whispercpp 1.3.5 → 1.3.6

data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-utils.c

@@ -1,1020 +0,0 @@
-#pragma clang diagnostic ignored "-Wunused-variable"
-#pragma clang diagnostic ignored "-Wunused-function"
-#pragma clang diagnostic ignored "-Wunused-but-set-variable"
-
-#ifdef HTP_DEBUG
-#    define FARF_HIGH 1
-#endif
-
-#include <HAP_farf.h>
-#include <HAP_mem.h>
-#include <HAP_perf.h>
-#include <HAP_ps.h>
-#include <hexagon_protos.h>
-#include <hexagon_types.h>
-#include <math.h>
-#include <string.h>
-
-#define GGML_COMMON_DECL_C
-#include "ggml-common.h"
-#include "hvx-utils.h"
-
-#define htp_binary_ops_preamble                                                                                \
-    int step_of_4 = num_elems >> 7;                                                                            \
-    int step_of_2 = (num_elems - step_of_4 * VLEN_FP32 * 4) >> 6;                                              \
-    int step_of_1 = (num_elems - step_of_4 * VLEN_FP32 * 4 - step_of_2 * VLEN_FP32 * 2) >> 5;                  \
-    int remaining = num_elems - step_of_4 * VLEN_FP32 * 4 - step_of_2 * VLEN_FP32 * 2 - step_of_1 * VLEN_FP32; \
-                                                                                                               \
-    const uint8_t * restrict src0_curr = src0;                                                                 \
-    const uint8_t * restrict src1_curr = src1;                                                                 \
-    uint8_t * restrict dst_curr        = dst;
-
-void hvx_mul_f32(const uint8_t * restrict src0,
-                 const uint8_t * restrict src1,
-                 uint8_t * restrict dst,
-                 const int num_elems) {
-    int left_over       = num_elems & (VLEN_FP32 - 1);
-    int num_elems_whole = num_elems - left_over;
-
-    int unaligned_addr = 0;
-    int unaligned_loop = 0;
-    if ((0 == htp_is_aligned((void *) src0, VLEN)) || (0 == htp_is_aligned((void *) src1, VLEN)) ||
-        (0 == htp_is_aligned((void *) dst, VLEN))) {
-        FARF(HIGH, "hvx_mul_f32: unaligned address in hvx op, possibly slower execution\n");
-        unaligned_addr = 1;
-    }
-
-    if ((1 == unaligned_addr) && (num_elems_whole != 0)) {
-        unaligned_loop = 1;
-        FARF(HIGH, "hvx_mul_f32: unaligned loop in hvx op, possibly slower execution\n");
-    }
-
-
-    bool handled_leftover = false;
-    if (0 == unaligned_loop) {
-        HVX_Vector * restrict vec_in1 = (HVX_Vector *) src0;
-        HVX_Vector * restrict vec_in2 = (HVX_Vector *) src1;
-        HVX_Vector * restrict vec_out = (HVX_Vector *) dst;
-
-        #pragma unroll(4)
-        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
-            HVX_Vector v = Q6_Vqf32_vmpy_VsfVsf(*vec_in1++, *vec_in2++);
-            *vec_out++   = Q6_Vsf_equals_Vqf32(v);
-        }
-    } else {
-        int step_of_1 = num_elems_whole >> 5;  // divby 32, because 32 float = 128 bytes per HVX vector
-        int leftover_size = left_over * sizeof(float);
-
-
-        HVX_Vector * restrict vec_in1 = (HVX_Vector *) src0;
-        HVX_Vector * restrict vec_in2 = (HVX_Vector *) src1;
-        HVX_UVector * restrict vec_out = (HVX_UVector *) dst;
-
-        HVX_Vector slinep;
-        HVX_Vector slinec;
-        HVX_Vector sline;
-        HVX_Vector sline2p;
-        HVX_Vector sline2c;
-        HVX_Vector sline2;
-
-        slinep  = *vec_in1++;
-        sline2p = *vec_in2++;
-        #pragma unroll(4)
-        for (int i = step_of_1 - 1; i > 0; i--) {
-            slinec  = *vec_in1++;
-            sline2c = *vec_in2++;
-            sline   = Q6_V_valign_VVR(slinec, slinep, (size_t) src0);
-            sline2  = Q6_V_valign_VVR(sline2c, sline2p, (size_t) src1);
-
-            *((HVX_UVector *) (vec_out++)) = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(sline, sline2));
-            slinep                         = slinec;
-            sline2p                        = sline2c;
-        }
-        if (step_of_1 > 1) {
-            slinec  = htp_is_aligned(vec_in1, VLEN) && left_over == 0 ? slinep : *vec_in1++;
-            sline2c = htp_is_aligned(vec_in2, VLEN) && left_over == 0 ? sline2p : *vec_in2++;
-
-            sline                          = Q6_V_valign_VVR(slinec, slinep, (size_t) src0);
-            sline2                         = Q6_V_valign_VVR(sline2c, sline2p, (size_t) src1);
-            *((HVX_UVector *) (vec_out++)) = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(sline, sline2));
-            slinep                         = slinec;
-            sline2p                        = sline2c;
-        }
-        if (left_over > 0) {
-            slinec = (is_in_one_chunk(vec_in1, leftover_size, VLEN) ? slinep : *vec_in1++);
-
-            sline   = Q6_V_valign_VVR(slinec, slinep, (size_t) src0);
-            sline2c = (is_in_one_chunk(vec_in2, leftover_size, VLEN) ? sline2p : *vec_in2++);
-            sline2  = Q6_V_valign_VVR(sline2c, sline2p, (size_t) src1);
-
-            HVX_Vector out = Q6_Vqf32_vmpy_VsfVsf(sline, sline2);
-            hvx_vec_store_u(vec_out, leftover_size, Q6_Vsf_equals_Vqf32(out));
-            handled_leftover = true;
-        }
-    }
-
-
-    if (left_over > 0 && !handled_leftover) {
-        const float * src0f = (const float *) src0 + num_elems_whole;
-        const float * src1f = (const float *) src1 + num_elems_whole;
-        float *       dstf  = (float *) dst + num_elems_whole;
-
-        HVX_Vector in1 = *(HVX_UVector *) src0f;
-        HVX_Vector in2 = *(HVX_UVector *) src1f;
-
-        HVX_Vector out = Q6_Vqf32_vmpy_VsfVsf(in1, in2);
-        hvx_vec_store_u((void *) dstf, left_over * SIZEOF_FP32, Q6_Vsf_equals_Vqf32(out));
-    }
-}
-
-void hvx_mul_f32_opt(const uint8_t * restrict src0,
-                     const uint8_t * restrict src1,
-                     uint8_t * restrict dst,
-                     const int num_elems) {
-    htp_binary_ops_preamble;
-
-    for (int i = 0; i < step_of_4; i++) {
-        HVX_Vector v1a = *(HVX_Vector *) src0_curr;
-
-        HVX_Vector v1b = *(HVX_Vector *) src1_curr;
-
-        HVX_Vector v2a = *(HVX_Vector *) (src0_curr + VLEN);
-
-        HVX_Vector v1 = Q6_Vqf32_vmpy_VsfVsf(v1a, v1b);
-
-        HVX_Vector v2b = *(HVX_Vector *) (src1_curr + VLEN);
-
-        HVX_Vector v3a = *(HVX_Vector *) (src0_curr + 2 * VLEN);
-
-        HVX_Vector v2 = Q6_Vqf32_vmpy_VsfVsf(v2a, v2b);
-
-        *(HVX_Vector *) dst_curr = Q6_Vsf_equals_Vqf32(v1);
-
-        HVX_Vector v3b = *(HVX_Vector *) (src1_curr + 2 * VLEN);
-
-        HVX_Vector v4a = *(HVX_Vector *) (src0_curr + 3 * VLEN);
-
-        src0_curr += 4 * VLEN;
-
-        HVX_Vector v3 = Q6_Vqf32_vmpy_VsfVsf(v3a, v3b);
-
-        *(HVX_Vector *) (dst_curr + VLEN) = Q6_Vsf_equals_Vqf32(v2);
-
-        HVX_Vector v4b = *(HVX_Vector *) (src1_curr + 3 * VLEN);
-
-        *(HVX_Vector *) (dst_curr + 2 * VLEN) = Q6_Vsf_equals_Vqf32(v3);
-
-        HVX_Vector v4 = Q6_Vqf32_vmpy_VsfVsf(v4a, v4b);
-
-        src1_curr += 4 * VLEN;
-
-        *(HVX_Vector *) (dst_curr + 3 * VLEN) = Q6_Vsf_equals_Vqf32(v4);
-
-        dst_curr += 4 * VLEN;
-    }
-
-    for (int i = 0; i < step_of_2; i++) {
-        HVX_Vector v1a = *(HVX_Vector *) src0_curr;
-
-        HVX_Vector v1b = *(HVX_Vector *) src1_curr;
-
-        HVX_Vector v2a = *(HVX_Vector *) (src0_curr + VLEN);
-
-        HVX_Vector v1 = Q6_Vqf32_vmpy_VsfVsf(v1a, v1b);
-
-        HVX_Vector v2b = *(HVX_Vector *) (src1_curr + VLEN);
-
-        *(HVX_Vector *) dst_curr = Q6_Vsf_equals_Vqf32(v1);
-
-        src0_curr += 2 * VLEN;
-
-        HVX_Vector v2 = Q6_Vqf32_vmpy_VsfVsf(v2a, v2b);
-
-        src1_curr += 2 * VLEN;
-
-        *(HVX_Vector *) (dst_curr + VLEN) = Q6_Vsf_equals_Vqf32(v2);
-
-        dst_curr += 2 * VLEN;
-    }
-
-    for (int i = 0; i < step_of_1; i++) {
-        HVX_Vector va = *(HVX_Vector *) src0_curr;
-
-        src0_curr += VLEN;
-
-        HVX_Vector vb = *(HVX_Vector *) src1_curr;
-
-        src1_curr += VLEN;
-
-        HVX_Vector v = Q6_Vqf32_vmpy_VsfVsf(va, vb);
-
-        *(HVX_Vector *) dst_curr = Q6_Vsf_equals_Vqf32(v);
-
-        dst_curr += VLEN;
-    }
-
-    if (remaining > 0) {
-        HVX_Vector v = Q6_Vqf32_vmpy_VsfVsf(*(HVX_Vector *) src0_curr, *(HVX_Vector *) src1_curr);
-        hvx_vec_store_u((void *) dst_curr, remaining * SIZEOF_FP32, Q6_Vsf_equals_Vqf32(v));
-    }
-}
-
-void hvx_mul_mul_f32_opt(const uint8_t * restrict src0,
-                         const uint8_t * restrict src1,
-                         const uint8_t * restrict src2,
-                         uint8_t * restrict dst,
-                         const int num_elems) {
-    const uint8_t * restrict src0_curr = src0;
-    const uint8_t * restrict src1_curr = src1;
-    const uint8_t * restrict src2_curr = src2;
-    uint8_t * restrict dst_curr        = dst;
-
-    int step_of_2 = num_elems >> 6;
-    int step_of_1 = (num_elems - step_of_2 * VLEN_FP32 * 2) >> 5;
-    int remaining = num_elems - step_of_2 * VLEN_FP32 * 2 - step_of_1 * VLEN_FP32;
-
-    for (int i = 0; i < step_of_2; i++) {
-        HVX_Vector v1a = *(HVX_Vector *) src0_curr;
-        HVX_Vector v1b = *(HVX_Vector *) src1_curr;
-        HVX_Vector v1c = *(HVX_Vector *) src2_curr;
-
-        HVX_Vector v2a = *(HVX_Vector *) (src0_curr + VLEN);
-
-        HVX_Vector v1_ = Q6_Vqf32_vmpy_VsfVsf(v1a, v1b);
-        HVX_Vector v1  = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(v1_), v1c);
-
-        HVX_Vector v2b = *(HVX_Vector *) (src1_curr + VLEN);
-
-        *(HVX_Vector *) dst_curr = Q6_Vsf_equals_Vqf32(v1);
-
-        HVX_Vector v2c = *(HVX_Vector *) (src2_curr + VLEN);
-
-        src0_curr += 2 * VLEN;
-
-        HVX_Vector v2_ = Q6_Vqf32_vmpy_VsfVsf(v2a, v2b);
-        HVX_Vector v2  = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(v2_), v2c);
-
-        src1_curr += 2 * VLEN;
-        src2_curr += 2 * VLEN;
-
-        *(HVX_Vector *) (dst_curr + VLEN) = Q6_Vsf_equals_Vqf32(v2);
-
-        dst_curr += 2 * VLEN;
-    }
-    for (int i = 0; i < step_of_1; i++) {
-        HVX_Vector va = *(HVX_Vector *) src0_curr;
-        src0_curr += VLEN;
-
-        HVX_Vector vb = *(HVX_Vector *) src1_curr;
-        src1_curr += VLEN;
-
-        HVX_Vector vc = *(HVX_Vector *) src2_curr;
-        src2_curr += VLEN;
-
-        HVX_Vector v1 = Q6_Vqf32_vmpy_VsfVsf(va, vb);
-        HVX_Vector v2 = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(v1), vc);
-
-        *(HVX_Vector *) dst_curr = Q6_Vsf_equals_Vqf32(v2);
-        dst_curr += VLEN;
-    }
-    if (remaining > 0) {
-        HVX_Vector v1 = Q6_Vqf32_vmpy_VsfVsf(*(HVX_Vector *) src0_curr, *(HVX_Vector *) src1_curr);
-        HVX_Vector v2 = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(v1), *(HVX_Vector *) src2_curr);
-        hvx_vec_store_u((void *) dst_curr, remaining * SIZEOF_FP32, Q6_Vsf_equals_Vqf32(v2));
-    }
-}
-
-void hvx_add_f32(const uint8_t * restrict src0,
-                 const uint8_t * restrict src1,
-                 uint8_t * restrict dst,
-                 const int num_elems) {
-    int left_over       = num_elems & (VLEN_FP32 - 1);
-    int num_elems_whole = num_elems - left_over;
-
-    int unaligned_addr = 0;
-    int unaligned_loop = 0;
-    if ((0 == htp_is_aligned((void *) src0, VLEN)) || (0 == htp_is_aligned((void *) src1, VLEN)) ||
-        (0 == htp_is_aligned((void *) dst, VLEN))) {
-        FARF(HIGH, "hvx_add_f32: unaligned address in hvx op, possibly slower execution\n");
-        unaligned_addr = 1;
-    }
-
-    if ((1 == unaligned_addr) && (num_elems_whole != 0)) {
-        unaligned_loop = 1;
-        FARF(HIGH, "hvx_add_f32: unaligned loop in hvx op, possibly slower execution\n");
-    }
-
-    if (0 == unaligned_loop) {
-        HVX_Vector * restrict vec_in1 = (HVX_Vector *) src0;
-        HVX_Vector * restrict vec_in2 = (HVX_Vector *) src1;
-        HVX_Vector * restrict vec_out = (HVX_Vector *) dst;
-
-        #pragma unroll(4)
-        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
-            HVX_Vector v = Q6_Vqf32_vadd_VsfVsf(*vec_in1++, *vec_in2++);
-            *vec_out++   = Q6_Vsf_equals_Vqf32(v);
-        }
-    } else {
-        #pragma unroll(4)
-        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
-            HVX_Vector in1 = *(HVX_UVector *) (src0 + i * SIZEOF_FP32);
-            HVX_Vector in2 = *(HVX_UVector *) (src1 + i * SIZEOF_FP32);
-
-            HVX_Vector out = Q6_Vqf32_vadd_VsfVsf(in1, in2);
-
-            *(HVX_UVector *) (dst + i * SIZEOF_FP32) = Q6_Vsf_equals_Vqf32(out);
-        }
-    }
-
-    if (left_over > 0) {
-        const float * src0f = (const float *) src0 + num_elems_whole;
-        const float * src1f = (const float *) src1 + num_elems_whole;
-        float *       dstf  = (float *) dst + num_elems_whole;
-
-        HVX_Vector in1 = *(HVX_UVector *) src0f;
-        HVX_Vector in2 = *(HVX_UVector *) src1f;
-
-        HVX_Vector out = Q6_Vqf32_vadd_VsfVsf(in1, in2);
-        hvx_vec_store_u((void *) dstf, left_over * SIZEOF_FP32, Q6_Vsf_equals_Vqf32(out));
-    }
-}
-
-void hvx_add_f32_opt(const uint8_t * restrict src0,
-                     const uint8_t * restrict src1,
-                     uint8_t * restrict dst,
-                     const int num_elems) {
-    htp_binary_ops_preamble;
-
-    for (int i = 0; i < step_of_4; i++) {
-        HVX_Vector v1a = *(HVX_Vector *) src0_curr;
-
-        HVX_Vector v1b = *(HVX_Vector *) src1_curr;
-
-        HVX_Vector v2a = *(HVX_Vector *) (src0_curr + VLEN);
-
-        HVX_Vector v1 = Q6_Vqf32_vadd_VsfVsf(v1a, v1b);
-
-        HVX_Vector v2b = *(HVX_Vector *) (src1_curr + VLEN);
-
-        HVX_Vector v3a = *(HVX_Vector *) (src0_curr + 2 * VLEN);
-
-        HVX_Vector v2 = Q6_Vqf32_vadd_VsfVsf(v2a, v2b);
-
-        *(HVX_Vector *) dst_curr = Q6_Vsf_equals_Vqf32(v1);
-
-        HVX_Vector v3b = *(HVX_Vector *) (src1_curr + 2 * VLEN);
-
-        HVX_Vector v4a = *(HVX_Vector *) (src0_curr + 3 * VLEN);
-
-        src0_curr += 4 * VLEN;
-
-        HVX_Vector v3 = Q6_Vqf32_vadd_VsfVsf(v3a, v3b);
-
-        *(HVX_Vector *) (dst_curr + VLEN) = Q6_Vsf_equals_Vqf32(v2);
-
-        HVX_Vector v4b = *(HVX_Vector *) (src1_curr + 3 * VLEN);
-
-        *(HVX_Vector *) (dst_curr + 2 * VLEN) = Q6_Vsf_equals_Vqf32(v3);
-
-        HVX_Vector v4 = Q6_Vqf32_vadd_VsfVsf(v4a, v4b);
-
-        src1_curr += 4 * VLEN;
-
-        *(HVX_Vector *) (dst_curr + 3 * VLEN) = Q6_Vsf_equals_Vqf32(v4);
-
-        dst_curr += 4 * VLEN;
-    }
-    for (int i = 0; i < step_of_2; i++) {
-        HVX_Vector v1a = *(HVX_Vector *) src0_curr;
-
-        HVX_Vector v1b = *(HVX_Vector *) src1_curr;
-
-        HVX_Vector v2a = *(HVX_Vector *) (src0_curr + VLEN);
-
-        HVX_Vector v1 = Q6_Vqf32_vadd_VsfVsf(v1a, v1b);
-
-        HVX_Vector v2b = *(HVX_Vector *) (src1_curr + VLEN);
-
-        *(HVX_Vector *) dst_curr = Q6_Vsf_equals_Vqf32(v1);
-
-        src0_curr += 2 * VLEN;
-
-        HVX_Vector v2 = Q6_Vqf32_vadd_VsfVsf(v2a, v2b);
-
-        src1_curr += 2 * VLEN;
-
-        *(HVX_Vector *) (dst_curr + VLEN) = Q6_Vsf_equals_Vqf32(v2);
-
-        dst_curr += 2 * VLEN;
-    }
-    for (int i = 0; i < step_of_1; i++) {
-        HVX_Vector va = *(HVX_Vector *) src0_curr;
-
-        src0_curr += VLEN;
-
-        HVX_Vector vb = *(HVX_Vector *) src1_curr;
-
-        src1_curr += VLEN;
-
-        HVX_Vector v = Q6_Vqf32_vadd_VsfVsf(va, vb);
-
-        *(HVX_Vector *) dst_curr = Q6_Vsf_equals_Vqf32(v);
-
-        dst_curr += VLEN;
-    }
-    if (remaining > 0) {
-        HVX_Vector v = Q6_Vqf32_vadd_VsfVsf(*(HVX_Vector *) src0_curr, *(HVX_Vector *) src1_curr);
-        hvx_vec_store_u((void *) dst_curr, remaining * SIZEOF_FP32, Q6_Vsf_equals_Vqf32(v));
-    }
-}
-
-void hvx_add_scalar_f32(const uint8_t * restrict src, const float val, uint8_t * restrict dst, const int num_elems) {
-    size_t left_over       = num_elems & (VLEN_FP32 - 1);
-    size_t num_elems_whole = num_elems - left_over;
-
-    int unaligned_addr = 0;
-    int unaligned_loop = 0;
-    if ((0 == htp_is_aligned((void *) src, VLEN)) || (0 == htp_is_aligned((void *) dst, VLEN))) {
-        FARF(HIGH, "hvx_add_scalar_f32: unaligned address in hvx op, possibly slower execution\n");
-        unaligned_addr = 1;
-    }
-
-    if ((1 == unaligned_addr) && (num_elems_whole != 0)) {
-        unaligned_loop = 1;
-        FARF(HIGH, "hvx_add_scalar_f32: unaligned loop in hvx op, possibly slower execution\n");
-    }
-
-    static const float kInf    = INFINITY;
-    const HVX_Vector   inf     = hvx_vec_splat_fp32(kInf);
-    HVX_Vector         val_vec = hvx_vec_splat_fp32(val);
-
-    if (0 == unaligned_loop) {
-        HVX_Vector * restrict vec_in1 = (HVX_Vector *) src;
-        HVX_Vector * restrict vec_out = (HVX_Vector *) dst;
-
-        #pragma unroll(4)
-        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
-            HVX_Vector           in       = *vec_in1++;
-            const HVX_VectorPred pred_inf = Q6_Q_vcmp_eq_VwVw(inf, in);
-            HVX_Vector           v        = Q6_Vqf32_vadd_VsfVsf(in, val_vec);
-            v                             = Q6_Vsf_equals_Vqf32(v);
-            v                             = Q6_V_vmux_QVV(pred_inf, inf, v);
-            *vec_out++                    = v;
-        }
-    } else {
-        #pragma unroll(4)
-        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
-            HVX_Vector in = *(HVX_UVector *) (src + i * SIZEOF_FP32);
-
-            const HVX_VectorPred pred_inf = Q6_Q_vcmp_eq_VwVw(inf, in);
-            HVX_Vector           out      = Q6_Vqf32_vadd_VsfVsf(in, val_vec);
-            out                           = Q6_Vsf_equals_Vqf32(out);
-            out                           = Q6_V_vmux_QVV(pred_inf, inf, out);
-
-            *(HVX_UVector *) (dst + i * SIZEOF_FP32) = out;
-        }
-    }
-
-    if (left_over > 0) {
-        const float * srcf = (const float *) src + num_elems_whole;
-        float *       dstf = (float *) dst + num_elems_whole;
-
-        HVX_Vector in = *(HVX_UVector *) srcf;
-
-        const HVX_VectorPred pred_inf = Q6_Q_vcmp_eq_VwVw(inf, in);
-        HVX_Vector           out      = Q6_Vqf32_vadd_VsfVsf(in, val_vec);
-        out                           = Q6_Vsf_equals_Vqf32(out);
-        out                           = Q6_V_vmux_QVV(pred_inf, inf, out);
-
-        hvx_vec_store_u((void *) dstf, left_over * SIZEOF_FP32, out);
-    }
-}
-
-void hvx_mul_scalar_f32(const uint8_t * restrict src, const float val, uint8_t * restrict dst, const int num_elems) {
-    size_t left_over       = num_elems & (VLEN_FP32 - 1);
-    size_t num_elems_whole = num_elems - left_over;
-
-    int unaligned_addr = 0;
-    int unaligned_loop = 0;
-    if ((0 == htp_is_aligned((void *) src, VLEN)) || (0 == htp_is_aligned((void *) dst, VLEN))) {
-        FARF(HIGH, "hvx_mul_scalar_f32: unaligned address in hvx op, possibly slower execution\n");
-        unaligned_addr = 1;
-    }
-
-    if ((1 == unaligned_addr) && (num_elems_whole != 0)) {
-        unaligned_loop = 1;
-        FARF(HIGH, "hvx_mul_scalar_f32: unaligned loop in hvx op, possibly slower execution\n");
-    }
-
-    HVX_Vector val_vec = hvx_vec_splat_fp32(val);
-    bool handled_leftover = false;
-    if (0 == unaligned_loop) {
-        HVX_Vector * restrict vec_in1 = (HVX_Vector *) src;
-        HVX_Vector * restrict vec_out = (HVX_Vector *) dst;
-
-        #pragma unroll(4)
-        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
-            HVX_Vector v = Q6_Vqf32_vmpy_VsfVsf(*vec_in1++, val_vec);
-            *vec_out++   = Q6_Vsf_equals_Vqf32(v);
-        }
-    } else {
-        int step_of_1 = num_elems >> 5;  // divby 32, because 32 float = 128 bytes per HVX vector
-        int leftover_size = left_over * sizeof(float);
-
-        HVX_Vector *  input_v_ptr  = (HVX_Vector *) src;
-        HVX_UVector * output_v_ptr = (HVX_UVector *) dst;
-
-        HVX_Vector slinep;
-        HVX_Vector slinec;
-        HVX_Vector sline;
-
-        slinep = *input_v_ptr++;
-
-        #pragma unroll(4)
-        for (int i = step_of_1 - 1; i > 0; i--) {
-            slinec                              = *input_v_ptr++;
-            sline                               = Q6_V_valign_VVR(slinec, slinep, (size_t) src);
-            *((HVX_UVector *) (output_v_ptr++)) = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(sline, val_vec));
-            /* Prepare slinep for next iteration */
-            slinep                              = slinec;
-        }
-
-        if (step_of_1 > 0) {
-            slinec = htp_is_aligned(input_v_ptr, VLEN) && left_over == 0 ? slinep : *input_v_ptr++;
-            sline  = Q6_V_valign_VVR(slinec, slinep, (size_t) src);
-            *((HVX_UVector *) (output_v_ptr++)) = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(sline, val_vec));
-
-            slinep = slinec;
-        }
-
-        if (leftover_size > 0) {
-            slinec = (is_in_one_chunk(input_v_ptr, leftover_size, VLEN) ? slinep : *input_v_ptr++);
-
-            sline = Q6_V_valign_VVR(slinec, slinep, (size_t) src);
-
-            HVX_Vector sout = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(sline, val_vec));
-            hvx_vec_store_u(output_v_ptr, leftover_size, sout);
-            handled_leftover = true;
-        }
-    }
-
-    if (left_over > 0 && !handled_leftover) {
-        const float * srcf = (const float *) src + num_elems_whole;
-        float *       dstf = (float *) dst + num_elems_whole;
-
-        HVX_Vector in = *(HVX_UVector *) srcf;
-
-        HVX_Vector out = Q6_Vqf32_vmpy_VsfVsf(in, val_vec);
-        hvx_vec_store_u((void *) dstf, left_over * SIZEOF_FP32, Q6_Vsf_equals_Vqf32(out));
-    }
-}
-
-void hvx_sub_f32(const uint8_t * restrict src0,
-                 const uint8_t * restrict src1,
-                 uint8_t * restrict dst,
-                 const int num_elems) {
-    size_t left_over       = num_elems & (VLEN_FP32 - 1);
-    size_t num_elems_whole = num_elems - left_over;
-
-    int unaligned_addr = 0;
-    int unaligned_loop = 0;
-    if ((0 == htp_is_aligned((void *) src0, VLEN)) || (0 == htp_is_aligned((void *) src1, VLEN)) ||
-        (0 == htp_is_aligned((void *) dst, VLEN))) {
-        FARF(HIGH, "hvx_sub_f32: unaligned address in hvx op, possibly slower execution\n");
-        unaligned_addr = 1;
-    }
-
-    if ((1 == unaligned_addr) && (num_elems_whole != 0)) {
-        unaligned_loop = 1;
-        FARF(HIGH, "hvx_sub_f32: unaligned loop in hvx op, possibly slower execution\n");
-    }
-
-    if (0 == unaligned_loop) {
-        HVX_Vector * restrict vec_in1 = (HVX_Vector *) src0;
-        HVX_Vector * restrict vec_in2 = (HVX_Vector *) src1;
-        HVX_Vector * restrict vec_out = (HVX_Vector *) dst;
-
-        #pragma unroll(4)
-        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
-            HVX_Vector v = Q6_Vqf32_vsub_VsfVsf(*vec_in1++, *vec_in2++);
-            *vec_out++   = Q6_Vsf_equals_Vqf32(v);
-        }
-    } else {
-        #pragma unroll(4)
-        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
-            HVX_Vector in1 = *(HVX_UVector *) (src0 + i * SIZEOF_FP32);
-            HVX_Vector in2 = *(HVX_UVector *) (src1 + i * SIZEOF_FP32);
-
-            HVX_Vector out = Q6_Vqf32_vsub_VsfVsf(in1, in2);
-
-            *(HVX_UVector *) (dst + i * SIZEOF_FP32) = Q6_Vsf_equals_Vqf32(out);
-        }
-    }
-
-    if (left_over > 0) {
-        const float * src0f = (const float *) src0 + num_elems_whole;
-        const float * src1f = (const float *) src1 + num_elems_whole;
-        float *       dstf  = (float *) dst + num_elems_whole;
-
-        HVX_Vector in1 = *(HVX_UVector *) src0f;
-        HVX_Vector in2 = *(HVX_UVector *) src1f;
-
-        HVX_Vector out = Q6_Vqf32_vsub_VsfVsf(in1, in2);
-        hvx_vec_store_u((void *) dstf, left_over * SIZEOF_FP32, Q6_Vsf_equals_Vqf32(out));
-    }
-}
-
-void hvx_sub_f32_opt(const uint8_t * restrict src0,
-                     const uint8_t * restrict src1,
-                     uint8_t * restrict dst,
-                     const int num_elems) {
-    htp_binary_ops_preamble;
-
-    for (int i = 0; i < step_of_4; i++) {
-        HVX_Vector v1a = *(HVX_Vector *) src0_curr;
-
-        HVX_Vector v1b = *(HVX_Vector *) src1_curr;
-
-        HVX_Vector v2a = *(HVX_Vector *) (src0_curr + VLEN);
-
-        HVX_Vector v1 = Q6_Vqf32_vsub_VsfVsf(v1a, v1b);
-
-        HVX_Vector v2b = *(HVX_Vector *) (src1_curr + VLEN);
-
-        HVX_Vector v3a = *(HVX_Vector *) (src0_curr + 2 * VLEN);
-
-        HVX_Vector v2 = Q6_Vqf32_vsub_VsfVsf(v2a, v2b);
-
-        *(HVX_Vector *) dst_curr = Q6_Vsf_equals_Vqf32(v1);
-
-        HVX_Vector v3b = *(HVX_Vector *) (src1_curr + 2 * VLEN);
-
-        HVX_Vector v4a = *(HVX_Vector *) (src0_curr + 3 * VLEN);
-
-        src0_curr += 4 * VLEN;
-
-        HVX_Vector v3 = Q6_Vqf32_vsub_VsfVsf(v3a, v3b);
-
-        *(HVX_Vector *) (dst_curr + VLEN) = Q6_Vsf_equals_Vqf32(v2);
-
-        HVX_Vector v4b = *(HVX_Vector *) (src1_curr + 3 * VLEN);
-
-        *(HVX_Vector *) (dst_curr + 2 * VLEN) = Q6_Vsf_equals_Vqf32(v3);
-
-        HVX_Vector v4 = Q6_Vqf32_vsub_VsfVsf(v4a, v4b);
-
-        src1_curr += 4 * VLEN;
-
-        *(HVX_Vector *) (dst_curr + 3 * VLEN) = Q6_Vsf_equals_Vqf32(v4);
-
-        dst_curr += 4 * VLEN;
-    }
-    for (int i = 0; i < step_of_2; i++) {
-        HVX_Vector v1a = *(HVX_Vector *) src0_curr;
-
-        HVX_Vector v1b = *(HVX_Vector *) src1_curr;
-
-        HVX_Vector v2a = *(HVX_Vector *) (src0_curr + VLEN);
-
-        HVX_Vector v1 = Q6_Vqf32_vsub_VsfVsf(v1a, v1b);
-
-        HVX_Vector v2b = *(HVX_Vector *) (src1_curr + VLEN);
-
-        *(HVX_Vector *) dst_curr = Q6_Vsf_equals_Vqf32(v1);
-
-        src0_curr += 2 * VLEN;
-
-        HVX_Vector v2 = Q6_Vqf32_vsub_VsfVsf(v2a, v2b);
-
-        src1_curr += 2 * VLEN;
-
-        *(HVX_Vector *) (dst_curr + VLEN) = Q6_Vsf_equals_Vqf32(v2);
-
-        dst_curr += 2 * VLEN;
-    }
-    for (int i = 0; i < step_of_1; i++) {
-        HVX_Vector va = *(HVX_Vector *) src0_curr;
-
-        src0_curr += VLEN;
-
-        HVX_Vector vb = *(HVX_Vector *) src1_curr;
-
-        src1_curr += VLEN;
-
-        HVX_Vector v = Q6_Vqf32_vsub_VsfVsf(va, vb);
-
-        *(HVX_Vector *) dst_curr = Q6_Vsf_equals_Vqf32(v);
-
-        dst_curr += VLEN;
-    }
-    if (remaining > 0) {
-        HVX_Vector v = Q6_Vqf32_vsub_VsfVsf(*(HVX_Vector *) src0_curr, *(HVX_Vector *) src1_curr);
-        hvx_vec_store_u((void *) dst_curr, remaining * SIZEOF_FP32, Q6_Vsf_equals_Vqf32(v));
-    }
-}
-
-void hvx_sub_scalar_f32(const uint8_t * restrict src, const float val, uint8_t * restrict dst, const int num_elems) {
-    size_t left_over       = num_elems & (VLEN_FP32 - 1);
-    size_t num_elems_whole = num_elems - left_over;
-
-    int unaligned_addr = 0;
-    int unaligned_loop = 0;
-    if ((0 == htp_is_aligned((void *) src, VLEN)) || (0 == htp_is_aligned((void *) dst, VLEN))) {
-        FARF(HIGH, "hvx_sub_scalar_f32: unaligned address in hvx op, possibly slower execution\n");
-        unaligned_addr = 1;
-    }
-
-    if ((1 == unaligned_addr) && (num_elems_whole != 0)) {
-        unaligned_loop = 1;
-        FARF(HIGH, "hvx_sub_scalar_f32: unaligned loop in hvx op, possibly slower execution\n");
-    }
-
-    HVX_Vector val_vec = hvx_vec_splat_fp32(val);
-
-    if (0 == unaligned_loop) {
-        HVX_Vector * restrict vec_in1 = (HVX_Vector *) src;
-        HVX_Vector * restrict vec_out = (HVX_Vector *) dst;
-
-        #pragma unroll(4)
-        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
-            HVX_Vector v = Q6_Vqf32_vsub_VsfVsf(*vec_in1++, val_vec);
-            *vec_out++   = Q6_Vsf_equals_Vqf32(v);
-        }
-    } else {
-        #pragma unroll(4)
-        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
-            HVX_Vector in = *(HVX_UVector *) (src + i * SIZEOF_FP32);
-
-            HVX_Vector out = Q6_Vqf32_vsub_VsfVsf(in, val_vec);
-
-            *(HVX_UVector *) (dst + i * SIZEOF_FP32) = Q6_Vsf_equals_Vqf32(out);
-        }
-    }
-
-    if (left_over > 0) {
-        const float * srcf = (const float *) src + num_elems_whole;
-        float *       dstf = (float *) dst + num_elems_whole;
-
-        HVX_Vector in = *(HVX_UVector *) srcf;
-
-        HVX_Vector out = Q6_Vqf32_vsub_VsfVsf(in, val_vec);
-        hvx_vec_store_u((void *) dstf, left_over * SIZEOF_FP32, Q6_Vsf_equals_Vqf32(out));
-    }
-}
-
-float hvx_sum_of_squares_f32(const uint8_t * restrict src, const int num_elems) {
-    int left_over       = num_elems & (VLEN_FP32 - 1);
-    int num_elems_whole = num_elems - left_over;
-
-    if (0 == htp_is_aligned((void *) src, VLEN)) {
-        FARF(HIGH, "hvx_sum_of_squares_f32: unaligned address in hvx op, possibly slower execution\n");
-    }
-
-    assert((1 == htp_is_aligned((void *) src, VLEN)) || (0 == num_elems_whole));
-
-    HVX_Vector * restrict vec_in1 = (HVX_Vector *) src;
-
-    HVX_Vector sum_vec_acc = Q6_V_vsplat_R(0x00000000);
-    HVX_Vector zero_vec    = Q6_V_vsplat_R(0x00000000);
-
-    #pragma unroll(4)
-    for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
-        HVX_Vector v = Q6_Vqf32_vmpy_VsfVsf(*vec_in1, *vec_in1);
-        sum_vec_acc  = Q6_Vqf32_vadd_Vqf32Vqf32(sum_vec_acc, v);
-        vec_in1++;
-    }
-
-    if (left_over > 0) {
-        const float * srcf = (const float *) src + num_elems_whole;
-
-        HVX_Vector vec_left = *(HVX_UVector *) srcf;
-
-        HVX_Vector vec_left_sq = Q6_Vqf32_vmpy_VsfVsf(vec_left, vec_left);
-        HVX_Vector vec_tmp     = Q6_V_valign_VVR(vec_left_sq, zero_vec, left_over * SIZEOF_FP32);
-
-        sum_vec_acc = Q6_Vqf32_vadd_Vqf32Vqf32(sum_vec_acc, vec_tmp);
-    }
-
-    HVX_Vector v = hvx_vec_qf32_reduce_sum(sum_vec_acc);
-    return hvx_vec_get_fp32(Q6_Vsf_equals_Vqf32(v));
-}
-
-float hvx_self_sum_f32(const uint8_t * restrict src, const int num_elems) {
-    int left_over       = num_elems & (VLEN_FP32 - 1);
-    int num_elems_whole = num_elems - left_over;
-
-    int unaligned_addr = 0;
-    int unaligned_loop = 0;
-    if (0 == htp_is_aligned((void *) src, VLEN)) {
-        FARF(HIGH, "hvx_self_sum_f32: unaligned address in hvx op, possibly slower execution\n");
-        unaligned_addr = 1;
-    }
-
-    if ((1 == unaligned_addr) && (num_elems_whole != 0)) {
-        unaligned_loop = 1;
-        FARF(HIGH, "hvx_self_sum_f32: unaligned loop in hvx op, possibly slower execution\n");
-    }
-
-    HVX_Vector sum_vec  = Q6_V_vsplat_R(0x00000000);
-    HVX_Vector zero_vec = Q6_V_vsplat_R(0x00000000);
-
-    if (0 == unaligned_loop) {
-        HVX_Vector * vec_in = (HVX_Vector *) src;
-
-        #pragma unroll(4)
-        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
-            // sum_vec = Q6_Vqf32_vadd_Vqf32Vsf(sum_vec, *vec_in++);
-            sum_vec = Q6_Vqf32_vadd_VsfVsf(Q6_Vsf_equals_Vqf32(sum_vec), *vec_in++);
-        }
-    } else {
-        #pragma unroll(4)
-        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
-            HVX_Vector in = *(HVX_UVector *) (src + i * SIZEOF_FP32);
-
-            sum_vec = Q6_Vqf32_vadd_VsfVsf(Q6_Vsf_equals_Vqf32(sum_vec), in);
-        }
-    }
-
-    if (left_over > 0) {
-        const float * srcf = (const float *) src + num_elems_whole;
-
-        HVX_Vector vec_left = *(HVX_UVector *) srcf;
-        HVX_Vector vec_tmp  = Q6_V_valign_VVR(vec_left, zero_vec, left_over * SIZEOF_FP32);
-        // sum_vec = Q6_Vqf32_vadd_Vqf32Vsf(sum_vec, vec_tmp);
-        sum_vec             = Q6_Vqf32_vadd_VsfVsf(Q6_Vsf_equals_Vqf32(sum_vec), vec_tmp);
-    }
-
-    HVX_Vector v = hvx_vec_qf32_reduce_sum(sum_vec);
-    return hvx_vec_get_fp32(Q6_Vsf_equals_Vqf32(v));
-}
-
-float hvx_self_max_f32(const uint8_t * restrict src, const int num_elems) {
-    int left_over       = num_elems & (VLEN_FP32 - 1);
-    int num_elems_whole = num_elems - left_over;
-
-    int unaligned_addr = 0;
-    int unaligned_loop = 0;
-    if (0 == htp_is_aligned((void *) src, VLEN)) {
-        FARF(HIGH, "hvx_self_max_f32: unaligned address in hvx op, possibly slower execution\n");
-        unaligned_addr = 1;
-    }
-
-    if ((1 == unaligned_addr) && (num_elems_whole != 0)) {
-        unaligned_loop = 1;
-        FARF(HIGH, "hvx_self_max_f32: unaligned loop in hvx op, possibly slower execution\n");
-    }
-
-    HVX_Vector vec_max   = hvx_vec_splat_fp32(((const float *) src)[0]);
-    HVX_Vector vec_first = hvx_vec_splat_fp32(((const float *) src)[0]);
-
-    if (0 == unaligned_loop) {
-        HVX_Vector * restrict vec_in = (HVX_Vector *) src;
-
-        #pragma unroll(4)
-        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
-            vec_max = Q6_Vsf_vmax_VsfVsf(vec_max, *vec_in++);
-        }
-    } else {
-        #pragma unroll(4)
-        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
-            HVX_Vector in = *(HVX_UVector *) (src + i * SIZEOF_FP32);
-
-            vec_max = Q6_Vsf_vmax_VsfVsf(vec_max, in);
-        }
-    }
-
-    if (left_over > 0) {
-        const float * srcf = (const float *) src + num_elems_whole;
-
-        HVX_Vector in = *(HVX_UVector *) srcf;
-
-        HVX_Vector temp = Q6_V_valign_VVR(in, vec_first, left_over * SIZEOF_FP32);
-        vec_max         = Q6_Vsf_vmax_VsfVsf(vec_max, temp);
-    }
-
-    HVX_Vector v = hvx_vec_reduce_max_fp32(vec_max);
-    return hvx_vec_get_fp32(v);
-}
-
-void hvx_min_scalar_f32(const uint8_t * restrict src, const float val, uint8_t * restrict dst, const int num_elems) {
-    size_t left_over       = num_elems & (VLEN_FP32 - 1);
-    size_t num_elems_whole = num_elems - left_over;
-    int unalign_address = 0;
-    if ((0 == htp_is_aligned((void *) src, VLEN)) || (0 == htp_is_aligned((void *) dst, VLEN))) {
-        FARF(HIGH, "hvx_min_scalar_f32: unaligned address in hvx op, possibly slower execution\n");
-        unalign_address = 1;
-    }
-
-    const float * src_f = (const float *) src;
-
-    HVX_Vector vec_min = hvx_vec_splat_fp32(val);
-
-    if(unalign_address == 0){
-        HVX_Vector * restrict vec_in  = (HVX_Vector *) src;
-        HVX_Vector * restrict vec_out = (HVX_Vector *) dst;
-
-        #pragma unroll(4)
-        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
-            HVX_Vector min_clamp    = Q6_Vsf_vmin_VsfVsf(vec_min, *vec_in++);
-            *vec_out++ = (min_clamp);
-        }
-    }else{
-        HVX_UVector * restrict vec_in  = (HVX_Vector *) src;
-        HVX_UVector * restrict vec_out = (HVX_Vector *) dst;
-
-        #pragma unroll(4)
-        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
-            HVX_Vector min_clamp     = Q6_Vsf_vmin_VsfVsf(vec_min, *vec_in++);
-            *vec_out++ = (min_clamp);
-        }
-    }
-
-    if (left_over > 0 ) {
-        const float * srcf = (const float *) src + num_elems_whole;
-        float *       dstf = (float *) dst + num_elems_whole;
-
-        HVX_UVector in = *(HVX_UVector *) srcf;
-
-        HVX_UVector min_clamp = Q6_Vsf_vmin_VsfVsf(vec_min, in);
-
-        hvx_vec_store_u((void *) dstf, left_over * SIZEOF_FP32, (min_clamp));
-    }
-}
-
-void hvx_clamp_scalar_f32(const uint8_t * restrict src,
-                          const float limit_left,
-                          const float limit_right,
-                          uint8_t * restrict dst,
-                          const int num_elems) {
-    size_t left_over       = num_elems & (VLEN_FP32 - 1);
-    size_t num_elems_whole = num_elems - left_over;
-
-    int unalign_address = 0;
-    if ((0 == htp_is_aligned((void *) src, VLEN)) || (0 == htp_is_aligned((void *) dst, VLEN))) {
-        FARF(HIGH, "hvx_clamp_scalar_f32: unaligned address in hvx op, possibly slower execution\n");
-        unalign_address = 1;
-    }
-
-    HVX_Vector range_left  = hvx_vec_splat_fp32(limit_left);
-    HVX_Vector range_right = hvx_vec_splat_fp32(limit_right);
-
-    if(unalign_address == 0){
-        HVX_Vector * restrict vec_in  = (HVX_Vector *) src;
-        HVX_Vector * restrict vec_out = (HVX_Vector *) dst;
-
-
-
-        #pragma unroll(4)
-        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
-            HVX_Vector in_vec = *vec_in++;
-            HVX_Vector temp_v = in_vec;
-
-            HVX_VectorPred pred_cap_right = Q6_Q_vcmp_gt_VsfVsf(in_vec, range_right);
-            HVX_VectorPred pred_cap_left  = Q6_Q_vcmp_gt_VsfVsf(range_left, in_vec);
-
-            in_vec = Q6_V_vmux_QVV(pred_cap_right, range_right, temp_v);
-            in_vec = Q6_V_vmux_QVV(pred_cap_left, range_left, in_vec);
-
-            *vec_out++ = in_vec;
-        }
-
-    }else{
-
-        HVX_UVector * restrict vec_in  = (HVX_UVector *) src;
-        HVX_UVector * restrict vec_out = (HVX_UVector *) dst;
-
-        #pragma unroll(4)
-        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
-            HVX_Vector in_vec = *vec_in++;
-            HVX_Vector temp_v = in_vec;
-
-            HVX_VectorPred pred_cap_right = Q6_Q_vcmp_gt_VsfVsf(in_vec, range_right);
-            HVX_VectorPred pred_cap_left  = Q6_Q_vcmp_gt_VsfVsf(range_left, in_vec);
-
-            in_vec = Q6_V_vmux_QVV(pred_cap_right, range_right, temp_v);
-            in_vec = Q6_V_vmux_QVV(pred_cap_left, range_left, in_vec);
-
-            *vec_out++ = in_vec;
-        }
-
-    }
-
-    if (left_over > 0) {
-        const float * srcf = (const float *) src + num_elems_whole;
-        float *       dstf = (float *) dst + num_elems_whole;
-
-        HVX_Vector in_vec = *(HVX_UVector *) srcf;
-
-        HVX_Vector temp_v = in_vec;
-
-        HVX_VectorPred pred_cap_right = Q6_Q_vcmp_gt_VsfVsf(in_vec, range_right);
-        HVX_VectorPred pred_cap_left  = Q6_Q_vcmp_gt_VsfVsf(range_left, in_vec);
-
-        in_vec = Q6_V_vmux_QVV(pred_cap_right, range_right, temp_v);
-        in_vec = Q6_V_vmux_QVV(pred_cap_left, range_left, in_vec);
-
-        hvx_vec_store_u((void *) dstf, left_over * SIZEOF_FP32, in_vec);
-    }
-}
-
-