floraachy
/
OpenBLAS
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge pull request #4887 from goplanid/develop 

Small GEMM improvements for AArch64 with SVE
This commit is contained in:
Martin Kroeker 2024-09-16 11:17:19 +02:00 committed by GitHub
parent 60c1519e01 4894c54055
commit fca86e359c
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 253 additions and 37 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

260
kernel/arm64/dgemm_small_kernel_tn_sve.c
View File

@ -211,6 +211,7 @@ CNAME(BLASLONG M,
const BLASLONG v_m1 = M & -v_size; const BLASLONG v_m1 = M & -v_size;
const BLASLONG n4 = N & -4; const BLASLONG n4 = N & -4;
const BLASLONG n2 = N & -2; const BLASLONG n2 = N & -2;
const BLASLONG n8 = N & -8;
const int pack_a = M >= v_size2 && N >= 8 && K >= 8 ? 1 : 0; const int pack_a = M >= v_size2 && N >= 8 && K >= 8 ? 1 : 0;
FLOAT* packed_a = FLOAT* packed_a =
@ -229,6 +230,145 @@ CNAME(BLASLONG M,
CREATE_A_POINTER(1, v_size); CREATE_A_POINTER(1, v_size);
BLASLONG j = 0; BLASLONG j = 0;
for (; j < n8; j += 8) {
CREATE_B_POINTER(0, 0);
CREATE_B_POINTER(1, 1);
CREATE_B_POINTER(2, 2);
CREATE_B_POINTER(3, 3);
CREATE_B_POINTER(4, 4);
CREATE_B_POINTER(5, 5);
CREATE_B_POINTER(6, 6);
CREATE_B_POINTER(7, 7);
UPDATE_B_POINTER(8);
BLASLONG k = 0;
DECLARE_RESULT_VECTOR(0, 0);
DECLARE_RESULT_VECTOR(0, 1);
DECLARE_RESULT_VECTOR(0, 2);
DECLARE_RESULT_VECTOR(0, 3);
DECLARE_RESULT_VECTOR(0, 4);
DECLARE_RESULT_VECTOR(0, 5);
DECLARE_RESULT_VECTOR(0, 6);
DECLARE_RESULT_VECTOR(0, 7);
DECLARE_RESULT_VECTOR(1, 0);
DECLARE_RESULT_VECTOR(1, 1);
DECLARE_RESULT_VECTOR(1, 2);
DECLARE_RESULT_VECTOR(1, 3);
DECLARE_RESULT_VECTOR(1, 4);
DECLARE_RESULT_VECTOR(1, 5);
DECLARE_RESULT_VECTOR(1, 6);
DECLARE_RESULT_VECTOR(1, 7);
if (LIKELY(packed_a != NULL)) {
if (j == 0) {
for (; k < K; k++) {
BROADCAST_LOAD_B(0, 0);
GATHER_LOAD_A(pg_true, 0, 0);
VECTOR_PACK_A(0, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
BROADCAST_LOAD_B(1, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 1, 0);
GATHER_LOAD_A(pg_true, 1, 0);
VECTOR_PACK_A(1, 0);
UPDATE_RESULT_VECTOR(pg_true, 1, 0, 0);
UPDATE_RESULT_VECTOR(pg_true, 1, 1, 0);
BROADCAST_LOAD_B(2, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 2, 0);
UPDATE_RESULT_VECTOR(pg_true, 1, 2, 0);
BROADCAST_LOAD_B(3, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 3, 0);
UPDATE_RESULT_VECTOR(pg_true, 1, 3, 0);
BROADCAST_LOAD_B(4, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 4, 0);
UPDATE_RESULT_VECTOR(pg_true, 1, 4, 0);
BROADCAST_LOAD_B(5, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 5, 0);
UPDATE_RESULT_VECTOR(pg_true, 1, 5, 0);
BROADCAST_LOAD_B(6, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 6, 0);
UPDATE_RESULT_VECTOR(pg_true, 1, 6, 0);
BROADCAST_LOAD_B(7, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 7, 0);
UPDATE_RESULT_VECTOR(pg_true, 1, 7, 0);
}
} else {
for (; k < K; k++) {
BROADCAST_LOAD_B(0, 0);
UNPACK_VECTOR_A(0, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
BROADCAST_LOAD_B(1, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 1, 0);
UNPACK_VECTOR_A(1, 0);
UPDATE_RESULT_VECTOR(pg_true, 1, 0, 0);
UPDATE_RESULT_VECTOR(pg_true, 1, 1, 0);
BROADCAST_LOAD_B(2, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 2, 0);
UPDATE_RESULT_VECTOR(pg_true, 1, 2, 0);
BROADCAST_LOAD_B(3, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 3, 0);
UPDATE_RESULT_VECTOR(pg_true, 1, 3, 0);
BROADCAST_LOAD_B(4, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 4, 0);
UPDATE_RESULT_VECTOR(pg_true, 1, 4, 0);
BROADCAST_LOAD_B(5, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 5, 0);
UPDATE_RESULT_VECTOR(pg_true, 1, 5, 0);
BROADCAST_LOAD_B(6, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 6, 0);
UPDATE_RESULT_VECTOR(pg_true, 1, 6, 0);
BROADCAST_LOAD_B(7, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 7, 0);
UPDATE_RESULT_VECTOR(pg_true, 1, 7, 0);
}
}
} else {
for (; k < K; k++) {
BROADCAST_LOAD_B(0, 0);
GATHER_LOAD_A(pg_true, 0, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
BROADCAST_LOAD_B(1, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 1, 0);
GATHER_LOAD_A(pg_true, 1, 0);
UPDATE_RESULT_VECTOR(pg_true, 1, 0, 0);
UPDATE_RESULT_VECTOR(pg_true, 1, 1, 0);
BROADCAST_LOAD_B(2, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 2, 0);
UPDATE_RESULT_VECTOR(pg_true, 1, 2, 0);
BROADCAST_LOAD_B(3, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 3, 0);
UPDATE_RESULT_VECTOR(pg_true, 1, 3, 0);
BROADCAST_LOAD_B(4, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 4, 0);
UPDATE_RESULT_VECTOR(pg_true, 1, 4, 0);
BROADCAST_LOAD_B(5, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 5, 0);
UPDATE_RESULT_VECTOR(pg_true, 1, 5, 0);
BROADCAST_LOAD_B(6, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 6, 0);
UPDATE_RESULT_VECTOR(pg_true, 1, 6, 0);
BROADCAST_LOAD_B(7, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 7, 0);
UPDATE_RESULT_VECTOR(pg_true, 1, 7, 0);
}
}
VECTOR_STORE(pg_true, 0, 0);
VECTOR_STORE(pg_true, 0, 1);
VECTOR_STORE(pg_true, 0, 2);
VECTOR_STORE(pg_true, 0, 3);
VECTOR_STORE(pg_true, 0, 4);
VECTOR_STORE(pg_true, 0, 5);
VECTOR_STORE(pg_true, 0, 6);
VECTOR_STORE(pg_true, 0, 7);
VECTOR_STORE(pg_true, 1, 0);
VECTOR_STORE(pg_true, 1, 1);
VECTOR_STORE(pg_true, 1, 2);
VECTOR_STORE(pg_true, 1, 3);
VECTOR_STORE(pg_true, 1, 4);
VECTOR_STORE(pg_true, 1, 5);
VECTOR_STORE(pg_true, 1, 6);
VECTOR_STORE(pg_true, 1, 7);
INCR_C_POINTER(0, 8);
INCR_C_POINTER(1, 8);
}
for (; j < n4; j += 4) { for (; j < n4; j += 4) {
CREATE_B_POINTER(0, 0); CREATE_B_POINTER(0, 0);
@ -248,27 +388,6 @@ CNAME(BLASLONG M,
DECLARE_RESULT_VECTOR(1, 3); DECLARE_RESULT_VECTOR(1, 3);
if (LIKELY(packed_a != NULL)) { if (LIKELY(packed_a != NULL)) {
if (j == 0) {
for (; k < K; k++) {
BROADCAST_LOAD_B(0, 0);
GATHER_LOAD_A(pg_true, 0, 0);
VECTOR_PACK_A(0, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
BROADCAST_LOAD_B(1, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 1, 0);
GATHER_LOAD_A(pg_true, 1, 0);
VECTOR_PACK_A(1, 0);
UPDATE_RESULT_VECTOR(pg_true, 1, 0, 0);
UPDATE_RESULT_VECTOR(pg_true, 1, 1, 0);
BROADCAST_LOAD_B(2, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 2, 0);
UPDATE_RESULT_VECTOR(pg_true, 1, 2, 0);
BROADCAST_LOAD_B(3, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 3, 0);
UPDATE_RESULT_VECTOR(pg_true, 1, 3, 0);
}
} else {
for (; k < K; k++) { for (; k < K; k++) {
BROADCAST_LOAD_B(0, 0); BROADCAST_LOAD_B(0, 0);
@ -286,7 +405,6 @@ CNAME(BLASLONG M,
UPDATE_RESULT_VECTOR(pg_true, 0, 3, 0); UPDATE_RESULT_VECTOR(pg_true, 0, 3, 0);
UPDATE_RESULT_VECTOR(pg_true, 1, 3, 0); UPDATE_RESULT_VECTOR(pg_true, 1, 3, 0);
} }
}
} else { } else {
for (; k < K; k++) { for (; k < K; k++) {
@ -405,6 +523,55 @@ CNAME(BLASLONG M,
CREATE_A_POINTER(0, 0); CREATE_A_POINTER(0, 0);
BLASLONG j = 0; BLASLONG j = 0;
for (; j < n8; j += 8) {
CREATE_B_POINTER(0, 0);
CREATE_B_POINTER(1, 1);
CREATE_B_POINTER(2, 2);
CREATE_B_POINTER(3, 3);
CREATE_B_POINTER(4, 4);
CREATE_B_POINTER(5, 5);
CREATE_B_POINTER(6, 6);
CREATE_B_POINTER(7, 7);
UPDATE_B_POINTER(8);
BLASLONG k = 0;
DECLARE_RESULT_VECTOR(0, 0);
DECLARE_RESULT_VECTOR(0, 1);
DECLARE_RESULT_VECTOR(0, 2);
DECLARE_RESULT_VECTOR(0, 3);
DECLARE_RESULT_VECTOR(0, 4);
DECLARE_RESULT_VECTOR(0, 5);
DECLARE_RESULT_VECTOR(0, 6);
DECLARE_RESULT_VECTOR(0, 7);
for (; k < K; k++) {
BROADCAST_LOAD_B(0, 0);
GATHER_LOAD_A(pg_true, 0, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
BROADCAST_LOAD_B(1, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 1, 0);
BROADCAST_LOAD_B(2, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 2, 0);
BROADCAST_LOAD_B(3, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 3, 0);
BROADCAST_LOAD_B(4, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 4, 0);
BROADCAST_LOAD_B(5, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 5, 0);
BROADCAST_LOAD_B(6, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 6, 0);
BROADCAST_LOAD_B(7, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 7, 0);
}
VECTOR_STORE(pg_true, 0, 0);
VECTOR_STORE(pg_true, 0, 1);
VECTOR_STORE(pg_true, 0, 2);
VECTOR_STORE(pg_true, 0, 3);
VECTOR_STORE(pg_true, 0, 4);
VECTOR_STORE(pg_true, 0, 5);
VECTOR_STORE(pg_true, 0, 6);
VECTOR_STORE(pg_true, 0, 7);
INCR_C_POINTER(0, 8);
}
for (; j < n4; j += 4) { for (; j < n4; j += 4) {
CREATE_B_POINTER(0, 0); CREATE_B_POINTER(0, 0);
@ -487,6 +654,55 @@ CNAME(BLASLONG M,
CREATE_A_POINTER(0, 0); CREATE_A_POINTER(0, 0);
BLASLONG j = 0; BLASLONG j = 0;
for (; j < n8; j += 8) {
CREATE_B_POINTER(0, 0);
CREATE_B_POINTER(1, 1);
CREATE_B_POINTER(2, 2);
CREATE_B_POINTER(3, 3);
CREATE_B_POINTER(4, 4);
CREATE_B_POINTER(5, 5);
CREATE_B_POINTER(6, 6);
CREATE_B_POINTER(7, 7);
UPDATE_B_POINTER(8);
BLASLONG k = 0;
DECLARE_RESULT_VECTOR(0, 0);
DECLARE_RESULT_VECTOR(0, 1);
DECLARE_RESULT_VECTOR(0, 2);
DECLARE_RESULT_VECTOR(0, 3);
DECLARE_RESULT_VECTOR(0, 4);
DECLARE_RESULT_VECTOR(0, 5);
DECLARE_RESULT_VECTOR(0, 6);
DECLARE_RESULT_VECTOR(0, 7);
for (; k < K; k++) {
BROADCAST_LOAD_B(0, 0);
GATHER_LOAD_A(pg_tail, 0, 0);
UPDATE_RESULT_VECTOR(pg_tail, 0, 0, 0);
BROADCAST_LOAD_B(1, 0);
UPDATE_RESULT_VECTOR(pg_tail, 0, 1, 0);
BROADCAST_LOAD_B(2, 0);
UPDATE_RESULT_VECTOR(pg_tail, 0, 2, 0);
BROADCAST_LOAD_B(3, 0);
UPDATE_RESULT_VECTOR(pg_tail, 0, 3, 0);
BROADCAST_LOAD_B(4, 0);
UPDATE_RESULT_VECTOR(pg_tail, 0, 4, 0);
BROADCAST_LOAD_B(5, 0);
UPDATE_RESULT_VECTOR(pg_tail, 0, 5, 0);
BROADCAST_LOAD_B(6, 0);
UPDATE_RESULT_VECTOR(pg_tail, 0, 6, 0);
BROADCAST_LOAD_B(7, 0);
UPDATE_RESULT_VECTOR(pg_tail, 0, 7, 0);
}
VECTOR_STORE(pg_tail, 0, 0);
VECTOR_STORE(pg_tail, 0, 1);
VECTOR_STORE(pg_tail, 0, 2);
VECTOR_STORE(pg_tail, 0, 3);
VECTOR_STORE(pg_tail, 0, 4);
VECTOR_STORE(pg_tail, 0, 5);
VECTOR_STORE(pg_tail, 0, 6);
VECTOR_STORE(pg_tail, 0, 7);
INCR_C_POINTER(0, 8);
}
for (; j < n4; j += 4) { for (; j < n4; j += 4) {
CREATE_B_POINTER(0, 0); CREATE_B_POINTER(0, 0);