#map0 = affine_map<(d0) -> (d0)>
#map1 = affine_map<(d0) -> (d0 ceildiv 256)>
module {
func private @print_memref_f32(memref<*xf32>)
// Creates and returns 4-D buffer of size (%s1, %s2, %s3, %s4) filled with the value %f
func @alloc_4d_filled_f32(%s1 : index, %s2 : index, %s3 : index, %s4 : index, %f : f32) -> memref<?x?x?x?xf32> {
%buf = memref.alloc(%s1, %s2, %s3, %s4) : memref<?x?x?x?xf32>
linalg.fill(%f, %buf) : f32, memref<?x?x?x?xf32>
return %buf : memref<?x?x?x?xf32>
}
func @conv_2d_nhwc_hwcf(%arg0: memref<?x?x?x?xf32>, %arg1: memref<?x?x?x?xf32>, %arg2: memref<?x?x?x?xf32>) {
linalg.conv_2d_nhwc_hwcf {dilations = dense<1> : tensor<2xi64>,
strides = dense<1> : tensor<2xi64>}
ins (%arg0, %arg1: memref<?x?x?x?xf32>, memref<?x?x?x?xf32>)
outs (%arg2: memref<?x?x?x?xf32>)
return
}
func @conv_2d_tensor(%input: tensor<?x?xi32>,
%filter: tensor<?x?xi32>,
%output: tensor<?x?xi32>) -> tensor<?x?xi32> {
%0 = linalg.conv_2d
ins (%input, %filter: tensor<?x?xi32>, tensor<?x?xi32>)
outs (%output: tensor<?x?xi32>) -> tensor<?x?xi32>
return %0 : tensor<?x?xi32>
}
func @conv_2d_memref(%arg0: memref<?x?xf32>, %arg1: memref<?x?xf32>, %arg2: memref<?x?xf32>) {
linalg.conv_2d ins (%arg0, %arg1: memref<?x?xf32>, memref<?x?xf32>)
outs (%arg2: memref<?x?xf32>)
return
}
func @cbsm_conv_inner_func(%arg0: memref<?x?xf32>, %arg1: memref<?x?xf32>, %arg2: memref<?x?xf32>) {
%c0 = arith.constant 0 : index
%c1 = arith.constant 1 : index
%0 = memref.dim %arg1, %c0 : memref<?x?xf32> // H
%1 = memref.dim %arg1, %c1 : memref<?x?xf32> // W
%2 = memref.dim %arg2, %c0 : memref<?x?xf32> //FH
%3 = memref.dim %arg2, %c1 : memref<?x?xf32> //FW
affine.for %arg3 = #map0(%c0) to #map0(%2) { // a3 : 0-fh
affine.for %arg4 = #map0(%c0) to #map0(%0) { // a4 : 0-h
affine.for %arg5 = #map0(%c0) to #map0(%1) { // a5 : 0-w
affine.for %arg6 = #map0(%c0) to #map1(%3) { // a6 : 0-up[fw/256]
// f4 = vector.load(filter[h,w])
%4 = affine.vector_load %arg1[%arg4, %arg5] : memref<?x?xf32>, vector<1xf32>
// vec.bcast(vector.load(filter[h,w])) 1-256
%5 = vector.broadcast %4 : vector<1xf32> to vector<256xf32>
// %6=vec.load256(img[fh+h, w+fw*256])
%6 = affine.vector_load %arg0[%arg3 + %arg4, %arg5 + %arg6 * 256] : memref<?x?xf32>, vector<256xf32>
// fi = vector.load(out[fw, up(fw/256)*256])
%7 = affine.vector_load %arg2[%arg3, %arg6 * 256] : memref<?x?xf32>, vector<256xf32>
// vec.fma(vec.load256(img[fh+h, w+fw*256])*vec.bcast(vector.load(filter[h,w]))
// +vec.load(out[fw, up(fw/256)*256]))
%8 = vector.fma %6, %5, %7 : vector<256xf32>
// out[fh,up(fw/256)*256]
affine.vector_store %8, %arg2[%arg3, %arg6 * 256] : memref<?x?xf32>, vector<256xf32>
}
}
}
}
return
}
func @pw_cbsm_conv2d_outer_func(%input: memref<?x?x?x?xf32>, %filter: memref<?x?x?x?xf32>, %output: memref<?x?x?x?xf32>) {
%c0 = arith.constant 0 : index
%c1 = arith.constant 1 : index
%c2 = arith.constant 2 : index
%c3 = arith.constant 3 : index
%KH = memref.dim %filter, %c0 : memref<?x?x?x?xf32> // FH
%KW = memref.dim %filter, %c1 : memref<?x?x?x?xf32> // FW
%KC = memref.dim %filter, %c2 : memref<?x?x?x?xf32> // FC
%ON = memref.dim %output, %c1 : memref<?x?x?x?xf32> // ON
%OH = memref.dim %output, %c1 : memref<?x?x?x?xf32> // OH
%OW = memref.dim %output, %c0 : memref<?x?x?x?xf32> // OW
%OF = memref.dim %output, %c2 : memref<?x?x?x?xf32> // OF
affine.for %on = #map0(%c0) to #map0(%ON) { // on : 0-on(batch)
affine.for %of = #map0(%c0) to #map0(%OF) { // of : 0-of
// out for adds out_adds_kc_tmp <1,OH,OW,1>
affine.for %kc = #map0(%c0) to #map0(%KC) { // kc : 0-kc (need to add)
// init kc_out_tmp[OH,OW]
%input_inner = memref.subview %input
%filter_inner = memref.subview %filter
call @conv_2d(%input_inner, %filter_inner, %output_inner) : (memref<?x?xf32>, memref<?x?xf32>, memref<?x?xf32>) -> ()
%out_adds_kc_tmp =+ %output_inner
}
// output.insert_stride_slice
}
}
}
func @pw_cbsm_conv2d_outer_func(%input: memref<?x?x?x?xf32>, %filter: memref<?x?x?x?xf32>, %output: memref<?x?x?x?xf32>) {
%c0 = arith.constant 0 : index
%c1 = arith.constant 1 : index
%c2 = arith.constant 2 : index
%c3 = arith.constant 3 : index
%KH = memref.dim %filter, %c0 : memref<?x?x?x?xf32> // FH
%KW = memref.dim %filter, %c1 : memref<?x?x?x?xf32> // FW
%KC = memref.dim %filter, %c2 : memref<?x?x?x?xf32> // FC
%ON = memref.dim %output, %c1 : memref<?x?x?x?xf32> // ON
%OH = memref.dim %output, %c1 : memref<?x?x?x?xf32> // OH
%OW = memref.dim %output, %c0 : memref<?x?x?x?xf32> // OW
%OF = memref.dim %output, %c2 : memref<?x?x?x?xf32> // OF
affine.for %on = #map0(%c0) to #map0(%ON) { // on : 0-on(batch)
affine.for %of = #map0(%c0) to #map0(%OF) { // of : 0-of
// out for adds out_adds_kc_tmp <1,OH,OW,1>
affine.for %kc = #map0(%c0) to #map0(%KC) { // kc : 0-kc (need to add)
// init kc_out_tmp[OH,OW]
%input_inner = memref.subview %input
%filter_inner = memref.subview %filter
call @conv_2d(%input_inner, %filter_inner, %output_inner) : (memref<?x?xf32>, memref<?x?xf32>, memref<?x?xf32>) -> ()
%out_adds_kc_tmp =+ %output_inner
}
// output.insert_stride_slice
}
}
}
func @pw_cbsm_conv2d_outer_func(%input: tensor<?x?x?x?xf32>, %filter: tensor<?x?x?x?xf32>) -> %output: tensor<?x?x?x?xf32> {
%c0 = arith.constant 0 : index
%c1 = arith.constant 1 : index
%c2 = arith.constant 2 : index
%c3 = arith.constant 3 : index
%KH = tensor.dim %filter, %c0 : tensor<?x?x?x?xf32> // FH
%KW = tensor.dim %filter, %c1 : tensor<?x?x?x?xf32> // FW
%KC = tensor.dim %filter, %c2 : tensor<?x?x?x?xf32> // FC
%ON = tensor.dim %output, %c1 : tensor<?x?x?x?xf32> // ON
%OH = tensor.dim %output, %c1 : tensor<?x?x?x?xf32> // OH
%OW = tensor.dim %output, %c0 : tensor<?x?x?x?xf32> // OW
%OF = tensor.dim %output, %c2 : tensor<?x?x?x?xf32> // OF
// out for adds out_adds_kc_tmp <1,OH,OW,1>
%output = arith.constant dense<0> : tensor<%ONx%OHx%OWx%OFxf32>
affine.for %on = #map0(%c0) to #map0(%ON) { // on : 0-on(batch)
affine.for %of = #map0(%c0) to #map0(%OF) { // of : 0-of
// out for adds out_adds_kc_tmp <1,OH,OW,1>
%out_adds_kc_tmp = arith.constant dense<0> : tensor<%ONx%OHx%OWx%OFxf32>
affine.for %kc = #map0(%c0) to #map0(%KC) { // kc : 0-kc (need to add)
// 1. init kc_out_tmp[OH,OW]
%output_inner = arith.constant dense<0> : tensor<%OHx%OWxf32>
// 2. silce input for cbsm
// input_inner = input[on,:,:,kc]
// filter_inner = filter[0,0,kc,of]
%input_inner = tensor.extract_slice %input[%on,0,0,%kc][1,%OH,%OW,1][1,1,1,1] : tensor<?x?x?x?xf32> to tensor<?x?xf32>
%filter_inner = tensor.extract_slice %filter[0,0,%kc,%of][1,1,1,1][1,1,1,1] : tensor<?x?x?x?xf32> to tensor<?x?xf32>
// 3. call conv_2d
call @conv_2d(%input_inner, %filter_inner, %output_inner) : (tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>) -> tensor<?x?xf32>
// 4. accmulate all kc with elementwise add to one of tensor
%out_adds_kc_tmp =+ %output_inner
}
// 5. insert added kc_out to one layer of real output using: output.insert_stride_slice
%output = tensor.insert_slice %out_adds_kc_tmp %into %output[%on,0,0,%of][1,%OH,%OW,1][1,1,1,1] : tensor<%OHx%OWxf32> into tensor<%ONx%OHx%OWxOFxf32>
}
}
return %output
}
func @pw_cbsm_conv2d(%input: memref<?x?x?x?xf32>, %filter: memref<?x?x?x?xf32>, %output: memref<?x?x?x?xf32>) {
%c0 = arith.constant 0 : index
%c1 = arith.constant 1 : index
%c2 = arith.constant 2 : index
%c3 = arith.constant 3 : index
%KH = memref.dim %filter, %c0 : memref<?x?x?x?xf32> // FH
%KW = memref.dim %filter, %c1 : memref<?x?x?x?xf32> // FW
%KC = memref.dim %filter, %c2 : memref<?x?x?x?xf32> // FC
%ON = memref.dim %output, %c1 : memref<?x?x?x?xf32> // ON
%OH = memref.dim %output, %c1 : memref<?x?x?x?xf32> // OH
%OW = memref.dim %output, %c0 : memref<?x?x?x?xf32> // OW
%OF = memref.dim %output, %c2 : memref<?x?x?x?xf32> // OF
// for (j:OF) {
// for (i:KC) {
// img_subview[:,:,:,kc]
// %output[i,of] += cbsm(img_subview[:,:,:,kc],filter[0,0,kc,of])
// }
// }
affine.for %on = #map0(%c0) to #map0(%ON) { // on : 0-on(batch)
affine.for %of = #map0(%c0) to #map0(%OF) { // of : 0-of
// output[n,h,w,f] = subview as output[:,:,:,f]
// output[n,h,w,f] = subview as output[:,:,:,0]
// output[n,h,w,f] = subview as output[:,:,:,1]
affine.for %kc = #map0(%c0) to #map0(%KC) { // kc : 0-kc (need to add)
// subview as output[:,:,:,f] += partial_output[:,:,kc,0]
affine.for %oh = #map0(%c0) to #map0(%OH) { // a3 : 0-oh
affine.for %ow_256 = #map0(%c0) to #map1(%OW) { // a6 : 0-up[ow/256]
// f4 = vector.load(filter[fh,fw,?fc,of]) (对于所有kc只有一个值of)
%4 = affine.vector_load %filter[0,0,%kc,%of] : memref<?x?x?x?xf32>, vector<1xf32>
// vec.bcast(vector.load(filter[fh,fw,kc,of])) 1-256
%5 = vector.broadcast %4 : vector<1xf32> to vector<256xf32>
// img[n,h,w,c] = subview as img[:,:,:,0]
// img[n,h,w,c] = subview as img[:,:,:,1]
// img[n,h,w,c] = subview as img[:,:,:,2]
// %6=vec.load256(img[on, fh+oh, ow+fw*256，kc])
%6 = affine.vector_load %input[%on, 0+%oh, 0+%ow_256*256, %kc] : memref<?x?x?x?xf32>, vector<256xf32>
// %66 = affine.vector_load %img[:,:,:,0][%on, 0+%oh, 0+%ow_256*256, %kc] : memref<?x?x?x?xf32>, vector<256xf32>
// %66 = affine.vector_load %img[:,:,:,1][%on, 0+%oh, 0+%ow_256*256, %kc] : memref<?x?x?x?xf32>, vector<256xf32>
// %66 = affine.vector_load %img[:,:,:,2][%on, 0+%oh, 0+%ow_256*256, %kc] : memref<?x?x?x?xf32>, vector<256xf32>
// img[n,h,w,c] = subview as img[:,:,:,0]
// %66 = vector.extract_strided_slice %6 {offsets = [0, 0, 0, 0], sizes = [2, 4,,], strides = [1, 1,,]}: vector<4x8x16xf32> to vector<2x4x16xf32>
// out[:,:,0,0]
// out[:,:,1,0]
// out[:,:,2,0]
// out[:,:,:,0] = reduce(dim[3],)
// fi = vector.load(out[on, oh, up(ow/256)*256], of]) [out fix with OF meaning add all kc to each of]
%7 = affine.vector_load %output[%on, %oh, %ow_256 * 256, %of] : memref<?x?x?x?xf32>, vector<256xf32>
// vec.fma(vec.load256(img[fh+oh, fw+ow*256])*vec.bcast(vector.load(filter[fh,fw]))
// +vec.load(out[ow, up(ow/256)*256]))
%8 = vector.fma %6, %5, %7 : vector<256xf32>
// out[oh,up(ow/256)*256]
affine.vector_store %8, %output[%on, %oh, %ow_256 * 256, %of] : memref<?x?x?x?xf32>, vector<256xf32>
// %66 = vector.insert_strided_slice %6 {offsets = [0, 0, 0, 0], sizes = [2, 4,,], strides = [1, 1,,]}: vector<4x8x16xf32> to vector<2x4x16xf32>
}
}
}
}
}
return
}
func @main() {
%c0 = arith.constant 0 : index
%c1 = arith.constant 1 : index
%c2 = arith.constant 2 : index
%c3 = arith.constant 3 : index
%c6 = arith.constant 6 : index
%c8 = arith.constant 8 : index
%f10 = arith.constant 10.00000e+00 : f32
%val = arith.constant 2.00000e+00 : f32
%zero = arith.constant 0.00000e+00 : f32
// normal_conv2d_test
// filter: 1,1,1,3
// in : 1,2,2,1
// out : 1,2,2,3
%filter2D_nhwc = call @alloc_4d_filled_f32(%c1, %c1, %c1, %c3, %val) :(index, index, index, index, f32) -> (memref<?x?x?x?xf32>)
%in2D_nhwc = call @alloc_4d_filled_f32(%c1, %c2, %c2, %c1, %val) : (index, index, index, index, f32) -> (memref<?x?x?x?xf32>)
%out2D_nhwc = call @alloc_4d_filled_f32(%c1, %c2, %c2, %c3, %zero) : (index, index, index, index, f32) -> (memref<?x?x?x?xf32>)
memref.store %f10, %in2D_nhwc[%c0, %c0, %c1, %c0] : memref<?x?x?x?xf32>
call @conv_2d_nhwc_hwcf(%in2D_nhwc, %filter2D_nhwc, %out2D_nhwc) : (memref<?x?x?x?xf32>, memref<?x?x?x?xf32>, memref<?x?x?x?xf32>) -> ()
%out2D_nhwc_ = memref.cast %out2D_nhwc : memref<?x?x?x?xf32> to memref<*xf32>
call @print_memref_f32(%out2D_nhwc_): (memref<*xf32>) -> ()
// pw_conv2d_test
// filter: 1,1,1,3
// in : 1,2,2,1
// out : 1,2,2,3
%filter2D_nhwc_pw = call @alloc_4d_filled_f32(%c1, %c1, %c1, %c3, %val) :(index, index, index, index, f32) -> (memref<?x?x?x?xf32>)
%in2D_nhwc_pw = call @alloc_4d_filled_f32(%c1, %c2, %c2, %c1, %val) : (index, index, index, index, f32) -> (memref<?x?x?x?xf32>)
%out2D_nhwc_pw = call @alloc_4d_filled_f32(%c1, %c2, %c2, %c3, %zero) : (index, index, index, index, f32) -> (memref<?x?x?x?xf32>)
memref.store %f10, %in2D_nhwc_pw[%c0, %c0, %c1, %c0] : memref<?x?x?x?xf32>
call @pw_cbsm_conv2d(%in2D_nhwc_pw, %filter2D_nhwc_pw, %out2D_nhwc_pw) : (memref<?x?x?x?xf32>, memref<?x?x?x?xf32>, memref<?x?x?x?xf32>) -> ()
%out2D_nhwc_pw_ = memref.cast %out2D_nhwc_pw : memref<?x?x?x?xf32> to memref<*xf32>
call @print_memref_f32(%out2D_nhwc_pw_): (memref<*xf32>) -> ()
// dealloc memref
memref.dealloc %filter2D_nhwc : memref<?x?x?x?xf32>
memref.dealloc %in2D_nhwc : memref<?x?x?x?xf32>
memref.dealloc %out2D_nhwc : memref<?x?x?x?xf32>
memref.dealloc %filter2D_nhwc_pw : memref<?x?x?x?xf32>
memref.dealloc %in2D_nhwc_pw : memref<?x?x?x?xf32>
memref.dealloc %out2D_nhwc_pw : memref<?x?x?x?xf32>
return
}
}