mirror of
https://github.com/tinygrad/tinygrad.git
synced 2026-04-07 03:00:26 -04:00
335 lines
16 KiB
Python
335 lines
16 KiB
Python
# type: ignore
|
|
|
|
from __future__ import annotations
|
|
import os
|
|
from tinygrad.llops.ops_gpu import GPUBuffer, CL, CLProgram, CLBuffer
|
|
from tinygrad.ops import ProcessingOps, ReduceOps, UnaryOps, BinaryOps
|
|
from tinygrad.helpers import prod, ConvArgs
|
|
from typing import List, Tuple, Optional, Dict, Set
|
|
import numpy as np
|
|
import pyopencl as cl
|
|
|
|
UNSAFE_FLOAT4 = int(os.getenv("UNSAFE_FLOAT4", 0))
|
|
NATIVE_EXPLOG = int(os.getenv("NATIVE_EXPLOG", 0)) # this is needed as a switch for the tests to pass
|
|
FLOAT16 = int(os.getenv("FLOAT16", 0))
|
|
|
|
import pathlib
|
|
def load(x):
|
|
with open(x) as f:
|
|
ret = f.read()
|
|
return ret
|
|
CONV_SRC = load(pathlib.Path(__file__).resolve().parent.parent.parent / 'accel/opencl/conv.cl')
|
|
MATMUL_SRC = load(pathlib.Path(__file__).resolve().parent.parent.parent / 'accel/opencl/matmul.cl')
|
|
|
|
class CLImage:
|
|
fmt = cl.ImageFormat(cl.channel_order.RGBA, cl.channel_type.HALF_FLOAT if FLOAT16 else cl.channel_type.FLOAT)
|
|
|
|
def __init__(self, shape):
|
|
self.max_hw = min(CL().cl_ctx.devices[0].image2d_max_width, CL.cl_ctx.devices[0].image2d_max_height)
|
|
self.shape = shape
|
|
self.n_tile = int(np.ceil(max(shape) / self.max_hw).item())
|
|
# if n_tile > 1, we can't fit the image into a CL image at native size,
|
|
# and need to internally store it as a set of disjoint tiles
|
|
if self.n_tile * min(shape) > self.max_hw:
|
|
raise Exception(f"shape {shape} exceeds Metal image limits, even after tiling")
|
|
if shape[0] >= shape[1]:
|
|
# wider than it is tall; extra tiles overflow on y
|
|
self.tile_axis, tiled_width, tiled_height = 1, min(shape[0], self.max_hw), self.n_tile * shape[1]
|
|
else:
|
|
# taller than it is wide; extra tiles overflow on x
|
|
self.tile_axis, tiled_width, tiled_height = 0, self.n_tile * shape[0], min(shape[1], self.max_hw)
|
|
self.cl = cl.Image(CL.cl_ctx, cl.mem_flags.READ_WRITE, CLImage.fmt, shape=(tiled_width, tiled_height))
|
|
CL.mem_used += self.cl.row_pitch * self.cl.height
|
|
|
|
def pos_to_sample_pos(self, l="l", check_bounds=True):
|
|
if self.n_tile == 1:
|
|
# happy path where no indexing ops are needed
|
|
return l
|
|
# sad tiled path; need to adjust indices, and manually check bounds for the tiled axis
|
|
if self.tile_axis == 1:
|
|
sample_pos = f"((int2)({l}.x % {self.max_hw}, ({l}.x / {self.max_hw}) * {self.shape[1]} + {l}.y))"
|
|
in_bounds = f"((0 <= {l}.y) && ({l}.y < {self.shape[1]}))"
|
|
else:
|
|
sample_pos = f"((int2)(({l}.y / {self.max_hw}) * {self.shape[0]} + {l}.x, {l}.y % {self.max_hw}))"
|
|
in_bounds = f"((0 <= {l}.x) && ({l}.x < {self.shape[0]}))"
|
|
if check_bounds:
|
|
return f"({in_bounds} ? {sample_pos} : (int2)(-1, -1))"
|
|
return sample_pos
|
|
|
|
def __del__(self):
|
|
if hasattr(self, "cl"):
|
|
CL.mem_used -= self.cl.row_pitch * self.cl.height
|
|
|
|
def get_replacements(prg_src:str, opencl_type:List[str]) -> Dict[str, str]:
|
|
middle_code = []
|
|
|
|
"""
|
|
vv = "xyzw"
|
|
for i in range(4):
|
|
acc = f"outputValues[i].{vv[i%4]}"
|
|
args = [x.split(" ")[-1].replace("*", "") for x in opencl_type]
|
|
args = [f"(outputRow * get_image_width(output) + outputLocation.x)*4+{i}", acc]+args
|
|
middle_code.append(f"{acc} = _ewop("+', '.join(args)+");\n")
|
|
"""
|
|
acc = "outputValues[i]"
|
|
args = [x.split(" ")[-1].replace("*", "") for x in opencl_type]
|
|
args = ["smp", "outputLocation", "(outputLocation.y * get_image_width(output) + outputLocation.x)*4", acc]+args
|
|
middle_code.append(f"{acc} = _ewop("+', '.join(args)+");\n")
|
|
|
|
replacements = {}
|
|
replacements["//PREFIX"] = prg_src
|
|
replacements["//BINOP"] = ''.join(middle_code)
|
|
if len(opencl_type) != 0:
|
|
replacements["//ARGS"] = ","+','.join(opencl_type)
|
|
return replacements
|
|
|
|
def get_getters(ewbufs, ret):
|
|
fakebufs = []
|
|
ewtypes = []
|
|
getters = []
|
|
for name, buf in ewbufs:
|
|
view, unfolded, _ = buf.contiguous_view_constant_fold(name)
|
|
if not unfolded:
|
|
getters.append(view)
|
|
fakebufs.append(name)
|
|
getters.append(f"inline float4 get4_{name}(int gid) {{"+
|
|
f"return (float4)(get_{name}(gid+0), get_{name}(gid+1), get_{name}(gid+2), get_{name}(gid+3)); }}")
|
|
elif buf.is_image() and buf.shape == ret.shape and buf.st.contiguous:
|
|
# use an image here
|
|
ewtypes.append(f"read_only image2d_t {name}_g")
|
|
getters.append(f"inline float4 get4_{name}(read_only image2d_t x, const sampler_t smp, int2 loc, int gid) {{ return read_imagef(x, smp, {buf._image.pos_to_sample_pos('loc')}); }}")
|
|
elif buf.st.contiguous:
|
|
# use float4
|
|
ewtypes.append(f"__global const float4 *{name}_g")
|
|
getters.append(f"inline float4 get4_{name}(__global const float4 *x, const sampler_t smp, int2 loc, int gid) {{ return x[gid/4]; }}")
|
|
elif UNSAFE_FLOAT4:
|
|
# aggressive constant folding
|
|
fakebufs.append(name)
|
|
prt = buf._backing.reshape((-1, 4))
|
|
cc = []
|
|
for ii in range(prt.shape[0]):
|
|
cc.append("(float4)(%ff, %ff, %ff, %ff)" % (prt[ii][0], prt[ii][1], prt[ii][2], prt[ii][3]))
|
|
getters.append(f"const __constant float4 const_{name}[] = {{"+', '.join(cc)+"};")
|
|
getters.append(f"inline float4 get4_{name}(int gid) {{"+
|
|
"int idx = gid;"+buf.st.expr()+";"+
|
|
f"return const_{name}[idx/4]; }}")
|
|
"""
|
|
# use float4 indexed (HACK!)
|
|
# TODO: work out when this is okay
|
|
ewtypes.append(f"__global const float4 *{name}_g")
|
|
getters.append(f"inline float4 get4_{name}(__global const float4 *x, const sampler_t smp, int2 loc, int gid) {{"+
|
|
"int valid = 1; int idx = gid;"+buf.st.expr()+";"+
|
|
f"return x[idx/4]; }}")
|
|
"""
|
|
else:
|
|
# fallback to float
|
|
getters.append(view)
|
|
ewtypes.append(f"__global const float *{name}_g")
|
|
getters.append(f"inline float4 get4_{name}(__global const float *x, const sampler_t smp, int2 loc, int gid) {{"+
|
|
f"return (float4)(get_{name}(x,gid+0), get_{name}(x,gid+1), get_{name}(x,gid+2), get_{name}(x,gid+3)); }}")
|
|
return fakebufs, ewtypes, getters
|
|
|
|
def roundup(x, n=4): return (x+(n-1))//n * n
|
|
class OpenCLBuffer(GPUBuffer):
|
|
code_for_op = {
|
|
UnaryOps.NOOP: "(A)", UnaryOps.NEG: "(-(A))", UnaryOps.RELU: "max(A, (float)0.)", UnaryOps.SIGN: "sign(A)",
|
|
UnaryOps.EXP: "native_exp(A)" if NATIVE_EXPLOG else "exp(A)",
|
|
UnaryOps.LOG: "native_log(A)" if NATIVE_EXPLOG else "log(A)",
|
|
UnaryOps.RECIPROCAL: "native_recip(A)" if NATIVE_EXPLOG else "((float)1.0/A)",
|
|
BinaryOps.ADD: "(A+B)", BinaryOps.SUB: "(A-B)", BinaryOps.MUL: "(A*B)", BinaryOps.DIV: "(A/B)", BinaryOps.POW: "pow(A,B)", BinaryOps.CMPEQ: "(A==B)",
|
|
ReduceOps.SUM: "(acc + A)", ReduceOps.MAX: "max(A, acc)"
|
|
}
|
|
def __init__(self, shape, hostbuf:Optional[OpenCLBuffer]=None, backing:Optional[np.ndarray]=None):
|
|
self._image = hostbuf._image if hostbuf is not None else None
|
|
super().__init__(shape, hostbuf, backing)
|
|
assert not (self._image and self._buf)
|
|
|
|
@staticmethod
|
|
def fromCPU(x): return OpenCLBuffer(x.shape, backing=x.view(np.ndarray).astype(np.float32).ravel())
|
|
|
|
def __repr__(self): return f"<OpenCLBuffer with shape {self.shape!r}>"
|
|
|
|
@property
|
|
def cl(self):
|
|
if self._buf is None:
|
|
if self._backing is not None:
|
|
self._buf = CLBuffer(4*roundup(prod(self._backing.shape)))
|
|
CL.enqueue_copy(self._buf.cl, self._backing, is_blocking=False)
|
|
elif self.st.contiguous:
|
|
self._buf = CLBuffer(4*roundup(prod(self.shape)))
|
|
|
|
if self._image is not None:
|
|
self._buf = CLBuffer(4*roundup(prod(self._image.shape)*4))
|
|
if self._backing is not None:
|
|
CL.enqueue_copy(self._buf.cl, self._backing, is_blocking=False)
|
|
#self._backing = None
|
|
#print(f"converting {self.shape} back to buffer, image shape is {self._image.shape}")
|
|
CLProgram("from_image", f"""
|
|
__kernel void from_image(
|
|
__global float4 *out,
|
|
read_only image2d_t in) {{
|
|
const sampler_t smp = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_CLAMP | CLK_FILTER_NEAREST;
|
|
int2 l;
|
|
l.y = get_global_id(1);
|
|
l.x = get_global_id(0);
|
|
int2 l_smp = {self._image.pos_to_sample_pos('l')};
|
|
int W = {str(self._image.shape[0])};
|
|
out[l.y*W + l.x] = read_imagef(in, smp, l_smp);
|
|
}}
|
|
""")(self._image.shape, None, self._buf.cl, self._image.cl)
|
|
self._image = None
|
|
return self._buf.cl
|
|
|
|
def is_image(self): return self._image is not None
|
|
|
|
@property
|
|
def image(self):
|
|
if self._image is None:
|
|
assert len(self.shape) == 3 and self.shape[2] == 4, f"bad shape for image {self.shape}"
|
|
self._image = CLImage(shape=(self.shape[1], self.shape[0]))
|
|
if self._buf is not None:
|
|
assert prod(self.shape) <= prod(self._image.cl.shape)*4
|
|
#print(f"converting {self.shape} to image with shape {self._image.shape}")
|
|
CLProgram("to_image", f"""
|
|
__kernel void to_image(
|
|
write_only image2d_t out,
|
|
__global const float4 *in) {{
|
|
int2 l;
|
|
l.y = get_global_id(1);
|
|
l.x = get_global_id(0);
|
|
int2 l_out = {self._image.pos_to_sample_pos('l', check_bounds=False)};
|
|
int W = {str(self._image.shape[0])};
|
|
write_imagef(out, l_out, in[l.y*W + l.x]);
|
|
}}
|
|
""")(self._image.shape, None, self._image.cl, self._buf.cl)
|
|
self._buf = None
|
|
return self._image.cl
|
|
|
|
SUPPORTS_PADDING = True
|
|
def processing_op(x, op:ProcessingOps, w:GPUBuffer, C:ConvArgs):
|
|
assert op == ProcessingOps.CONV, f"{op} isn't supported"
|
|
return type(x)(C.out_shape)._processing_op([("input", x.contiguous_op()), ("weight", w.contiguous_op())], "acc", C)
|
|
|
|
def contiguous_view_constant_fold(x, name:str, reduce:Optional[int]=None) -> Tuple[str, Optional[str], str]:
|
|
# this will only be for convs, for reduce we have to fall back to cl
|
|
if x.is_image() and reduce is None:
|
|
#print("is image")
|
|
return f"""inline float get_{name}(const sampler_t smp, read_only image2d_t x, int gid) {{
|
|
int valid = 1; int idx = gid; {x.st.expr().replace('//', '/')};
|
|
int2 l;
|
|
int W = {str(x._image.shape[0])};
|
|
l.y = idx / (W*4);
|
|
l.x = (idx/4) % W;
|
|
int idx4 = idx % 4;
|
|
int2 l_smp = {x._image.pos_to_sample_pos('l')};
|
|
float4 dat = read_imagef(x, smp, l_smp);
|
|
return valid ? (idx4 == 0 ? dat.x : (idx4 == 1 ? dat.y : (idx4 == 2 ? dat.z : dat.w))) : 0.0;
|
|
}}""", f"read_only image2d_t {name}_g", f"get_{name}(smp, {name}_g, gid);"
|
|
#ewtypes.append(f"read_only image2d_t {name}_g")
|
|
return super().contiguous_view_constant_fold(name, reduce)
|
|
|
|
def _processing_op(ret, bufs: List[Tuple[str, OpenCLBuffer]]=[], code:str="acc", C=None, op=ReduceOps.SUM, reduce_shape=None, earlybufs:Set[str]=set(), earlycode:str="acc"):
|
|
if C is None or earlycode != "acc":
|
|
# TODO: handle an opencl conv without the conv part
|
|
return super()._processing_op(bufs, code, C, op, reduce_shape, earlybufs, earlycode)
|
|
assert earlycode == "acc"
|
|
|
|
x = [x for x in bufs if x[0] == "input"][0][1]
|
|
w = [x for x in bufs if x[0] == "weight"][0][1]
|
|
ewbufs = [x for x in bufs if x[0] not in ["input", "weight"]]
|
|
|
|
# remove fakebufs
|
|
fakebufs, ewtypes, getters = get_getters(ewbufs, ret)
|
|
ewbufs = [x for x in ewbufs if x[0] not in fakebufs]
|
|
|
|
elementwise_prefix = '\n'.join(getters)+ \
|
|
"\n\ninline float4 _ewop("+','.join(["const sampler_t smp", "int2 loc", "int gid", "float4 acc"]+ewtypes)+") {\n"+ \
|
|
''.join([f"float4 {name} = get4_{name}(gid);\n" for name in fakebufs])+ \
|
|
''.join([f"float4 {name} = get4_{name}({name}_g, smp, loc, gid);\n" for name, _ in ewbufs])+ \
|
|
f"return {code}; }}"
|
|
|
|
replacements = get_replacements(elementwise_prefix, ewtypes)
|
|
|
|
(x.image, w.image, ret.image)
|
|
# fix sampling
|
|
replacements["INPUT_LOCATION"] = x._image.pos_to_sample_pos("inputLocation")
|
|
replacements["WEIGHT_LOCATION"] = w._image.pos_to_sample_pos("weightLocation")
|
|
replacements["OUTPUT_LOCATION"] = ret._image.pos_to_sample_pos("outputLocation", check_bounds=False)
|
|
# fix widths
|
|
replacements["get_image_width(output)"] = f"({ret._image.shape[0]})"
|
|
|
|
x, w = x.contiguous_op(), w.contiguous_op()
|
|
options = []
|
|
if C.bs > 1:
|
|
options.append("-DBATCH")
|
|
assert C.py == 0, "batched conv doesn't work with y-padding"
|
|
if C.sx == 1 and C.sy == 1 and C.dx == 1 and C.dy == 1 and C.cin == 1:
|
|
options.append("-DDEPTHWISE_UNSTRIDED")
|
|
elif C.cin == 1:
|
|
options.append("-DDEPTHWISE")
|
|
if C.groups == 1 and C.H == 1 and C.W == 1 and C.iy == 1 and C.ix == 1 and C.oy == 1 and C.ox == 1 and C.sx == 1 and C.sy == 1 and C.dx == 1 and C.dy == 1 and C.bs == 1:
|
|
options.append("-DMATMUL")
|
|
# NOTE: this is not actually a matmul, it's a vector * matrix
|
|
|
|
conv_args = []
|
|
conv_short_names = ["numPackedInputChannelsForGroup", "totalNumPackedInputChannels", "numPackedOutputChannelsForGroup", "totalNumPackedOutputChannels", "numOutputColumns", "numOutputRows", "numInputRows"]
|
|
conv_shorts = [max(1, C.cin//4), C.groups*C.cin//4, max(1, C.rcout//4), C.cout//4, C.ox, C.oy, C.iy]
|
|
|
|
conv_src = MATMUL_SRC
|
|
replacements["//SHORTS"] = ''.join([f"short {name} = {val};" for name,val in zip(conv_short_names, conv_shorts)])
|
|
if "//BINOP" in replacements:
|
|
replacements["//BINOP"] = replacements["//BINOP"].replace("outputValues[i]", "outputValues")
|
|
for k,v in replacements.items():
|
|
conv_src = conv_src.replace(k, v)
|
|
|
|
#print(conv_src)
|
|
conv_prg = CLProgram("matmul", conv_src,
|
|
options=tuple(options),
|
|
argdtypes=tuple([None, None, None, None] + [np.int16]*len(conv_args) + [None]*len(ewbufs))
|
|
)
|
|
global_work_size = [4, 16, C.cout//4]
|
|
|
|
# must be even
|
|
lw = CL.cl_ctx.devices[0].max_work_group_size // (global_work_size[0] * global_work_size[1])
|
|
while global_work_size[2] % lw != 0:
|
|
lw -= 1
|
|
local_work_size = [4, global_work_size[1], lw]
|
|
|
|
#print(global_work_size, local_work_size)
|
|
conv_prg(global_work_size, local_work_size, ret.image, cl.LocalMemory(4 * local_work_size[0] * local_work_size[1] * lw), x.image, w.image, *conv_args, *[buf.image if 'image2d_t' in typ else buf.cl for typ, (_, buf) in zip(ewtypes, ewbufs)])
|
|
return ret
|
|
|
|
# this option is unused
|
|
if C.H == 1 and C.W == 1:
|
|
options.append("-DONLY_1X1_CONV")
|
|
|
|
assert C.cout%4 == 0
|
|
conv_src = CONV_SRC
|
|
conv_short_names = ["filterSizeX", "filterSizeY", "paddingX", "paddingY", "strideX", "strideY", "dilationX", "dilationY"]
|
|
conv_shorts = [C.W, C.H, C.px, C.py, C.sx, C.sy, C.dx, C.dy]
|
|
conv_arg_names = ["numPackedInputChannelsForGroup", "totalNumPackedInputChannels", "numPackedOutputChannelsForGroup", "totalNumPackedOutputChannels", "numOutputColumns", "numOutputRows", "numInputRows"]
|
|
conv_args = [max(1, C.cin//4), C.groups*C.cin//4, max(1, C.rcout//4), C.cout//4, C.ox, C.oy, C.iy]
|
|
|
|
NUM_OUTPUTS = 4
|
|
options.append(f"-DNUM_OUTPUTS={NUM_OUTPUTS}")
|
|
|
|
# comment out for args
|
|
conv_short_names += conv_arg_names
|
|
conv_shorts += conv_args
|
|
conv_args = []
|
|
options.append("-DNOARGS")
|
|
|
|
replacements["//SHORTS"] = ''.join([f"short {name} = {val};" for name,val in zip(conv_short_names, conv_shorts)])
|
|
for k,v in replacements.items():
|
|
conv_src = conv_src.replace(k, v)
|
|
#print(conv_src)
|
|
conv_prg = CLProgram("image_conv", conv_src,
|
|
options=tuple(options),
|
|
argdtypes=tuple([None, None, None] + [np.int16]*len(conv_args) + [None]*len(ewbufs))
|
|
)
|
|
global_work_size = [C.cout//4, (C.ox+NUM_OUTPUTS-1)//NUM_OUTPUTS, C.bs*C.oy]
|
|
conv_prg(global_work_size, None, ret.image, x.image, w.image, *conv_args, *[buf.image if 'image2d_t' in typ else buf.cl for typ, (_, buf) in zip(ewtypes, ewbufs)])
|
|
return ret
|
|
|
|
GPUBuffer = OpenCLBuffer
|