Failing to build riscV/rvv

[irieger]

Hello,

I'm trying to build some code I written and successfully run on SSE4.2, AVX2, arm neon for riscV too. (riscv64 on Orange Pi RV2). So as a minimum viable example, I tried to start with the example from https://xsimd.readthedocs.io/en/latest/api/dispatching.html

In my actual code, I'd like to use the same basic principle. Currently, I have a dispatching file where I use an #if defined .../#elif ... etc. to define my architecture list based on what I build for. And then I have a simple impl file for each of the archs that I then call with the required arch information for the compiler as suggested in the example. Here I just copied the example from the doc into a subfolder of the xsimd checkout (tag for 14.0.0) with adaptation to target rvv.

Source code

So what I have now - just with a few attempts for some debugging output - I have based on the mentioned example.

sum.hpp:

#ifndef _SUM_HPP
#define _SUM_HPP
#include "xsimd/xsimd.hpp"
#include <iostream>

// functor with a call method that depends on `Arch`
struct sum
{
    // It's critical not to use an in-class definition here.
    // In-class and inline definition bypass extern template mechanism.
    template <class Arch, class T>
    T operator()(Arch, T const* data, unsigned size);
};

template <class Arch, class T>
T sum::operator()(Arch, T const* data, unsigned size)
{
    using batch = xsimd::batch<T, Arch>;

std::cout << "Vector size: " << batch::size << '\n';

    batch acc(static_cast<T>(0));
    const unsigned n = size / batch::size * batch::size;
    for (unsigned i = 0; i != n; i += batch::size)
        acc += batch::load_unaligned(data + i);
    //T star_acc = xsimd::reduce_add(acc);
    std::array<T, 128> out_values;
    acc.store_unaligned(out_values.data());
    T star_acc = 0;
    for (unsigned i = 0; i < batch::size; ++i)
        star_acc += out_values[i];
    for (unsigned i = n; i < size; ++i)
        star_acc += data[i];
    return star_acc;
}

// Inform the compiler that sse2 and avx2 implementation are to be found in another compilation unit.
extern template float sum::operator()<xsimd::rvv<128>, float>(xsimd::rvv<128>, float const*, unsigned);
//extern template float sum::operator()<xsimd::sse2, float>(xsimd::sse2, float const*, unsigned);
#endif

sum_rvv128.cpp (this I normally would compile as a seprate object, but to my understanding this shouldn't make a difference in this use case). For this I'd probably do a rvv128 and rvv256 if my assumption is correct and dispatch with priority for 256 and then 128.

#include "sum.hpp"
template float sum::operator()<xsimd::rvv<128>, float>(xsimd::rvv<128>, float const*, unsigned);

Is that understanding correct, that for riscv the length of the target vector is templated and needs to be the number of bits? I didn't see a clear example for it but that is what I condensed from looking at the includes.

sum.cpp (The dispatch, for simplicity with a main to quickly test once it is compiling):

#include "sum.hpp"

#include <iostream>
#include <vector>

// Create the dispatching function, specifying the architecture we want to
// target.
static auto dispatched = xsimd::dispatch<xsimd::arch_list<xsimd::rvv<128>>>(sum{});

// Call the appropriate implementation based on runtime information.
int main(int argc, char* argv[])
{
    std::vector<float> data;

    size_t num_entries = 17;
    for (size_t idx = 0; idx < num_entries; ++idx)
    {
        data.push_back(idx * 2.3456f);
    }

    float res = dispatched(data.data(), num_entries);
    std::cout << "Sum of vector\n";

    return 0;
}

Build attempt

Trying for simplicity to just compile it manually, normally I'd use cmake but wanted to keep it simple.

g++-14 -o sum.o -std=c++20 -I../include sum.cpp sum_rvv128.cpp -march=rv64gcv_zvl128b_zba_zbb_zbs -mrvv-vector-bits=zvl

The build error

Here some excerpt from the from the first errors provided, it goes on like this for a while as anyway usual for templates. This one is huge but if it helps I can provide the full log.

In file included from ../include/xsimd/memory/../config/../types/xsimd_all_registers.hpp:47,
                 from ../include/xsimd/memory/../config/xsimd_arch.hpp:19,
                 from ../include/xsimd/memory/xsimd_aligned_allocator.hpp:27,
                 from ../include/xsimd/xsimd.hpp:63,
                 from sum.hpp:3,
                 from sum.cpp:1:
../include/xsimd/memory/../config/../types/xsimd_rvv_register.hpp:238:27: error: explicit specialization in non-namespace scope ‘struct xsimd::types::detail::rvv_semiblob<T, divisor>’
  238 |                 template <>
      |                           ^
../include/xsimd/memory/../config/../types/xsimd_rvv_register.hpp:239:24: error: too few template-parameter-lists
  239 |                 struct semitype<2>
      |                        ^~~~~~~~~~~
../include/xsimd/memory/../config/../types/xsimd_rvv_register.hpp:243:27: error: explicit specialization in non-namespace scope ‘struct xsimd::types::detail::rvv_semiblob<T, divisor>’
  243 |                 template <>
      |                           ^
../include/xsimd/memory/../config/../types/xsimd_rvv_register.hpp:244:24: error: too few template-parameter-lists
  244 |                 struct semitype<4>
      |                        ^~~~~~~~~~~
../include/xsimd/memory/../config/../types/xsimd_rvv_register.hpp:248:27: error: explicit specialization in non-namespace scope ‘struct xsimd::types::detail::rvv_semiblob<T, divisor>’
  248 |                 template <>
      |                           ^
../include/xsimd/memory/../config/../types/xsimd_rvv_register.hpp:249:24: error: too few template-parameter-lists
  249 |                 struct semitype<8>
      |                        ^~~~~~~~~~~
../include/xsimd/memory/../config/../types/xsimd_rvv_register.hpp:260:27: error: explicit specialization in non-namespace scope ‘struct xsimd::types::detail::rvv_semiblob<T, divisor>’
  260 |                 template <>
      |                           ^
../include/xsimd/memory/../config/../types/xsimd_rvv_register.hpp:261:28: error: template-id ‘get_bytes<2>’ in declaration of primary template
  261 |                 vuint8m1_t get_bytes<2>() const { return __riscv_vlmul_ext_v_u8mf2_u8m1(value); }
      |                            ^~~~~~~~~~~~
../include/xsimd/memory/../config/../types/xsimd_rvv_register.hpp:262:27: error: explicit specialization in non-namespace scope ‘struct xsimd::types::detail::rvv_semiblob<T, divisor>’
  262 |                 template <>
../include/xsimd/memory/../config/../types/xsimd_rvv_register.hpp:263:28: error: template-id ‘get_bytes<4>’ in declaration of primary template
  263 |                 vuint8m1_t get_bytes<4>() const { return __riscv_vlmul_ext_v_u8mf4_u8m1(value); }
      |                            ^~~~~~~~~~~~
../include/xsimd/memory/../config/../types/xsimd_rvv_register.hpp:263:28: error: ‘vuint8m1_t xsimd::types::detail::rvv_semiblob<T, divisor>::get_bytes() const’ cannot be overloaded with ‘vuint8m1_t xsimd::types::detail::rvv_semiblob<T, divisor>::get_bytes() const’
../include/xsimd/memory/../config/../types/xsimd_rvv_register.hpp:261:28: note: previous declaration ‘vuint8m1_t xsimd::types::detail::rvv_semiblob<T, divisor>::get_bytes() const’
  261 |                 vuint8m1_t get_bytes<2>() const { return __riscv_vlmul_ext_v_u8mf2_u8m1(value); }
      |                            ^~~~~~~~~~~~
../include/xsimd/memory/../config/../types/xsimd_rvv_register.hpp:264:27: error: explicit specialization in non-namespace scope ‘struct xsimd::types::detail::rvv_semiblob<T, divisor>’
  264 |                 template <>
      |                           ^
../include/xsimd/memory/../config/../types/xsimd_rvv_register.hpp:265:28: error: template-id ‘get_bytes<8>’ in declaration of primary template
  265 |                 vuint8m1_t get_bytes<8>() const { return __riscv_vlmul_ext_v_u8mf8_u8m1(value); }
      |                            ^~~~~~~~~~~~
../include/xsimd/memory/../config/../types/xsimd_rvv_register.hpp:265:28: error: ‘vuint8m1_t xsimd::types::detail::rvv_semiblob<T, divisor>::get_bytes() const’ cannot be overloaded with ‘vuint8m1_t xsimd::types::detail::rvv_semiblob<T, divisor>::get_bytes() const’
../include/xsimd/memory/../config/../types/xsimd_rvv_register.hpp:261:28: note: previous declaration ‘vuint8m1_t xsimd::types::detail::rvv_semiblob<T, divisor>::get_bytes() const’
  261 |                 vuint8m1_t get_bytes<2>() const { return __riscv_vlmul_ext_v_u8mf2_u8m1(value); }
      |                            ^~~~~~~~~~~~
../include/xsimd/memory/../config/../types/xsimd_rvv_register.hpp:273:27: error: explicit specialization in non-namespace scope ‘struct xsimd::types::detail::rvv_semiblob<T, divisor>’
  273 |                 template <>
      |                           ^
../include/xsimd/memory/../config/../types/xsimd_rvv_register.hpp:274:22: error: template-id ‘set_bytes<2>’ in declaration of primary template
  274 |                 void set_bytes<2>(vuint8m1_t v) { value = __riscv_vlmul_trunc_v_u8m1_u8mf2(v); }
      |                      ^~~~~~~~~~~~
../include/xsimd/memory/../config/../types/xsimd_rvv_register.hpp:275:27: error: explicit specialization in non-namespace scope ‘struct xsimd::types::detail::rvv_semiblob<T, divisor>’
  275 |                 template <>
      |                           ^
../include/xsimd/memory/../config/../types/xsimd_rvv_register.hpp:276:22: error: template-id ‘set_bytes<4>’ in declaration of primary template
  276 |                 void set_bytes<4>(vuint8m1_t v) { value = __riscv_vlmul_trunc_v_u8m1_u8mf4(v); }
      |                      ^~~~~~~~~~~~
../include/xsimd/memory/../config/../types/xsimd_rvv_register.hpp:276:22: error: ‘void xsimd::types::detail::rvv_semiblob<T, divisor>::set_bytes(vuint8m1_t)’ cannot be overloaded with ‘void xsimd::types::detail::rvv_semiblob<T, divisor>::set_bytes(vuint8m1_t)’
../include/xsimd/memory/../config/../types/xsimd_rvv_register.hpp:274:22: note: previous declaration ‘void xsimd::types::detail::rvv_semiblob<T, divisor>::set_bytes(vuint8m1_t)’
  274 |                 void set_bytes<2>(vuint8m1_t v) { value = __riscv_vlmul_trunc_v_u8m1_u8mf2(v); }
      |                      ^~~~~~~~~~~~
../include/xsimd/memory/../config/../types/xsimd_rvv_register.hpp:277:27: error: explicit specialization in non-namespace scope ‘struct xsimd::types::detail::rvv_semiblob<T, divisor>’
  277 |                 template <>
      |                           ^
../include/xsimd/memory/../config/../types/xsimd_rvv_register.hpp:278:22: error: template-id ‘set_bytes<8>’ in declaration of primary template
  278 |                 void set_bytes<8>(vuint8m1_t v) { value = __riscv_vlmul_trunc_v_u8m1_u8mf8(v); }
      |                      ^~~~~~~~~~~~
../include/xsimd/memory/../config/../types/xsimd_rvv_register.hpp:278:22: error: ‘void xsimd::types::detail::rvv_semiblob<T, divisor>::set_bytes(vuint8m1_t)’ cannot be overloaded with ‘void xsimd::types::detail::rvv_semiblob<T, divisor>::set_bytes(vuint8m1_t)’
../include/xsimd/memory/../config/../types/xsimd_rvv_register.hpp:274:22: note: previous declaration ‘void xsimd::types::detail::rvv_semiblob<T, divisor>::set_bytes(vuint8m1_t)’
  274 |                 void set_bytes<2>(vuint8m1_t v) { value = __riscv_vlmul_trunc_v_u8m1_u8mf2(v); }
      |                      ^~~~~~~~~~~~
In file included from ../include/xsimd/types/../arch/xsimd_isa.hpp:128,
                 from ../include/xsimd/types/xsimd_batch.hpp:520,
                 from ../include/xsimd/xsimd.hpp:77:
../include/xsimd/types/../arch/./xsimd_rvv.hpp:157:15: warning: declaration of ‘using xsimd::kernel::detail::rvvid_cruft::impl<signed char, Ret(First, Args ...)>::ctx = struct xsimd::kernel::detail::rvvid_cruft::ctx<signed char>’ changes meaning of ‘ctx’ [-Wchanges-meaning]
  157 |         using ctx = ctx<KEY>;                                        \
      |               ^~~
../include/xsimd/types/../arch/./xsimd_rvv.hpp:221:5: note: in expansion of macro ‘XSIMD_RVV_WRAPPER_DROP_1ST’
  221 |     XSIMD_RVV_WRAPPER##variant(int8_t, XSIMD_RVV_IDENTIFIER(i, 8, name), __VA_ARGS__)   \
      |     ^~~~~~~~~~~~~~~~~
../include/xsimd/types/../arch/./xsimd_rvv.hpp:260:5: note: in expansion of macro ‘XSIMD_RVV_OVERLOAD_i’
  260 |     XSIMD_RVV_OVERLOAD_i(name_i, variant, __VA_ARGS__)                     \
      |     ^~~~~~~~~~~~~~~~~~~~
../include/xsimd/types/../arch/./xsimd_rvv.hpp:265:48: note: in expansion of macro ‘XSIMD_RVV_OVERLOAD3’

So what do I need to do for RVV targets? What am I getting wrong? Do I need to provide something special here?

[AntoinePrv]

I think you may want to compile with -mrvv-vector-bits=128 instead.
Unfortunately, xsimd currently has no support for dynamically sized register.

Also, as of 14.0.0, dynamic dispatch will never detect RVV (fixed in #1291)