adds vector_imm instructions to vector_functions, makes size of all involved registers a template parameter

This commit is contained in:
2025-02-12 20:19:25 +01:00
parent 6ce0d97e81
commit 51f3802394
3 changed files with 110 additions and 42 deletions

View File

@ -71,9 +71,12 @@ uint64_t vector_load_store_index(void* core, std::function<bool(void*, uint64_t,
uint16_t VLEN, uint8_t XLEN, uint8_t addressed_register, uint8_t index_register, uint64_t base_addr,
uint64_t vl, uint64_t vstart, vtype_t vtype, bool vm, uint8_t elem_size_byte, uint64_t elem_count,
uint8_t segment_size, bool ordered);
template <unsigned VLEN, typename elem_t>
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, uint64_t vl, uint64_t vstart, vtype_t vtype, bool vm, unsigned vs2, unsigned vs1,
unsigned vd);
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, uint64_t vl, uint64_t vstart, vtype_t vtype, bool vm, unsigned vs2,
typename std::make_signed<src1_elem_t>::type imm, unsigned vd);
} // namespace softvector
#include "vm/vector_functions.hpp"
#endif /* _VM_VECTOR_FUNCTIONS_H_ */

View File

@ -64,49 +64,50 @@ template <unsigned VLEN> vmask_view read_vmask(uint8_t* V, uint16_t elem_count,
assert(mask_start + elem_count / 8 <= V + VLEN * RFS / 8);
return {mask_start, elem_count};
}
template <typename elem_t> std::function<elem_t(elem_t, elem_t)> get_funct(unsigned 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(src2_elem_t, src1_elem_t)> get_funct(unsigned funct) {
switch(funct) {
case 0b000000: {
// VADD
return [](elem_t vs2, elem_t vs1) { return vs2 + vs1; };
}
case 0b000010: {
// VSUB
}
case 0b000100: {
// VMINU
}
case 0b000101: {
// VMIN
}
case 0b000110: {
// VMAXU
}
case 0b000111: {
// VMAX
}
case 0b001001: {
// VAND
}
case 0b001010: {
// VOR
}
case 0b001011: {
// VXOR
}
case 0b000000: // VADD
case 0b110000: // VWADDU
case 0b110100: // VWADDU.W
return [](src2_elem_t vs2, src1_elem_t vs1) { return vs2 + vs1; };
case 0b000010: // VSUB
case 0b110010: // VWSUBU
case 0b110110: // VWSUBU.W
return [](src2_elem_t vs2, src1_elem_t vs1) { return vs2 - vs1; };
case 0b000011: // VRSUB
return [](src2_elem_t vs2, src1_elem_t vs1) { return vs1 - vs2; };
case 0b001001: // VAND
return [](src2_elem_t vs2, src1_elem_t vs1) { return vs1 & vs2; };
case 0b001010: // VOR
return [](src2_elem_t vs2, src1_elem_t vs1) { return vs1 | vs2; };
case 0b001011: // VXOR
return [](src2_elem_t vs2, src1_elem_t vs1) { return vs1 ^ vs2; };
case 0b110001: // VWADD
case 0b110101: // VWADD.W
return [](src2_elem_t vs2, src1_elem_t vs1) {
return static_cast<typename std::make_signed_t<dest_elem_t>>(static_cast<typename std::make_signed_t<src2_elem_t>>(vs2) +
static_cast<typename std::make_signed_t<src1_elem_t>>(vs1));
};
case 0b110011: // VWSUB
case 0b110111: // VWSUB.W
return [](src2_elem_t vs2, src1_elem_t vs1) {
return static_cast<typename std::make_signed_t<dest_elem_t>>(static_cast<typename std::make_signed_t<src2_elem_t>>(vs2) -
static_cast<typename std::make_signed_t<src1_elem_t>>(vs1));
};
default:
throw new std::runtime_error("Uknown funct in get_funct");
}
}
template <unsigned VLEN, typename elem_t>
template <unsigned VLEN, typename dest_elem_t, typename src2_elem_t, typename src1_elem_t>
void vector_vector_op(uint8_t* V, unsigned funct6, uint64_t vl, uint64_t vstart, vtype_t vtype, bool vm, unsigned vs2, unsigned vs1,
unsigned vd) {
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, elem_t>(V, vs1, elem_count);
auto vs2_view = get_vreg<VLEN, elem_t>(V, vs2, elem_count);
auto vd_view = get_vreg<VLEN, elem_t>(V, vd, elem_count);
auto fn = get_funct<elem_t>(funct6);
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_funct<dest_elem_t, src2_elem_t, src1_elem_t>(funct6);
// 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++) {
@ -119,20 +120,20 @@ void vector_vector_op(uint8_t* V, unsigned funct6, uint64_t vl, uint64_t vstart,
}
// elements w/ index larger than elem_count are in the tail (fractional LMUL)
// elements w/ index larger than vl are in the tail
unsigned maximum_elems = VLEN * vtype.lmul() / (sizeof(elem_t) * 8);
unsigned maximum_elems = VLEN * vtype.lmul() / (sizeof(dest_elem_t) * 8);
for(unsigned idx = std::min(elem_count, vl); idx < maximum_elems; idx++) {
vd_view[idx] = vtype.vta() ? vd_view[idx] : vd_view[idx];
}
return;
}
template <unsigned VLEN, typename elem_t>
template <unsigned VLEN, typename dest_elem_t, typename src2_elem_t, typename src1_elem_t>
void vector_imm_op(uint8_t* V, unsigned funct6, uint64_t vl, uint64_t vstart, vtype_t vtype, bool vm, unsigned vs2,
typename std::make_signed<elem_t>::type imm, unsigned vd) {
typename std::make_signed<src1_elem_t>::type imm, unsigned vd) {
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, elem_t>(V, vs2, elem_count);
auto vd_view = get_vreg<VLEN, elem_t>(V, vd, elem_count);
auto fn = get_funct<elem_t>(funct6);
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_funct<dest_elem_t, src2_elem_t, src1_elem_t>(funct6);
// 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++) {
@ -145,7 +146,7 @@ void vector_imm_op(uint8_t* V, unsigned funct6, uint64_t vl, uint64_t vstart, vt
}
// elements w/ index larger than elem_count are in the tail (fractional LMUL)
// elements w/ index larger than vl are in the tail
unsigned maximum_elems = VLEN * vtype.lmul() / (sizeof(elem_t) * 8);
unsigned maximum_elems = VLEN * vtype.lmul() / (sizeof(dest_elem_t) * 8);
for(unsigned idx = std::min(elem_count, vl); idx < maximum_elems; idx++) {
vd_view[idx] = vtype.vta() ? vd_view[idx] : vd_view[idx];
}