subgraph: validate filter shape dims[1..3] in conv/deconv/depthwise-conv

[tinhien11]

subgraph: validate filter shape dims[1..3] in conv/deconv/depthwise-conv

Summary

xnn_define_convolution_2d validated only filter_value->shape.dim[0] (== group_output_channels * groups) but not dim[1], dim[2], dim[3] against kernel_height, kernel_width, and group_input_channels. When a filter tensor is declared with dims[1..3] smaller than the convolution's kernel parameters, weight packing (xnn_pack_f32_conv_goki_w) reads nc * ks * kc elements from the filter buffer — where ks = kernel_height * kernel_width comes from the convolution parameters, not the filter's declared shape — causing a heap out-of-bounds read during xnn_create_runtime_v2.

xnn_define_deconvolution_2d and xnn_define_depthwise_convolution_2d had no filter shape validation at all (not even dim[0]).

This PR adds full 4D filter shape validation to all three operators, rejecting mismatched shapes at define-time with xnn_status_invalid_parameter.

Root Cause

Filter shape layout (NHWC convention):

• Conv/Deconv: [groups * group_output_channels, kernel_height, kernel_width, group_input_channels]
• Depthwise: [1, kernel_height, kernel_width, input_channels * depth_multiplier]

Weight packing reads nc * ks * kc floats from the filter buffer, where:

• nc = group_output_channels (from conv params)
• ks = kernel_height * kernel_width (from conv params)
• kc = group_input_channels (from conv params)

The packed-weights destination buffer is sized from convolution parameters (safe — no OOB write), but the source filter buffer is sized from the declared filter shape. When dims[1..3] are smaller than the conv params, the read overflows the filter buffer.

PoC

Filter declared as {32, 3, 3, 3} = 864 floats. Convolution defined with kernel_height=5, kernel_width=5, group_input_channels=3, group_output_channels=32, groups=1. The dim[0] check passed (32 == 32 * 1). Weight packing reads 32 * 25 * 3 = 2400 floats from the 864-float buffer → 1536-float (6144-byte) heap OOB read.

ASan output (before fix, HEAD 6c48650)

==51790==ERROR: AddressSanitizer: heap-buffer-overflow on address 0x61f000000e90
READ of size 4 at 0x61f000000e90 thread T0
    #0 __asan_memmove+0x2a4
    #1 xnn_pack_f32_conv_goki_w packing.cc:3553
    #2 create_igemm convolution-nhwc.c:452
    #3 create_convolution2d_nhwc convolution-nhwc.c:806
    #4 xnn_create_convolution2d_nhwc_pf32 convolution-nhwc.c:2559
    #5 create_convolution_operator convolution-2d.c:261
    #6 xnn_create_runtime_v4 runtime.c:727
    #7 xnn_create_runtime_v2 runtime.c:219

0x61f000000e90 is located 144 bytes after 3456-byte region [0x61f000000080,0x61f000000e00)
SUMMARY: AddressSanitizer: heap-buffer-overflow in xnn_pack_f32_conv_goki_w packing.cc:3553

After fix

[*] calling xnn_define_convolution_2d (kernel 5x5, filter shape {32,3,3,3})...
xnn_define_convolution_2d failed: 2  (xnn_status_invalid_parameter)

Bug blocked at define-time, no crash at runtime.

Testing

• convolution-2d-test: 11/11 PASSED
• deconvolution-2d-test: 8/8 PASSED
• depthwise-convolution-2d-test: 8/8 PASSED

Built with -fsanitize=address -g -O1, all tests pass under ASan.

Fixes #11118