happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Revert "opt: use qf32 internal precision for vec_dot_f16_f32" 

This reverts commit 8600ecd20d6c902fe16271d6af1e59504eff4a27.
This commit is contained in:
chraac 2025-12-22 12:35:46 +08:00
parent cb0a8ff4e7
commit 2058f28b3e
1 changed files with 17 additions and 21 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

38
ggml/src/ggml-hexagon/htp/matmul-ops.c
View File

@ -915,21 +915,18 @@ static void vec_dot_f16_f32(const int n, float * restrict s, const void * restri
uint32_t nv0 = n / VLEN_FP16; // num full fp16 hvx vectors uint32_t nv0 = n / VLEN_FP16; // num full fp16 hvx vectors
uint32_t nv1 = n % VLEN_FP16; // leftover elements uint32_t nv1 = n % VLEN_FP16; // leftover elements
const HVX_Vector one = Q6_Vh_vsplat_R(0x3C00); // 1.0 in fp16 const HVX_Vector zero = Q6_Vh_vsplat_R(0x3C00); // 1.0 in fp16
const HVX_Vector zero = Q6_V_vsplat_R(0); // 0.0 in fp16
HVX_Vector rsum = Q6_V_vsplat_R(0); HVX_Vector rsum = Q6_V_vsplat_R(0);
uint32_t i = 0; uint32_t i = 0;
#pragma unroll(2) #pragma unroll(2)
for (i = 0; i < nv0; i++) { for (i = 0; i < nv0; i++) {
HVX_VectorPair yp = vy[i]; HVX_VectorPair yp = vy[i];
HVX_Vector x = vx[i]; HVX_Vector x = vx[i];
HVX_VectorPair xp = Q6_Wqf32_vmpy_VhfVhf(Q6_Vh_vshuff_Vh(x), one); // mul by 1.0 HVX_VectorPair xp = Q6_Wqf32_vmpy_VhfVhf(Q6_Vh_vshuff_Vh(x), zero); // mul by 1.0
HVX_Vector y_hi = Q6_Vqf32_vadd_VsfVsf(Q6_V_hi_W(yp), zero); // convert to qf32
HVX_Vector y_lo = Q6_Vqf32_vadd_VsfVsf(Q6_V_lo_W(yp), zero); // convert to qf32
HVX_Vector hi = Q6_Vqf32_vmpy_Vqf32Vqf32(Q6_V_hi_W(xp), y_hi); HVX_Vector hi = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(Q6_V_hi_W(xp)), Q6_V_hi_W(yp));
HVX_Vector lo = Q6_Vqf32_vmpy_Vqf32Vqf32(Q6_V_lo_W(xp), y_lo); HVX_Vector lo = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(Q6_V_lo_W(xp)), Q6_V_lo_W(yp));
HVX_Vector sum = Q6_Vqf32_vadd_Vqf32Vqf32(hi, lo); HVX_Vector sum = Q6_Vqf32_vadd_Vqf32Vqf32(hi, lo);
rsum = Q6_Vqf32_vadd_Vqf32Vqf32(rsum, sum); rsum = Q6_Vqf32_vadd_Vqf32Vqf32(rsum, sum);
@ -938,25 +935,24 @@ static void vec_dot_f16_f32(const int n, float * restrict s, const void * restri
if (nv1) { if (nv1) {
HVX_VectorPair yp = vy[i]; HVX_VectorPair yp = vy[i];
HVX_Vector x = vx[i]; HVX_Vector x = vx[i];
HVX_VectorPair xp = Q6_Wqf32_vmpy_VhfVhf(Q6_Vh_vshuff_Vh(x), one); // mul by 1.0 HVX_VectorPair xp = Q6_Wqf32_vmpy_VhfVhf(Q6_Vh_vshuff_Vh(x), zero); // mul by 1.0
HVX_Vector leftover_x; HVX_Vector l_x;
HVX_Vector leftover_y; HVX_Vector l_y;
if (nv1 >= VLEN_FP32) { if (nv1 >= VLEN_FP32) {
HVX_Vector y_lo = Q6_Vqf32_vadd_VsfVsf(Q6_V_lo_W(yp), zero); // convert to qf32 HVX_Vector hi = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(Q6_V_lo_W(xp)), Q6_V_lo_W(yp));
HVX_Vector lo = Q6_Vqf32_vmpy_Vqf32Vqf32(Q6_V_lo_W(xp), y_lo); rsum = Q6_Vqf32_vadd_Vqf32Vqf32(rsum, hi);
rsum = Q6_Vqf32_vadd_Vqf32Vqf32(rsum, lo);
nv1 -= VLEN_FP32; nv1 -= VLEN_FP32;
leftover_x = Q6_V_hi_W(xp); l_x = Q6_V_hi_W(xp);
leftover_y = Q6_Vqf32_vadd_VsfVsf(Q6_V_hi_W(yp), zero); // convert to qf32 l_y = Q6_V_hi_W(yp);
} else { } else {
leftover_x = Q6_V_lo_W(xp); l_x = Q6_V_lo_W(xp);
leftover_y = Q6_Vqf32_vadd_VsfVsf(Q6_V_lo_W(yp), zero); // convert to qf32 l_y = Q6_V_lo_W(yp);
} }
if (nv1) { if (nv1) {
HVX_Vector lo = Q6_Vqf32_vmpy_Vqf32Vqf32(leftover_x, leftover_y); HVX_Vector lo = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(l_x), l_y);
HVX_Vector sum = Q6_V_valign_VVR(lo, zero, nv1 * sizeof(float)); HVX_Vector sum = Q6_V_valign_VVR(lo, Q6_V_vzero(), nv1 * sizeof(float));
rsum = Q6_Vqf32_vadd_Vqf32Vqf32(rsum, sum); rsum = Q6_Vqf32_vadd_Vqf32Vqf32(rsum, sum);
} }