OSchip
/
OpenBLAS
Not watched
1
0
Fork 0
Code
Releases 66 Wiki evaluate Activity
Issues 0 Pull Requests 0 Datasets Model Cloudbrain HPC
Browse Source

Update sbgemm_tcopy_4_neoversev1 kernel to use standard C types

pull/5319/head
Ian McInerney 5 months ago
parent
4e6da5ed34
commit
badef1d32e
1 changed files with 16 additions and 16 deletions
Unified View
Diff Options
Show Stats Download Patch File Download Diff File
  1. +16
    -16
      kernel/arm64/sbgemm_tcopy_4_neoversev1.c

+ 16
- 16
kernel/arm64/sbgemm_tcopy_4_neoversev1.c View File

@@ -52,16 +52,16 @@ int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b) {
svbool_t pg16_first_8 = svwhilelt_b16(0, 8); svbool_t pg16_first_8 = svwhilelt_b16(0, 8);


svbool_t pg64_first_4 = svwhilelt_b64(0, 4); svbool_t pg64_first_4 = svwhilelt_b64(0, 4);
u_int32_t sizeof_u64 = 8;
u_int64_t _st_offsets_0[4] = {
uint32_t sizeof_u64 = 8;
uint64_t _st_offsets_0[4] = {
0 * sizeof_u64, 0 * sizeof_u64,
1 * sizeof_u64, 1 * sizeof_u64,
4 * sizeof_u64, 4 * sizeof_u64,
5 * sizeof_u64, 5 * sizeof_u64,
}; };


u_int64_t _st_offsets_1[4] = {
uint64_t _st_offsets_1[4] = {
2 * sizeof_u64, 2 * sizeof_u64,
3 * sizeof_u64, 3 * sizeof_u64,
6 * sizeof_u64, 6 * sizeof_u64,
@@ -108,13 +108,13 @@ int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b) {
m01 = svzip1_u32(svreinterpret_u32_u16(t2), svreinterpret_u32_u16(t3)); m01 = svzip1_u32(svreinterpret_u32_u16(t2), svreinterpret_u32_u16(t3));
m11 = svzip2_u32(svreinterpret_u32_u16(t2), svreinterpret_u32_u16(t3)); m11 = svzip2_u32(svreinterpret_u32_u16(t2), svreinterpret_u32_u16(t3));


svst1_scatter_u64offset_u64(pg64_first_4, (u_int64_t *)b_offset0,
svst1_scatter_u64offset_u64(pg64_first_4, (uint64_t *)b_offset0,
st_offsets_0, svreinterpret_u64_u32(m00)); st_offsets_0, svreinterpret_u64_u32(m00));
svst1_scatter_u64offset_u64(pg64_first_4, (u_int64_t *)b_offset0,
svst1_scatter_u64offset_u64(pg64_first_4, (uint64_t *)b_offset0,
st_offsets_1, svreinterpret_u64_u32(m01)); st_offsets_1, svreinterpret_u64_u32(m01));
svst1_scatter_u64offset_u64(pg64_first_4, (u_int64_t *)b_offset1,
svst1_scatter_u64offset_u64(pg64_first_4, (uint64_t *)b_offset1,
st_offsets_0, svreinterpret_u64_u32(m10)); st_offsets_0, svreinterpret_u64_u32(m10));
svst1_scatter_u64offset_u64(pg64_first_4, (u_int64_t *)b_offset1,
svst1_scatter_u64offset_u64(pg64_first_4, (uint64_t *)b_offset1,
st_offsets_1, svreinterpret_u64_u32(m11)); st_offsets_1, svreinterpret_u64_u32(m11));


a_offset0 += 8 * lda; a_offset0 += 8 * lda;
@@ -150,13 +150,13 @@ int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b) {
m01 = svzip1_u32(svreinterpret_u32_u16(t2), svreinterpret_u32_u16(t3)); m01 = svzip1_u32(svreinterpret_u32_u16(t2), svreinterpret_u32_u16(t3));
m11 = svzip2_u32(svreinterpret_u32_u16(t2), svreinterpret_u32_u16(t3)); m11 = svzip2_u32(svreinterpret_u32_u16(t2), svreinterpret_u32_u16(t3));


svst1_scatter_u64offset_u64(pg64_first_4, (u_int64_t *)b_offset0,
svst1_scatter_u64offset_u64(pg64_first_4, (uint64_t *)b_offset0,
st_offsets_0, svreinterpret_u64_u32(m00)); st_offsets_0, svreinterpret_u64_u32(m00));
svst1_scatter_u64offset_u64(pg64_first_4, (u_int64_t *)b_offset0,
svst1_scatter_u64offset_u64(pg64_first_4, (uint64_t *)b_offset0,
st_offsets_1, svreinterpret_u64_u32(m01)); st_offsets_1, svreinterpret_u64_u32(m01));
svst1_scatter_u64offset_u64(pg64_first_4, (u_int64_t *)b_offset1,
svst1_scatter_u64offset_u64(pg64_first_4, (uint64_t *)b_offset1,
st_offsets_0, svreinterpret_u64_u32(m10)); st_offsets_0, svreinterpret_u64_u32(m10));
svst1_scatter_u64offset_u64(pg64_first_4, (u_int64_t *)b_offset1,
svst1_scatter_u64offset_u64(pg64_first_4, (uint64_t *)b_offset1,
st_offsets_1, svreinterpret_u64_u32(m11)); st_offsets_1, svreinterpret_u64_u32(m11));
} }
} }
@@ -194,9 +194,9 @@ int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b) {


m00 = svzip1_u32(svreinterpret_u32_u16(t0), svreinterpret_u32_u16(t1)); m00 = svzip1_u32(svreinterpret_u32_u16(t0), svreinterpret_u32_u16(t1));
m01 = svzip1_u32(svreinterpret_u32_u16(t2), svreinterpret_u32_u16(t3)); m01 = svzip1_u32(svreinterpret_u32_u16(t2), svreinterpret_u32_u16(t3));
svst1_scatter_u64offset_u64(pg64_first_4, (u_int64_t *)b_offset0,
svst1_scatter_u64offset_u64(pg64_first_4, (uint64_t *)b_offset0,
st_offsets_0, svreinterpret_u64_u32(m00)); st_offsets_0, svreinterpret_u64_u32(m00));
svst1_scatter_u64offset_u64(pg64_first_4, (u_int64_t *)b_offset0,
svst1_scatter_u64offset_u64(pg64_first_4, (uint64_t *)b_offset0,
st_offsets_1, svreinterpret_u64_u32(m01)); st_offsets_1, svreinterpret_u64_u32(m01));


a_offset0 += 8 * lda; a_offset0 += 8 * lda;
@@ -229,9 +229,9 @@ int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b) {
m00 = svzip1_u32(svreinterpret_u32_u16(t0), svreinterpret_u32_u16(t1)); m00 = svzip1_u32(svreinterpret_u32_u16(t0), svreinterpret_u32_u16(t1));
m01 = svzip1_u32(svreinterpret_u32_u16(t2), svreinterpret_u32_u16(t3)); m01 = svzip1_u32(svreinterpret_u32_u16(t2), svreinterpret_u32_u16(t3));


svst1_scatter_u64offset_u64(pg64_first_4, (u_int64_t *)b_offset0,
svst1_scatter_u64offset_u64(pg64_first_4, (uint64_t *)b_offset0,
st_offsets_0, svreinterpret_u64_u32(m00)); st_offsets_0, svreinterpret_u64_u32(m00));
svst1_scatter_u64offset_u64(pg64_first_4, (u_int64_t *)b_offset0,
svst1_scatter_u64offset_u64(pg64_first_4, (uint64_t *)b_offset0,
st_offsets_1, svreinterpret_u64_u32(m01)); st_offsets_1, svreinterpret_u64_u32(m01));
} }
} }


Write Preview
Loading…
Cancel
Save