Operations and types supported by the target.

target { tgt.i2f(%x) : (i32) -> f32, (vector<8x128xi32>) -> vector<8x128xf32>; tgt.f2i(%x) : (f32) -> i32, (vector<8x128xf32>) -> vector<8x128xi32>; tgt.neg(%x) : vector<8x128xf32>, f32; # vector<128xf32>

tgt.add(%x, %y) : vector<128xf32>, vector<8x128xf32>, f32; tgt.sub(%x, %y) : vector<128xf32>, vector<8x128xf32>, f32; tgt.mul(%x, %y) : vector<128xf32>, vector<8x128xf32>, f32;

tgt.matmul(%x, %y) : (vector<128x128xf32>, vector<128x128xf32>) -> vector<128x128xf32>; }

Map LAVM ops to previously defined target ops or their combinations.

current restriction is that lavm ops on the lhs should appear with same arg names and count

map { lavm.neg(%a) : tgt.neg(%a); lavm.neg(%a) : tgt.neg(tgt.i2f(%a));

lavm.add(%a, %b) : tgt.add(%a, %b), tgt.add(tgt.i2f(%a), tgt.i2f(%b));

lavm.sub(%a, %b) : lavm.add(%a, lavm.neg(%b));

lavm.matmul(%a, %b) : tgt.matmul(%a, %b); }

Target memory description.

memory {

TPU-like memory:

HBM : size = 16G; HBM : garbage = ignored; VMEM : size = 16M; SMEM : size = 16K; CMEM : size = 16M; HBM -> VMEM; VMEM -> HBM; HBM -> SMEM; SMEM -> HBM; HBM -> CMEM -> VMEM; VMEM -> CMEM -> HBM;

GPU-like memory:

GLOBAL : size = 8G; SHARED : size = 16M; LOCAL : size = 1M;

CPU-like memory:

MEMORY : size = 64GB; $L1 : size = 512K; $L2 : size = 4M; MEMORY -> $L2 -> $L1; $L1 -> $L2 -> MEMORY; }