int8/uint8 support (#837)

* feat: int8 support

* feat: uint8 support

* feat: int8 tests

* fix: fix uint8 on clang

* feat: test casting between int8/uint8/float16/float32

* clean: way cleaner dtype tests

* feat: preprocess_imagenet using the correct dtype

* feat: add test for overflow between uint8 and int8

tinygrad/codegen/cstyle.py

@@ -132,7 +132,7 @@ def uops_to_cstyle(uops:List[UOp], bufs:List[Union[LocalBuffer,LazyBuffer]], lan
       if bufs[args.i] is not None and isinstance(bufs[args.i].realized, RawConst):
         assert newvar.ltype == LocalTypes.float, "const can't be float4"
         # nan? inf?
-        val = f"{bufs[args.i].realized._buf}f"
+        val = f"{bufs[args.i].realized._buf}" + ("f" if bufs[args.i].dtype not in (dtypes.int8, dtypes.uint8) else "")
       elif isinstance(bufs[args.i].dtype, ImageDType):
         assert newvar.ltype == LocalTypes.float4, "image must be float4"
         prekernel.add("const sampler_t smp = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_CLAMP | CLK_FILTER_NEAREST;\n")

tinygrad/codegen/linearizer.py

@@ -201,7 +201,7 @@ class Linearizer:
     store_offset: Dict[Tuple[int, ...], Token] = dict(zip(self.shape_offsets(i), store))
 
     # float4 grouping (optional)
-    should_upcast = self.supports_float4 and self.bufs[i].dtype != dtypes.float16 and len(self.float4_axis(i)) == 1
+    should_upcast = self.supports_float4 and (self.bufs[i].dtype not in (dtypes.float16, dtypes.int8, dtypes.uint8)) and len(self.float4_axis(i)) == 1
     if should_upcast:
       store_offset_new = {}
       for k,out_tokens in self._group_float4(i, store_offset).items():

tinygrad/codegen/llvmir.py

@@ -38,7 +38,7 @@ def uops_to_llvm_ir(uops:List[UOp], bufs:List[LazyBuffer]) -> str:
   module = ir.Module(name=__file__)
 
   # create llvm function
-  func_dtypes = [{dtypes.float16:ir.HalfType(), dtypes.float32:ir.FloatType()}[buf.dtype] for buf in bufs]
+  func_dtypes = [{dtypes.float16:ir.HalfType(), dtypes.float32:ir.FloatType(), dtypes.int8:ir.IntType(8), dtypes.uint8:ir.IntType(8)}[buf.dtype] for buf in bufs]
   func = ir.Function(module, ir.FunctionType(ir.VoidType(), [x.as_pointer() for x in func_dtypes]), name='exec')
 
   # force llvmlite to allow us to add function attribute then add the attribute