Skip to content

Commit

Permalink
Merge pull request #457 from howjmay/vmlal_high_laneq
Browse files Browse the repository at this point in the history 
feat: Add vmlal_high_laneq_[s16|s32|u16|u32]
  • Loading branch information
@howjmay
howjmay authored Jul 25, 2024
2 parents 7b8df8d + 38cc54b commit b5f1a82
Show file tree
Hide file tree
Showing 3 changed files with 132 additions and 12 deletions.
28 changes: 24 additions & 4 deletions neon2rvv.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -8965,13 +8965,33 @@ FORCE_INLINE uint64x2_t vmlal_laneq_u32(uint64x2_t a, uint32x2_t b, uint32x4_t c
return __riscv_vlmul_trunc_v_u64m2_u64m1(__riscv_vwmaccu_vv_u64m2(__riscv_vlmul_ext_v_u64m1_u64m2(a), b, c_dup, 2));
}

// FORCE_INLINE int32x4_t vmlal_high_laneq_s16(int32x4_t a, int16x8_t b, int16x8_t v, const int lane);
FORCE_INLINE int32x4_t vmlal_high_laneq_s16(int32x4_t a, int16x8_t b, int16x8_t c, const int lane) {
vint16m1_t b_high = __riscv_vslidedown_vx_i16m1(b, 4, 8);
vint16m1_t c_dup = __riscv_vrgather_vx_i16m1(c, lane, 4);
return __riscv_vlmul_trunc_v_i32m2_i32m1(
__riscv_vwmacc_vv_i32m2(__riscv_vlmul_ext_v_i32m1_i32m2(a), b_high, c_dup, 4));
}

// FORCE_INLINE int64x2_t vmlal_high_laneq_s32(int64x2_t a, int32x4_t b, int32x4_t v, const int lane);
FORCE_INLINE int64x2_t vmlal_high_laneq_s32(int64x2_t a, int32x4_t b, int32x4_t c, const int lane) {
vint32m1_t b_high = __riscv_vslidedown_vx_i32m1(b, 2, 4);
vint32m1_t c_dup = __riscv_vrgather_vx_i32m1(c, lane, 2);
return __riscv_vlmul_trunc_v_i64m2_i64m1(
__riscv_vwmacc_vv_i64m2(__riscv_vlmul_ext_v_i64m1_i64m2(a), b_high, c_dup, 2));
}

// FORCE_INLINE uint32x4_t vmlal_high_laneq_u16(uint32x4_t a, uint16x8_t b, uint16x8_t v, const int lane);
FORCE_INLINE uint32x4_t vmlal_high_laneq_u16(uint32x4_t a, uint16x8_t b, uint16x8_t c, const int lane) {
vuint16m1_t b_high = __riscv_vslidedown_vx_u16m1(b, 4, 8);
vuint16m1_t c_dup = __riscv_vrgather_vx_u16m1(c, lane, 4);
return __riscv_vlmul_trunc_v_u32m2_u32m1(
__riscv_vwmaccu_vv_u32m2(__riscv_vlmul_ext_v_u32m1_u32m2(a), b_high, c_dup, 4));
}

// FORCE_INLINE uint64x2_t vmlal_high_laneq_u32(uint64x2_t a, uint32x4_t b, uint32x4_t v, const int lane);
FORCE_INLINE uint64x2_t vmlal_high_laneq_u32(uint64x2_t a, uint32x4_t b, uint32x4_t c, const int lane) {
vuint32m1_t b_high = __riscv_vslidedown_vx_u32m1(b, 2, 4);
vuint32m1_t c_dup = __riscv_vrgather_vx_u32m1(c, lane, 2);
return __riscv_vlmul_trunc_v_u64m2_u64m1(
__riscv_vwmaccu_vv_u64m2(__riscv_vlmul_ext_v_u64m1_u64m2(a), b_high, c_dup, 2));
}

FORCE_INLINE int32x4_t vqdmlal_lane_s16(int32x4_t a, int16x4_t b, int16x4_t c, const int __d) {
vint16m1_t c_dup = __riscv_vrgather_vx_i16m1(c, __d, 4);
Expand Down
108 changes: 104 additions & 4 deletions tests/impl.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -31723,13 +31723,113 @@ result_t test_vmlal_laneq_u32(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) {
#endif // ENABLE_TEST_ALL
}

result_t test_vmlal_high_laneq_s16(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) { return TEST_UNIMPL; }
result_t test_vmlal_high_laneq_s16(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) {
#ifdef ENABLE_TEST_ALL
const int32_t *_a = (int32_t *)impl.test_cases_int_pointer1;
const int16_t *_b = (int16_t *)impl.test_cases_int_pointer2;
const int16_t *_c = (int16_t *)impl.test_cases_int_pointer3;
int32_t _d[8];
int32x4_t a = vld1q_s32(_a);
int16x8_t b = vld1q_s16(_b);
int16x8_t c = vld1q_s16(_c);
int32x4_t d;

#define TEST_IMPL(IDX) \
for (int i = 0; i < 4; i++) { \
_d[i] = _a[i] + (int32_t)_b[i + 4] * (int32_t)_c[IDX]; \
} \
d = vmlal_high_laneq_s16(a, b, c, IDX); \
CHECK_RESULT(validate_int32(d, _d[0], _d[1], _d[2], _d[3]))

IMM_4_ITER
#undef TEST_IMPL

return TEST_SUCCESS;
#else
return TEST_UNIMPL;
#endif // ENABLE_TEST_ALL
}

result_t test_vmlal_high_laneq_s32(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) {
#ifdef ENABLE_TEST_ALL
const int64_t *_a = (int64_t *)impl.test_cases_int_pointer1;
const int32_t *_b = (int32_t *)impl.test_cases_int_pointer2;
const int32_t *_c = (int32_t *)impl.test_cases_int_pointer3;
int64_t _d[4];
int64x2_t a = vld1q_s64(_a);
int32x4_t b = vld1q_s32(_b);
int32x4_t c = vld1q_s32(_c);
int64x2_t d;

#define TEST_IMPL(IDX) \
for (int i = 0; i < 2; i++) { \
_d[i] = _a[i] + (int64_t)_b[i + 2] * (int64_t)_c[IDX]; \
} \
d = vmlal_high_laneq_s32(a, b, c, IDX); \
CHECK_RESULT(validate_int64(d, _d[0], _d[1]))

IMM_2_ITER
#undef TEST_IMPL

return TEST_SUCCESS;
#else
return TEST_UNIMPL;
#endif // ENABLE_TEST_ALL
}

result_t test_vmlal_high_laneq_u16(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) {
#ifdef ENABLE_TEST_ALL
const uint32_t *_a = (uint32_t *)impl.test_cases_int_pointer1;
const uint16_t *_b = (uint16_t *)impl.test_cases_int_pointer2;
const uint16_t *_c = (uint16_t *)impl.test_cases_int_pointer3;
uint32_t _d[8];
uint32x4_t a = vld1q_u32(_a);
uint16x8_t b = vld1q_u16(_b);
uint16x8_t c = vld1q_u16(_c);
uint32x4_t d;

#define TEST_IMPL(IDX) \
for (int i = 0; i < 4; i++) { \
_d[i] = _a[i] + (uint32_t)_b[i + 4] * (uint32_t)_c[IDX]; \
} \
d = vmlal_high_laneq_u16(a, b, c, IDX); \
CHECK_RESULT(validate_uint32(d, _d[0], _d[1], _d[2], _d[3]))

IMM_4_ITER
#undef TEST_IMPL

return TEST_SUCCESS;
#else
return TEST_UNIMPL;
#endif // ENABLE_TEST_ALL
}

result_t test_vmlal_high_laneq_u32(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) {
#ifdef ENABLE_TEST_ALL
const uint64_t *_a = (uint64_t *)impl.test_cases_int_pointer1;
const uint32_t *_b = (uint32_t *)impl.test_cases_int_pointer2;
const uint32_t *_c = (uint32_t *)impl.test_cases_int_pointer3;
uint64_t _d[4];
uint64x2_t a = vld1q_u64(_a);
uint32x4_t b = vld1q_u32(_b);
uint32x4_t c = vld1q_u32(_c);
uint64x2_t d;

result_t test_vmlal_high_laneq_s32(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) { return TEST_UNIMPL; }
#define TEST_IMPL(IDX) \
for (int i = 0; i < 2; i++) { \
_d[i] = _a[i] + (uint64_t)_b[i + 2] * (uint64_t)_c[IDX]; \
} \
d = vmlal_high_laneq_u32(a, b, c, IDX); \
CHECK_RESULT(validate_uint64(d, _d[0], _d[1]))

result_t test_vmlal_high_laneq_u16(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) { return TEST_UNIMPL; }
IMM_2_ITER
#undef TEST_IMPL

result_t test_vmlal_high_laneq_u32(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) { return TEST_UNIMPL; }
return TEST_SUCCESS;
#else
return TEST_UNIMPL;
#endif // ENABLE_TEST_ALL
}

result_t test_vqdmlal_lane_s16(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) {
#ifdef ENABLE_TEST_ALL
Expand Down
8 changes: 4 additions & 4 deletions tests/impl.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1942,10 +1942,10 @@
_(vmlal_laneq_s32) \
_(vmlal_laneq_u16) \
_(vmlal_laneq_u32) \
/*_(vmlal_high_laneq_s16) */ \
/*_(vmlal_high_laneq_s32) */ \
/*_(vmlal_high_laneq_u16) */ \
/*_(vmlal_high_laneq_u32) */ \
_(vmlal_high_laneq_s16) \
_(vmlal_high_laneq_s32) \
_(vmlal_high_laneq_u16) \
_(vmlal_high_laneq_u32) \
_(vqdmlal_lane_s16) \
_(vqdmlal_lane_s32) \
/*_(vqdmlalh_lane_s16) */ \
Expand Down

0 comments on commit b5f1a82

Please sign in to comment.