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

disambiguate whilelt calls

This commit is contained in:
Chris Sidebottom 2024-07-25 00:50:27 +01:00
parent a5b9a0d358
commit 3c3d439721
2 changed files with 8 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
kernel/arm64/gemv_n_sve.c
View File

@ -95,7 +95,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
} }
for (; i < M; i += v_size) { for (; i < M; i += v_size) {
svbool_t pg = SV_WHILE(i, M); svbool_t pg = SV_WHILE((uint64_t)i, (uint64_t)M);
SV_TYPE a_vec1 = svld1(pg, a + i); SV_TYPE a_vec1 = svld1(pg, a + i);
SV_TYPE a_vec2 = svld1(pg, a + i + lda); SV_TYPE a_vec2 = svld1(pg, a + i + lda);
SV_TYPE a_vec3 = svld1(pg, a + i + lda * 2); SV_TYPE a_vec3 = svld1(pg, a + i + lda * 2);
@ -138,7 +138,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
} }
for (; i < M; i += v_size) { for (; i < M; i += v_size) {
svbool_t pg = SV_WHILE(i, M); svbool_t pg = SV_WHILE((uint64_t)i, (uint64_t)M);
SV_TYPE a_vec1 = svld1(pg, a + i); SV_TYPE a_vec1 = svld1(pg, a + i);
SV_TYPE a_vec2 = svld1(pg, a + i + lda); SV_TYPE a_vec2 = svld1(pg, a + i + lda);
SV_TYPE a_vec3 = svld1(pg, a + i + lda * 2); SV_TYPE a_vec3 = svld1(pg, a + i + lda * 2);
@ -174,7 +174,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
svst1(pg_true, y + i, y_vec); svst1(pg_true, y + i, y_vec);
} }
for (; i < M; i += v_size) { for (; i < M; i += v_size) {
svbool_t pg = SV_WHILE(i, M); svbool_t pg = SV_WHILE((uint64_t)i, (uint64_t)M);
SV_TYPE a_vec1 = svld1(pg, a + i); SV_TYPE a_vec1 = svld1(pg, a + i);
SV_TYPE a_vec2 = svld1(pg, a + i + lda); SV_TYPE a_vec2 = svld1(pg, a + i + lda);
SV_TYPE a_vec3 = svld1(pg, a + i + lda * 2); SV_TYPE a_vec3 = svld1(pg, a + i + lda * 2);
@ -203,7 +203,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
svst1(pg_true, y + i, y_vec); svst1(pg_true, y + i, y_vec);
} }
for (; i < M; i += v_size) { for (; i < M; i += v_size) {
svbool_t pg = SV_WHILE(i, M); svbool_t pg = SV_WHILE((uint64_t)i, (uint64_t)M);
SV_TYPE a_vec1 = svld1(pg, a + i); SV_TYPE a_vec1 = svld1(pg, a + i);
SV_TYPE a_vec2 = svld1(pg, a + i + lda); SV_TYPE a_vec2 = svld1(pg, a + i + lda);
SV_TYPE y_vec = svld1(pg, y + i); SV_TYPE y_vec = svld1(pg, y + i);
@ -230,7 +230,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
svst1(pg_true, y + i, y_vec); svst1(pg_true, y + i, y_vec);
} }
for (; i < M; i += v_size) { for (; i < M; i += v_size) {
svbool_t pg = SV_WHILE(i, M); svbool_t pg = SV_WHILE((uint64_t)i, (uint64_t)M);
SV_TYPE a_vec = svld1(pg, a + i); SV_TYPE a_vec = svld1(pg, a + i);
SV_TYPE y_vec = svld1(pg, y + i); SV_TYPE y_vec = svld1(pg, y + i);
y_vec = svmla_x(pg, y_vec, temp_vec1, a_vec); y_vec = svmla_x(pg, y_vec, temp_vec1, a_vec);

6
kernel/arm64/gemv_t_sve.c
View File

@ -82,7 +82,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
} }
for (; i < M; i += v_size) { for (; i < M; i += v_size) {
svbool_t pg = SV_WHILE(i, M); svbool_t pg = SV_WHILE((uint64_t)i, (uint64_t)M);
SV_TYPE a_vec1 = svld1(pg, a + i); SV_TYPE a_vec1 = svld1(pg, a + i);
SV_TYPE a_vec2 = svld1(pg, a + i + lda); SV_TYPE a_vec2 = svld1(pg, a + i + lda);
SV_TYPE a_vec3 = svld1(pg, a + i + lda * 2); SV_TYPE a_vec3 = svld1(pg, a + i + lda * 2);
@ -120,7 +120,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
temp_vec2 = svmla_x(pg_true, temp_vec2, a_vec2, x_vec); temp_vec2 = svmla_x(pg_true, temp_vec2, a_vec2, x_vec);
} }
for (; i < M; i += v_size) { for (; i < M; i += v_size) {
svbool_t pg = SV_WHILE(i, M); svbool_t pg = SV_WHILE((uint64_t)i, (uint64_t)M);
SV_TYPE a_vec1 = svld1(pg, a + i); SV_TYPE a_vec1 = svld1(pg, a + i);
SV_TYPE a_vec2 = svld1(pg, a + lda + i); SV_TYPE a_vec2 = svld1(pg, a + lda + i);
SV_TYPE x_vec = svld1(pg, x + i); SV_TYPE x_vec = svld1(pg, x + i);
@ -156,7 +156,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
temp_vec = svmla_x(pg_true, temp_vec, a_vec, x_vec); temp_vec = svmla_x(pg_true, temp_vec, a_vec, x_vec);
} }
for (; i < M; i += v_size) { for (; i < M; i += v_size) {
svbool_t pg = SV_WHILE(i, M); svbool_t pg = SV_WHILE((uint64_t)i, (uint64_t)M);
SV_TYPE a_vec = svld1(pg, a + i); SV_TYPE a_vec = svld1(pg, a + i);
SV_TYPE x_vec = svld1(pg, x + i); SV_TYPE x_vec = svld1(pg, x + i);
temp_vec = svmla_x(pg, temp_vec, a_vec, x_vec); temp_vec = svmla_x(pg, temp_vec, a_vec, x_vec);