strange type on bitcast UNREACHABLE executed at clspv/lib/ReplacePointerBitcastPass.cpp:752! about clspv HOT 3 CLOSED

I don't get the same results as you.

I have the ARM compute library at commit db0e610fa90a4c7c534049792ee7fe143d46f72f

Convolution layer:

It compiles for me, but does not validate. Validation error is:

info: 172: Duplicate non-aggregate type declarations are not allowed. Opcode: 21 error: 1251: The Object type (OpTypePointer) in OpCompositeInsert does not match the type that results from indexing into the Composite (OpTypeRuntimeArray).

Normalization layer:

Asserts out around line 835 of ReplacePointerBitcastPass.cpp.

The offending instruction is %12 = bitcast i8 addrspace(1)* %.elt to float addrspace(1)*
That actually occurs twice, once in normalization_layer_cross_map, and another time in normalization_layer_in_map_1D.

Tracing the *map_1D case, there is a vload4 from a pointer derived from the input argument, which is declared as a TENSOR3D_DECLARATION. In helpers.h that's defined as:

#define TENSOR3D_DECLARATION(name)   \
    __global uchar *name##_ptr,      \
    uint        name##_stride_x, \
    uint        name##_step_x,   \
    uint        name##_stride_y, \
    uint        name##_step_y,   \
    uint        name##_stride_z, \
    uint        name##_step_z,   \
    uint        name##_offset_first_element_in_bytes

The problem is that the base pointer is a pointer-to-uchar.

Vulkan does not support 8-bit loads and stores. Loads and stores to 16bit types are only supported by the new VK_KHR_16bit_storage extension.

I'll mark this as "Invalid" since Clspv can't generate functioning code anyway. But I will grant that the error message is TERRIBLE.

from clspv.

I filed #13 for improving error messages.

from clspv.

Closing in favour of the narrower but still valid #14

from clspv.