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

Further tweak small GEMM for AArch64

Browse Source
This commit is contained in:
Chris Sidebottom
2024-06-24 10:47:47 +01:00
parent 7a6fa699f2
commit 8c472ef7e3
9 changed files with 772 additions and 2286 deletions
Download Patch File Download Diff File Expand all files Collapse all files
kernel/arm64/sgemm_small_kernel_nn_sve.c
+76 -597
View File
@@ -46,15 +46,29 @@ THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
})
#endif
#define A_ELEMENT_K(m, offset_k) A[(i + (m)) + (k + offset_k) * lda]
#define RESET_A_POINTER() a_offset = A;
#define CREATE_A_POINTER(m, scale) FLOAT* a_offset##m = a_offset + scale;
#define UPDATE_A_POINTER(scale) a_offset = a_offset + scale;
#define A_ELEMENT_K(m, offset_k) *(a_offset##m + (k + offset_k) * lda)
#define A_ELEMENT(m) A_ELEMENT_K(m, 0)
#define B_ELEMENT_K(n, offset_k) B[(k + offset_k) + (j + (n)) * ldb]
#define RESET_B_POINTER() b_offset = B;
#define CREATE_B_POINTER(n, scale) FLOAT* b_offset##n = b_offset + scale * ldb;
#define UPDATE_B_POINTER(scale) b_offset = b_offset + scale * ldb;
#define B_ELEMENT_K(n, offset_k) *(b_offset##n + (k + offset_k))
#define B_ELEMENT(n) B_ELEMENT_K(n, 0)
#define C_ELEMENT(m, n) C[(i + (m)) + (j + (n)) * ldc]
#define CREATE_C_POINTER(n, scale) FLOAT* c_offset##n = c_offset + scale * ldc;
#define INCR_C_POINTER(m, incr) // c_offset ## m += incr;
#define UPDATE_C_POINTER(scale) c_offset = c_offset + scale * ldc;
#define C_ELEMENT(m, n) *(c_offset##n + ((m * v_size) + i))
#define PACK_ELEMENT_K(n, offset_k) packed_b[(k + offset_k) * 8 + n]
// #undef C_ELEMENT
// #define C_ELEMENT(m, n) C[(i+(m))+(j+(n))*ldc]
#define PACK_ELEMENT_K(n, offset_k) packed_b[(k + offset_k) * 4 + n]
#define PACK_ELEMENT(n) PACK_ELEMENT_K(n, 0)
// ASIMD
@@ -141,8 +155,7 @@ THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define BROADCAST_LOAD_B(n, offset_k) \
svfloat32_t b##s##n##_k##offset_k = svdup_f32(B_ELEMENT_K(n, offset_k));
#define VECTOR_LOAD_A(pg, m, offset_k) \
svfloat32_t a##s##m##_k##offset_k = \
svld1(pg, &A_ELEMENT_K(v_size * m, offset_k));
svfloat32_t a##s##m##_k##offset_k = svld1(pg, &A_ELEMENT_K(m, offset_k));
#define QUADWORD_LOAD_B(n, offset_k) \
svfloat32_t b##s##n##_k##offset_k = \
svld1rq(pg_true, &B_ELEMENT_K(n, offset_k));
@@ -169,26 +182,23 @@ THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#ifdef B0
#define VECTOR_STORE(pg, m, n) \
result##m##n = svmul_m(pg, result##m##n, alpha_vec); \
svst1(pg, &C_ELEMENT(v_size* m, n), result##m##n);
svst1(pg, &C_ELEMENT(m, n), result##m##n);
#define SCATTER_STORE(pg, m, n) \
result##m##n = svmul_m(pg, result##m##n, alpha_vec); \
svst1_scatter_index( \
pg, &C_ELEMENT(v_size* m, n), svindex_u32(0LL, ldc), result##m##n);
svst1_scatter_index(pg, &C_ELEMENT(m, n), ldc_vec, result##m##n);
#else
#define VECTOR_STORE(pg, m, n) \
result##m##n = svmul_m(pg, result##m##n, alpha_vec); \
result##m##n = \
svmla_m(pg, result##m##n, svld1(pg, &C_ELEMENT(v_size * m, n)), beta_vec); \
svst1(pg, &C_ELEMENT(v_size* m, n), result##m##n);
svmla_m(pg, result##m##n, svld1(pg, &C_ELEMENT(m, n)), beta_vec); \
svst1(pg, &C_ELEMENT(m, n), result##m##n);
#define SCATTER_STORE(pg, m, n) \
result##m##n = svmul_m(pg, result##m##n, alpha_vec); \
result##m##n = svmla_m( \
pg, \
result##m##n, \
svld1_gather_index(pg, &C_ELEMENT(v_size * m, n), svindex_u32(0LL, ldc)), \
beta_vec); \
svst1_scatter_index( \
pg, &C_ELEMENT(v_size* m, n), svindex_u32(0LL, ldc), result##m##n);
result##m##n = svmla_m(pg, \
result##m##n, \
svld1_gather_index(pg, &C_ELEMENT(m, n), ldc_vec), \
beta_vec); \
svst1_scatter_index(pg, &C_ELEMENT(m, n), ldc_vec, result##m##n);
#endif
#ifndef LIKELY
@@ -198,13 +208,6 @@ THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define LIKELY(x) (x)
#endif
#endif
#ifndef UNLIKELY
#ifdef __GNUC__
#define UNLIKELY(x) __builtin_expect(!!(x), 0)
#else
#define UNLIKELY(x) (x)
#endif
#endif
#ifdef B0
int
@@ -234,7 +237,6 @@ CNAME(BLASLONG M,
#endif
{
const uint64_t v_size = svcntw();
const uint64_t v_size2 = v_size * 2;
const svbool_t pg_true = svptrue_b32();
const svbool_t pg_quad = svwhilelt_b32(0, 4);
const svbool_t pg_first = svwhilelt_b32(0, 1);
@@ -242,39 +244,41 @@ CNAME(BLASLONG M,
#ifndef B0
const svfloat32_t beta_vec = svdup_f32(beta);
#endif
const BLASLONG n8 = N & -8;
const BLASLONG n4 = N & -4;
const BLASLONG v_m2 = M & -v_size2;
const BLASLONG v_m1 = M & -v_size;
const BLASLONG k4 = K & -4;
const int pack_b = M >= v_size2 && N >= 8 && K >= 8 ? 1 : 0;
const int pack_b = M >= v_size && N >= 8 && K >= 8 ? 1 : 0;
FLOAT* packed_b =
(pack_b) ? packed_b = (FLOAT*)malloc(K * 8 * sizeof(FLOAT)) : NULL;
(pack_b) ? packed_b = (FLOAT*)malloc(K * 4 * sizeof(FLOAT)) : NULL;
FLOAT* b_offset = B;
FLOAT* a_offset = A;
FLOAT* c_offset = C;
BLASLONG j = 0;
for (; j < n8; j += 8) {
for (; j < n4; j += 4) {
CREATE_C_POINTER(0, 0);
CREATE_C_POINTER(1, 1);
CREATE_C_POINTER(2, 2);
CREATE_C_POINTER(3, 3);
CREATE_B_POINTER(0, 0);
CREATE_B_POINTER(1, 1);
CREATE_B_POINTER(2, 2);
CREATE_B_POINTER(3, 3);
BLASLONG i = 0;
for (; i < v_m2; i += v_size2) {
for (; i < v_m1; i += v_size) {
CREATE_A_POINTER(0, 0);
UPDATE_A_POINTER(v_size);
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_b != NULL)) {
if (i == 0) {
@@ -310,68 +314,6 @@ CNAME(BLASLONG M,
UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 3);
VECTOR_LOAD_B_K4(4, 0);
VECTOR_LOAD_B_K4(5, 0);
VECTOR_LOAD_B_K4(6, 0);
VECTOR_LOAD_B_K4(7, 0);
TRANSPOSE_B4_K4(4, 5, 6, 7, 0, 1, 2, 3);
SCALE_B4_K4(4, 0, 1, 2, 3);
VECTOR_PACK_B4(4, 0);
VECTOR_PACK_B4(4, 1);
VECTOR_PACK_B4(4, 2);
VECTOR_PACK_B4(4, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 4, 4, 0, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 5, 4, 1, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 6, 4, 2, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 7, 4, 3, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 4, 4, 0, 1);
UPDATE_RESULT_VECTOR_QUADWORD(0, 5, 4, 1, 1);
UPDATE_RESULT_VECTOR_QUADWORD(0, 6, 4, 2, 1);
UPDATE_RESULT_VECTOR_QUADWORD(0, 7, 4, 3, 1);
UPDATE_RESULT_VECTOR_QUADWORD(0, 4, 4, 0, 2);
UPDATE_RESULT_VECTOR_QUADWORD(0, 5, 4, 1, 2);
UPDATE_RESULT_VECTOR_QUADWORD(0, 6, 4, 2, 2);
UPDATE_RESULT_VECTOR_QUADWORD(0, 7, 4, 3, 2);
UPDATE_RESULT_VECTOR_QUADWORD(0, 4, 4, 0, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 5, 4, 1, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 6, 4, 2, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 7, 4, 3, 3);
VECTOR_LOAD_A(pg_true, 1, 0);
UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 0);
UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 0);
UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 0, 2, 0);
UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 0, 3, 0);
UPDATE_RESULT_VECTOR_QUADWORD(1, 4, 4, 0, 0);
UPDATE_RESULT_VECTOR_QUADWORD(1, 5, 4, 1, 0);
UPDATE_RESULT_VECTOR_QUADWORD(1, 6, 4, 2, 0);
UPDATE_RESULT_VECTOR_QUADWORD(1, 7, 4, 3, 0);
VECTOR_LOAD_A(pg_true, 1, 1);
UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 1);
UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 1);
UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 0, 2, 1);
UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 0, 3, 1);
UPDATE_RESULT_VECTOR_QUADWORD(1, 4, 4, 0, 1);
UPDATE_RESULT_VECTOR_QUADWORD(1, 5, 4, 1, 1);
UPDATE_RESULT_VECTOR_QUADWORD(1, 6, 4, 2, 1);
UPDATE_RESULT_VECTOR_QUADWORD(1, 7, 4, 3, 1);
VECTOR_LOAD_A(pg_true, 1, 2);
UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 2);
UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 2);
UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 0, 2, 2);
UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 0, 3, 2);
UPDATE_RESULT_VECTOR_QUADWORD(1, 4, 4, 0, 2);
UPDATE_RESULT_VECTOR_QUADWORD(1, 5, 4, 1, 2);
UPDATE_RESULT_VECTOR_QUADWORD(1, 6, 4, 2, 2);
UPDATE_RESULT_VECTOR_QUADWORD(1, 7, 4, 3, 2);
VECTOR_LOAD_A(pg_true, 1, 3);
UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 3);
UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 3);
UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 0, 2, 3);
UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 0, 3, 3);
UPDATE_RESULT_VECTOR_QUADWORD(1, 4, 4, 0, 3);
UPDATE_RESULT_VECTOR_QUADWORD(1, 5, 4, 1, 3);
UPDATE_RESULT_VECTOR_QUADWORD(1, 6, 4, 2, 3);
UPDATE_RESULT_VECTOR_QUADWORD(1, 7, 4, 3, 3);
}
for (; k < K; k++) {
@@ -382,33 +324,12 @@ CNAME(BLASLONG M,
BROADCAST_LOAD_B(1, 0);
PACK_B(1, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 1, 0);
VECTOR_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);
PACK_B(2, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 2, 0);
UPDATE_RESULT_VECTOR(pg_true, 1, 2, 0);
BROADCAST_LOAD_B(3, 0);
PACK_B(3, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 3, 0);
UPDATE_RESULT_VECTOR(pg_true, 1, 3, 0);
BROADCAST_LOAD_B(4, 0);
PACK_B(4, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 4, 0);
UPDATE_RESULT_VECTOR(pg_true, 1, 4, 0);
BROADCAST_LOAD_B(5, 0);
PACK_B(5, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 5, 0);
UPDATE_RESULT_VECTOR(pg_true, 1, 5, 0);
BROADCAST_LOAD_B(6, 0);
PACK_B(6, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 6, 0);
UPDATE_RESULT_VECTOR(pg_true, 1, 6, 0);
BROADCAST_LOAD_B(7, 0);
PACK_B(7, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 7, 0);
UPDATE_RESULT_VECTOR(pg_true, 1, 7, 0);
}
} else {
for (; k < K; k++) {
@@ -419,20 +340,6 @@ CNAME(BLASLONG M,
UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 0);
UNPACK_QUADWORD_B(4, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 4, 4, 0, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 5, 4, 1, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 6, 4, 2, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 7, 4, 3, 0);
VECTOR_LOAD_A(pg_true, 1, 0);
UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 0);
UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 0);
UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 0, 2, 0);
UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 0, 3, 0);
UPDATE_RESULT_VECTOR_QUADWORD(1, 4, 4, 0, 0);
UPDATE_RESULT_VECTOR_QUADWORD(1, 5, 4, 1, 0);
UPDATE_RESULT_VECTOR_QUADWORD(1, 6, 4, 2, 0);
UPDATE_RESULT_VECTOR_QUADWORD(1, 7, 4, 3, 0);
}
}
} else {
@@ -464,190 +371,6 @@ CNAME(BLASLONG M,
UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 3);
VECTOR_LOAD_B_K4(4, 0);
VECTOR_LOAD_B_K4(5, 0);
VECTOR_LOAD_B_K4(6, 0);
VECTOR_LOAD_B_K4(7, 0);
TRANSPOSE_B4_K4(4, 5, 6, 7, 0, 1, 2, 3);
SCALE_B4_K4(4, 0, 1, 2, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 4, 4, 0, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 5, 4, 1, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 6, 4, 2, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 7, 4, 3, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 4, 4, 0, 1);
UPDATE_RESULT_VECTOR_QUADWORD(0, 5, 4, 1, 1);
UPDATE_RESULT_VECTOR_QUADWORD(0, 6, 4, 2, 1);
UPDATE_RESULT_VECTOR_QUADWORD(0, 7, 4, 3, 1);
UPDATE_RESULT_VECTOR_QUADWORD(0, 4, 4, 0, 2);
UPDATE_RESULT_VECTOR_QUADWORD(0, 5, 4, 1, 2);
UPDATE_RESULT_VECTOR_QUADWORD(0, 6, 4, 2, 2);
UPDATE_RESULT_VECTOR_QUADWORD(0, 7, 4, 3, 2);
UPDATE_RESULT_VECTOR_QUADWORD(0, 4, 4, 0, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 5, 4, 1, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 6, 4, 2, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 7, 4, 3, 3);
VECTOR_LOAD_A(pg_true, 1, 0);
UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 0);
UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 0);
UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 0, 2, 0);
UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 0, 3, 0);
UPDATE_RESULT_VECTOR_QUADWORD(1, 4, 4, 0, 0);
UPDATE_RESULT_VECTOR_QUADWORD(1, 5, 4, 1, 0);
UPDATE_RESULT_VECTOR_QUADWORD(1, 6, 4, 2, 0);
UPDATE_RESULT_VECTOR_QUADWORD(1, 7, 4, 3, 0);
VECTOR_LOAD_A(pg_true, 1, 1);
UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 1);
UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 1);
UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 0, 2, 1);
UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 0, 3, 1);
UPDATE_RESULT_VECTOR_QUADWORD(1, 4, 4, 0, 1);
UPDATE_RESULT_VECTOR_QUADWORD(1, 5, 4, 1, 1);
UPDATE_RESULT_VECTOR_QUADWORD(1, 6, 4, 2, 1);
UPDATE_RESULT_VECTOR_QUADWORD(1, 7, 4, 3, 1);
VECTOR_LOAD_A(pg_true, 1, 2);
UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 2);
UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 2);
UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 0, 2, 2);
UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 0, 3, 2);
UPDATE_RESULT_VECTOR_QUADWORD(1, 4, 4, 0, 2);
UPDATE_RESULT_VECTOR_QUADWORD(1, 5, 4, 1, 2);
UPDATE_RESULT_VECTOR_QUADWORD(1, 6, 4, 2, 2);
UPDATE_RESULT_VECTOR_QUADWORD(1, 7, 4, 3, 2);
VECTOR_LOAD_A(pg_true, 1, 3);
UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 3);
UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 3);
UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 0, 2, 3);
UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 0, 3, 3);
UPDATE_RESULT_VECTOR_QUADWORD(1, 4, 4, 0, 3);
UPDATE_RESULT_VECTOR_QUADWORD(1, 5, 4, 1, 3);
UPDATE_RESULT_VECTOR_QUADWORD(1, 6, 4, 2, 3);
UPDATE_RESULT_VECTOR_QUADWORD(1, 7, 4, 3, 3);
}
for (; k < K; k++) {
BROADCAST_LOAD_B(0, 0);
VECTOR_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);
VECTOR_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);
}
for (; i < v_m1; i += v_size) {
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);
if (LIKELY(packed_b != NULL)) {
for (; k < K; k++) {
UNPACK_QUADWORD_B(0, 0);
VECTOR_LOAD_A(pg_true, 0, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 0);
UNPACK_QUADWORD_B(4, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 4, 4, 0, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 5, 4, 1, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 6, 4, 2, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 7, 4, 3, 0);
}
} else {
for (; k < k4; k += 4) {
VECTOR_LOAD_B_K4(0, 0);
VECTOR_LOAD_B_K4(1, 0);
VECTOR_LOAD_B_K4(2, 0);
VECTOR_LOAD_B_K4(3, 0);
TRANSPOSE_B4_K4(0, 1, 2, 3, 0, 1, 2, 3);
SCALE_B4_K4(0, 0, 1, 2, 3);
VECTOR_LOAD_A(pg_true, 0, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 0);
VECTOR_LOAD_A(pg_true, 0, 1);
UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 1);
UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 1);
UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 1);
UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 1);
VECTOR_LOAD_A(pg_true, 0, 2);
UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 2);
UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 2);
UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 2);
UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 2);
VECTOR_LOAD_A(pg_true, 0, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 3);
VECTOR_LOAD_B_K4(4, 0);
VECTOR_LOAD_B_K4(5, 0);
VECTOR_LOAD_B_K4(6, 0);
VECTOR_LOAD_B_K4(7, 0);
TRANSPOSE_B4_K4(4, 5, 6, 7, 0, 1, 2, 3);
SCALE_B4_K4(4, 0, 1, 2, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 4, 4, 0, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 5, 4, 1, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 6, 4, 2, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 7, 4, 3, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 4, 4, 0, 1);
UPDATE_RESULT_VECTOR_QUADWORD(0, 5, 4, 1, 1);
UPDATE_RESULT_VECTOR_QUADWORD(0, 6, 4, 2, 1);
UPDATE_RESULT_VECTOR_QUADWORD(0, 7, 4, 3, 1);
UPDATE_RESULT_VECTOR_QUADWORD(0, 4, 4, 0, 2);
UPDATE_RESULT_VECTOR_QUADWORD(0, 5, 4, 1, 2);
UPDATE_RESULT_VECTOR_QUADWORD(0, 6, 4, 2, 2);
UPDATE_RESULT_VECTOR_QUADWORD(0, 7, 4, 3, 2);
UPDATE_RESULT_VECTOR_QUADWORD(0, 4, 4, 0, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 5, 4, 1, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 6, 4, 2, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 7, 4, 3, 3);
}
for (; k < K; k++) {
@@ -660,37 +383,27 @@ CNAME(BLASLONG M,
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, v_size);
INCR_C_POINTER(1, v_size);
INCR_C_POINTER(2, v_size);
INCR_C_POINTER(3, v_size);
}
for (; i < M; i += v_size) {
const svbool_t pg_tail = svwhilelt_b32((uint32_t)i, (uint32_t)(M));
CREATE_A_POINTER(0, 0);
UPDATE_A_POINTER(0);
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);
if (LIKELY(packed_b != NULL)) {
for (; k < K; k++) {
@@ -701,11 +414,6 @@ CNAME(BLASLONG M,
UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 0);
UNPACK_QUADWORD_B(4, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 4, 4, 0, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 5, 4, 1, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 6, 4, 2, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 7, 4, 3, 0);
}
} else {
for (; k < k4; k += 4) {
@@ -736,28 +444,6 @@ CNAME(BLASLONG M,
UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 3);
VECTOR_LOAD_B_K4(4, 0);
VECTOR_LOAD_B_K4(5, 0);
VECTOR_LOAD_B_K4(6, 0);
VECTOR_LOAD_B_K4(7, 0);
TRANSPOSE_B4_K4(4, 5, 6, 7, 0, 1, 2, 3);
SCALE_B4_K4(4, 0, 1, 2, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 4, 4, 0, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 5, 4, 1, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 6, 4, 2, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 7, 4, 3, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 4, 4, 0, 1);
UPDATE_RESULT_VECTOR_QUADWORD(0, 5, 4, 1, 1);
UPDATE_RESULT_VECTOR_QUADWORD(0, 6, 4, 2, 1);
UPDATE_RESULT_VECTOR_QUADWORD(0, 7, 4, 3, 1);
UPDATE_RESULT_VECTOR_QUADWORD(0, 4, 4, 0, 2);
UPDATE_RESULT_VECTOR_QUADWORD(0, 5, 4, 1, 2);
UPDATE_RESULT_VECTOR_QUADWORD(0, 6, 4, 2, 2);
UPDATE_RESULT_VECTOR_QUADWORD(0, 7, 4, 3, 2);
UPDATE_RESULT_VECTOR_QUADWORD(0, 4, 4, 0, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 5, 4, 1, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 6, 4, 2, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 7, 4, 3, 3);
}
for (; k < K; k++) {
@@ -770,248 +456,33 @@ CNAME(BLASLONG M,
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, 0);
INCR_C_POINTER(1, 0);
INCR_C_POINTER(2, 0);
INCR_C_POINTER(3, 0);
}
}
for (; j < n4; j += 4) {
BLASLONG i = 0;
for (; i < v_m2; i += v_size2) {
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(1, 0);
DECLARE_RESULT_VECTOR(1, 1);
DECLARE_RESULT_VECTOR(1, 2);
DECLARE_RESULT_VECTOR(1, 3);
for (; k < k4; k += 4) {
VECTOR_LOAD_B_K4(0, 0);
VECTOR_LOAD_B_K4(1, 0);
VECTOR_LOAD_B_K4(2, 0);
VECTOR_LOAD_B_K4(3, 0);
TRANSPOSE_B4_K4(0, 1, 2, 3, 0, 1, 2, 3);
SCALE_B4_K4(0, 0, 1, 2, 3);
VECTOR_LOAD_A(pg_true, 0, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 0);
VECTOR_LOAD_A(pg_true, 0, 1);
UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 1);
UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 1);
UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 1);
UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 1);
VECTOR_LOAD_A(pg_true, 0, 2);
UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 2);
UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 2);
UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 2);
UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 2);
VECTOR_LOAD_A(pg_true, 0, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 3);
VECTOR_LOAD_A(pg_true, 1, 0);
UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 0);
UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 0);
UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 0, 2, 0);
UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 0, 3, 0);
VECTOR_LOAD_A(pg_true, 1, 1);
UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 1);
UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 1);
UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 0, 2, 1);
UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 0, 3, 1);
VECTOR_LOAD_A(pg_true, 1, 2);
UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 2);
UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 2);
UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 0, 2, 2);
UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 0, 3, 2);
VECTOR_LOAD_A(pg_true, 1, 3);
UPDATE_RESULT_VECTOR_QUADWORD(1, 0, 0, 0, 3);
UPDATE_RESULT_VECTOR_QUADWORD(1, 1, 0, 1, 3);
UPDATE_RESULT_VECTOR_QUADWORD(1, 2, 0, 2, 3);
UPDATE_RESULT_VECTOR_QUADWORD(1, 3, 0, 3, 3);
}
for (; k < K; k++) {
BROADCAST_LOAD_B(0, 0);
VECTOR_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);
VECTOR_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);
}
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, 1, 0);
VECTOR_STORE(pg_true, 1, 1);
VECTOR_STORE(pg_true, 1, 2);
VECTOR_STORE(pg_true, 1, 3);
}
for (; i < v_m1; i += v_size) {
BLASLONG k = 0;
DECLARE_RESULT_VECTOR(0, 0);
DECLARE_RESULT_VECTOR(0, 1);
DECLARE_RESULT_VECTOR(0, 2);
DECLARE_RESULT_VECTOR(0, 3);
for (; k < k4; k += 4) {
VECTOR_LOAD_B_K4(0, 0);
VECTOR_LOAD_B_K4(1, 0);
VECTOR_LOAD_B_K4(2, 0);
VECTOR_LOAD_B_K4(3, 0);
TRANSPOSE_B4_K4(0, 1, 2, 3, 0, 1, 2, 3);
SCALE_B4_K4(0, 0, 1, 2, 3);
VECTOR_LOAD_A(pg_true, 0, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 0);
VECTOR_LOAD_A(pg_true, 0, 1);
UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 1);
UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 1);
UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 1);
UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 1);
VECTOR_LOAD_A(pg_true, 0, 2);
UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 2);
UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 2);
UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 2);
UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 2);
VECTOR_LOAD_A(pg_true, 0, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 3);
}
for (; k < K; k++) {
BROADCAST_LOAD_B(0, 0);
VECTOR_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);
}
VECTOR_STORE(pg_true, 0, 0);
VECTOR_STORE(pg_true, 0, 1);
VECTOR_STORE(pg_true, 0, 2);
VECTOR_STORE(pg_true, 0, 3);
}
for (; i < M; i += v_size) {
const svbool_t pg_tail = svwhilelt_b32((uint32_t)i, (uint32_t)(M));
BLASLONG k = 0;
DECLARE_RESULT_VECTOR(0, 0);
DECLARE_RESULT_VECTOR(0, 1);
DECLARE_RESULT_VECTOR(0, 2);
DECLARE_RESULT_VECTOR(0, 3);
for (; k < k4; k += 4) {
VECTOR_LOAD_B_K4(0, 0);
VECTOR_LOAD_B_K4(1, 0);
VECTOR_LOAD_B_K4(2, 0);
VECTOR_LOAD_B_K4(3, 0);
TRANSPOSE_B4_K4(0, 1, 2, 3, 0, 1, 2, 3);
SCALE_B4_K4(0, 0, 1, 2, 3);
VECTOR_LOAD_A(pg_tail, 0, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 0);
UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 0);
VECTOR_LOAD_A(pg_tail, 0, 1);
UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 1);
UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 1);
UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 1);
UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 1);
VECTOR_LOAD_A(pg_tail, 0, 2);
UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 2);
UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 2);
UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 2);
UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 2);
VECTOR_LOAD_A(pg_tail, 0, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 0, 0, 0, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 1, 0, 1, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 2, 0, 2, 3);
UPDATE_RESULT_VECTOR_QUADWORD(0, 3, 0, 3, 3);
}
for (; k < K; k++) {
BROADCAST_LOAD_B(0, 0);
VECTOR_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);
}
VECTOR_STORE(pg_tail, 0, 0);
VECTOR_STORE(pg_tail, 0, 1);
VECTOR_STORE(pg_tail, 0, 2);
VECTOR_STORE(pg_tail, 0, 3);
}
UPDATE_B_POINTER(4);
RESET_A_POINTER();
UPDATE_C_POINTER(4);
}
for (; j < N; j++) {
CREATE_C_POINTER(0, 0);
CREATE_B_POINTER(0, 0);
BLASLONG i = 0;
for (; i < v_m2; i += v_size2) {
BLASLONG k = 0;
DECLARE_RESULT_VECTOR(0, 0);
DECLARE_RESULT_VECTOR(1, 0);
for (; k < K; k++) {
BROADCAST_LOAD_B(0, 0);
VECTOR_LOAD_A(pg_true, 0, 0);
UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
VECTOR_LOAD_A(pg_true, 1, 0);
UPDATE_RESULT_VECTOR(pg_true, 1, 0, 0);
}
VECTOR_STORE(pg_true, 0, 0);
VECTOR_STORE(pg_true, 1, 0);
}
for (; i < v_m1; i += v_size) {
CREATE_A_POINTER(0, 0);
UPDATE_A_POINTER(v_size);
BLASLONG k = 0;
DECLARE_RESULT_VECTOR(0, 0);
@@ -1022,9 +493,12 @@ CNAME(BLASLONG M,
UPDATE_RESULT_VECTOR(pg_true, 0, 0, 0);
}
VECTOR_STORE(pg_true, 0, 0);
INCR_C_POINTER(0, v_size);
}
for (; i < M; i += v_size) {
const svbool_t pg_tail = svwhilelt_b32((uint32_t)i, (uint32_t)(M));
CREATE_A_POINTER(0, 0);
UPDATE_A_POINTER(0);
BLASLONG k = 0;
DECLARE_RESULT_VECTOR(0, 0);
@@ -1036,11 +510,16 @@ CNAME(BLASLONG M,
UPDATE_RESULT_VECTOR(pg_tail, 0, 0, 0);
}
VECTOR_STORE(pg_tail, 0, 0);
INCR_C_POINTER(0, 0);
}
UPDATE_B_POINTER(1);
RESET_A_POINTER();
UPDATE_C_POINTER(1);
}
if (pack_b)
free(packed_b);
return 0;
}
}