0
0
Fork 0
mirror of https://github.com/matrix-construct/construct synced 2024-11-25 16:22:35 +01:00
construct/ircd/gpt_cpu.cc
2022-06-19 20:14:22 -07:00

821 lines
21 KiB
C++

// Matrix Construct Is All You Need Is All You Need Is AllĊĊĊĊĊĊĊĊ
//
// Copyright (C) Matrix Construct Developers, Authors & Contributors
// Copyright (C) 2016-2021 Jason Volk <jason@zemos.net>
//
// Permission to use, copy, modify, and/or distribute this software for any
// purpose with or without fee is hereby granted, provided that the above
// copyright notice and this permission notice is present in all copies. The
// full license for this software is available in the LICENSE file.
#pragma clang fp exceptions(ignore)
#pragma clang fp reassociate(on)
#pragma clang fp contract(fast)
namespace ircd::gpt
{
static void adamw(const opts &, const u32, const f32, const uint, f32 *, f32 *, f32 *);
static void backprop(const opts &, const u32, const f32, model::norm &, model::norm &, model::norm &);
static void backprop(const opts &, const u32, const f32, model::attn &, model::attn &, model::attn &);
static void backprop(const opts &, const u32, const f32, model::ffnn &, model::ffnn &, model::ffnn &);
static void backprop(const opts &, const u32, const f32, model::block &, model::block &, model::block &);
static void backprop(const opts &, const u32, const f32, model::embed &, model::embed &, model::embed &);
static void backprop(const opts &, const u32, const f32, model::decoder &, model::decoder &, model::decoder &);
extern void backprop(const opts &, const u32, const f32, model::decoder &, f32 *const __restrict__ [2]) noexcept;
template<class T>
static void fmma(T *out, const T *in, const T *bias, const T *weight, const math::fmma_opts &);
static void gelu(f32x4 &, const f32x4 &);
static void gelu(f32x4 *, const f32x4 *);
static void norm(f32x4 *, const f32x4 *, const f32x4 *, const f32x4 *, const f32);
static void vals(float (&)[12][1024][64], const float (&)[12][1024][1024], const float (&)[3][1024][12][64], const size_t);
static void pare(float (&)[12][1024][1024], const float (&)[3][1024][12][64], const size_t);
static void mask(float (&)[12][1024][1024], const float (&)[12][1024][1024], const size_t);
static void smax(float (&)[12][1024][1024], const float (&)[12][1024][1024], const size_t);
static void attn(float (&)[3][1024][12][64], const float *const, const size_t, const model::decoder &, const uint layer);
static void ffnn(float *, const float *, const model::decoder &, const uint layer);
static void coil(float *, const size_t, const model::decoder &);
static void logitsmax(float *, const float *, const size_t);
static void logits(float *, const float *, const model::decoder &);
static void tail(float *, const float *, const model::decoder &);
static u16 argmax(const float *, const opts &);
static void embed(float *, const u16 token, const u16 position, const model::decoder &);
static f32
logit alignas(64) [65536],
embeds alignas(64) [1024 * 768],
scratch alignas(64) [1024 * 768];
}
void
ircd::gpt::embed(float *const out,
const u16 token,
const u16 position,
const model::decoder &model)
{
const auto &wpe
{
model.embed.pos[position]
};
const auto &wte
{
model.embed.token[token]
};
for(uint j(0); j < 768; ++j)
out[j] = wte.elem[j] + wpe.elem[j];
}
uint16_t
ircd::gpt::argmax(const float *const __restrict__ logit,
const opts &opts)
{
static const auto max
{
32U
};
const auto top
{
std::clamp(opts.top_k, 1U, max - 1)
};
u16 best[max] {0};
for(uint j(0); j < vocab::tokens; ++j)
{
best[top] = j;
std::sort(begin(best), begin(best) + top + 1, [&logit]
(const auto &a, const auto &b)
{
return logit[a] > logit[b];
});
}
const auto x
{
top > 1?
rand::integer(0, top - 1):
0
};
return best[x];
}
[[gnu::noinline]]
void
ircd::gpt::tail(float *const __restrict__ logit,
const float *const __restrict__ state,
const model::decoder &d)
{
constexpr float lnf_epsilon
{
0.00001
};
static float
buf alignas(64) [1][768];
for(uint i(0); i < 768; ++i)
buf[0][i] = state[i];
norm((f32x4 *)buf[0], (const f32x4 *)state, (const f32x4 *)d.embed.norm.bias.elem, (const f32x4 *)d.embed.norm.weight.elem, lnf_epsilon);
logits(logit, buf[0], d);
//logitsmax(logit, logit, vocab::tokens);
}
void
ircd::gpt::logits(float *const __restrict__ out,
const float *const __restrict__ in,
const model::decoder &d)
{
for(uint j(0); j < vocab::tokens; ++j)
out[j] = 0;
for(uint j(0); j < vocab::tokens; ++j)
for(uint k(0); k < 768; ++k)
out[j] += in[k] * d.embed.token[j].elem[k];
}
[[gnu::noinline]]
void
ircd::gpt::logitsmax(float *const out,
const float *const in,
const size_t num)
{
static f64x4
exps alignas(4096) [2][65536 / 4];
math::smax<f32x4, f64x4>
(
{(f32x4 *)out, num / 4},
{(const f32x4 *)in, num / 4},
exps[0],
exps[1]
);
}
[[gnu::noinline]]
void
ircd::gpt::coil(float *__restrict__ accum,
const size_t tokens,
const model::decoder &decoder)
{
static float
qkv alignas(4096) [3][1024][12][64],
state alignas(4096) [12][1024][1024],
attns alignas(4096) [12][1024][64];
for(uint i(0); i < 12; ++i)
{
const auto &layer
{
decoder.layer[i]
};
attn(qkv, accum, tokens, decoder, i);
pare(state, qkv, tokens);
mask(state, state, tokens);
smax(state, state, tokens);
vals(attns, state, qkv, tokens);
static f32 a alignas(64) [1024][768];
memset(a, 0x0, 768 * tokens * sizeof(float));
for(uint j(0); j < tokens; j++)
{
for(uint k(0); k < 12; k++)
for(uint l(0); l < 64; l++)
a[j][k * 64 + l] = attns[k][j][l];
}
static const math::fmma_opts fmma_opts
{
768, 768, 2U
};
for(uint j(0); j < tokens; ++j)
fmma((f32x4 *)(accum + j * 768), (const f32x4 *)(a[j]), (const f32x4 *)layer.attn.proj_bias.elem, (const f32x4 *)layer.attn.proj_weight, fmma_opts);
for(uint j(0); j < tokens; ++j)
ffnn(accum + j * 768, accum + j * 768, decoder, i);
}
}
void
ircd::gpt::attn(float (&__restrict__ out)[3][1024][12][64],
const float *const __restrict__ in,
const size_t num,
const model::decoder &decoder,
const uint laynum)
{
constexpr float ln1_epsilon
{
0.00001
};
const auto &layer
{
decoder.layer[laynum]
};
float
(&__restrict__ qry)[1024][12][64] { out[0] },
(&__restrict__ key)[1024][12][64] { out[1] },
(&__restrict__ val)[1024][12][64] { out[2] };
for(uint i(0); i < num; ++i)
{
static float
buf alignas(64) [768],
proj alignas(64) [2304];
norm((f32x4 *)buf, (const f32x4 *)(in + i * 768), (const f32x4 *)layer.attn.norm.bias.elem, (const f32x4 *)layer.attn.norm.weight.elem, ln1_epsilon);
static const math::fmma_opts fmma_opts
{
768, 2304, 2U,
};
memset(proj, 0x0, sizeof(proj));
fmma((f32x4 *)proj, (const f32x4 *)buf, (const f32x4 *)layer.attn.fcon_bias.fcon, (const f32x4 *)layer.attn.fcon_weight, fmma_opts);
#pragma clang loop unroll (disable)
for(uint j(0); j < 12; ++j)
for(uint k(0); k < 64; ++k)
qry[i][j][k] = proj[768 * 0 + j * 64 + k];
#pragma clang loop unroll (disable)
for(uint j(0); j < 12; ++j)
for(uint k(0); k < 64; ++k)
key[i][j][k] = proj[768 * 1 + j * 64 + k];
#pragma clang loop unroll (disable)
for(uint j(0); j < 12; ++j)
for(uint k(0); k < 64; ++k)
val[i][j][k] = proj[768 * 2 + j * 64 + k];
}
}
void
ircd::gpt::pare(float (&__restrict__ out)[12][1024][1024],
const float (&__restrict__ qkv)[3][1024][12][64],
const size_t num)
{
const float
(&__restrict__ qry)[1024][12][64] { qkv[0] },
(&__restrict__ key)[1024][12][64] { qkv[1] },
(&__restrict__ val)[1024][12][64] { qkv[2] };
#pragma clang loop unroll (disable)
for(uint j(0); j < 12; ++j)
for(uint k(0); k < num; ++k)
for(uint l(0); l < num; ++l)
out[j][k][l] = 0;
#pragma clang loop unroll (disable)
for(uint j(0); j < 12; ++j)
for(uint k(0); k < num; ++k)
for(uint l(0); l < num; ++l)
for(uint m(0); m < 64; ++m)
out[j][k][l] += qry[k][j][m] * key[l][j][m];
#pragma clang loop unroll (disable)
for(uint j(0); j < 12; ++j)
for(uint k(0); k < num; ++k)
for(uint l(0); l < num; ++l)
out[j][k][l] /= 8.0;
}
void
ircd::gpt::mask(float (&__restrict__ out)[12][1024][1024],
const float (&__restrict__ in)[12][1024][1024],
const size_t num)
{
static const float masked
{
-10000.0
};
#pragma clang loop unroll (disable)
for(uint j(0); j < 12; ++j)
for(uint k(0); k < num; ++k)
for(uint l(0); l < num; ++l)
out[j][k][l] = (k < l)? in[j][k][l]: masked;
}
void
ircd::gpt::smax(float (&__restrict__ out)[12][1024][1024],
const float (&__restrict__ in)[12][1024][1024],
const size_t num)
{
static f64
tmp alignas(4096) [2][1024];
#pragma clang loop unroll (disable)
for(uint j(0); j < 12; ++j)
for(uint k(0); k < num; ++k)
math::smax<f32, f64>
(
out[j][k], { in[j][k], num }, tmp[0], tmp[1]
);
}
void
ircd::gpt::vals(float (&__restrict__ out)[12][1024][64],
const float (&__restrict__ in)[12][1024][1024],
const float (&__restrict__ qkv)[3][1024][12][64],
const size_t num)
{
const float
(&__restrict__ val)[1024][12][64] { qkv[2] };
#pragma clang loop unroll (disable)
for(uint j(0); j < 12; ++j)
for(uint k(0); k < num; ++k)
for(uint l(0); l < 64; ++l)
out[j][k][l] = 0;
#pragma clang loop unroll (disable)
for(uint j(0); j < 12; ++j)
for(uint k(0); k < num; ++k)
for(uint l(0); l < num; ++l)
for(uint m(0); m < 64; ++m)
out[j][k][m] += in[j][k][l] * val[l][j][m];
}
void
ircd::gpt::ffnn(float *const out,
const float *const in,
const model::decoder &decoder,
const uint laynum)
{
static const math::fmma_opts fmma3_opts
{
768, 3072, 2U,
};
static const math::fmma_opts fmma4_opts
{
3072, 768, 2U,
};
constexpr float ln2_epsilon
{
0.00001
};
const auto &layer
{
decoder.layer[laynum]
};
static float
buf alignas(64) [768],
buf2 alignas(64) [3072];
memset(buf2, 0x0, sizeof(buf2));
norm((f32x4 *)buf, (const f32x4 *)in, (const f32x4 *)layer.ffnn.norm.bias.elem, (const f32x4 *)layer.ffnn.norm.weight.elem, ln2_epsilon);
fmma((f32x4 *)buf2, (const f32x4 *)buf, (const f32x4 *)layer.ffnn.fcon_bias.fcon, (const f32x4 *)layer.ffnn.fcon_weight, fmma3_opts);
gelu((f32x4 *)buf2, (const f32x4 *)buf2);
fmma((f32x4 *)out, (const f32x4 *)buf2, (const f32x4 *)layer.ffnn.proj_bias.elem, (const f32x4 *)layer.ffnn.proj_weight, fmma4_opts);
}
void
ircd::gpt::norm(f32x4 *const __restrict__ out,
const f32x4 *const __restrict__ in,
const f32x4 *const __restrict__ bias,
const f32x4 *const __restrict__ weight,
const float epsilon)
{
static f64x4
tmp alignas(64) [768 / 4];
math::norm<f32x4, f64x4>
(
{out, 192}, {in, 192}, epsilon, tmp
);
for(uint j(0); j < 768 / 4; ++j)
out[j] = out[j] * weight[j] + bias[j];
}
template<class T>
void
ircd::gpt::fmma(T *const __restrict__ out,
const T *const __restrict__ in,
const T *const __restrict__ bias,
const T *const __restrict__ weight,
const math::fmma_opts &opts)
{
for(uint i(0); i < opts.rows / simd::lanes<T>(); ++i)
out[i] += bias[i];
math::fmma(out, in, weight, opts);
}
void
ircd::gpt::gelu(f32x4 *const out,
const f32x4 *const in)
{
for(uint j(0); j < 3072 / 4; ++j)
gelu(out[j], in[j]);
}
void
ircd::gpt::gelu(f32x4 &out,
const f32x4 &in)
{
out = 0.5 * in * (1.0 + tanh(in * f32(0.7978845608) * (1.0 + f32(0.044715) * in * in)));
}
//
// backside
//
[[gnu::noinline]]
void
ircd::gpt::backprop(const opts &opts,
const u32 step,
const f32 grad,
model::decoder &__restrict__ param,
f32 *const __restrict__ buf[2])
noexcept
{
model::decoder *const __restrict__ moment[2]
{
reinterpret_cast<model::decoder *>(buf[0]),
reinterpret_cast<model::decoder *>(buf[1]),
};
backprop(opts, step, grad, param, *moment[0], *moment[1]);
}
void
ircd::gpt::backprop(const opts &opts,
const u32 step,
const f32 grad,
model::decoder &__restrict__ param,
model::decoder &__restrict__ moment0,
model::decoder &__restrict__ moment1)
{
fpe::errors_handle eh;
assume(opts.attn_rank > 0);
assume(opts.layers > 0);
const auto eln
{
opts.layers - 1 // step % opts.layers
};
for(int i(opts.layers - 1); i >= int(opts.layers - eln - 1); --i)
{
assert(i >= 0 && i < int(opts.layers));
backprop(opts, step, grad, param.layer[i], moment0.layer[i], moment1.layer[i]);
}
backprop(opts, step, grad, param.embed, moment0.embed, moment1.embed);
auto pending(eh.pending());
eh.clear_pending();
pending &= ~pending & FE_INEXACT;
if(unlikely(pending))
fpe::throw_errors(pending);
}
void
ircd::gpt::backprop(const opts &opts,
const u32 step,
const f32 grad,
model::embed &__restrict__ param,
model::embed &__restrict__ moment0,
model::embed &__restrict__ moment1)
{
backprop(opts, step, grad, param.norm, moment0.norm, moment1.norm);
assume(opts.context_tokens > 0);
for(uint i(0); i < opts.context_tokens; ++i)
adamw(opts, step, grad, 768, param.pos[i].elem, moment0.pos[i].elem, moment1.pos[i].elem);
assume(opts.logits > 0);
for(uint i(0); i < opts.logits; ++i)
adamw(opts, step, grad, 768, param.token[i].elem, moment0.token[i].elem, moment1.token[i].elem);
}
void
ircd::gpt::backprop(const opts &opts,
const u32 step,
const f32 grad,
model::block &__restrict__ param,
model::block &__restrict__ moment0,
model::block &__restrict__ moment1)
{
backprop(opts, step, grad, param.attn.norm, moment0.attn.norm, moment1.attn.norm);
backprop(opts, step, grad, param.attn, moment0.attn, moment1.attn);
backprop(opts, step, grad, param.ffnn.norm, moment0.ffnn.norm, moment1.ffnn.norm);
backprop(opts, step, grad, param.ffnn, moment0.ffnn, moment1.ffnn);
}
void
ircd::gpt::backprop(const opts &opts,
const u32 step,
const f32 grad,
model::attn &__restrict__ param,
model::attn &__restrict__ moment0,
model::attn &__restrict__ moment1)
{
adamw(opts, step, grad, 2304, param.fcon_bias.fcon, moment0.fcon_bias.fcon, moment1.fcon_bias.fcon);
for(uint i(0); i < 768; ++i)
adamw(opts, step, grad, 2304, param.fcon_weight[i].fcon, moment0.fcon_weight[i].fcon, moment1.fcon_weight[i].fcon);
adamw(opts, step, grad, 768, param.proj_bias.elem, moment0.proj_bias.elem, moment1.proj_bias.elem);
for(uint i(0); i < 768; ++i)
adamw(opts, step, grad, 768, param.proj_weight[i].elem, moment0.proj_weight[i].elem, moment1.proj_weight[i].elem);
}
void
ircd::gpt::backprop(const opts &opts,
const u32 step,
const f32 grad,
model::ffnn &__restrict__ param,
model::ffnn &__restrict__ moment0,
model::ffnn &__restrict__ moment1)
{
adamw(opts, step, grad, 3072, param.fcon_bias.fcon, moment0.fcon_bias.fcon, moment1.fcon_bias.fcon);
for(uint i(0); i < 768; ++i)
adamw(opts, step, grad, 3072, param.fcon_weight[i].fcon, moment0.fcon_weight[i].fcon, moment1.fcon_weight[i].fcon);
adamw(opts, step, grad, 768, param.proj_bias.elem, moment0.proj_bias.elem, moment1.proj_bias.elem);
for(uint i(0); i < 3072; ++i)
adamw(opts, step, grad, 768, param.proj_weight[i].elem, moment0.proj_weight[i].elem, moment1.proj_weight[i].elem);
}
void
ircd::gpt::backprop(const opts &opts,
const u32 step,
const f32 grad,
model::norm &__restrict__ param,
model::norm &__restrict__ moment0,
model::norm &__restrict__ moment1)
{
adamw(opts, step, grad, 768, param.bias.elem, moment0.bias.elem, moment1.bias.elem);
adamw(opts, step, grad, 768, param.weight.elem, moment0.weight.elem, moment1.weight.elem);
}
namespace ircd::gpt
{
static f32x4 adamw_moment(const f32x4, const f32, const f32);
static f32x4 adamw_numer(const f32x4, const f32, const u32);
static f32x4 adamw_denom(const f32x4, const f32, const u32);
static f32x4 adamw_delta(const f32x4, const f32x4, const f32, const f32, const f32, const u32);
static void adamw(f32x4 &, f32x4 &, f32x4 &, const f32, const f32, const f32, const f32, const u32);
static void adamw(const opts &, const u32, const f32, const u32, f32 *, f32 *, f32 *);
}
void
ircd::gpt::adamw(const opts &opts,
const u32 step,
const f32 grad,
const u32 num,
f32 *const __restrict__ param,
f32 *const __restrict__ moment0,
f32 *const __restrict__ moment1)
{
f32x4 *const __restrict__ val[3]
{
reinterpret_cast<f32x4 *>(param),
reinterpret_cast<f32x4 *>(moment0),
reinterpret_cast<f32x4 *>(moment1),
};
const auto n
{
num / 4
};
assume(0 < n);
for(uint i(0); i < n; ++i)
adamw
(
val[0][i],
val[1][i],
val[2][i],
grad,
opts.alpha,
opts.beta[0],
opts.beta[1],
step + 1
);
}
void
ircd::gpt::adamw(f32x4 &__restrict__ param,
f32x4 &__restrict__ moment0,
f32x4 &__restrict__ moment1,
const f32 grad_,
const f32 alpha_,
const f32 beta0,
const f32 beta1,
const u32 step)
{
const f32 alpha
{
grad_ < 0? -alpha_ : alpha_
};
const f32 grad
{
grad_ < 0? -grad_ : grad_
};
const f32 grad_grad
{
grad * grad
};
const f32x4 moment[]
{
adamw_moment(moment0, grad, beta0),
adamw_moment(moment1, grad_grad, beta1)
};
const f32x4 delta
{
adamw_delta(moment[0], moment[1], alpha, beta0, beta1, step)
};
const f32x4 update
{
param - delta
};
if((false))
for(uint i(0); i < 4; ++i)
printf("%-15p p[%11.8lf] m[%11.8lf][%11.8lf] g[%11.8lf] d[%11.8lf] p[%11.8lf] m[%11.8lf][%11.8lf]\n",
((f32 *)&param) + i,
param[i],
moment0[i],
moment1[i],
grad,
delta[i],
update[i],
moment[0][i],
moment[1][i]);
assert(std::isnormal(update[0]));
assert(std::isnormal(update[1]));
assert(std::isnormal(update[2]));
assert(std::isnormal(update[3]));
assert(std::isnormal(moment[0][0]));
assert(std::isnormal(moment[0][1]));
assert(std::isnormal(moment[0][2]));
assert(std::isnormal(moment[0][3]));
assert(std::isnormal(moment[1][0]));
assert(std::isnormal(moment[1][1]));
assert(std::isnormal(moment[1][2]));
assert(std::isnormal(moment[1][3]));
param = update;
//__builtin_nontemporal_store(update, &param);
moment0 = moment[0];
moment1 = moment[1];
//__builtin_nontemporal_store(moment[0], &moment0);
//__builtin_nontemporal_store(moment[1], &moment1);
}
ircd::f32x4
ircd::gpt::adamw_delta(const f32x4 moment0,
const f32x4 moment1,
const f32 alpha,
const f32 beta0,
const f32 beta1,
const u32 step)
{
static const f32 epsilon
{
FLT_EPSILON
};
const f32x4 denom
{
adamw_denom(moment1, beta1, step) + epsilon
};
const f32x4 decay
{
adamw_numer(moment0, beta0, step)
};
const f32x4 smooth
{
alpha * decay
};
assert(std::isnormal(denom[0]));
assert(std::isnormal(denom[1]));
assert(std::isnormal(denom[2]));
assert(std::isnormal(denom[3]));
const f32x4 delta
{
smooth / denom
};
return delta;
}
ircd::f32x4
ircd::gpt::adamw_denom(const f32x4 moment,
const f32 beta,
const u32 step)
{
static const f32x4 one
{
1.0f, 1.0f, 1.0f, 1.0f,
};
assert(step > 0);
const f32x4 decay
{
one - powf(beta, step)
};
assert(std::isnormal(decay[0]));
assert(std::isnormal(decay[1]));
assert(std::isnormal(decay[2]));
assert(std::isnormal(decay[3]));
const f32x4 bias
{
moment / decay
};
const f32x4 denom
{
sqrtf(bias)
};
return denom;
}
ircd::f32x4
ircd::gpt::adamw_numer(const f32x4 moment,
const f32 beta,
const u32 step)
{
static const f32x4 one
{
1.0f, 1.0f, 1.0f, 1.0f,
};
assert(step > 0);
const f32x4 decay
{
one - powf(beta, step)
};
assert(std::isnormal(decay[0]));
assert(std::isnormal(decay[1]));
assert(std::isnormal(decay[2]));
assert(std::isnormal(decay[3]));
const f32x4 bias
{
moment / decay
};
return bias;
}
ircd::f32x4
ircd::gpt::adamw_moment(const f32x4 moment,
const f32 grad,
const f32 beta)
{
static const f32x4 one
{
1.0f, 1.0f, 1.0f, 1.0f,
};
const f32x4 rate
{
one - beta
};
const f32x4 avg
{
moment * beta
};
const f32x4 dot
{
rate * grad + avg
};
return dot;
}