Compiler fails when vectorizing loop with type conversion

overnite81 — Fri, 21 Apr 2023 00:43:27 GMT

Hi to all.

I am having the following issue with the Intel CPU Runtime for OpenCL version 23.1.46319:

This code

__kernel void logisticMap(__global const float *params, __global const float *initialValues, __global float *result, __global unsigned int *bits, const int numIters, const int k) { int i = get_global_id(0); int j = get_global_id(1); int ti = get_local_id(0); int n = get_global_size(0); int nt = get_local_size(0); int nb = n / nt; const float kpow = pown(10.0f, k); float r, s; unsigned int w; r = initialValues[i]; s = params[i]; int cnt = 0; w = 0; // __attribute__((opencl_unroll_hint)) for (int ic = 0; ic < numIters; ic++) { r = s * r * (1.0f - r); float t = r * kpow; float u = t - floor(t); unsigned int x = (unsigned int) step(0.5f, u); x ^= 1; w |= x << (ic % 32); if (ic % 32 == 31) { const int index = i * (numIters / 32) + cnt++; bits[index] = w; w = 0; } } result[i] = r; }

outputs

**Internal compiler error** Cannot select: 0x11c67196648: v4f32 = X86ISD::VRNDSCALE contract 0x11c67196568, TargetConstant:i32<1>
0x11c67196568: v4f32 = fmul 0x11c671969c8, 0x11c67195ed8
0x11c671969c8: v4f32,ch = CopyFromReg 0x11c5e7119a8, Register:v4f32 %12
0x11c67196108: v4f32 = Register %12
0x11c67195ed8: v4f32 = fmul 0x11c6719a1f8, 0x11c6719d878
0x11c6719a1f8: v4f32 = fmul 0x11c67195b58, 0x11c6719a498
0x11c67195b58: v4f32,ch = CopyFromReg 0x11c5e7119a8, Register:v4f32 %19
0x11c6719ddb8: v4f32 = Register %19
0x11c6719a498: v4f32,ch = CopyFromReg 0x11c5e7119a8, Register:v4f32 %20
0x11c6719db88: v4f32 = Register %20
0x11c6719d878: v4f32 = fsub 0x11c6719d9c8, 0x11c6719a498
0x11c6719d9c8: v4f32,ch = load<(load (s128) from constant-pool)> 0x11c5e7119a8, 0x11c67195e68, undef:i64
0x11c67195e68: i64 = X86ISD::Wrapper TargetConstantPool:i64<> 0
0x11c67196958: i64 = TargetConstantPool<> 0
0x11c6719daa8: i64 = undef
0x11c6719a498: v4f32,ch = CopyFromReg 0x11c5e7119a8, Register:v4f32 %20
0x11c6719db88: v4f32 = Register %20
0x11c67199fc8: i32 = TargetConstant<1>
In function: logisticMap

after attempt to compile.

Re:Compiler fails when vectorizing loop with type conversion

cw_intel — Mon, 08 May 2023 03:07:17 GMT

Hi,

Can you provide the command you used to compile the kernel code?

BTW, have you used other versions of OpenCL CPU RT before? If so, please tell us the OpenCL CPU RT version that can work successfully.

Thanks

Re:Compiler fails when vectorizing loop with type conversion

cw_intel — Mon, 29 May 2023 02:23:05 GMT

Hi,

We haven't heard back from you for a long time so we are assuming that you have found a solution on your own and we will no longer respond to this thread. If you require additional assistance from Intel, please start a new thread. Any further interaction in this thread will be considered community only.

Thanks,

Chen

topic Re:Compiler fails when vectorizing loop with type conversion in OpenCL* for CPU

Compiler fails when vectorizing loop with type conversion

Re:Compiler fails when vectorizing loop with type conversion

Re:Compiler fails when vectorizing loop with type conversion