Fuse Average Pooling with Convolution

Fuse average pooling with convolution

@77 = gpu::code_object[code_object=9344,symbol_name=pad_kernel,global=262848,local=1024,](@57,@76) -> float_type, {1, 192, 37, 37}, {262848, 1369, 37, 1}
@78 = load[offset=705600,end=1646400](@1) -> float_type, {1, 192, 35, 35}, {235200, 1225, 35, 1}
@79 = gpu::pooling[mode=average,padding={0, 0, 0, 0},padding_mode=0,stride={1, 1},lengths={3, 3},dilations={1, 1},ceil_mode=0,lp_order=2,dyn_global=0](@77,@78) -> float_type, {1, 192, 35, 35}, {235200, 1225, 35, 1}
@80 = hip::hip_copy_literal[id=main:@literal:147] -> float_type, {32}, {1}
@81 = hip::hip_copy_literal[id=main:@literal:148] -> float_type, {32, 192, 1, 1}, {192, 1, 1, 1}
@82 = load[offset=1960000,end=2116800](@1) -> float_type, {1, 32, 35, 35}, {39200, 1225, 35, 1}
@83 = broadcast[axis=1,out_lens={1, 32, 35, 35}](@80) -> float_type, {1, 32, 35, 35}, {0, 1, 0, 0}
@84 = gpu::code_object[code_object=7312,symbol_name=mlir_convolution_add_relu,global=9856,local=64,](@83,@79,@81,@82) -> float_type, {1, 32, 35, 35}, {39200, 1225, 35, 1}

We could convert the average pooling to a convolution(and get rid of the pad_kernel). Then we will have a convolution into another convolution. Convolution is not commutative like gemms, but I believe we could rewrite it as a backwards convolution where we apply convolution to the weights.

Related to inceptionV3 and other models