luo jiyin
9a22905fb7
Reuse each widened weight chunk across up to four activation rows. Use a scalar path when the CPU lacks AVX2 or FMA. The blocked kernel reduces scoring time for both locked models. Keep a focused test for row blocks, input tails, and optional bias.
71 lines
2.8 KiB
C++
71 lines
2.8 KiB
C++
#include "neural/int8_linear_avx2.h"
|
|
|
|
#include <algorithm>
|
|
#include <cmath>
|
|
#include <cstdint>
|
|
#include <cstdio>
|
|
#include <vector>
|
|
|
|
int main()
|
|
{
|
|
if (!neural::detail::supports_avx2_fma())
|
|
{
|
|
std::puts("Skipped: AVX2 and FMA are not available.");
|
|
return 0;
|
|
}
|
|
|
|
for (std::size_t rows : {1u, 2u, 3u, 4u, 5u, 8u, 9u, 16u})
|
|
{
|
|
for (std::size_t inputs : {1u, 7u, 8u, 9u, 15u, 16u, 17u, 192u, 448u})
|
|
{
|
|
for (std::size_t outputs : {1u, 3u, 11u})
|
|
{
|
|
std::vector<float> x(rows * inputs);
|
|
std::vector<std::int8_t> weights(inputs * outputs);
|
|
std::vector<float> scales(outputs);
|
|
std::vector<float> bias(outputs);
|
|
std::vector<float> actual(rows * outputs);
|
|
for (std::size_t i = 0; i < x.size(); ++i)
|
|
{
|
|
x[i] = std::sin(static_cast<float>(i) * 0.071f);
|
|
}
|
|
for (std::size_t i = 0; i < weights.size(); ++i)
|
|
{
|
|
weights[i] = static_cast<std::int8_t>(static_cast<int>(i * 17u % 255u) - 127);
|
|
}
|
|
for (std::size_t i = 0; i < outputs; ++i)
|
|
{
|
|
scales[i] = 0.002f + static_cast<float>(i % 9) * 0.001f;
|
|
bias[i] = std::sin(static_cast<float>(i) * 0.03f);
|
|
}
|
|
|
|
for (bool with_bias : {false, true})
|
|
{
|
|
neural::detail::linear_int8_avx2(x.data(), weights.data(), scales.data(),
|
|
with_bias ? bias.data() : nullptr, rows, inputs, outputs,
|
|
actual.data());
|
|
for (std::size_t row = 0; row < rows; ++row)
|
|
{
|
|
for (std::size_t output = 0; output < outputs; ++output)
|
|
{
|
|
float total = 0.0f;
|
|
for (std::size_t input = 0; input < inputs; ++input)
|
|
{
|
|
total += x[row * inputs + input] * static_cast<float>(weights[output * inputs + input]);
|
|
}
|
|
const float expected = total * scales[output] + (with_bias ? bias[output] : 0.0f);
|
|
if (!(std::fabs(actual[row * outputs + output] - expected) <= 1e-4f))
|
|
{
|
|
std::fprintf(stderr, "int8 linear mismatch: rows=%zu inputs=%zu outputs=%zu\n", rows,
|
|
inputs, outputs);
|
|
return 1;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
return 0;
|
|
}
|