Commit c8b105a7 authored by Andreas Marek's avatar Andreas Marek
Browse files

Reactivate abort on error in kernels

parent a56303be
...@@ -207,14 +207,14 @@ void single_hh_trafo_complex_avx_avx2_1hv_single(float complex* q, float complex ...@@ -207,14 +207,14 @@ void single_hh_trafo_complex_avx_avx2_1hv_single(float complex* q, float complex
for (i = 0; i < nq-10; i+=12) for (i = 0; i < nq-10; i+=12)
{ {
hh_trafo_complex_kernel_12_AVX_1hv_double(&q[i], hh, nb, ldq); hh_trafo_complex_kernel_12_AVX_1hv_double(&q[i], hh, nb, ldq);
worked_on += i; worked_on += 12;
} }
#endif #endif
#ifdef SINGLE_PRECISION_COMPLEX #ifdef SINGLE_PRECISION_COMPLEX
for (i = 0; i < nq-20; i+=24) for (i = 0; i < nq-20; i+=24)
{ {
hh_trafo_complex_kernel_24_AVX_1hv_single(&q[i], hh, nb, ldq); hh_trafo_complex_kernel_24_AVX_1hv_single(&q[i], hh, nb, ldq);
worked_on += i; worked_on += 24;
} }
#endif #endif
if (nq == i) if (nq == i)
...@@ -295,8 +295,8 @@ void single_hh_trafo_complex_avx_avx2_1hv_single(float complex* q, float complex ...@@ -295,8 +295,8 @@ void single_hh_trafo_complex_avx_avx2_1hv_single(float complex* q, float complex
} }
#endif #endif
if (worked_on != nq) { if (worked_on != nq) {
//printf("Error in complex avx-avx2 BLOCK 1 kernel \n"); printf("Error in complex avx-avx2 BLOCK 1 kernel \n");
//abort(); abort();
} }
} }
......
...@@ -219,14 +219,14 @@ void double_hh_trafo_complex_avx_avx2_2hv_single(float complex* q, float complex ...@@ -219,14 +219,14 @@ void double_hh_trafo_complex_avx_avx2_2hv_single(float complex* q, float complex
for (i = 0; i < nq-6; i+=8) for (i = 0; i < nq-6; i+=8)
{ {
hh_trafo_complex_kernel_8_AVX_2hv_double(&q[i], hh, nb, ldq, ldh, s); hh_trafo_complex_kernel_8_AVX_2hv_double(&q[i], hh, nb, ldq, ldh, s);
worked_on += i; worked_on += 8;
} }
#endif #endif
#ifdef SINGLE_PRECISION_COMPLEX #ifdef SINGLE_PRECISION_COMPLEX
for (i = 0; i < nq-12; i+=16) for (i = 0; i < nq-12; i+=16)
{ {
hh_trafo_complex_kernel_16_AVX_2hv_single(&q[i], hh, nb, ldq, ldh, s , s); hh_trafo_complex_kernel_16_AVX_2hv_single(&q[i], hh, nb, ldq, ldh, s , s);
worked_on += i; worked_on += 16;
} }
#endif #endif
if (nq-i == 0) { if (nq-i == 0) {
...@@ -270,8 +270,8 @@ void double_hh_trafo_complex_avx_avx2_2hv_single(float complex* q, float complex ...@@ -270,8 +270,8 @@ void double_hh_trafo_complex_avx_avx2_2hv_single(float complex* q, float complex
} }
#endif #endif
if (worked_on != nq) { if (worked_on != nq) {
//printf("Error in complex avx-avx2 BLOCK 2 kernel \n"); printf("Error in complex avx-avx2 BLOCK 2 kernel \n");
//abort(); abort();
} }
} }
......
...@@ -173,7 +173,7 @@ void single_hh_trafo_complex_avx512_1hv_single(float complex* q, float complex* ...@@ -173,7 +173,7 @@ void single_hh_trafo_complex_avx512_1hv_single(float complex* q, float complex*
for (i = 0; i < nq-20; i+=24) for (i = 0; i < nq-20; i+=24)
{ {
hh_trafo_complex_kernel_24_AVX512_1hv_double(&q[i], hh, nb, ldq); hh_trafo_complex_kernel_24_AVX512_1hv_double(&q[i], hh, nb, ldq);
worked_on += i; worked_on += 24;
} }
#endif #endif
...@@ -181,7 +181,7 @@ void single_hh_trafo_complex_avx512_1hv_single(float complex* q, float complex* ...@@ -181,7 +181,7 @@ void single_hh_trafo_complex_avx512_1hv_single(float complex* q, float complex*
for (i = 0; i < nq-40; i+=48) for (i = 0; i < nq-40; i+=48)
{ {
hh_trafo_complex_kernel_48_AVX512_1hv_single(&q[i], hh, nb, ldq); hh_trafo_complex_kernel_48_AVX512_1hv_single(&q[i], hh, nb, ldq);
worked_on += i; worked_on += 48;
} }
#endif #endif
if (nq == i) if (nq == i)
...@@ -270,8 +270,8 @@ void single_hh_trafo_complex_avx512_1hv_single(float complex* q, float complex* ...@@ -270,8 +270,8 @@ void single_hh_trafo_complex_avx512_1hv_single(float complex* q, float complex*
#endif #endif
if (worked_on != nq) if (worked_on != nq)
{ {
// printf("Error in complex AVX512 BLOCK 1 kernel \n"); printf("Error in complex AVX512 BLOCK 1 kernel \n");
// abort(); abort();
} }
} }
......
...@@ -179,14 +179,14 @@ void double_hh_trafo_complex_avx512_2hv_single(float complex* q, float complex* ...@@ -179,14 +179,14 @@ void double_hh_trafo_complex_avx512_2hv_single(float complex* q, float complex*
for (i = 0; i < nq-12; i+=16) for (i = 0; i < nq-12; i+=16)
{ {
hh_trafo_complex_kernel_16_AVX512_2hv_double(&q[i], hh, nb, ldq, ldh, s); hh_trafo_complex_kernel_16_AVX512_2hv_double(&q[i], hh, nb, ldq, ldh, s);
worked_on += i; worked_on += 16;
} }
#endif #endif
#ifdef SINGLE_PRECISION_COMPLEX #ifdef SINGLE_PRECISION_COMPLEX
for (i = 0; i < nq-24; i+=32) for (i = 0; i < nq-24; i+=32)
{ {
hh_trafo_complex_kernel_32_AVX512_2hv_single(&q[i], hh, nb, ldq, ldh, s); hh_trafo_complex_kernel_32_AVX512_2hv_single(&q[i], hh, nb, ldq, ldh, s);
worked_on += i; worked_on += 32;
} }
#endif #endif
...@@ -240,8 +240,8 @@ void double_hh_trafo_complex_avx512_2hv_single(float complex* q, float complex* ...@@ -240,8 +240,8 @@ void double_hh_trafo_complex_avx512_2hv_single(float complex* q, float complex*
if (worked_on != nq) if (worked_on != nq)
{ {
// printf("Error in complex AVX512 BLOCK 2 kernel \n"); printf("Error in complex AVX512 BLOCK 2 kernel \n");
// abort(); abort();
} }
} }
......
...@@ -214,14 +214,14 @@ void double_hh_trafo_real_avx_avx2_2hv_single(float* q, float* hh, int* pnb, int ...@@ -214,14 +214,14 @@ void double_hh_trafo_real_avx_avx2_2hv_single(float* q, float* hh, int* pnb, int
for (i = 0; i < nq-20; i+=24) for (i = 0; i < nq-20; i+=24)
{ {
hh_trafo_kernel_24_AVX_2hv_double(&q[i], hh, nb, ldq, ldh, s); hh_trafo_kernel_24_AVX_2hv_double(&q[i], hh, nb, ldq, ldh, s);
worked_on += i; worked_on += 24;
} }
#endif #endif
#ifdef SINGLE_PRECISION_REAL #ifdef SINGLE_PRECISION_REAL
for (i = 0; i < nq-40; i+=48) for (i = 0; i < nq-40; i+=48)
{ {
hh_trafo_kernel_48_AVX_2hv_single(&q[i], hh, nb, ldq, ldh, s); hh_trafo_kernel_48_AVX_2hv_single(&q[i], hh, nb, ldq, ldh, s);
worked_on += i; worked_on += 48;
} }
#endif #endif
...@@ -312,8 +312,8 @@ void double_hh_trafo_real_avx_avx2_2hv_single(float* q, float* hh, int* pnb, int ...@@ -312,8 +312,8 @@ void double_hh_trafo_real_avx_avx2_2hv_single(float* q, float* hh, int* pnb, int
#endif #endif
if (worked_on != nq) if (worked_on != nq)
{ {
//printf("Error in real avx/avx2 BLOCK 2 kernel \n"); printf("Error in real avx/avx2 BLOCK 2 kernel \n");
//abort(); abort();
} }
} }
......
...@@ -250,14 +250,14 @@ void quad_hh_trafo_real_avx_avx2_4hv_single(float* q, float* hh, int* pnb, int* ...@@ -250,14 +250,14 @@ void quad_hh_trafo_real_avx_avx2_4hv_single(float* q, float* hh, int* pnb, int*
for (i = 0; i < nq-8; i+=12) for (i = 0; i < nq-8; i+=12)
{ {
hh_trafo_kernel_12_AVX_4hv_double(&q[i], hh, nb, ldq, ldh, s_1_2, s_1_3, s_2_3, s_1_4, s_2_4, s_3_4); hh_trafo_kernel_12_AVX_4hv_double(&q[i], hh, nb, ldq, ldh, s_1_2, s_1_3, s_2_3, s_1_4, s_2_4, s_3_4);
worked_on += i; worked_on += 12;
} }
#endif #endif
#ifdef SINGLE_PRECISION_REAL #ifdef SINGLE_PRECISION_REAL
for (i = 0; i < nq-16; i+=24) for (i = 0; i < nq-16; i+=24)
{ {
hh_trafo_kernel_24_AVX_4hv_single(&q[i], hh, nb, ldq, ldh, s_1_2, s_1_3, s_2_3, s_1_4, s_2_4, s_3_4); hh_trafo_kernel_24_AVX_4hv_single(&q[i], hh, nb, ldq, ldh, s_1_2, s_1_3, s_2_3, s_1_4, s_2_4, s_3_4);
worked_on += i; worked_on += 24;
} }
#endif #endif
if (nq == i) if (nq == i)
...@@ -299,8 +299,8 @@ void quad_hh_trafo_real_avx_avx2_4hv_single(float* q, float* hh, int* pnb, int* ...@@ -299,8 +299,8 @@ void quad_hh_trafo_real_avx_avx2_4hv_single(float* q, float* hh, int* pnb, int*
if (worked_on != nq) if (worked_on != nq)
{ {
//printf("Error in real AVX/AVX2 BLOCK4 kernel \n"); printf("Error in real AVX/AVX2 BLOCK4 kernel \n");
//abort(); abort();
} }
} }
/** /**
......
...@@ -307,14 +307,14 @@ void hexa_hh_trafo_real_avx_avx2_6hv_single(float* q, float* hh, int* pnb, int* ...@@ -307,14 +307,14 @@ void hexa_hh_trafo_real_avx_avx2_6hv_single(float* q, float* hh, int* pnb, int*
for (i = 0; i < nq-4; i+=8) for (i = 0; i < nq-4; i+=8)
{ {
hh_trafo_kernel_8_AVX_6hv_double(&q[i], hh, nb, ldq, ldh, scalarprods); hh_trafo_kernel_8_AVX_6hv_double(&q[i], hh, nb, ldq, ldh, scalarprods);
worked_on += i; worked_on += 8;
} }
#endif #endif
#ifdef SINGLE_PRECISION_REAL #ifdef SINGLE_PRECISION_REAL
for (i = 0; i < nq-8; i+=16) for (i = 0; i < nq-8; i+=16)
{ {
hh_trafo_kernel_16_AVX_6hv_single(&q[i], hh, nb, ldq, ldh, scalarprods); hh_trafo_kernel_16_AVX_6hv_single(&q[i], hh, nb, ldq, ldh, scalarprods);
worked_on += i; worked_on += 16;
} }
#endif #endif
if (nq == i) if (nq == i)
...@@ -337,8 +337,8 @@ void hexa_hh_trafo_real_avx_avx2_6hv_single(float* q, float* hh, int* pnb, int* ...@@ -337,8 +337,8 @@ void hexa_hh_trafo_real_avx_avx2_6hv_single(float* q, float* hh, int* pnb, int*
#endif #endif
if (worked_on != nq) if (worked_on != nq)
{ {
//printf("Error in real AVX/AVX2 BLOCK6 kernel \n"); printf("Error in real AVX/AVX2 BLOCK6 kernel \n");
//abort(); abort();
} }
} }
......
...@@ -176,14 +176,14 @@ void double_hh_trafo_real_avx512_2hv_single(float* q, float* hh, int* pnb, int* ...@@ -176,14 +176,14 @@ void double_hh_trafo_real_avx512_2hv_single(float* q, float* hh, int* pnb, int*
for (i = 0; i < nq-24; i+=32) for (i = 0; i < nq-24; i+=32)
{ {
hh_trafo_kernel_32_AVX512_2hv_double(&q[i], hh, nb, ldq, ldh, s); hh_trafo_kernel_32_AVX512_2hv_double(&q[i], hh, nb, ldq, ldh, s);
worked_on += i; worked_on += 32;
} }
#endif #endif
#ifdef SINGLE_PRECISION_REAL #ifdef SINGLE_PRECISION_REAL
for (i = 0; i < nq-48; i+=64) for (i = 0; i < nq-48; i+=64)
{ {
hh_trafo_kernel_64_AVX512_2hv_single(&q[i], hh, nb, ldq, ldh, s); hh_trafo_kernel_64_AVX512_2hv_single(&q[i], hh, nb, ldq, ldh, s);
worked_on += i; worked_on += 64;
} }
#endif #endif
if (nq == i) if (nq == i)
...@@ -234,8 +234,8 @@ void double_hh_trafo_real_avx512_2hv_single(float* q, float* hh, int* pnb, int* ...@@ -234,8 +234,8 @@ void double_hh_trafo_real_avx512_2hv_single(float* q, float* hh, int* pnb, int*
#endif #endif
if (worked_on != nq) if (worked_on != nq)
{ {
// printf("Error in AVX512 real BLOCK 2 kernel \n"); printf("Error in AVX512 real BLOCK 2 kernel \n");
// abort(); abort();
} }
} }
......
...@@ -214,14 +214,14 @@ void quad_hh_trafo_real_avx512_4hv_single(float* q, float* hh, int* pnb, int* pn ...@@ -214,14 +214,14 @@ void quad_hh_trafo_real_avx512_4hv_single(float* q, float* hh, int* pnb, int* pn
for (i = 0; i < nq-24; i+=32) for (i = 0; i < nq-24; i+=32)
{ {
hh_trafo_kernel_32_AVX512_4hv_double(&q[i], hh, nb, ldq, ldh, s_1_2, s_1_3, s_2_3, s_1_4, s_2_4, s_3_4); hh_trafo_kernel_32_AVX512_4hv_double(&q[i], hh, nb, ldq, ldh, s_1_2, s_1_3, s_2_3, s_1_4, s_2_4, s_3_4);
worked_on += i; worked_on += 32;
} }
#endif #endif
#ifdef SINGLE_PRECISION_REAL #ifdef SINGLE_PRECISION_REAL
for (i = 0; i < nq-48; i+=64) for (i = 0; i < nq-48; i+=64)
{ {
hh_trafo_kernel_64_AVX512_4hv_single(&q[i], hh, nb, ldq, ldh, s_1_2, s_1_3, s_2_3, s_1_4, s_2_4, s_3_4); hh_trafo_kernel_64_AVX512_4hv_single(&q[i], hh, nb, ldq, ldh, s_1_2, s_1_3, s_2_3, s_1_4, s_2_4, s_3_4);
worked_on += i; worked_on += 64;
} }
#endif #endif
if (nq == i) if (nq == i)
...@@ -278,8 +278,8 @@ void quad_hh_trafo_real_avx512_4hv_single(float* q, float* hh, int* pnb, int* pn ...@@ -278,8 +278,8 @@ void quad_hh_trafo_real_avx512_4hv_single(float* q, float* hh, int* pnb, int* pn
if (worked_on != nq) if (worked_on != nq)
{ {
// printf("Error in AVX512 real BLOCK 2 kernel \n"); printf("Error in AVX512 real BLOCK 2 kernel \n");
// abort(); abort();
} }
} }
......
...@@ -279,14 +279,14 @@ void hexa_hh_trafo_real_avx512_6hv_single(float* q, float* hh, int* pnb, int* pn ...@@ -279,14 +279,14 @@ void hexa_hh_trafo_real_avx512_6hv_single(float* q, float* hh, int* pnb, int* pn
for (i = 0; i < nq-24; i+=32) for (i = 0; i < nq-24; i+=32)
{ {
hh_trafo_kernel_32_AVX512_6hv_double(&q[i], hh, nb, ldq, ldh, scalarprods); hh_trafo_kernel_32_AVX512_6hv_double(&q[i], hh, nb, ldq, ldh, scalarprods);
worked_on += i; worked_on += 32;
} }
#endif #endif
#ifdef SINGLE_PRECISION_REAL #ifdef SINGLE_PRECISION_REAL
for (i = 0; i < nq-48; i+=64) for (i = 0; i < nq-48; i+=64)
{ {
hh_trafo_kernel_64_AVX512_6hv_single(&q[i], hh, nb, ldq, ldh, scalarprods); hh_trafo_kernel_64_AVX512_6hv_single(&q[i], hh, nb, ldq, ldh, scalarprods);
worked_on += i; worked_on += 64;
} }
#endif #endif
if (nq == i) if (nq == i)
...@@ -342,8 +342,8 @@ void hexa_hh_trafo_real_avx512_6hv_single(float* q, float* hh, int* pnb, int* pn ...@@ -342,8 +342,8 @@ void hexa_hh_trafo_real_avx512_6hv_single(float* q, float* hh, int* pnb, int* pn
#endif #endif
if (worked_on != nq) if (worked_on != nq)
{ {
// printf("ERROR in avx512 kernel\n"); printf("ERROR in avx512 kernel\n");
// abort(); abort();
} }
} }
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment