You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@tvm.apache.org by GitBox <gi...@apache.org> on 2019/11/22 03:34:56 UTC
[GitHub] [incubator-tvm] jackwish commented on a change in pull request
#3859: [TOPI][AutoTVM] NHWC conv2d templates for ARM
jackwish commented on a change in pull request #3859: [TOPI][AutoTVM] NHWC conv2d templates for ARM
URL: https://github.com/apache/incubator-tvm/pull/3859#discussion_r349421208
##########
File path: topi/python/topi/arm_cpu/conv2d_spatial_pack.py
##########
@@ -196,3 +196,160 @@ def schedule_conv2d_spatial_pack_nchw(cfg, s, data_vec, kernel_vec,
s[kernel_vec].parallel(co)
return s
+
+def conv2d_spatial_pack_nhwc(cfg, data, kernel, strides, padding, dilation, out_dtype):
+ """Spatial pack compute for Conv2d NHWC"""
+ out_dtype = out_dtype or data.dtype
+
+ N, IH, IW, IC = get_const_tuple(data.shape)
+ assert len(kernel.shape) == 4, "AlterOpLayout not enabled for NHWC yet"
+ KH, KW, _, OC = get_const_tuple(kernel.shape)
+
+ if isinstance(dilation, int):
+ dilation_h = dilation_w = dilation
+ else:
+ dilation_h, dilation_w = dilation
+
+ dilated_kernel_h = (KH - 1) * dilation_h + 1
+ dilated_kernel_w = (KW - 1) * dilation_w + 1
+
+ pad_top, pad_left, pad_down, pad_right = \
+ get_pad_tuple(padding, (dilated_kernel_h, dilated_kernel_w))
+ HSTR, WSTR = strides if isinstance(strides, (tuple, list)) else (strides, strides)
+
+ OH = (IH + pad_top + pad_down - dilated_kernel_h) // HSTR + 1
+ OW = (IW + pad_left + pad_right - dilated_kernel_w) // WSTR + 1
+ data_pad = nn.pad(data, [0, pad_top, pad_left, 0], [0, pad_down, pad_right, 0])
+
+ # ==================== define configuration space ====================
+ n, oc, oh, ow = cfg.axis(N), cfg.axis(OC), cfg.axis(OH), cfg.axis(OW)
+ ic, kh, kw = cfg.reduce_axis(IC), cfg.reduce_axis(KH), cfg.reduce_axis(KW)
+
+ oco, oci = cfg.define_split('tile_co', oc, num_outputs=2)
+ oho, ohi = cfg.define_split('tile_oh', oh, num_outputs=2)
+ owo, owi = cfg.define_split('tile_ow', ow, num_outputs=2)
+
+ cfg.define_reorder('reorder_conv',
+ [n, oho, owo, oco, kh, kw, ic, ohi, owi, oci],
+ policy='candidate', candidate=[
+ [n, oho, owo, oco, kh, kw, ic, ohi, owi, oci],
+ [n, oho, owo, oco, ohi, kh, kw, ic, owi, oci],
+ [n, oho, owo, oco, ohi, kh, kw, owi, ic, oci],
+ [n, oho, owo, ohi, oco, kh, kw, owi, ic, oci]])
+
+ cfg.define_annotate("ann_reduce", [kh, kw], policy='try_unroll')
+ cfg.define_annotate("ann_spatial", [ohi, owi, oci], policy='try_unroll_vec')
+ # ====================================================================
+
+ OCI = cfg['tile_co'].size[-1]
+ OHI = cfg['tile_oh'].size[-1]
+ OWI = cfg['tile_ow'].size[-1]
+ OCO = OC // OCI
+ OHO = OH // OHI
+ OWO = OW // OWI
+
+ kvshape = (OCO, KH, KW, IC, OCI)
+ ovshape = (N, OHO, OWO, OCO, OHI, OWI, OCI)
+ oshape = (N, OH, OW, OC)
+
+ if dilation_h != 1 or dilation_w != 1:
+ # undilate input data
+ dvshape = (N, OHO, OWO, KH, KW, IC, OHI, OWI)
+ data_vec = tvm.compute(dvshape, lambda n, oho, owo, kh, kw, ic, ohi, owi:
+ data_pad[n][(oho*OHI+ohi)*HSTR+kh*dilation_h]
+ [(owo*OWI+owi)*WSTR+kw*dilation_w][ic],
+ name='data_vec_undilated')
+ else:
+ dvshape = (N, OHO, OWO, KH + (OHI-1)*HSTR, KW + (OWI-1)*WSTR, IC)
+ data_vec = tvm.compute(dvshape, lambda n, oho, owo, ohi, owi, ic:
+ data_pad[n][oho*OHI*HSTR+ohi][owo*OWI*WSTR+owi][ic],
+ name='data_vec')
+ kernel_vec = tvm.compute(kvshape, lambda oco, kh, kw, ic, oci: \
+ kernel[kh][kw][ic][oco*OCI+oci],
+ name='kernel_vec')
+
+ ic = tvm.reduce_axis((0, IC), name='ic')
+ kh = tvm.reduce_axis((0, KH), name='kh')
+ kw = tvm.reduce_axis((0, KW), name='kw')
+
+ if dilation_h != 1 or dilation_w != 1:
+ conv = tvm.compute(ovshape, lambda n, oho, owo, oco, ohi, owi, oci: \
+ tvm.sum(data_vec[n, oho, owo, kh, kw, ohi, owi, ic].astype(out_dtype) *
+ kernel_vec[oco, kh, kw, ic, oci].astype(out_dtype),
+ axis=[ic, kh, kw]), name='conv')
+ else:
+ conv = tvm.compute(ovshape, lambda n, oho, owo, oco, ohi, owi, oci: \
+ tvm.sum(data_vec[n, oho, owo, ohi*HSTR+kh, owi*WSTR+kw, ic].astype(out_dtype) *
+ kernel_vec[oco, kh, kw, ic, oci].astype(out_dtype),
+ axis=[ic, kh, kw]), name='conv')
+
+ output = tvm.compute(oshape, lambda n, oho, owo, oc:
+ conv[n][oho//OHI][owo//OWI][oc//OCI][oho%OHI][owo%OWI][oc%OCI],
+ name='output_unpack', tag='spatial_conv_output_NHWC')
+ return output
+
+def schedule_conv2d_spatial_pack_nhwc(cfg, s, op, output):
+ """Spatial Pack schedule for Conv2d NHWC"""
+ unpack = op.output(0)
+ conv = unpack.op.input_tensors[0]
+ data_vec = conv.op.input_tensors[0]
+ kernel_vec = conv.op.input_tensors[1]
+ data_pad = data_vec.op.input_tensors[0]
+ OHI = cfg['tile_oh'].size[-1]
+ OWI = cfg['tile_ow'].size[-1]
+ OCI = cfg['tile_co'].size[-1]
+
+ # schedule unpack/output
+ if output != unpack:
+ s[unpack].compute_inline()
+ n, oh, ow, oc = s[output].op.axis
+ oco, oci = cfg['tile_co'].apply(s, output, oc)
+ oho, ohi = cfg['tile_oh'].apply(s, output, oh)
+ owo, owi = cfg['tile_ow'].apply(s, output, ow)
+ s[output].reorder(n, oho, owo, oco, ohi, owi, oci)
+ cfg['ann_spatial'].apply(s, output, [ohi, owi, oci], axis_lens=[OHI, OWI, OCI],
+ max_unroll=32, cfg=cfg)
+ cfg.define_knob('compat', [0, 1, 2])
+ if cfg['compat'].val < 2:
+ compat_axis = [owo, oco][cfg['compat'].val]
+ s[conv].compute_at(s[output], compat_axis)
+ paxis = s[output].fuse(n, oho)
+ s[output].parallel(paxis)
+
+ # schedule conv
+ n, oho, owo, oco, ohi, owi, oci = s[conv].op.axis
+ ic, kh, kw = s[conv].op.reduce_axis
+ cfg['reorder_conv'].apply(s, conv, [n, oho, owo, oco, kh, kw, ohi, owi, ic, oci])
+ cfg['ann_reduce'].apply(s, conv, [kh, kw],
+ axis_lens=[get_const_int(kh.dom.extent),
+ get_const_int(kw.dom.extent)],
+ max_unroll=32,
+ cfg=cfg)
+ cfg['ann_spatial'].apply(s, conv, [ohi, owi, oci], axis_lens=[OHI, OWI, OCI],
+ max_unroll=32, cfg=cfg)
+ if cfg['compat'].val < 2:
+ compat_axis = [owo, oco][cfg['compat'].val]
+ s[kernel_vec].compute_at(s[conv], compat_axis)
+ s[data_vec].compute_at(s[conv], compat_axis)
+
+ # schedule kernel pack
+ oco, kh, kw, ic, oci = kernel_vec.op.axis
+ s[kernel_vec].vectorize(oci)
+ s[kernel_vec].unroll(ic)
+ if cfg['compat'].val == 2:
+ s[kernel_vec].parallel(oco)
Review comment:
Attaching `data_vec` and `kernel_vec` compute to `conv` when `cfg['compat'].val < 2`, in which case parallel them will lead to *nested parallel* error - `conv` has already been parallelized.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
With regards,
Apache Git Services