Skip to content

Commit

Permalink
Power
Browse files Browse the repository at this point in the history
  • Loading branch information
@eshoguli @alvoron
eshoguli authored and alvoron committed Sep 21, 2023
1 parent 30630b4 commit ac8b2fc
Show file tree
Hide file tree
Showing 7 changed files with 368 additions and 93 deletions.
275 changes: 267 additions & 8 deletions src/plugins/intel_cpu/src/emitters/aarch64/jit_eltwise_emitters.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -196,12 +196,249 @@ void jit_power_emitter::emit_impl(const std::vector<size_t>& in_vec_idxs, const
}
}


extern "C" float my_function(float v1, float v2);
float my_function(float v1, float v2) {
return pow(v1, v2);
}

namespace {
// void preamble(dnnl::impl::cpu::aarch64::jit_generator* h) {
// using namespace Xbyak_aarch64::util;
// uint64_t sveLen = get_sve_length();

// h->stp(h->x29, h->x30, pre_ptr(h->sp, -16));
// /* x29 is a frame pointer. */
// h->mov(h->x29, h->sp);
// h->sub(h->sp, h->sp, static_cast<int64_t>(preserved_stack_size) - 16);

// /* x9 can be used as a temporal register. */
// h->mov(h->x9, h->sp);

// // if (vreg_to_preserve) {
// // st4((v8.d - v11.d)[0], post_ptr(x9, vreg_len_preserve * 4));
// // st4((v12.d - v15.d)[0], post_ptr(x9, vreg_len_preserve * 4));
// // }
// for (size_t i = 0; i < num_abi_save_gpr_regs; i += 2) {
// stp(Xbyak_aarch64::XReg(abi_save_gpr_regs[i]),
// Xbyak_aarch64::XReg(abi_save_gpr_regs[i + 1]),
// post_ptr(x9, xreg_len * 2));
// }

// // if (sveLen) { /* SVE is available. */
// // ptrue(P_ALL_ONE.b);
// // pfalse(P_ALL_ZERO.b);
// // }
// // if (sveLen >= SVE_256) {
// // ptrue(P_NOT_128.b, Xbyak_aarch64::VL16);
// // not_(P_NOT_128.b, P_ALL_ONE / Xbyak_aarch64::T_z, P_NOT_128.b);
// // }
// // if (sveLen >= SVE_512) {
// // ptrue(P_NOT_256.b, Xbyak_aarch64::VL32);
// // not_(P_NOT_256.b, P_ALL_ONE / Xbyak_aarch64::T_z, P_NOT_256.b);
// // }

// h->mov(h->X_SP, sp);
// h->sub_imm(h->X_TRANSLATOR_STACK, h->X_SP, translator_stack_offset, h->X_TMP_0);
// }

// void postamble(dnnl::impl::cpu::aarch64::jit_generator* h) {
// using namespace Xbyak_aarch64::util;

// h->mov(h->x9, h->sp);

// if (vreg_to_preserve) {
// ld4((v8.d - v11.d)[0], post_ptr(x9, vreg_len_preserve * 4));
// ld4((v12.d - v15.d)[0], post_ptr(x9, vreg_len_preserve * 4));
// }

// for (size_t i = 0; i < num_abi_save_gpr_regs; i += 2) {
// ldp(Xbyak_aarch64::XReg(abi_save_gpr_regs[i]),
// Xbyak_aarch64::XReg(abi_save_gpr_regs[i + 1]),
// post_ptr(x9, xreg_len * 2));
// }

// h->add(h->sp, h->sp, static_cast<int64_t>(preserved_stack_size) - 16);
// h->ldp(h->x29, h->x30, post_ptr(h->sp, 16));
// h->ret();
// }
} // namespace

template <dnnl::impl::cpu::aarch64::cpu_isa_t isa>
void jit_power_emitter::emit_isa(const std::vector<size_t> &in_vec_idxs, const std::vector<size_t> &out_vec_idxs) const {
if (exec_prc_ != Precision::FP32) {
IE_THROW() << "unsupported precision: " << exec_prc_;
}

// {
// auto func = powf;
// auto res = func(2.f, 3.f);
// res = res + 1.f;
// if (res == 8.f) {
// std::cout << res << std::endl;
// }
// }

// {
// Xbyak_aarch64::XReg x8(8);

// //h->str(x8, pre_ptr(h->X_SP, 0x4));
// //h->ldr(x8, pre_ptr(h->X_SP, 0x4));

// //h->str(x8, reinterpret_cast<uintptr_t>(powf));
// // auto func_addr = reinterpret_cast<uintptr_t>(powf);
// auto func_addr = reinterpret_cast<uintptr_t>(my_function);
// h->mov(x8, func_addr);
// //h->ldr(x8, pre_ptr(h->X_SP, reinterpret_cast<uintptr_t>(powf)));
// //h->ldr(x8, pre_ptr(h->X_SP, 4));
// // h->ldr(x8, post_ptr(h->sp, 4));

// // Xbyak_aarch64::SReg s0(0);
// // h->fmov(s0, 2.f);

// // Xbyak_aarch64::SReg s1(0);
// // h->fmov(s1, 3.f);

// Xbyak_aarch64::SReg x0(0);
// h->fmov(x0, 2.0);

// Xbyak_aarch64::SReg x1(1);
// h->fmov(x1, 3.0);

// h->blr(x8);

// //Xbyak_aarch64::XReg x0(0);
// //h->ldr(x0, pre_ptr(h->X_SP, 0x4));
// //h->str(x0, reinterpret_cast<uintptr_t>(reinterpret_cast<uintptr_t>(powf) - 4));
// //h->ldr(x0, reinterpret_cast<uintptr_t>(reinterpret_cast<uintptr_t>(powf) - 4));
// }

// // h->mov(h->, reinterpret_cast<uintptr_t>(powf));
// //h->b(Xbyak_aarch64::AL, loop_label);



//h->preamble();

// {
// const size_t xreg_len = 8;
// const size_t vreg_len_preserve = 8; // Only bottom 8byte must be preserved.
// const size_t vreg_to_preserve = 8; // VREG8 - VREG15
// const size_t num_abi_save_gpr_regs = 12;

// const size_t preserved_stack_size = xreg_len * (2 + num_abi_save_gpr_regs)
// + vreg_len_preserve * vreg_to_preserve;

// // const size_t size_of_abi_save_regs = num_abi_save_gpr_regs * x0.getBit() / 8
// // + vreg_to_preserve * vreg_len_preserve;

// using namespace Xbyak_aarch64::util;
// uint64_t sveLen = 0; //get_sve_length();

// //h->stp(h->x29, h->x30, pre_ptr(h->sp, -16));
// /* x29 is a frame pointer. */
// //h->mov(h->x29, h->sp);
// h->sub(h->sp, h->sp, static_cast<int64_t>(preserved_stack_size) - 16);

// /* x9 can be used as a temporal register. */
// h->mov(h->x9, h->sp);

// if (vreg_to_preserve) {
// h->st4((h->v8.d - h->v11.d)[0], post_ptr(h->x9, vreg_len_preserve * 4));
// h->st4((h->v12.d - h->v15.d)[0], post_ptr(h->x9, vreg_len_preserve * 4));
// }
// for (size_t i = 0; i < num_abi_save_gpr_regs; i += 2) {
// h->stp(Xbyak_aarch64::XReg(dnnl::impl::cpu::aarch64::abi_save_gpr_regs[i]),
// Xbyak_aarch64::XReg(dnnl::impl::cpu::aarch64::abi_save_gpr_regs[i + 1]),
// post_ptr(h->x9, xreg_len * 2));
// }

// // if (sveLen) { /* SVE is available. */
// // h->ptrue(P_ALL_ONE.b);
// // h->pfalse(P_ALL_ZERO.b);
// // }
// // if (sveLen >= SVE_256) {
// // h->ptrue(P_NOT_128.b, Xbyak_aarch64::VL16);
// // h->not_(P_NOT_128.b, P_ALL_ONE / Xbyak_aarch64::T_z, P_NOT_128.b);
// // }
// // if (sveLen >= SVE_512) {
// // h->ptrue(h->P_NOT_256.b, Xbyak_aarch64::VL32);
// // h->not_(h->P_NOT_256.b, h->P_ALL_ONE / Xbyak_aarch64::T_z, P_NOT_256.b);
// // }

// h->mov(h->X_SP, h->sp);
// h->sub_imm(h->X_TRANSLATOR_STACK, h->X_SP, dnnl::impl::cpu::aarch64::jit_generator::translator_stack_offset, h->X_TMP_0);
// }



// // TODO: debug
// std::cout << "in_vec_idxs[0]=" << in_vec_idxs[0] << std::endl;
// std::cout << "out_vec_idxs[0]=" << out_vec_idxs[0] << std::endl;

// using TReg = typename dnnl::impl::cpu::aarch64::cpu_isa_traits<isa>::TReg;
// TReg src = TReg(in_vec_idxs[0]);
// TReg dst = TReg(out_vec_idxs[0]);
// Xbyak_aarch64::XReg x8(8);

// // preamble
// //stp(x29, x30, pre_ptr(sp, -16));
// //h->sub(h->sp, h->sp, static_cast<int64_t>(64 + 2 * 128));

// auto func_addr = reinterpret_cast<uintptr_t>(my_function);
// h->mov(x8, func_addr);

// Xbyak_aarch64::SReg s0(0);
// Xbyak_aarch64::SReg s1(1);
// //h->fmov(s1, power);

// for (auto i = 0; i < 4; i++) {
// h->mov(s0, src.s[i]);
// // TODO: move out of loop
// h->fmov(s1, power);
// h->blr(x8);

// Xbyak_aarch64::WReg w0(0);
// h->fmov(w0, s0);
// h->mov(dst.s[i], w0);
// }

// postamble
//h->add(h->sp, h->sp, static_cast<int64_t>(64 + 2 * 128));




// h->postamble();

// {
// using namespace Xbyak_aarch64::util;

// h->mov(h->x9, h->sp);

// const size_t xreg_len = 8;
// const size_t vreg_len_preserve = 8; // Only bottom 8byte must be preserved.
// const size_t vreg_to_preserve = 8; // VREG8 - VREG15
// const size_t num_abi_save_gpr_regs = 12;

// const size_t preserved_stack_size = xreg_len * (2 + num_abi_save_gpr_regs) + vreg_len_preserve * vreg_to_preserve;

// if (vreg_to_preserve) {
// h->ld4((h->v8.d - h->v11.d)[0], post_ptr(h->x9, vreg_len_preserve * 4));
// h->ld4((h->v12.d - h->v15.d)[0], post_ptr(h->x9, vreg_len_preserve * 4));
// }

// for (size_t i = 0; i < num_abi_save_gpr_regs; i += 2) {
// h->ldp(Xbyak_aarch64::XReg(dnnl::impl::cpu::aarch64::abi_save_gpr_regs[i]),
// Xbyak_aarch64::XReg(dnnl::impl::cpu::aarch64::abi_save_gpr_regs[i + 1]),
// post_ptr(h->x9, xreg_len * 2));
// }

// h->add(h->sp, h->sp, static_cast<int64_t>(preserved_stack_size) - 16);
// //h->ldp(h->x29, h->x30, post_ptr(h->sp, 16));
// }


using TReg = typename dnnl::impl::cpu::aarch64::cpu_isa_traits<isa>::TReg;
TReg src = TReg(in_vec_idxs[0]);
TReg dst = TReg(out_vec_idxs[0]);
Expand All @@ -218,17 +455,39 @@ void jit_power_emitter::emit_isa(const std::vector<size_t> &in_vec_idxs, const s
return;
}

h->fmov(dst.s, 1.);
if (std::floor(power) == power && power != 0) {
h->fmov(dst.s, 1.);

auto current_power = static_cast<size_t>(power);
while (current_power > 0) {
if (current_power & 1) {
h->fmul(dst.s, dst.s, src.s);
auto current_power = static_cast<size_t>(power);
while (current_power > 0) {
if (current_power & 1) {
h->fmul(dst.s, dst.s, src.s);
}
if (current_power > 1) {
h->fmul(src.s, src.s, src.s);
}
current_power = current_power >> 1;
}
if (current_power > 1) {
h->fmul(src.s, src.s, src.s);
} else {
auto func_addr = reinterpret_cast<uintptr_t>(my_function);

Xbyak_aarch64::XReg x8(8);
h->mov(x8, func_addr);

Xbyak_aarch64::SReg s0(0);
Xbyak_aarch64::SReg s1(1);
//h->fmov(s1, power);

for (auto i = 0; i < 4; i++) {
h->mov(s0, src.s[i]);
// TODO: move out of loop
h->fmov(s1, power);
h->blr(x8);

Xbyak_aarch64::WReg w0(0);
h->fmov(w0, s0);
h->mov(dst.s[i], w0);
}
current_power = current_power >> 1;
}
}

Expand Down
Loading

0 comments on commit ac8b2fc

Please sign in to comment.