[Beignet] [PATCH version 3] add 64-bit version of "sub_sat"
Homer Hsing
homer.xing at intel.com
Sat Sep 21 23:18:03 PDT 2013
passed PIGLIT test cases:
bin/cl-program-tester generated_tests/cl/builtin/int/builtin-long-sub_sat-1.0.generated.cl
bin/cl-program-tester generated_tests/cl/builtin/int/builtin-ulong-sub_sat-1.0.generated.cl
version 2:
temp flag register is allocated by RA now
version 3:
subnr of temp flag reg is divided by typesize
Signed-off-by: Homer Hsing <homer.xing at intel.com>
---
backend/src/backend/gen_context.cpp | 48 +++++++++++++++++++++-
backend/src/backend/gen_context.hpp | 1 +
.../src/backend/gen_insn_gen7_schedule_info.hxx | 1 +
backend/src/backend/gen_insn_selection.cpp | 21 ++++++++++
backend/src/backend/gen_insn_selection.hxx | 1 +
backend/src/ocl_stdlib.tmpl.h | 9 +++-
6 files changed, 79 insertions(+), 2 deletions(-)
diff --git a/backend/src/backend/gen_context.cpp b/backend/src/backend/gen_context.cpp
index 3d4f616..345d37f 100644
--- a/backend/src/backend/gen_context.cpp
+++ b/backend/src/backend/gen_context.cpp
@@ -983,6 +983,52 @@ namespace gbe
storeBottomHalf(dst, b);
}
+ void GenContext::emitI64SATSUBInstruction(const SelectionInstruction &insn) {
+ GenRegister x = ra->genReg(insn.src(0));
+ GenRegister y = ra->genReg(insn.src(1));
+ GenRegister dst = ra->genReg(insn.dst(0));
+ GenRegister a = ra->genReg(insn.dst(1));
+ GenRegister b = ra->genReg(insn.dst(2));
+ GenRegister c = ra->genReg(insn.dst(3));
+ GenRegister d = ra->genReg(insn.dst(4));
+ GenRegister e = ra->genReg(insn.dst(5));
+ GenRegister flagReg = ra->genReg(insn.dst(6));
+ loadTopHalf(a, x);
+ loadBottomHalf(b, x);
+ loadTopHalf(c, y);
+ loadBottomHalf(d, y);
+ if(dst.is_signed_int())
+ p->SHR(e, a, GenRegister::immud(31));
+ subWithBorrow(b, b, d);
+ subWithBorrow(a, a, d);
+ subWithBorrow(a, a, c);
+ p->ADD(c, c, d);
+ p->push();
+ p->curr.predicate = GEN_PREDICATE_NONE;
+ p->curr.useFlag(flagReg.flag_nr(), flagReg.flag_subnr());
+ if(! dst.is_signed_int()) {
+ p->CMP(GEN_CONDITIONAL_NZ, c, GenRegister::immud(0));
+ p->curr.predicate = GEN_PREDICATE_NORMAL;
+ p->MOV(a, GenRegister::immud(0));
+ p->MOV(b, GenRegister::immud(0));
+ } else {
+ p->CMP(GEN_CONDITIONAL_EQ, e, GenRegister::immud(1));
+ p->curr.predicate = GEN_PREDICATE_NORMAL;
+ p->CMP(GEN_CONDITIONAL_L, a, GenRegister::immud(0x80000000u));
+ p->MOV(a, GenRegister::immud(0x80000000u));
+ p->MOV(b, GenRegister::immud(0));
+ p->curr.predicate = GEN_PREDICATE_NONE;
+ p->CMP(GEN_CONDITIONAL_EQ, e, GenRegister::immud(0));
+ p->curr.predicate = GEN_PREDICATE_NORMAL;
+ p->CMP(GEN_CONDITIONAL_GE, a, GenRegister::immud(0x80000000u));
+ p->MOV(a, GenRegister::immud(0x7FFFFFFFu));
+ p->MOV(b, GenRegister::immud(0xFFFFFFFFu));
+ }
+ p->pop();
+ storeTopHalf(dst, a);
+ storeBottomHalf(dst, b);
+ }
+
void GenContext::loadTopHalf(GenRegister dest, GenRegister src) {
int execWidth = p->curr.execWidth;
src = src.top_half();
@@ -1070,11 +1116,11 @@ namespace gbe
int execWidth = p->curr.execWidth;
GenRegister acc0 = GenRegister::retype(GenRegister::acc(), GEN_TYPE_D);
p->push();
- p->curr.predicate = GEN_PREDICATE_NONE;
p->curr.execWidth = 8;
p->SUBB(dest, src0, src1);
p->MOV(src1, acc0);
if (execWidth == 16) {
+ p->curr.quarterControl = 1;
p->SUBB(GenRegister::suboffset(dest, 8),
GenRegister::suboffset(src0, 8),
GenRegister::suboffset(src1, 8));
diff --git a/backend/src/backend/gen_context.hpp b/backend/src/backend/gen_context.hpp
index fa43571..e24c126 100644
--- a/backend/src/backend/gen_context.hpp
+++ b/backend/src/backend/gen_context.hpp
@@ -107,6 +107,7 @@ namespace gbe
void emitI64ShiftInstruction(const SelectionInstruction &insn);
void emitI64CompareInstruction(const SelectionInstruction &insn);
void emitI64SATADDInstruction(const SelectionInstruction &insn);
+ void emitI64SATSUBInstruction(const SelectionInstruction &insn);
void emitI64ToFloatInstruction(const SelectionInstruction &insn);
void emitCompareInstruction(const SelectionInstruction &insn);
void emitJumpInstruction(const SelectionInstruction &insn);
diff --git a/backend/src/backend/gen_insn_gen7_schedule_info.hxx b/backend/src/backend/gen_insn_gen7_schedule_info.hxx
index 1ff10de..b94f235 100644
--- a/backend/src/backend/gen_insn_gen7_schedule_info.hxx
+++ b/backend/src/backend/gen_insn_gen7_schedule_info.hxx
@@ -36,3 +36,4 @@ DECL_GEN7_SCHEDULE(GetImageInfo, 20, 4, 2)
DECL_GEN7_SCHEDULE(Atomic, 80, 1, 1)
DECL_GEN7_SCHEDULE(I64MUL, 20, 4, 2)
DECL_GEN7_SCHEDULE(I64SATADD, 20, 4, 2)
+DECL_GEN7_SCHEDULE(I64SATSUB, 20, 4, 2)
diff --git a/backend/src/backend/gen_insn_selection.cpp b/backend/src/backend/gen_insn_selection.cpp
index bc3c884..83b4f1b 100644
--- a/backend/src/backend/gen_insn_selection.cpp
+++ b/backend/src/backend/gen_insn_selection.cpp
@@ -485,6 +485,8 @@ namespace gbe
void I64CMP(uint32_t conditional, Reg src0, Reg src1, GenRegister tmp[3]);
/*! Saturated addition of 64-bit integer */
void I64SATADD(Reg dst, Reg src0, Reg src1, GenRegister tmp[6]);
+ /*! Saturated subtraction of 64-bit integer */
+ void I64SATSUB(Reg dst, Reg src0, Reg src1, GenRegister tmp[6]);
/*! Encode a barrier instruction */
void BARRIER(GenRegister src);
/*! Encode a barrier instruction */
@@ -1094,6 +1096,15 @@ namespace gbe
insn->dst(i + 1) = tmp[i];
}
+ void Selection::Opaque::I64SATSUB(Reg dst, Reg src0, Reg src1, GenRegister tmp[6]) {
+ SelectionInstruction *insn = this->appendInsn(SEL_OP_I64SATSUB, 7, 2);
+ insn->dst(0) = dst;
+ insn->src(0) = src0;
+ insn->src(1) = src1;
+ for(int i=0; i<6; i++)
+ insn->dst(i + 1) = tmp[i];
+ }
+
void Selection::Opaque::CONVI64_TO_F(Reg dst, Reg src, GenRegister tmp[4]) {
SelectionInstruction *insn = this->appendInsn(SEL_OP_CONVI64_TO_F, 5, 1);
insn->dst(0) = dst;
@@ -1684,6 +1695,16 @@ namespace gbe
sel.ADD(dst, src0, GenRegister::negate(src1));
break;
case OP_SUBSAT:
+ if (type == Type::TYPE_U64 || type == Type::TYPE_S64) {
+ GenRegister tmp[6];
+ for(int i=0; i<5; i++) {
+ tmp[i] = sel.selReg(sel.reg(FAMILY_DWORD));
+ tmp[i].type = GEN_TYPE_UD;
+ }
+ tmp[5] = sel.selReg(sel.reg(FAMILY_BOOL));
+ sel.I64SATSUB(dst, src0, src1, tmp);
+ break;
+ }
sel.push();
sel.curr.saturate = GEN_MATH_SATURATE_SATURATE;
sel.ADD(dst, src0, GenRegister::negate(src1));
diff --git a/backend/src/backend/gen_insn_selection.hxx b/backend/src/backend/gen_insn_selection.hxx
index 9814cc0..89b878d 100644
--- a/backend/src/backend/gen_insn_selection.hxx
+++ b/backend/src/backend/gen_insn_selection.hxx
@@ -30,6 +30,7 @@ DECL_SELECTION_IR(ADD, BinaryInstruction)
DECL_SELECTION_IR(I64ADD, BinaryWithTempInstruction)
DECL_SELECTION_IR(I64SATADD, I64SATADDInstruction)
DECL_SELECTION_IR(I64SUB, BinaryWithTempInstruction)
+DECL_SELECTION_IR(I64SATSUB, I64SATSUBInstruction)
DECL_SELECTION_IR(MUL, BinaryInstruction)
DECL_SELECTION_IR(I64MUL, I64MULInstruction)
DECL_SELECTION_IR(ATOMIC, AtomicInstruction)
diff --git a/backend/src/ocl_stdlib.tmpl.h b/backend/src/ocl_stdlib.tmpl.h
index bc0eb2c..f7dbd3a 100644
--- a/backend/src/ocl_stdlib.tmpl.h
+++ b/backend/src/ocl_stdlib.tmpl.h
@@ -199,7 +199,14 @@ INLINE_OVERLOADABLE long add_sat(long x, long y) {
return ocl_sadd_sat(x, y);
}
OVERLOADABLE long ocl_ssub_sat(long x, long y);
-INLINE_OVERLOADABLE long sub_sat(long x, long y) { return ocl_ssub_sat(x, y); }
+INLINE_OVERLOADABLE long sub_sat(long x, long y) {
+ union {long l; uint i[2];} ux, uy;
+ ux.l = x;
+ uy.l = y;
+ if((ux.i[1] ^ uy.i[1]) & 0x80000000u)
+ return ocl_ssub_sat(x, y);
+ return x - y;
+}
#define UDEF(TYPE) \
OVERLOADABLE TYPE ocl_uadd_sat(TYPE x, TYPE y); \
OVERLOADABLE TYPE ocl_usub_sat(TYPE x, TYPE y); \
--
1.8.1.2
More information about the Beignet
mailing list