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

adds Integer Reduction Instructions

Browse Source
This commit is contained in:
Eyck-Alexander Jentzsch 2025-02-19 14:36:25 +01:00
parent 28ac169cfe
commit f049d8cbb3
3 changed files with 81 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
gen_input/templates/interp/CORENAME.cpp.gtl
View File

@ -467,6 +467,20 @@ if(vector != null) {%>
throw new std::runtime_error("Unsupported sew bit value");
}
}
void vector_red_op(uint8_t* V, uint8_t funct6, uint8_t funct3, uint64_t vl, uint64_t vstart, softvector::vtype_t vtype, bool vm, uint8_t vd, uint8_t vs2, uint8_t vs1, uint8_t sew_val){
switch(sew_val){
case 0b000:
return softvector::vector_red_op<${vlen}, uint8_t>(V, funct6, funct3, vl, vstart, vtype, vm, vd, vs2, vs1);
case 0b001:
return softvector::vector_red_op<${vlen}, uint16_t>(V, funct6, funct3, vl, vstart, vtype, vm, vd, vs2, vs1);
case 0b010:
return softvector::vector_red_op<${vlen}, uint32_t>(V, funct6, funct3, vl, vstart, vtype, vm, vd, vs2, vs1);
case 0b011:
return softvector::vector_red_op<${vlen}, uint64_t>(V, funct6, funct3, vl, vstart, vtype, vm, vd, vs2, vs1);
default:
throw new std::runtime_error("Unsupported sew bit value");
}
}
<%}%>
uint64_t fetch_count{0};
uint64_t tval{0};

3
src/vm/vector_functions.h
View File

@ -97,6 +97,9 @@ bool sat_vector_vector_op(uint8_t* V, unsigned funct6, unsigned funct3, uint64_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<src1_elem_t>::type imm);
template <unsigned VLEN, typename dest_elem_t, typename src_elem_t = dest_elem_t>
void vector_red_op(uint8_t* V, unsigned funct6, unsigned funct3, uint64_t vl, uint64_t vstart, vtype_t vtype, bool vm, unsigned vd,
unsigned vs2, unsigned vs1);
} // namespace softvector
#include "vm/vector_functions.hpp"
#endif /* _VM_VECTOR_FUNCTIONS_H_ */

74
src/vm/vector_functions.hpp
View File

@ -147,21 +147,11 @@ std::function<dest_elem_t(dest_elem_t, src2_elem_t, src1_elem_t)> get_funct(unsi
return [](dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) {
return static_cast<std::make_signed_t<src2_elem_t>>(vs2) >> (vs1 & shift_mask<src2_elem_t>());
};
// case 0b110000: // VWREDSUMU
// case 0b110001: // VWREDSUM
default:
throw new std::runtime_error("Unknown funct6 in get_funct");
}
else if(funct3 == OPMVV || funct3 == OPMVX)
switch(funct6) {
// case 0b000000: // VREDSUM
// case 0b000001: // VREDAND
// case 0b000010: // VREDOR
// case 0b000011: // VREDXOR
// case 0b000100: // VREDMINU
// case 0b000101: // VREDMIN
// case 0b000110: // VREDMAXU
// case 0b000111: // VREDMAX
// case 0b001110: // VSLID1EUP
// case 0b001111: // VSLIDE1DOWN
// case 0b010111: // VCOMPRESS
@ -804,4 +794,68 @@ bool sat_vector_imm_op(uint8_t* V, unsigned funct6, unsigned funct3, uint64_t vl
}
return saturated;
}
template <typename dest_elem_t, typename src_elem_t>
std::function<void(dest_elem_t&, src_elem_t)> get_red_funct(unsigned funct6, unsigned funct3) {
if(funct3 == OPIVV || funct3 == OPIVX || funct3 == OPIVI)
switch(funct6) {
// case 0b110000: // VWREDSUMU
// case 0b110001: // VWREDSUM
default:
throw new std::runtime_error("Unknown funct6 in get_funct");
}
else if(funct3 == OPMVV || funct3 == OPMVX)
switch(funct6) {
case 0b000000: // VREDSUM
return [](dest_elem_t& running_total, src_elem_t vs2) { return running_total += vs2; };
case 0b000001: // VREDAND
return [](dest_elem_t& running_total, src_elem_t vs2) { return running_total &= vs2; };
case 0b000010: // VREDOR
return [](dest_elem_t& running_total, src_elem_t vs2) { return running_total |= vs2; };
case 0b000011: // VREDXOR
return [](dest_elem_t& running_total, src_elem_t vs2) { running_total ^= vs2; };
case 0b000100: // VREDMINU
return [](dest_elem_t& running_total, src_elem_t vs2) { running_total = std::min(running_total, vs2); };
case 0b000101: // VREDMIN
return [](dest_elem_t& running_total, src_elem_t vs2) {
running_total =
std::min(static_cast<std::make_signed_t<dest_elem_t>>(running_total), static_cast<std::make_signed_t<src_elem_t>>(vs2));
};
case 0b000110: // VREDMAXU
return [](dest_elem_t& running_total, src_elem_t vs2) { running_total = std::max(running_total, vs2); };
case 0b000111: // VREDMAX
return [](dest_elem_t& running_total, src_elem_t vs2) {
running_total =
std::max(static_cast<std::make_signed_t<dest_elem_t>>(running_total), static_cast<std::make_signed_t<src_elem_t>>(vs2));
};
default:
throw new std::runtime_error("Unknown funct6 in get_funct");
}
else
throw new std::runtime_error("Unknown funct3 in get_funct");
}
template <unsigned VLEN, typename dest_elem_t, typename src_elem_t>
void vector_red_op(uint8_t* V, unsigned funct6, unsigned funct3, uint64_t vl, uint64_t vstart, vtype_t vtype, bool vm, unsigned vd,
unsigned vs2, unsigned vs1) {
if(vl == 0)
return;
uint64_t elem_count = VLEN * vtype.lmul() / vtype.sew();
vmask_view mask_reg = read_vmask<VLEN>(V, elem_count);
auto vs1_elem = get_vreg<VLEN, src_elem_t>(V, vs1, elem_count)[0];
auto vs2_view = get_vreg<VLEN, src_elem_t>(V, vs2, elem_count);
auto vd_view = get_vreg<VLEN, dest_elem_t>(V, vd, elem_count);
auto fn = get_red_funct<dest_elem_t, src_elem_t>(funct6, funct3);
dest_elem_t& running_total = {vs1_elem};
for(unsigned idx = vstart; idx < std::min(elem_count, vl); idx++) {
bool mask_active = vm ? 1 : mask_reg[idx];
if(mask_active) {
fn(running_total, vs2_view[idx]);
}
}
vd_view[0] = running_total;
// the tail is all elements of the destination register beyond the first one
for(unsigned idx = 1; idx < VLEN / (vtype.sew() * RFS); idx++) {
vd_view[idx] = vtype.vta() ? vd_view[idx] : vd_view[idx];
}
return;
}
} // namespace softvector