nihui
/
ncnn
Not watched
1
0
Fork 0
Code
Releases 46 Wiki evaluate Activity
Issues 0 Pull Requests 0 Datasets Model Cloudbrain HPC
Browse Source

Riscv64 c906 d1 (#3159) 

* Use RVV spec 0.7.1 for C906.

* Fix code style issue.

* Update convolution_sgemm_packn_fp16s.h

RVV_SPEC_0_7 update

* apply code-format changes

Co-authored-by: Zhang Xianyi <xianyi@perfxlab.com>
Co-authored-by: yaobyPerfxlab <yaobyPerfxlab@users.noreply.github.com>
tags/20211122
yaobyPerfxlab GitHub 5 years ago
parent
b0f8290657
commit
ec561736a5
No known key found for this signature in database GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 57 additions and 26 deletions
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. +57
    -26
      src/layer/riscv/convolution_sgemm_packn_fp16s.h

+ 57
- 26
src/layer/riscv/convolution_sgemm_packn_fp16s.h View File

@@ -55,20 +55,35 @@ static void im2col_sgemm_packn_fp16sa_rvv(const Mat& bottom_im2col, Mat& top_blo
for (int k = 0; k < maxk; k++)
{
#if RVV_SPEC_0_7
for (int l = 0; l < packn; l++)
{
tmpptr[0] = img0[l];
tmpptr[1] = img0[l + packn];
tmpptr[2] = img0[l + packn * 2];
tmpptr[3] = img0[l + packn * 3];
tmpptr[4] = img0[l + packn * 4];
tmpptr[5] = img0[l + packn * 5];
tmpptr[6] = img0[l + packn * 6];
tmpptr[7] = img0[l + packn * 7];
tmpptr += 8;
}
asm volatile(
"mv t3, %[LEN] \n\t"
"mv t1, %[SRC] \n\t"
"mv t2, %[TMP] \n\t"
"slli t3, t3, 1 \n\t"
"vle.v v0, (t1) \n\t"
"add t1, t1, t3 \n\t"
"vle.v v1, (t1) \n\t"
"add t1, t1, t3 \n\t"
"vle.v v2, (t1) \n\t"
"add t1, t1, t3 \n\t"
"vle.v v3, (t1) \n\t"
"add t1, t1, t3 \n\t"
"vle.v v4, (t1) \n\t"
"add t1, t1, t3 \n\t"
"vle.v v5, (t1) \n\t"
"add t1, t1, t3 \n\t"
"vle.v v6, (t1) \n\t"
"add t1, t1, t3 \n\t"
"vle.v v7, (t1) \n\t"
"add t1, t1, t3 \n\t"
"vsseg8e.v v0, (t2) \n\t"
:
: [LEN] "r"(packn), [SRC] "r"(img0), [TMP] "r"(tmpptr)
: "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "t1", "t2", "t3");

img0 += size * packn;
tmpptr += packn * 8;

#else
vfloat16m1_t _val0 = vle16_v_f16m1(img0, vl);
vfloat16m1_t _val1 = vle16_v_f16m1(img0 + packn, vl);
@@ -104,16 +119,25 @@ static void im2col_sgemm_packn_fp16sa_rvv(const Mat& bottom_im2col, Mat& top_blo
for (int k = 0; k < maxk; k++)
{
#if RVV_SPEC_0_7
for (int l = 0; l < packn; l++)
{
tmpptr[0] = img0[l];
tmpptr[1] = img0[l + packn];
tmpptr[2] = img0[l + packn * 2];
tmpptr[3] = img0[l + packn * 3];
tmpptr += 4;
}
asm volatile(
"mv t3, %[LEN] \n\t"
"mv t1, %[SRC] \n\t"
"mv t2, %[TMP] \n\t"
"slli t3, t3, 1 \n\t"
"vle.v v0, (t1) \n\t"
"add t1, t1, t3 \n\t"
"vle.v v1, (t1) \n\t"
"add t1, t1, t3 \n\t"
"vle.v v2, (t1) \n\t"
"add t1, t1, t3 \n\t"
"vle.v v3, (t1) \n\t"
"vsseg4e.v v0, (t2) \n\t"
:
: [LEN] "r"(packn), [SRC] "r"(img0), [TMP] "r"(tmpptr)
: "cc", "memory", "v0", "v1", "v2", "v3", "t1", "t2", "t3");

img0 += size * packn;
tmpptr += packn * 4;
#else
vfloat16m1_t _val0 = vle16_v_f16m1(img0, vl);
vfloat16m1_t _val1 = vle16_v_f16m1(img0 + packn, vl);
@@ -146,14 +170,21 @@ static void im2col_sgemm_packn_fp16sa_rvv(const Mat& bottom_im2col, Mat& top_blo
for (int k = 0; k < maxk; k++)
{
#if RVV_SPEC_0_7
for (int l = 0; l < packn; l++)
{
tmpptr[0] = img0[l];
tmpptr[1] = img0[l + packn];
tmpptr += 2;
}
asm volatile(
"mv t3, %[LEN] \n\t"
"mv t1, %[SRC] \n\t"
"mv t2, %[TMP] \n\t"
"slli t3, t3, 1 \n\t"
"vle.v v0, (t1) \n\t"
"add t1, t1, t3 \n\t"
"vle.v v1, (t1) \n\t"
"add t1, t1, t3 \n\t"
:
: [LEN] "r"(packn), [SRC] "r"(img0), [TMP] "r"(tmpptr)
: "cc", "memory", "v0", "v1", "t1", "t2", "t3");

img0 += size * packn;
tmpptr += packn * 2;
#else
vfloat16m1_t _val0 = vle16_v_f16m1(img0, vl);
vfloat16m1_t _val1 = vle16_v_f16m1(img0 + packn, vl);


Write Preview
Loading…
Cancel
Save