Skip to content

GitLab

Commit a6235503 authored by Andreas Marek's avatar Andreas Marek
Browse files
Options

Allocate memory with different alignment for AVX-512

parent 40a4b6e9
Hide whitespace changes
Inline Side-by-side
Showing with 91 additions and 16 deletions
src/elpa2_trans_ev_tridi_to_band_template.X90
View file @ a6235503
......@@ -345,21 +345,58 @@
#if REALCASE == 1
#ifdef DOUBLE_PRECISION_REAL
stripe_width = ((stripe_width+3)/4)*4 ! Must be a multiple of 4 because of AVX/SSE memory alignment of 32 bytes
! (4 * sizeof(double) == 32)
if (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_AVX512_BLOCK2 .or. &
THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_AVX512_BLOCK4 .or. &
THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_AVX512_BLOCK6) then
stripe_width = ((stripe_width+7)/8)*8 ! Must be a multiple of 8 because of AVX-512 memory alignment of 64 bytes
! (8 * sizeof(double) == 64)
else
stripe_width = ((stripe_width+3)/4)*4 ! Must be a multiple of 4 because of AVX/SSE memory alignment of 32 bytes
! (4 * sizeof(double) == 32)
endif
#else
stripe_width = ((stripe_width+7)/8)*8 ! Must be a multiple of 8 because of AVX/SSE memory alignment of 32 bytes
! (8 * sizeof(float) == 32)
if (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_AVX512_BLOCK2 .or. &
THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_AVX512_BLOCK4 .or. &
THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_AVX512_BLOCK6) then
stripe_width = ((stripe_width+15)/16)*16 ! Must be a multiple of 16 because of AVX-512 memory alignment of 64 bytes
! (16 * sizeof(float) == 64)
else
stripe_width = ((stripe_width+7)/8)*8 ! Must be a multiple of 8 because of AVX/SSE memory alignment of 32 bytes
! (8 * sizeof(float) == 32)
endif
#endif
#endif /* REALCASE */
#if COMPLEXCASE == 1
#ifdef DOUBLE_PRECISION_COMPLEX
stripe_width = ((stripe_width+3)/4)*4 ! Must be a multiple of 2 because of AVX/SSE memory alignment of 32 bytes
! (2 * sizeof(double complex) == 32)
if (THIS_COMPLEX_ELPA_KERNEL .eq. COMPLEX_ELPA_KERNEL_AVX512_BLOCK1 .or. &
THIS_COMPLEX_ELPA_KERNEL .eq. COMPLEX_ELPA_KERNEL_AVX512_BLOCK2) then
stripe_width = ((stripe_width+3)/4)*4 ! Must be a multiple of 4 because of AVX-512 memory alignment of 64 bytes
! (4 * sizeof(double complex) == 64)
else
stripe_width = ((stripe_width+3)/4)*4 ! Must be a multiple of 2 because of AVX/SSE memory alignment of 32 bytes
! (2 * sizeof(double complex) == 32)
endif
#else
stripe_width = ((stripe_width+3)/4)*4 ! Must be a multiple of 4 because of AVX/SSE memory alignment of 32 bytes
! (4 * sizeof(float complex) == 32)
if (THIS_COMPLEX_ELPA_KERNEL .eq. COMPLEX_ELPA_KERNEL_AVX512_BLOCK1 .or. &
THIS_COMPLEX_ELPA_KERNEL .eq. COMPLEX_ELPA_KERNEL_AVX512_BLOCK2) then
stripe_width = ((stripe_width+7)/8)*8 ! Must be a multiple of 8 because of AVX-512 memory alignment of 64 bytes
! (8 * sizeof(float complex) == 64)
else
stripe_width = ((stripe_width+3)/4)*4 ! Must be a multiple of 4 because of AVX/SSE memory alignment of 32 bytes
! (4 * sizeof(float complex) == 32)
endif
#endif
#endif /* COMPLEXCASE */
......@@ -401,21 +438,59 @@
#if REALCASE == 1
#ifdef DOUBLE_PRECISION_REAL
stripe_width = ((stripe_width+3)/4)*4 ! Must be a multiple of 4 because of AVX/SSE memory alignment of 32 bytes
! (4 * sizeof(double) == 32)
if (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_AVX512_BLOCK2 .or. &
THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_AVX512_BLOCK4 .or. &
THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_AVX512_BLOCK6) then
stripe_width = ((stripe_width+7)/8)*8 ! Must be a multiple of 8 because of AVX-512 memory alignment of 64 bytes
! (8 * sizeof(double) == 64)
else
stripe_width = ((stripe_width+3)/4)*4 ! Must be a multiple of 4 because of AVX/SSE memory alignment of 32 bytes
! (4 * sizeof(double) == 32)
endif
#else
stripe_width = ((stripe_width+7)/8)*8 ! Must be a multiple of 8 because of AVX/SSE memory alignment of 32 bytes
! (8 * sizeof(float) == 32)
if (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_AVX512_BLOCK2 .or. &
THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_AVX512_BLOCK4 .or. &
THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_AVX512_BLOCK6) then
stripe_width = ((stripe_width+15)/16)*16 ! Must be a multiple of 16 because of AVX-512 memory alignment of 64 bytes
! (16 * sizeof(float) == 64)
else
stripe_width = ((stripe_width+7)/8)*8 ! Must be a multiple of 8 because of AVX/SSE memory alignment of 32 bytes
! (8 * sizeof(float) == 32)
endif
#endif
#endif /* REALCASE */
#if COMPLEXCASE == 1
#ifdef DOUBLE_PRECISION_COMPLEX
stripe_width = ((stripe_width+3)/4)*4 ! Must be a multiple of 2 because of AVX/SSE memory alignment of 32 bytes
! (2 * sizeof(double complex) == 32)
if (THIS_COMPLEX_ELPA_KERNEL .eq. COMPLEX_ELPA_KERNEL_AVX512_BLOCK1 .or. &
THIS_COMPLEX_ELPA_KERNEL .eq. COMPLEX_ELPA_KERNEL_AVX512_BLOCK2) then
stripe_width = ((stripe_width+3)/4)*4 ! Must be a multiple of 4 because of AVX-512 memory alignment of 64 bytes
! (4 * sizeof(double complex) == 64)
else
stripe_width = ((stripe_width+3)/4)*4 ! Must be a multiple of 2 because of AVX/SSE memory alignment of 32 bytes
! (2 * sizeof(double complex) == 32)
endif
#else
stripe_width = ((stripe_width+3)/4)*4 ! Must be a multiple of 4 because of AVX/SSE memory alignment of 32 bytes
! (4 * sizeof(float complex) == 32)
if (THIS_COMPLEX_ELPA_KERNEL .eq. COMPLEX_ELPA_KERNEL_AVX512_BLOCK1 .or. &
THIS_COMPLEX_ELPA_KERNEL .eq. COMPLEX_ELPA_KERNEL_AVX512_BLOCK2) then
stripe_width = ((stripe_width+7)/8)*8 ! Must be a multiple of 8 because of AVX-512 memory alignment of 64 bytes
! (8 * sizeof(float complex) == 64)
else
stripe_width = ((stripe_width+3)/4)*4 ! Must be a multiple of 4 because of AVX/SSE memory alignment of 32 bytes
! (4 * sizeof(float complex) == 32)
endif
#endif
#endif /* COMPLEXCASE */
endif ! useGPU
......
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