File i16u64_to_f32_core.h¶

#ifndef DP_I16U64_TO_F32_CORE_H
#define DP_I16U64_TO_F32_CORE_H

#include "doppler/clib_common.h"
#include "doppler/jm_perf.h"
#ifdef __cplusplus
extern "C" {
#endif

typedef struct {
    float iscale; /* 1.0f / scale, pre-computed for single-multiply step */
} dp_i16u64_to_f32_state_t;

dp_i16u64_to_f32_state_t *dp_i16u64_to_f32_create(float scale);

void dp_i16u64_to_f32_destroy(dp_i16u64_to_f32_state_t *state);

void dp_i16u64_to_f32_reset(dp_i16u64_to_f32_state_t *state);

JM_FORCEINLINE JM_HOT float
dp_i16u64_to_f32_step(const dp_i16u64_to_f32_state_t *state, uint64_t x)
{
    /* Extract lower 16 bits as signed int16, then scale to float. */
    int16_t v = (int16_t)(uint16_t)(x & 0xFFFFull);
    return (float)v * state->iscale;
}

void dp_i16u64_to_f32_steps(
    dp_i16u64_to_f32_state_t *state,
    const uint64_t    *input,
    float          *output,
    size_t               n);





#ifdef __cplusplus
}
#endif

#endif /* I16U64_TO_F32_CORE_H */