[Beignet] [PATCH V2] GBE: use native exp instruction when enough precision
Guo Yejun
yejun.guo at intel.com
Sun Jan 19 16:38:23 PST 2014
for the input data with enough precision, use the native exp instruction,
otherwise, use the software path to emulate the exp function.
Signed-off-by: Guo Yejun <yejun.guo at intel.com>
---
backend/src/backend/gen_insn_selection.cpp | 1 +
backend/src/ir/instruction.hpp | 2 ++
backend/src/ir/instruction.hxx | 1 +
backend/src/llvm/llvm_gen_backend.cpp | 2 ++
backend/src/llvm/llvm_gen_ocl_function.hxx | 1 +
backend/src/ocl_stdlib.tmpl.h | 10 ++++++++--
6 files changed, 15 insertions(+), 2 deletions(-)
diff --git a/backend/src/backend/gen_insn_selection.cpp b/backend/src/backend/gen_insn_selection.cpp
index 3d2ebcc..3d35365 100644
--- a/backend/src/backend/gen_insn_selection.cpp
+++ b/backend/src/backend/gen_insn_selection.cpp
@@ -1563,6 +1563,7 @@ namespace gbe
case ir::OP_COS: sel.MATH(dst, GEN_MATH_FUNCTION_COS, src); break;
case ir::OP_SIN: sel.MATH(dst, GEN_MATH_FUNCTION_SIN, src); break;
case ir::OP_LOG: sel.MATH(dst, GEN_MATH_FUNCTION_LOG, src); break;
+ case ir::OP_EXP: sel.MATH(dst, GEN_MATH_FUNCTION_EXP, src); break;
case ir::OP_SQR: sel.MATH(dst, GEN_MATH_FUNCTION_SQRT, src); break;
case ir::OP_RSQ: sel.MATH(dst, GEN_MATH_FUNCTION_RSQ, src); break;
case ir::OP_RCP: sel.MATH(dst, GEN_MATH_FUNCTION_INV, src); break;
diff --git a/backend/src/ir/instruction.hpp b/backend/src/ir/instruction.hpp
index c827b9a..248a802 100644
--- a/backend/src/ir/instruction.hpp
+++ b/backend/src/ir/instruction.hpp
@@ -575,6 +575,8 @@ namespace ir {
Instruction ABS(Type type, Register dst, Register src);
/*! log.type dst src */
Instruction LOG(Type type, Register dst, Register src);
+ /*! exp.type dst src */
+ Instruction EXP(Type type, Register dst, Register src);
/*! sqr.type dst src */
Instruction SQR(Type type, Register dst, Register src);
/*! rsq.type dst src */
diff --git a/backend/src/ir/instruction.hxx b/backend/src/ir/instruction.hxx
index baaaca2..bb5229a 100644
--- a/backend/src/ir/instruction.hxx
+++ b/backend/src/ir/instruction.hxx
@@ -29,6 +29,7 @@ DECL_INSN(MOV, UnaryInstruction)
DECL_INSN(COS, UnaryInstruction)
DECL_INSN(SIN, UnaryInstruction)
DECL_INSN(LOG, UnaryInstruction)
+DECL_INSN(EXP, UnaryInstruction)
DECL_INSN(SQR, UnaryInstruction)
DECL_INSN(RSQ, UnaryInstruction)
DECL_INSN(RCP, UnaryInstruction)
diff --git a/backend/src/llvm/llvm_gen_backend.cpp b/backend/src/llvm/llvm_gen_backend.cpp
index c468b02..6d61f90 100644
--- a/backend/src/llvm/llvm_gen_backend.cpp
+++ b/backend/src/llvm/llvm_gen_backend.cpp
@@ -2035,6 +2035,7 @@ namespace gbe
case GEN_OCL_SQR:
case GEN_OCL_RSQ:
case GEN_OCL_LOG:
+ case GEN_OCL_EXP:
case GEN_OCL_POW:
case GEN_OCL_RCP:
case GEN_OCL_ABS:
@@ -2314,6 +2315,7 @@ namespace gbe
case GEN_OCL_COS: this->emitUnaryCallInst(I,CS,ir::OP_COS); break;
case GEN_OCL_SIN: this->emitUnaryCallInst(I,CS,ir::OP_SIN); break;
case GEN_OCL_LOG: this->emitUnaryCallInst(I,CS,ir::OP_LOG); break;
+ case GEN_OCL_EXP: this->emitUnaryCallInst(I,CS,ir::OP_EXP); break;
case GEN_OCL_SQR: this->emitUnaryCallInst(I,CS,ir::OP_SQR); break;
case GEN_OCL_RSQ: this->emitUnaryCallInst(I,CS,ir::OP_RSQ); break;
case GEN_OCL_RCP: this->emitUnaryCallInst(I,CS,ir::OP_RCP); break;
diff --git a/backend/src/llvm/llvm_gen_ocl_function.hxx b/backend/src/llvm/llvm_gen_ocl_function.hxx
index c9e634c..de2890c 100644
--- a/backend/src/llvm/llvm_gen_ocl_function.hxx
+++ b/backend/src/llvm/llvm_gen_ocl_function.hxx
@@ -25,6 +25,7 @@ DECL_LLVM_GEN_FUNCTION(SIN, __gen_ocl_sin)
DECL_LLVM_GEN_FUNCTION(SQR, __gen_ocl_sqrt)
DECL_LLVM_GEN_FUNCTION(RSQ, __gen_ocl_rsqrt)
DECL_LLVM_GEN_FUNCTION(LOG, __gen_ocl_log)
+DECL_LLVM_GEN_FUNCTION(EXP, __gen_ocl_exp)
DECL_LLVM_GEN_FUNCTION(POW, __gen_ocl_pow)
DECL_LLVM_GEN_FUNCTION(RCP, __gen_ocl_rcp)
DECL_LLVM_GEN_FUNCTION(RNDZ, __gen_ocl_rndz)
diff --git a/backend/src/ocl_stdlib.tmpl.h b/backend/src/ocl_stdlib.tmpl.h
index bbd056f..d191b8e 100755
--- a/backend/src/ocl_stdlib.tmpl.h
+++ b/backend/src/ocl_stdlib.tmpl.h
@@ -688,6 +688,7 @@ PURE CONST float __gen_ocl_cos(float x);
PURE CONST float __gen_ocl_sqrt(float x);
PURE CONST float __gen_ocl_rsqrt(float x);
PURE CONST float __gen_ocl_log(float x);
+PURE CONST float __gen_ocl_exp(float x);
PURE CONST float __gen_ocl_pow(float x, float y);
PURE CONST float __gen_ocl_rcp(float x);
PURE CONST float __gen_ocl_rndz(float x);
@@ -2247,7 +2248,7 @@ INLINE_OVERLOADABLE float native_tan(float x) {
INLINE_OVERLOADABLE float __gen_ocl_internal_tanpi(float x) {
return native_tan(x * M_PI_F);
}
-INLINE_OVERLOADABLE float native_exp(float x) { return __gen_ocl_pow(M_E_F, x); }
+INLINE_OVERLOADABLE float native_exp(float x) { return __gen_ocl_exp(x); }
INLINE_OVERLOADABLE float native_exp2(float x) { return __gen_ocl_pow(2, x); }
INLINE_OVERLOADABLE float native_exp10(float x) { return __gen_ocl_pow(10, x); }
INLINE_OVERLOADABLE float __gen_ocl_internal_cbrt(float x) {
@@ -2619,7 +2620,12 @@ INLINE_OVERLOADABLE float __gen_ocl_internal_rint(float x) {
}
INLINE_OVERLOADABLE float __gen_ocl_internal_exp(float x) {
- //return native_exp(x);
+ //use native instruction when it has enough precision
+ if (x > 128 || x < -128)
+ {
+ return native_exp(x);
+ }
+
float o_threshold = 8.8721679688e+01, /* 0x42b17180 */
u_threshold = -1.0397208405e+02, /* 0xc2cff1b5 */
twom100 = 7.8886090522e-31, /* 2**-100=0x0d800000 */
--
1.7.9.5
More information about the Beignet
mailing list