DBT-RISE/DBT-RISE-TGC
12
0
Fork 0
Code Issues 1 Pull Requests 1 Projects Releases Wiki Activity

adds narrowing fixed point instructions

Browse Source
This commit is contained in:
Eyck-Alexander Jentzsch 2025-02-19 10:10:41 +01:00
parent a6f24db83a
commit 28ac169cfe
3 changed files with 104 additions and 55 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

26
gen_input/templates/interp/CORENAME.cpp.gtl
View File

@ -441,6 +441,32 @@ if(vector != null) {%>
throw new std::runtime_error("Unsupported sew bit value");
}
}
bool sat_vector_vector_vw(uint8_t* V, uint8_t funct6, uint8_t funct3, uint64_t vl, uint64_t vstart, softvector::vtype_t vtype, uint64_t vxrm, bool vm, uint8_t vd, uint8_t vs2, uint8_t vs1, uint8_t sew_val){
switch(sew_val){
case 0b000:
return softvector::sat_vector_vector_op<${vlen}, uint8_t, uint16_t>(V, funct6, funct3, vl, vstart, vtype, vxrm, vm, vd, vs2, vs1);
case 0b001:
return softvector::sat_vector_vector_op<${vlen}, uint16_t, uint32_t>(V, funct6, funct3, vl, vstart, vtype, vxrm, vm, vd, vs2, vs1);
case 0b010:
return softvector::sat_vector_vector_op<${vlen}, uint32_t, uint64_t>(V, funct6, funct3, vl, vstart, vtype, vxrm, vm, vd, vs2, vs1);
case 0b011: // would require 128 bits vs2 value
default:
throw new std::runtime_error("Unsupported sew bit value");
}
}
bool sat_vector_imm_vw(uint8_t* V, uint8_t funct6, uint8_t funct3, uint64_t vl, uint64_t vstart, softvector::vtype_t vtype, uint64_t vxrm, bool vm, uint8_t vd, uint8_t vs2, int64_t imm, uint8_t sew_val){
switch(sew_val){
case 0b000:
return softvector::sat_vector_imm_op<${vlen}, uint8_t, uint16_t>(V, funct6, funct3, vl, vstart, vtype, vxrm, vm, vd, vs2, imm);
case 0b001:
return softvector::sat_vector_imm_op<${vlen}, uint16_t, uint32_t>(V, funct6, funct3, vl, vstart, vtype, vxrm, vm, vd, vs2, imm);
case 0b010:
return softvector::sat_vector_imm_op<${vlen}, uint32_t, uint64_t>(V, funct6, funct3, vl, vstart, vtype, vxrm, vm, vd, vs2, imm);
case 0b011: // would require 128 bits vs2 value
default:
throw new std::runtime_error("Unsupported sew bit value");
}
}
<%}%>
uint64_t fetch_count{0};
uint64_t tval{0};

6
src/vm/vector_functions.h
View File

@ -91,12 +91,12 @@ void carry_vector_vector_op(uint8_t* V, unsigned funct, uint64_t vl, uint64_t vs
template <unsigned VLEN, typename elem_t>
void carry_vector_imm_op(uint8_t* V, unsigned funct, uint64_t vl, uint64_t vstart, vtype_t vtype, bool vm, unsigned vd, unsigned vs2,
typename std::make_signed<elem_t>::type imm);
template <unsigned VLEN, typename dest_elem_t, typename src_elem_t = dest_elem_t>
template <unsigned VLEN, typename dest_elem_t, typename src2_elem_t = dest_elem_t, typename src1_elem_t = dest_elem_t>
bool sat_vector_vector_op(uint8_t* V, unsigned funct6, unsigned funct3, uint64_t vl, uint64_t vstart, vtype_t vtype, int64_t vxrm, bool vm,
unsigned vd, unsigned vs2, unsigned vs1);
template <unsigned VLEN, typename dest_elem_t, typename src_elem_t = dest_elem_t>
template <unsigned VLEN, typename dest_elem_t, typename src2_elem_t = dest_elem_t, typename src1_elem_t = dest_elem_t>
bool sat_vector_imm_op(uint8_t* V, unsigned funct6, unsigned funct3, uint64_t vl, uint64_t vstart, vtype_t vtype, int64_t vxrm, bool vm,
unsigned vd, unsigned vs2, typename std::make_signed<src_elem_t>::type imm);
unsigned vd, unsigned vs2, typename std::make_signed<src1_elem_t>::type imm);
} // namespace softvector
#include "vm/vector_functions.hpp"
#endif /* _VM_VECTOR_FUNCTIONS_H_ */

127
src/vm/vector_functions.hpp
View File

@ -606,15 +606,15 @@ template <typename T> T roundoff(T v, uint64_t d, int64_t vxrm) {
unsigned r = get_rounding_increment(v, d, vxrm);
return (v >> d) + r;
}
template <typename dest_elem_t, typename src_elem_t = dest_elem_t>
std::function<bool(uint64_t, vtype_t, dest_elem_t&, src_elem_t, src_elem_t)> get_sat_funct(unsigned funct6, unsigned funct3) {
template <typename dest_elem_t, typename src2_elem_t = dest_elem_t, typename src1_elem_T = dest_elem_t>
std::function<bool(uint64_t, vtype_t, dest_elem_t&, src2_elem_t, src1_elem_T)> get_sat_funct(unsigned funct6, unsigned funct3) {
if(funct3 == OPIVV || funct3 == OPIVX || funct3 == OPIVI)
switch(funct6) {
case 0b100000: // VSADDU
return [](uint64_t vxrm, vtype_t vtype, dest_elem_t& vd, src_elem_t vs2, src_elem_t vs1) {
auto res = static_cast<twice_t<src_elem_t>>(vs2) + static_cast<twice_t<src_elem_t>>(vs1);
if(res > std::numeric_limits<src_elem_t>::max()) {
vd = std::numeric_limits<src_elem_t>::max();
return [](uint64_t vxrm, vtype_t vtype, dest_elem_t& vd, src2_elem_t vs2, src1_elem_T vs1) {
auto res = static_cast<twice_t<src1_elem_T>>(vs2) + static_cast<twice_t<src1_elem_T>>(vs1);
if(res > std::numeric_limits<dest_elem_t>::max()) {
vd = std::numeric_limits<dest_elem_t>::max();
return 1;
} else {
vd = res;
@ -622,14 +622,14 @@ std::function<bool(uint64_t, vtype_t, dest_elem_t&, src_elem_t, src_elem_t)> get
}
};
case 0b100001: // VSADD
return [](uint64_t vxrm, vtype_t vtype, dest_elem_t& vd, src_elem_t vs2, src_elem_t vs1) {
auto res = static_cast<twice_t<std::make_signed_t<src_elem_t>>>(static_cast<std::make_signed_t<src_elem_t>>(vs2)) +
static_cast<twice_t<std::make_signed_t<src_elem_t>>>(static_cast<std::make_signed_t<src_elem_t>>(vs1));
if(res < std::numeric_limits<std::make_signed_t<src_elem_t>>::min()) {
vd = std::numeric_limits<std::make_signed_t<src_elem_t>>::min();
return [](uint64_t vxrm, vtype_t vtype, dest_elem_t& vd, src2_elem_t vs2, src1_elem_T vs1) {
auto res = static_cast<twice_t<std::make_signed_t<src2_elem_t>>>(static_cast<std::make_signed_t<src2_elem_t>>(vs2)) +
static_cast<twice_t<std::make_signed_t<src1_elem_T>>>(static_cast<std::make_signed_t<src1_elem_T>>(vs1));
if(res < std::numeric_limits<std::make_signed_t<dest_elem_t>>::min()) {
vd = std::numeric_limits<std::make_signed_t<dest_elem_t>>::min();
return 1;
} else if(res > std::numeric_limits<std::make_signed_t<src_elem_t>>::max()) {
vd = std::numeric_limits<std::make_signed_t<src_elem_t>>::max();
} else if(res > std::numeric_limits<std::make_signed_t<dest_elem_t>>::max()) {
vd = std::numeric_limits<std::make_signed_t<dest_elem_t>>::max();
return 1;
} else {
vd = res;
@ -637,7 +637,7 @@ std::function<bool(uint64_t, vtype_t, dest_elem_t&, src_elem_t, src_elem_t)> get
}
};
case 0b100010: // VSSUBU
return [](uint64_t vxrm, vtype_t vtype, dest_elem_t& vd, src_elem_t vs2, src_elem_t vs1) {
return [](uint64_t vxrm, vtype_t vtype, dest_elem_t& vd, src2_elem_t vs2, src1_elem_T vs1) {
if(vs2 < vs1) {
vd = 0;
return 1;
@ -647,14 +647,14 @@ std::function<bool(uint64_t, vtype_t, dest_elem_t&, src_elem_t, src_elem_t)> get
}
};
case 0b100011: // VSSUB
return [](uint64_t vxrm, vtype_t vtype, dest_elem_t& vd, src_elem_t vs2, src_elem_t vs1) {
auto res = static_cast<twice_t<std::make_signed_t<src_elem_t>>>(static_cast<std::make_signed_t<src_elem_t>>(vs2)) -
static_cast<twice_t<std::make_signed_t<src_elem_t>>>(static_cast<std::make_signed_t<src_elem_t>>(vs1));
if(res < std::numeric_limits<std::make_signed_t<src_elem_t>>::min()) {
vd = std::numeric_limits<std::make_signed_t<src_elem_t>>::min();
return [](uint64_t vxrm, vtype_t vtype, dest_elem_t& vd, src2_elem_t vs2, src1_elem_T vs1) {
auto res = static_cast<twice_t<std::make_signed_t<src2_elem_t>>>(static_cast<std::make_signed_t<src2_elem_t>>(vs2)) -
static_cast<twice_t<std::make_signed_t<src1_elem_T>>>(static_cast<std::make_signed_t<src1_elem_T>>(vs1));
if(res < std::numeric_limits<std::make_signed_t<dest_elem_t>>::min()) {
vd = std::numeric_limits<std::make_signed_t<dest_elem_t>>::min();
return 1;
} else if(res > std::numeric_limits<std::make_signed_t<src_elem_t>>::max()) {
vd = std::numeric_limits<std::make_signed_t<src_elem_t>>::max();
} else if(res > std::numeric_limits<std::make_signed_t<dest_elem_t>>::max()) {
vd = std::numeric_limits<std::make_signed_t<dest_elem_t>>::max();
return 1;
} else {
vd = res;
@ -662,15 +662,15 @@ std::function<bool(uint64_t, vtype_t, dest_elem_t&, src_elem_t, src_elem_t)> get
}
};
case 0b100111: // VSMUL
return [](uint64_t vxrm, vtype_t vtype, dest_elem_t& vd, src_elem_t vs2, src_elem_t vs1) {
auto big_val = static_cast<twice_t<std::make_signed_t<src_elem_t>>>(static_cast<std::make_signed_t<src_elem_t>>(vs2)) *
static_cast<twice_t<std::make_signed_t<src_elem_t>>>(static_cast<std::make_signed_t<src_elem_t>>(vs1));
return [](uint64_t vxrm, vtype_t vtype, dest_elem_t& vd, src2_elem_t vs2, src1_elem_T vs1) {
auto big_val = static_cast<twice_t<std::make_signed_t<src2_elem_t>>>(static_cast<std::make_signed_t<src2_elem_t>>(vs2)) *
static_cast<twice_t<std::make_signed_t<src1_elem_T>>>(static_cast<std::make_signed_t<src1_elem_T>>(vs1));
auto res = roundoff(big_val, vtype.sew() - 1, vxrm);
if(res < std::numeric_limits<std::make_signed_t<src_elem_t>>::min()) {
vd = std::numeric_limits<std::make_signed_t<src_elem_t>>::min();
if(res < std::numeric_limits<std::make_signed_t<dest_elem_t>>::min()) {
vd = std::numeric_limits<std::make_signed_t<dest_elem_t>>::min();
return 1;
} else if(res > std::numeric_limits<std::make_signed_t<src_elem_t>>::max()) {
vd = std::numeric_limits<std::make_signed_t<src_elem_t>>::max();
} else if(res > std::numeric_limits<std::make_signed_t<dest_elem_t>>::max()) {
vd = std::numeric_limits<std::make_signed_t<dest_elem_t>>::max();
return 1;
} else {
vd = res;
@ -678,45 +678,68 @@ std::function<bool(uint64_t, vtype_t, dest_elem_t&, src_elem_t, src_elem_t)> get
}
};
case 0b101010: // VSSRL
return [](uint64_t vxrm, vtype_t vtype, dest_elem_t& vd, src_elem_t vs2, src_elem_t vs1) {
vd = roundoff(vs2, vs1 & shift_mask<src_elem_t>(), vxrm);
return [](uint64_t vxrm, vtype_t vtype, dest_elem_t& vd, src2_elem_t vs2, src1_elem_T vs1) {
vd = roundoff(vs2, vs1 & shift_mask<src1_elem_T>(), vxrm);
return 0;
};
case 0b101011: // VSSRA
return [](uint64_t vxrm, vtype_t vtype, dest_elem_t& vd, src_elem_t vs2, src_elem_t vs1) {
vd = roundoff(static_cast<std::make_signed_t<src_elem_t>>(vs2), vs1 & shift_mask<src_elem_t>(), vxrm);
return [](uint64_t vxrm, vtype_t vtype, dest_elem_t& vd, src2_elem_t vs2, src1_elem_T vs1) {
vd = roundoff(static_cast<std::make_signed_t<src2_elem_t>>(vs2), vs1 & shift_mask<src1_elem_T>(), vxrm);
return 0;
};
// case 0b101110: // VNCLIPU
// case 0b101111: // VNCLIP
case 0b101110: // VNCLIPU
return [](uint64_t vxrm, vtype_t vtype, dest_elem_t& vd, src2_elem_t vs2, src1_elem_T vs1) {
auto res = roundoff(vs2, vs1 & shift_mask<src2_elem_t>(), vxrm);
if(res > std::numeric_limits<dest_elem_t>::max()) {
vd = std::numeric_limits<dest_elem_t>::max();
return 1;
} else {
vd = res;
return 0;
}
};
case 0b101111: // VNCLIP
return [](uint64_t vxrm, vtype_t vtype, dest_elem_t& vd, src2_elem_t vs2, src1_elem_T vs1) {
auto res = roundoff(static_cast<std::make_signed_t<src2_elem_t>>(vs2), vs1 & shift_mask<src2_elem_t>(), vxrm);
if(res < std::numeric_limits<std::make_signed_t<dest_elem_t>>::min()) {
vd = std::numeric_limits<std::make_signed_t<dest_elem_t>>::min();
return 1;
} else if(res > std::numeric_limits<std::make_signed_t<dest_elem_t>>::max()) {
vd = std::numeric_limits<std::make_signed_t<dest_elem_t>>::max();
return 1;
} else {
vd = res;
return 0;
}
};
default:
throw new std::runtime_error("Unknown funct6 in get_sat_funct");
}
else if(funct3 == OPMVV || funct3 == OPMVX)
switch(funct6) {
case 0b001000: // VAADDU
return [](uint64_t vxrm, vtype_t vtype, dest_elem_t& vd, src_elem_t vs2, src_elem_t vs1) {
auto res = static_cast<twice_t<src_elem_t>>(vs2) + static_cast<twice_t<src_elem_t>>(vs1);
return [](uint64_t vxrm, vtype_t vtype, dest_elem_t& vd, src2_elem_t vs2, src1_elem_T vs1) {
auto res = static_cast<twice_t<src2_elem_t>>(vs2) + static_cast<twice_t<src1_elem_T>>(vs1);
vd = roundoff(res, 1, vxrm);
return 0;
};
case 0b001001: // VAADD
return [](uint64_t vxrm, vtype_t vtype, dest_elem_t& vd, src_elem_t vs2, src_elem_t vs1) {
auto res = static_cast<twice_t<std::make_signed_t<src_elem_t>>>(static_cast<std::make_signed_t<src_elem_t>>(vs2)) +
static_cast<twice_t<std::make_signed_t<src_elem_t>>>(static_cast<std::make_signed_t<src_elem_t>>(vs1));
return [](uint64_t vxrm, vtype_t vtype, dest_elem_t& vd, src2_elem_t vs2, src1_elem_T vs1) {
auto res = static_cast<twice_t<std::make_signed_t<src2_elem_t>>>(static_cast<std::make_signed_t<src2_elem_t>>(vs2)) +
static_cast<twice_t<std::make_signed_t<src1_elem_T>>>(static_cast<std::make_signed_t<src1_elem_T>>(vs1));
vd = roundoff(res, 1, vxrm);
return 0;
};
case 0b001010: // VASUBU
return [](uint64_t vxrm, vtype_t vtype, dest_elem_t& vd, src_elem_t vs2, src_elem_t vs1) {
auto res = static_cast<twice_t<src_elem_t>>(vs2) - static_cast<twice_t<src_elem_t>>(vs1);
return [](uint64_t vxrm, vtype_t vtype, dest_elem_t& vd, src2_elem_t vs2, src1_elem_T vs1) {
auto res = static_cast<twice_t<src2_elem_t>>(vs2) - static_cast<twice_t<src1_elem_T>>(vs1);
vd = roundoff(res, 1, vxrm);
return 0;
};
case 0b001011: // VASUB
return [](uint64_t vxrm, vtype_t vtype, dest_elem_t& vd, src_elem_t vs2, src_elem_t vs1) {
auto res = static_cast<twice_t<std::make_signed_t<src_elem_t>>>(static_cast<std::make_signed_t<src_elem_t>>(vs2)) -
static_cast<twice_t<std::make_signed_t<src_elem_t>>>(static_cast<std::make_signed_t<src_elem_t>>(vs1));
return [](uint64_t vxrm, vtype_t vtype, dest_elem_t& vd, src2_elem_t vs2, src1_elem_T vs1) {
auto res = static_cast<twice_t<std::make_signed_t<src2_elem_t>>>(static_cast<std::make_signed_t<src2_elem_t>>(vs2)) -
static_cast<twice_t<std::make_signed_t<src1_elem_T>>>(static_cast<std::make_signed_t<src1_elem_T>>(vs1));
vd = roundoff(res, 1, vxrm);
return 0;
};
@ -726,16 +749,16 @@ std::function<bool(uint64_t, vtype_t, dest_elem_t&, src_elem_t, src_elem_t)> get
else
throw new std::runtime_error("Unknown funct3 in get_sat_funct");
}
template <unsigned VLEN, typename dest_elem_t, typename src_elem_t>
template <unsigned VLEN, typename dest_elem_t, typename src2_elem_t, typename src1_elem_T>
bool sat_vector_vector_op(uint8_t* V, unsigned funct6, unsigned funct3, uint64_t vl, uint64_t vstart, vtype_t vtype, int64_t vxrm, bool vm,
unsigned vd, unsigned vs2, unsigned vs1) {
uint64_t elem_count = VLEN * vtype.lmul() / vtype.sew();
bool saturated = false;
vmask_view mask_reg = read_vmask<VLEN>(V, elem_count);
auto vs1_view = get_vreg<VLEN, src_elem_t>(V, vs1, elem_count);
auto vs2_view = get_vreg<VLEN, src_elem_t>(V, vs2, elem_count);
auto vs1_view = get_vreg<VLEN, src1_elem_T>(V, vs1, elem_count);
auto vs2_view = get_vreg<VLEN, src2_elem_t>(V, vs2, elem_count);
auto vd_view = get_vreg<VLEN, dest_elem_t>(V, vd, elem_count);
auto fn = get_sat_funct<dest_elem_t, src_elem_t>(funct6, funct3);
auto fn = get_sat_funct<dest_elem_t, src2_elem_t, src1_elem_T>(funct6, funct3);
// elements w/ index smaller than vstart are in the prestart and get skipped
// body is from vstart to min(elem_count, vl)
for(unsigned idx = vstart; idx < std::min(elem_count, vl); idx++) {
@ -754,15 +777,15 @@ bool sat_vector_vector_op(uint8_t* V, unsigned funct6, unsigned funct3, uint64_t
}
return saturated;
}
template <unsigned VLEN, typename dest_elem_t, typename src_elem_t>
template <unsigned VLEN, typename dest_elem_t, typename src2_elem_t, typename src1_elem_T>
bool sat_vector_imm_op(uint8_t* V, unsigned funct6, unsigned funct3, uint64_t vl, uint64_t vstart, vtype_t vtype, int64_t vxrm, bool vm,
unsigned vd, unsigned vs2, typename std::make_signed<src_elem_t>::type imm) {
unsigned vd, unsigned vs2, typename std::make_signed<src1_elem_T>::type imm) {
uint64_t elem_count = VLEN * vtype.lmul() / vtype.sew();
bool saturated = false;
vmask_view mask_reg = read_vmask<VLEN>(V, elem_count);
auto vs2_view = get_vreg<VLEN, src_elem_t>(V, vs2, elem_count);
auto vs2_view = get_vreg<VLEN, src2_elem_t>(V, vs2, elem_count);
auto vd_view = get_vreg<VLEN, dest_elem_t>(V, vd, elem_count);
auto fn = get_sat_funct<dest_elem_t, src_elem_t>(funct6, funct3);
auto fn = get_sat_funct<dest_elem_t, src2_elem_t, src1_elem_T>(funct6, funct3);
// elements w/ index smaller than vstart are in the prestart and get skipped
// body is from vstart to min(elem_count, vl)
for(unsigned idx = vstart; idx < std::min(elem_count, vl); idx++) {