hexagon: vectorize softplus with HVX (#30435)

softplus_f32 and tile_softplus_f32 ran a scalar logf(1.0f + expf(x)) loop.
Replace both with a shared hvx_softplus_f32_aa() helper built from the
backend's existing hvx_vec_exp_f32 / hvx_vec_log_f32 primitives, so no new
numerical approximation is introduced.

hvx_vec_exp_f32 clamps its input to (-88, 88) before exponentiating, so a
naive log(1 + exp(x)) composition would saturate to ~88.72 for large x
instead of returning x. The Q6_V_vmux_QVV select keeps the scalar
reference's x > 20 guard (ggml_compute_softplus_f32 in ggml-impl.h),
which is what makes the vector path match the CPU result.

Assisted-by: Claude
This commit is contained in:
wes authored and GitHub committed 2026-10-16 15:20:43 -07:00
1 parent 2f2763db6e
commit 27c7a769da
1 file changed
+37 -16
+37 -16
View File
@@ -690,22 +690,48 @@ static void tri_f32(const void * restrict src,
}
}
// softplus(x) = log(1 + exp(x))
// Match CPU reference: return x for x > 20 to avoid overflow.
static inline void hvx_softplus_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
const HVX_Vector * restrict v_src = (const HVX_Vector *) src;
HVX_Vector * restrict v_dst = (HVX_Vector *) dst;
const HVX_Vector twenty = hvx_vec_splat_f32(20.0f);
const HVX_Vector one = hvx_vec_splat_f32(1.0f);
const uint32_t nvec = n / VLEN_FP32;
const uint32_t nloe = n % VLEN_FP32;
#pragma unroll(4)
for (uint32_t i = 0; i < nvec; i++) {
HVX_Vector x = v_src[i];
HVX_VectorPred pred_gt20 = Q6_Q_vcmp_gt_VsfVsf(x, twenty);
HVX_Vector expx = hvx_vec_exp_f32(x);
HVX_Vector logv = hvx_vec_log_f32(hvx_vec_add_f32_f32(one, expx));
v_dst[i] = Q6_V_vmux_QVV(pred_gt20, x, logv);
}
if (nloe > 0) {
HVX_Vector x = v_src[nvec];
HVX_VectorPred pred_gt20 = Q6_Q_vcmp_gt_VsfVsf(x, twenty);
HVX_Vector expx = hvx_vec_exp_f32(x);
HVX_Vector logv = hvx_vec_log_f32(hvx_vec_add_f32_f32(one, expx));
HVX_Vector sp = Q6_V_vmux_QVV(pred_gt20, x, logv);
hvx_vec_store_a((void *) &v_dst[nvec], nloe * sizeof(float), sp);
}
}
static void softplus_f32(const void * restrict src,
void * restrict dst,
const uint32_t num_rows,
const struct htp_unary_context * uctx) {
htp_unary_op_preamble;
// softplus(x) = log(1 + exp(x))
// Match CPU reference: ggml_compute_softplus_f32() in ggml-impl.h
for (uint32_t ir = 0; ir < num_rows; ir++) {
const float * restrict src_f = (const float *)((const uint8_t *)src + (ir * src0_row_size_aligned));
float * restrict dst_f = (float *)((uint8_t *)dst + (ir * dst_row_size_aligned));
for (uint32_t i = 0; i < ne0; i++) {
float x = src_f[i];
// For x > 20: softplus(x) ~ x (avoids exp overflow)
dst_f[i] = (x > 20.0f) ? x : logf(1.0f + expf(x));
}
for (uint32_t ir = 0; ir < num_rows; ir++) {
const uint8_t * restrict src_local = (const uint8_t *)src + (ir * src0_row_size_aligned);
uint8_t * restrict dst_local = (uint8_t *)dst + (ir * dst_row_size_aligned);
hvx_softplus_f32_aa(dst_local, src_local, ne0);
}
}
@@ -866,12 +892,7 @@ static void tile_gelu_erf_f32(void * restrict dst, const void * restrict src, ui
static void tile_softplus_f32(void * restrict dst, const void * restrict src, uint32_t tw, const struct htp_unary_context * uctx) {
(void) uctx;
const float * restrict sf = (const float *) src;
float * restrict df = (float *) dst;
for (uint32_t i = 0; i < tw; i++) {
float x = sf[i];
df[i] = (x > 20.0f) ? x : logf(1.0f + expf(x));
}
hvx_softplus_f32_aa((uint8_t *) dst, (const uint8_t *) src, tw);
}
static void tile_tanh_f32(void * restrict dst, const void * restrict src, uint32_t tw, const struct htp_unary_context * uctx) {