wehub-resource-sync
69 lines
2.3 KiB
C++
69 lines
2.3 KiB
C++
#include <riscv_vector.h>
|
|
|
|
void MNNAsyQuantFunc_RVV(int8_t* dst, const float* src, float* qscale, float* qbias, const size_t* info) {
|
|
auto blockNum = info[0];
|
|
auto EP = info[1];
|
|
auto LP = info[2];
|
|
auto kernelsize = info[5];
|
|
auto blockLU = info[6];
|
|
|
|
auto stride0 = blockNum * blockLU * EP * LP;
|
|
auto stride1 = blockLU * EP * LP;
|
|
|
|
const int int8_min = -128;
|
|
const int int8_max = 127;
|
|
const size_t RV_RNU = 0x4;
|
|
|
|
const int EP_TILE = 4;
|
|
|
|
for (int n = 0; n < kernelsize; ++n) {
|
|
for (int bk = 0; bk < blockNum; ++bk) {
|
|
const float* scalePtr = qscale + bk * EP;
|
|
const float* biasPtr = qbias + bk * EP;
|
|
|
|
for (int k = 0; k < blockLU; ++k) {
|
|
int base = n * stride0 + bk * stride1 + k * EP * LP;
|
|
const float* srcBase = src + base;
|
|
int8_t* dstBase = dst + base;
|
|
|
|
for (int i = 0; i < EP; i += EP_TILE) {
|
|
int tile = (i + EP_TILE <= EP) ? EP_TILE : (EP - i);
|
|
|
|
__builtin_prefetch(srcBase + (i + EP_TILE) * LP, 0, 1);
|
|
|
|
for (int t = 0; t < tile; ++t) {
|
|
float scaleVal = scalePtr[i + t];
|
|
float biasVal = biasPtr[i + t];
|
|
|
|
const float* srcZ = srcBase + (i + t) * LP;
|
|
int8_t* dstZ = dstBase + (i + t) * LP;
|
|
|
|
int j = 0;
|
|
|
|
while (j < LP) {
|
|
size_t vl = __riscv_vsetvl_e32m4(LP - j);
|
|
|
|
vfloat32m4_t v = __riscv_vle32_v_f32m4(srcZ + j, vl);
|
|
|
|
v = __riscv_vfmul_vf_f32m4(v, scaleVal, vl);
|
|
v = __riscv_vfadd_vf_f32m4(v, biasVal, vl);
|
|
|
|
vint32m4_t vi = __riscv_vfcvt_x_f_v_i32m4_rm(v, RV_RNU, vl);
|
|
|
|
vi = __riscv_vmax_vx_i32m4(vi, int8_min, vl);
|
|
vi = __riscv_vmin_vx_i32m4(vi, int8_max, vl);
|
|
|
|
vint16m2_t v16 = __riscv_vncvt_x_x_w_i16m2(vi, vl);
|
|
vint8m1_t v8 = __riscv_vncvt_x_x_w_i8m1(v16, vl);
|
|
|
|
__riscv_vse8_v_i8m1(dstZ + j, v8, vl);
|
|
|
|
j += vl;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|