tests/vmax_vv_test.cpp

#include <limits.h>
#include <riscv_vector.h>
#include <springbok.h>
#include <stdio.h>
#include <stdlib.h>

#include <bit>
#include <tuple>

#include "pw_unit_test/framework.h"
#include "test_v_helpers.h"

namespace vmax_vv_test {
namespace {

using namespace test_v_helpers;

uint8_t test_vector_1[MAXVL_BYTES];
uint8_t test_vector_2[MAXVL_BYTES];

template <typename T>
static std::tuple<int, int> vmax_vv_test_setup(VLMUL lmul, int32_t avl) {
  // Clear all vector registers
  zero_vector_registers();

  // Initialize test_vector_1 and determine vl, vlmax
  uint32_t bw = std::__bit_width(sizeof(T));
  VSEW sew = static_cast<VSEW>(bw - 1);
  int vlmax = get_vsetvlmax_intrinsic(sew, lmul);
  if (avl > vlmax) {
    avl = vlmax;
  }
  memset(test_vector_1, 0, MAXVL_BYTES);
  memset(test_vector_2, 0, MAXVL_BYTES);
  int vl = set_vsetvl_intrinsic(sew, lmul, avl);
  EXPECT_EQ(avl, vl);

  return std::make_tuple(vlmax, vl);
}

class VmaxVvTest : public ::testing::Test {
 protected:
  void SetUp() override { zero_vector_registers(); }
  void TearDown() override { zero_vector_registers(); }
};

// Below is a non-macro version of the test for more convenient debugging.
// Remove the "DISABLED_" prefix to enable this test for debugging.
TEST_F(VmaxVvTest, DISABLED_vmax_vv_demo) {
  for (int i = 0; i < AVL_COUNT; i++) {
    int32_t avl = AVLS[i];
    int vlmax;
    int vl;
    std::tie(vlmax, vl) = vmax_vv_test_setup<int16_t>(VLMUL::LMUL_M1, avl);
    if (avl > vlmax) {
      continue;
    }
    int16_t *ptr_vec_1 = reinterpret_cast<int16_t *>(test_vector_1);
    int16_t *ptr_vec_2 = reinterpret_cast<int16_t *>(test_vector_2);

    const int16_t start_value = INT16_MAX;

    // set up values for vs2 of vmax.vv
    for (int idx = 0; idx < vl; idx++) {
      // restrict values to valid int8_t range
      ptr_vec_1[idx] = (int16_t)(idx + start_value);
    }

    __asm__ volatile("vle16.v v8, (%0)" : : "r"(ptr_vec_1));

    // set up values for vs2 of vmax.vv
    for (int idx = 0; idx < vl; idx++) {
      // offset sequence for ptr_vec_2
      ptr_vec_2[idx] = (int16_t)(idx - start_value);
    }

    __asm__ volatile("vle16.v v16, (%0)" : : "r"(ptr_vec_2));

    __asm__ volatile("vmax.vv v24, v16, v8");

    // emulate operation in C
    for (int idx = 0; idx < vl; idx++) {
      ptr_vec_1[idx] =
          (ptr_vec_1[idx] > ptr_vec_2[idx]) ? ptr_vec_1[idx] : ptr_vec_2[idx];
    }

    __asm__ volatile("vse16.v v24, (%0)" : : "r"(ptr_vec_2));
    assert_vec_elem_eq<int16_t>(vlmax, test_vector_1, test_vector_2);
  }
}

#define DEFINE_TEST_VMAX_VV(_SEW_, _LMUL_, START_VALUE)                        \
  TEST_F(VmaxVvTest, vmax_vv##_SEW_##m##_LMUL_) {                              \
    for (int i = 0; i < AVL_COUNT; i++) {                                      \
      int32_t avl = AVLS[i];                                                   \
      int vlmax;                                                               \
      int vl;                                                                  \
      std::tie(vlmax, vl) =                                                    \
          vmax_vv_test_setup<int##_SEW_##_t>(VLMUL::LMUL_M##_LMUL_, avl);      \
      if (avl > vlmax) {                                                       \
        continue;                                                              \
      }                                                                        \
      int##_SEW_##_t *ptr_vec_1 =                                              \
          reinterpret_cast<int##_SEW_##_t *>(test_vector_1);                   \
      int##_SEW_##_t *ptr_vec_2 =                                              \
          reinterpret_cast<int##_SEW_##_t *>(test_vector_2);                   \
      const int##_SEW_##_t start_value = START_VALUE;                          \
      for (int idx = 0; idx < vl; idx++) {                                     \
        ptr_vec_1[idx] = (int##_SEW_##_t)(idx + start_value);                  \
      }                                                                        \
      for (int idx = 0; idx < vl; idx++) {                                     \
        ptr_vec_1[idx] = (int##_SEW_##_t)(idx - start_value);                  \
      }                                                                        \
      __asm__ volatile("vle" #_SEW_ ".v v8, (%0)" : : "r"(ptr_vec_1));         \
      __asm__ volatile("vle" #_SEW_ ".v v16, (%0)" : : "r"(ptr_vec_2));        \
      __asm__ volatile("vmax.vv v24, v16, v8");                                \
      for (int idx = 0; idx < vl; idx++) {                                     \
        ptr_vec_1[idx] = (ptr_vec_1[idx] > ptr_vec_2[idx]) ? ptr_vec_1[idx]    \
                                                           : ptr_vec_2[idx];   \
      }                                                                        \
      __asm__ volatile("vse" #_SEW_ ".v v24, (%0)" : : "r"(ptr_vec_2));        \
      assert_vec_elem_eq<int##_SEW_##_t>(vlmax, test_vector_1, test_vector_2); \
    }                                                                          \
  }

// TODO(gkielian): modify macro to permit more than one test per sew/lmul pair
DEFINE_TEST_VMAX_VV(8, 1, INT8_MAX)
DEFINE_TEST_VMAX_VV(8, 2, INT8_MIN)
DEFINE_TEST_VMAX_VV(8, 4, 120)
DEFINE_TEST_VMAX_VV(8, 8, 2)

DEFINE_TEST_VMAX_VV(16, 1, INT16_MAX)
DEFINE_TEST_VMAX_VV(16, 2, INT16_MIN)
DEFINE_TEST_VMAX_VV(16, 4, 16000)
DEFINE_TEST_VMAX_VV(16, 8, 2)

DEFINE_TEST_VMAX_VV(32, 1, INT32_MAX)
DEFINE_TEST_VMAX_VV(32, 2, INT32_MIN)
DEFINE_TEST_VMAX_VV(32, 4, 1000000000)
DEFINE_TEST_VMAX_VV(32, 8, 2)

}  // namespace
}  // namespace vmax_vv_test