DBT-RISE/DBT-RISE-TGC
12
0
Fork 0
Code Issues1 Pull Requests1 Projects Releases Wiki Activity

adds merge and move instructions

Browse Source
This commit is contained in:
Eyck-Alexander Jentzsch 2025-02-17 18:03:42 +01:00
parent 6852d1d299
commit 77807fec01
3 changed files with 59 additions and 7 deletions

28
gen_input/templates/interp/CORENAME.cpp.gtl

@ -385,6 +385,34 @@ if(vector != null) {%>
throw new std::runtime_error("Unsupported sew bit value");
}
}
void vector_vector_merge(uint8_t* V, 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_vector_op<${vlen}, uint8_t>(V, 0, 0, vl, vstart, vtype, vm, vd, vs2, vs1, softvector::carry_t::NO_CARRY, true);
case 0b001:
return softvector::vector_vector_op<${vlen}, uint16_t>(V, 0, 0, vl, vstart, vtype, vm, vd, vs2, vs1, softvector::carry_t::NO_CARRY, true);
case 0b010:
return softvector::vector_vector_op<${vlen}, uint32_t>(V, 0, 0, vl, vstart, vtype, vm, vd, vs2, vs1, softvector::carry_t::NO_CARRY, true);
case 0b011:
return softvector::vector_vector_op<${vlen}, uint64_t>(V, 0, 0, vl, vstart, vtype, vm, vd, vs2, vs1, softvector::carry_t::NO_CARRY, true);
default:
throw new std::runtime_error("Unsupported sew bit value");
}
}
void vector_imm_merge(uint8_t* V, uint64_t vl, uint64_t vstart, softvector::vtype_t vtype, bool vm, uint8_t vd, uint8_t vs2, int64_t imm, uint8_t sew_val){
switch(sew_val){
case 0b000:
return softvector::vector_imm_op<${vlen}, uint8_t>(V, 0, 0, vl, vstart, vtype, vm, vd, vs2, imm, softvector::carry_t::NO_CARRY, true);
case 0b001:
return softvector::vector_imm_op<${vlen}, uint16_t>(V, 0, 0, vl, vstart, vtype, vm, vd, vs2, imm, softvector::carry_t::NO_CARRY, true);
case 0b010:
return softvector::vector_imm_op<${vlen}, uint32_t>(V, 0, 0, vl, vstart, vtype, vm, vd, vs2, imm, softvector::carry_t::NO_CARRY, true);
case 0b011:
return softvector::vector_imm_op<${vlen}, uint64_t>(V, 0, 0, vl, vstart, vtype, vm, vd, vs2, imm, softvector::carry_t::NO_CARRY, true);
default:
throw new std::runtime_error("Unsupported sew bit value");
}
}
<%}%>
uint64_t fetch_count{0};
uint64_t tval{0};

4
src/vm/vector_functions.h

@ -74,10 +74,10 @@ uint64_t vector_load_store_index(void* core, std::function<bool(void*, uint64_t,
uint8_t segment_size, bool ordered);
template <unsigned VLEN, typename dest_elem_t, typename src2_elem_t = dest_elem_t, typename src1_elem_t = src2_elem_t>
void vector_vector_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, carry_t carry = carry_t::NO_CARRY);
unsigned vs2, unsigned vs1, carry_t carry = carry_t::NO_CARRY, bool merge = false);
template <unsigned VLEN, typename dest_elem_t, typename src2_elem_t = dest_elem_t, typename src1_elem_t = src2_elem_t>
void vector_imm_op(uint8_t* V, unsigned funct6, unsigned funct3, uint64_t vl, uint64_t vstart, vtype_t vtype, bool vm, unsigned vd,
unsigned vs2, typename std::make_signed<src1_elem_t>::type imm, carry_t carry = carry_t::NO_CARRY);
unsigned vs2, typename std::make_signed<src1_elem_t>::type imm, carry_t carry = carry_t::NO_CARRY, bool merge = false);
template <unsigned VLEN, typename dest_elem_t, typename src2_elem_t = dest_elem_t>
void vector_unary_op(uint8_t* V, unsigned unary_op, uint64_t vl, uint64_t vstart, vtype_t vtype, bool vm, unsigned vd, unsigned vs2);
template <unsigned VLEN, typename elem_t>

34
src/vm/vector_functions.hpp

@ -131,7 +131,6 @@ std::function<dest_elem_t(dest_elem_t, src2_elem_t, src1_elem_t)> get_funct(unsi
return static_cast<std::make_signed_t<dest_elem_t>>(static_cast<std::make_signed_t<src2_elem_t>>(vs2) -
static_cast<std::make_signed_t<src1_elem_t>>(vs1));
};
// case 0b010111: // VMERGE / VMV
// case 0b100000: // VSADDU
// case 0b100001: // VSADD
// case 0b100010: // VSSUBU
@ -308,9 +307,17 @@ std::function<dest_elem_t(dest_elem_t, src2_elem_t, src1_elem_t)> get_funct(unsi
else
throw new std::runtime_error("Unknown funct3 in get_funct");
}
template <typename dest_elem_t, typename src2_elem_t = dest_elem_t, typename src1_elem_t = dest_elem_t>
std::function<dest_elem_t(bool, dest_elem_t, src2_elem_t, src1_elem_t)> get_merge_funct(bool vm) {
if(vm) { // VMV
return [](bool vm, dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) { return vs1; };
} else { // VMERGE
return [](bool vm, dest_elem_t vd, src2_elem_t vs2, src1_elem_t vs1) { return vm ? vs1 : vs2; };
}
};
template <unsigned VLEN, typename dest_elem_t, typename src2_elem_t, typename src1_elem_t>
void vector_vector_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, carry_t carry) {
unsigned vs2, unsigned vs1, carry_t carry, bool merge) {
uint64_t elem_count = VLEN * vtype.lmul() / vtype.sew();
vmask_view mask_reg = read_vmask<VLEN>(V, elem_count);
auto vs1_view = get_vreg<VLEN, src1_elem_t>(V, vs1, elem_count);
@ -319,7 +326,16 @@ void vector_vector_op(uint8_t* V, unsigned funct6, unsigned funct3, uint64_t vl,
auto fn = get_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)
if(carry == carry_t::NO_CARRY) {
if(merge) {
for(unsigned idx = vstart; idx < std::min(elem_count, vl); idx++) {
auto merge_fn = get_merge_funct<dest_elem_t, src2_elem_t, src1_elem_t>(vm);
auto cur_mask = mask_reg[idx];
auto vd_val = vd_view[idx];
auto vs2_val = vs2_view[idx];
auto vs1_val = vs1_view[idx];
vd_view[idx] = merge_fn(mask_reg[idx], vd_view[idx], vs2_view[idx], vs1_view[idx]);
}
} else if(carry == carry_t::NO_CARRY) {
for(unsigned idx = vstart; idx < std::min(elem_count, vl); idx++) {
bool mask_active = vm ? 1 : mask_reg[idx];
if(mask_active) {
@ -347,7 +363,7 @@ void vector_vector_op(uint8_t* V, unsigned funct6, unsigned funct3, uint64_t vl,
}
template <unsigned VLEN, typename dest_elem_t, typename src2_elem_t, typename src1_elem_t>
void vector_imm_op(uint8_t* V, unsigned funct6, unsigned funct3, uint64_t vl, uint64_t vstart, vtype_t vtype, bool vm, unsigned vd,
unsigned vs2, typename std::make_signed<src1_elem_t>::type imm, carry_t carry) {
unsigned vs2, typename std::make_signed<src1_elem_t>::type imm, carry_t carry, bool merge) {
uint64_t elem_count = VLEN * vtype.lmul() / vtype.sew();
vmask_view mask_reg = read_vmask<VLEN>(V, elem_count);
auto vs2_view = get_vreg<VLEN, src2_elem_t>(V, vs2, elem_count);
@ -355,7 +371,15 @@ void vector_imm_op(uint8_t* V, unsigned funct6, unsigned funct3, uint64_t vl, ui
auto fn = get_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)
if(carry == carry_t::NO_CARRY) {
if(merge) {
for(unsigned idx = vstart; idx < std::min(elem_count, vl); idx++) {
auto cur_mask = mask_reg[idx];
auto vd_val = vd_view[idx];
auto vs2_val = vs2_view[idx];
auto merge_fn = get_merge_funct<dest_elem_t, src2_elem_t, src1_elem_t>(vm);
vd_view[idx] = merge_fn(mask_reg[idx], vd_view[idx], vs2_view[idx], imm);
}
} else if(carry == carry_t::NO_CARRY) {
for(unsigned idx = vstart; idx < std::min(elem_count, vl); idx++) {
bool mask_active = vm ? 1 : mask_reg[idx];
if(mask_active) {