); }
__DEVICE__
inline
double __dsqrt_rn(double x) { return __ocml_sqrt_f64(x); }
double __dsqrt_rn(double x) { return __ocml_sqrt_rte_f64(x); }
__DEVICE__
inline
double __dsqrt_ru(double x) { return __ocml_sqrt_f64(x); }
double __dsqrt_ru(double x) { return __ocml_sqrt_rtp_f64(x); }
__DEVICE__
inline
double __dsqrt_rz(double x) { return __ocml_sqrt_f64(x); }
double __dsqrt_rz(double x) { return __ocml_sqrt_rtz_f64(x); }
__DEVICE__
inline
double __dsub_rd(double x, double y) { return __ocml_sub_rtp_f64(x, y); }
double __dsub_rd(double x, double y) { return __ocml_sub_rtn_f64(x, y); }
__DEVICE__
inline
double __dsub_rn(double x, double y) { return __ocml_sub_rte_f64(x, y); }
__DEVICE__
inline
double __dsub_ru(double x, double y) { return __ocml_sub_rtn_f64(x, y); }
double __dsub_ru(double x, double y) { return __ocml_sub_rtp_f64(x, y); }
__DEVICE__
inline
double __dsub_rz(double x, double y) { return __ocml_sub_rtz_f64(x, y); }
@@ -1110,7 +1115,7 @@ __DEVICE__
inline
double __fma_rd(double x, double y, double z)
{
return __ocml_fma_rtp_f64(x, y, z);
return __ocml_fma_rtn_f64(x, y, z);
}
__DEVICE__
inline
@@ -1122,7 +1127,7 @@ __DEVICE__
inline
double __fma_ru(double x, double y, double z)
{
return __ocml_fma_rtn_f64(x, y, z);
return __ocml_fma_rtp_f64(x, y, z);
}
__DEVICE__
inline
@@ -1130,6 +1135,7 @@ double __fma_rz(double x, double y, double z)
{
return __ocml_fma_rtz_f64(x, y, z);
}
#endif
// END INTRINSICS
// END DOUBLE
projects/hip/include/hip/hcc_detail/math_fwd.h
+49 -1
View File
@@ -288,6 +288,30 @@ __attribute__((const))
float __ocml_mul_rtz_f32(float, float);
__device__
__attribute__((const))
float __ocml_div_rte_f32(float, float);
__device__
__attribute__((const))
float __ocml_div_rtn_f32(float, float);
__device__
__attribute__((const))
float __ocml_div_rtp_f32(float, float);
__device__
__attribute__((const))
float __ocml_div_rtz_f32(float, float);
__device__
__attribute__((const))
float __ocml_sqrt_rte_f32(float, float);
__device__
__attribute__((const))
float __ocml_sqrt_rtn_f32(float, float);
__device__
__attribute__((const))
float __ocml_sqrt_rtp_f32(float, float);
__device__
__attribute__((const))
float __ocml_sqrt_rtz_f32(float, float);
__device__
__attribute__((const))
float __ocml_fma_rte_f32(float, float, float);
__device__
__attribute__((const))
@@ -572,6 +596,30 @@ __attribute__((const))
double __ocml_mul_rtz_f64(double, double);
__device__
__attribute__((const))
double __ocml_div_rte_f64(double, double);
__device__
__attribute__((const))
double __ocml_div_rtn_f64(double, double);
__device__
__attribute__((const))
double __ocml_div_rtp_f64(double, double);
__device__
__attribute__((const))
double __ocml_div_rtz_f64(double, double);
__device__
__attribute__((const))
double __ocml_sqrt_rte_f64(double, double);
__device__
__attribute__((const))
double __ocml_sqrt_rtn_f64(double, double);
__device__
__attribute__((const))
double __ocml_sqrt_rtp_f64(double, double);
__device__
__attribute__((const))
double __ocml_sqrt_rtz_f64(double, double);
__device__
__attribute__((const))
double __ocml_fma_rte_f64(double, double, double);
__device__
__attribute__((const))
@@ -594,4 +642,4 @@ double __llvm_amdgcn_rsq_f64(double) __asm("llvm.amdgcn.rsq.f64");
#if defined(__cplusplus)
} // extern "C"
#endif
#endif
projects/hip/include/hip/hcc_detail/program_state.hpp
+2
View File
@@ -99,6 +99,8 @@ const std::unordered_map<std::uintptr_t, std::vector<std::pair<hsa_agent_t, Kern
functions(bool rebuild = false);
const std::unordered_map<std::uintptr_t, std::string>& function_names(bool rebuild = false);
std::unordered_map<std::string, void*>& globals(bool rebuild = false);
std::unordered_map<
std::string, std::vector<std::pair<std::size_t, std::size_t>>>& kernargs();
hsa_executable_t load_executable(const std::string& file, hsa_executable_t executable,
hsa_agent_t agent);