Skip to content

Commit

Permalink
no #define _At_
Browse files Browse the repository at this point in the history 
[skip ci] fix typo

[skip ci] fix again
  • Loading branch information
@archibate
archibate committed Mar 28, 2020
1 parent a40ac4f commit 44b3c2f
Showing 1 changed file with 143 additions and 30 deletions.
173 changes: 143 additions & 30 deletions taichi/codegen/codegen_opengl.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -212,52 +212,167 @@ class KernelGen : public IRVisitor {
"#define _ext_ns_f32(x) _extr_f32_[(x) >> 0]\n"
"#define _ext_ns_f64(x) _extr_f64_[(x) >> 0]\n";
}
kernel_header +=
"#define _Ax_(x) x\n"
"#define _At_(x) _Ax_(_at_##x(x))\n";
if (used.atomic_float) { // {{{
kernel_header +=
"\
#define _Atmf_Def(Add, _f_, _o_, mem, _32, float) \
float atomic##Add##_##mem##_f##_32(int addr, float rhs) \
float atomicAdd_mem_f32(int addr, float rhs) \
{ \
int old, new, ret; \
do { \
old = _##mem##_i##_32(addr); \
new = floatBitsToInt(_f_(intBitsToFloat(old) _o_ rhs)); \
} while (old != atomicCompSwap(_Ax_(_##mem##_i##_32(addr)), old, new)); \
old = _mem_i32(addr); \
new = floatBitsToInt((intBitsToFloat(old) + rhs)); \
} while (old != atomicCompSwap((_mem_i32(addr)), old, new)); \
return intBitsToFloat(old); \
}\n\
float atomicSub_mem_f32(int addr, float rhs) \
{ \
int old, new, ret; \
do { \
old = _mem_i32(addr); \
new = floatBitsToInt((intBitsToFloat(old) - rhs)); \
} while (old != atomicCompSwap((_mem_i32(addr)), old, new)); \
return intBitsToFloat(old); \
}\n\
float atomicMax_mem_f32(int addr, float rhs) \
{ \
int old, new, ret; \
do { \
old = _mem_i32(addr); \
new = floatBitsToInt(max(intBitsToFloat(old), rhs)); \
} while (old != atomicCompSwap((_mem_i32(addr)), old, new)); \
return intBitsToFloat(old); \
}\n\
float atomicMin_mem_f32(int addr, float rhs) \
{ \
int old, new, ret; \
do { \
old = _mem_i32(addr); \
new = floatBitsToInt(min(intBitsToFloat(old), rhs)); \
} while (old != atomicCompSwap((_mem_i32(addr)), old, new)); \
return intBitsToFloat(old); \
}\n\
#define _Acma_ ,\n\
#define _Atm_(func, at, x, rhs) _Ax_(func##at(x, rhs))\n\
_Atmf_Def(Add,, +, mem, 32, float)\n\
_Atmf_Def(Sub,, -, mem, 32, float)\n\
_Atmf_Def(Max, max, _Acma_, mem, 32, float)\n\
_Atmf_Def(Min, min, _Acma_, mem, 32, float)\n\
"
#ifdef _GLSL_INT64
"\
_Atmf_Def(Add,, +, 64, double)\n\
_Atmf_Def(Sub,, -, 64, double)\n\
_Atmf_Def(Max, max, _Acma_, 64, double)\n\
_Atmf_Def(Min, min, _Acma_, 64, double)\n\
double atomicAdd_mem_f64(int addr, double rhs) \
{ \
int old, new, ret; \
do { \
old = _mem_i64(addr); \
new = floatBitsToInt((intBitsToFloat(old) + rhs)); \
} while (old != atomicCompSwap((_mem_i64(addr)), old, new)); \
return intBitsToFloat(old); \
}\n\
double atomicSub_mem_f64(int addr, double rhs) \
{ \
int old, new, ret; \
do { \
old = _mem_i64(addr); \
new = floatBitsToInt((intBitsToFloat(old) - rhs)); \
} while (old != atomicCompSwap((_mem_i64(addr)), old, new)); \
return intBitsToFloat(old); \
}\n\
double atomicMax_mem_f64(int addr, double rhs) \
{ \
int old, new, ret; \
do { \
old = _mem_i64(addr); \
new = floatBitsToInt(max(intBitsToFloat(old), rhs)); \
} while (old != atomicCompSwap((_mem_i64(addr)), old, new)); \
return intBitsToFloat(old); \
}\n\
double atomicMin_mem_f64(int addr, double rhs) \
{ \
int old, new, ret; \
do { \
old = _mem_i64(addr); \
new = floatBitsToInt(min(intBitsToFloat(old), rhs)); \
} while (old != atomicCompSwap((_mem_i64(addr)), old, new)); \
return intBitsToFloat(old); \
}\n\
"
#endif
"\n"; // discussion:
// https://github.com/taichi-dev/taichi/pull/495#issuecomment-590074123
if (used.global_temp) {
kernel_header +=
"_Atmf_Def(Add,, +, gtx, 32, float)\n"
"_Atmf_Def(Sub,, -, gtx, 32, float)\n"
"_Atmf_Def(Max, max, _Acma_, gtx, 32, float)\n"
"_Atmf_Def(Min, min, _Acma_, gtx, 32, float)\n";
kernel_header += "\
float atomicAdd_gtx_f32(int addr, float rhs) \
{ \
int old, new, ret; \
do { \
old = _gtx_i32(addr); \
new = floatBitsToInt((intBitsToFloat(old) + rhs)); \
} while (old != atomicCompSwap((_gtx_i32(addr)), old, new)); \
return intBitsToFloat(old); \
}\n\
float atomicSub_gtx_f32(int addr, float rhs) \
{ \
int old, new, ret; \
do { \
old = _gtx_i32(addr); \
new = floatBitsToInt((intBitsToFloat(old) - rhs)); \
} while (old != atomicCompSwap((_gtx_i32(addr)), old, new)); \
return intBitsToFloat(old); \
}\n\
float atomicMax_gtx_f32(int addr, float rhs) \
{ \
int old, new, ret; \
do { \
old = _gtx_i32(addr); \
new = floatBitsToInt(max(intBitsToFloat(old), rhs)); \
} while (old != atomicCompSwap((_gtx_i32(addr)), old, new)); \
return intBitsToFloat(old); \
}\n\
float atomicMin_gtx_f32(int addr, float rhs) \
{ \
int old, new, ret; \
do { \
old = _gtx_i32(addr); \
new = floatBitsToInt(min(intBitsToFloat(old), rhs)); \
} while (old != atomicCompSwap((_gtx_i32(addr)), old, new)); \
return intBitsToFloat(old); \
}\n\
";
}
if (used.external_ptr) {
kernel_header +=
"_Atmf_Def(Add,, +, ext_ns, 32, float)\n"
"_Atmf_Def(Sub,, -, ext_ns, 32, float)\n"
"_Atmf_Def(Max, max, _Acma_, ext_ns, 32, float)\n"
"_Atmf_Def(Min, min, _Acma_, ext_ns, 32, float)\n";
kernel_header += "\
float atomicAdd_ext_ns_f32(int addr, float rhs) \
{ \
int old, new, ret; \
do { \
old = _ext_ns_i32(addr); \
new = floatBitsToInt((intBitsToFloat(old) + rhs)); \
} while (old != atomicCompSwap((_ext_ns_i32(addr)), old, new)); \
return intBitsToFloat(old); \
}\n\
float atomicSub_ext_ns_f32(int addr, float rhs) \
{ \
int old, new, ret; \
do { \
old = _ext_ns_i32(addr); \
new = floatBitsToInt((intBitsToFloat(old) - rhs)); \
} while (old != atomicCompSwap((_ext_ns_i32(addr)), old, new)); \
return intBitsToFloat(old); \
}\n\
float atomicMax_ext_ns_f32(int addr, float rhs) \
{ \
int old, new, ret; \
do { \
old = _ext_ns_i32(addr); \
new = floatBitsToInt(max(intBitsToFloat(old), rhs)); \
} while (old != atomicCompSwap((_ext_ns_i32(addr)), old, new)); \
return intBitsToFloat(old); \
}\n\
float atomicMin_ext_ns_f32(int addr, float rhs) \
{ \
int old, new, ret; \
do { \
old = _ext_ns_i32(addr); \
new = floatBitsToInt(min(intBitsToFloat(old), rhs)); \
} while (old != atomicCompSwap((_ext_ns_i32(addr)), old, new)); \
return intBitsToFloat(old); \
}\n\
";
}
} // }}}
if (used.random) { // TODO(archibate): random in different offloads should
Expand Down Expand Up @@ -390,7 +505,6 @@ int _rand_i32()\n\

void map_stmt_ptr_signat(Stmt *stmt, std::string signat) {
ptr_signats[stmt->id] = signat;
emit("#define _at_{} {}", stmt->raw_name(), signat);
}

void visit(GetChStmt *stmt) override {
Expand All @@ -403,7 +517,6 @@ int _rand_i32()\n\

void visit(GlobalStoreStmt *stmt) override {
TI_ASSERT(stmt->width() == 1);
//emit("_At_({}) = {};", stmt->ptr->raw_name(), stmt->data->raw_name());
emit("{}({}) = {};", ptr_signats.at(stmt->ptr->id), // throw out_of_range if not a pointer
stmt->ptr->raw_name(), stmt->data->raw_name());
}
Expand Down

0 comments on commit 44b3c2f

Please sign in to comment.