freedreno/a6xx: Clover support
Some basic things work, missing things to get more of test_basic to pass:
-
hostptr support (ie. CTS hostptr) would require kernel support (SVM) for userptr memory, or work on clover to better emulate.. punting on this for now -
scalarize vector load_global_ir3/store_global_ir3(fpmath_floatN,int_mathN) -
lower 64b load_global_ir3/store_global_ir3to 32b (intmath_long*) -
lower 64b phi ( local_kernel_scope) -
better 8b support ( hiloeo).. we also might need 8b/16b push constants, ie. support to disableSP_MODE_CONTROL.CONSTANT_DEMOTION_ENABLE) -
Add CI job(s).. looks like lavapipe already has some opencl ci which we could copy
Probably some other things I'm missing, I didn't analyze all the failures yet.. but those look like they should account for the majority of fails.
Current state of test_basic:
-
hostptr -
fpmath_float -
fpmath_float2 -
fpmath_float4 -
intmath_int -
intmath_int2 -
intmath_int4 -
intmath_long -
intmath_long2 -
intmath_long4 -
hiloeo -
if -
sizeof -
loop -
pointer_cast -
local_arg_def -
local_kernel_def -
local_kernel_scope -
constant -
constant_source -
readimage -
readimage_int16 -
readimage_fp32 -
writeimage -
writeimage_int16 -
writeimage_fp32 -
mri_one -
mri_multiple -
image_r8 -
barrier -
int2float -
float2int -
imagereadwrite -
imagereadwrite3d -
readimage3d -
readimage3d_int16 -
readimage3d_fp32 -
bufferreadwriterect -
arrayreadwrite -
arraycopy -
imagearraycopy -
imagearraycopy3d -
imagecopy -
imagecopy3d -
imagerandomcopy -
arrayimagecopy -
arrayimagecopy3d -
imagenpot -
vload_global -
vload_local -
vload_constant -
vload_private -
vstore_global -
vstore_local -
vstore_private -
createkernelsinprogram -
imagedim_pow2 -
imagedim_non_pow2 -
image_param -
image_multipass_integer_coord -
image_multipass_float_coord -
explicit_s2v_char -
explicit_s2v_uchar -
explicit_s2v_short -
explicit_s2v_ushort -
explicit_s2v_int -
explicit_s2v_uint -
explicit_s2v_long -
explicit_s2v_ulong -
explicit_s2v_float -
explicit_s2v_double -
enqueue_map_buffer -
enqueue_map_image -
work_item_functions -
astype -
async_copy_global_to_local -
async_copy_local_to_global -
async_strided_copy_global_to_local -
async_strided_copy_local_to_global -
async_copy_global_to_local2D -
async_copy_local_to_global2D -
async_copy_global_to_local3D -
async_copy_local_to_global3D -
async_work_group_copy_fence_import_after_export_aliased_local -
async_work_group_copy_fence_import_after_export_aliased_global -
async_work_group_copy_fence_import_after_export_aliased_global_and_local -
async_work_group_copy_fence_export_after_import_aliased_local -
async_work_group_copy_fence_export_after_import_aliased_global -
async_work_group_copy_fence_export_after_import_aliased_global_and_local -
prefetch -
kernel_call_kernel_function -
host_numeric_constants -
kernel_numeric_constants -
kernel_limit_constants -
kernel_preprocessor_macros -
parameter_types -
vector_creation -
vector_swizzle -
vec_type_hint -
kernel_memory_alignment_local -
kernel_memory_alignment_global -
kernel_memory_alignment_constant -
kernel_memory_alignment_private -
global_work_offsets -
get_global_offset -
simple_read_image_pitch -
simple_write_image_pitch
Note that some OpenCL-CTS kernels trigger hangcheck timeout because they run for a long time, to workaround:
echo 500000 > /sys/kernel/debug/dri/0/hangcheck_period_ms