D991 - A device-addressed dispatch is traced on the host with its images

Status: decided Date: 2026-10-25 known_by: measured

orbistoun-cli trace-dispatch runs a device-addressed capture (D980) on the host evaluator and compares what it stores with what the device stores, over the same inputs. The host trace reads the capture's guest ranges where they lay, and its images through wave_trace_with_images. A device run places the guest ranges again and compares them as guest-<n>.bin. Stored images are compared texel by texel. A capture that noted its source supplies its own program and context, so --kept is no longer needed.

What the host trace models. Reads and writes:

• image_bvh_intersect_ray and its 64-bit form, through the node test the translator emits (model::bvh), in host arithmetic;
• image_load, image_load_mip, image_sample_lz and image_sample_l at level zero, of the captured image a descriptor names. The image is matched by its extent, its layers and its format;
• image_store and image_store_mip, recorded as texels (WaveTrace::texels);
• buffer_load_format_x through a descriptor of 32-bit elements, read as a word.

It also computes:

• src_scc as a source;
• inline 1/(2*pi);
• an output clamp, under the shader's DX10_CLAMP mode (LaneProgram::clamping);
• the GLSL.std.450 operations, the transcendentals among them, in a trace only. A run that forms addresses still leaves them unknown.

Several groups, or all. --group x,y repeats. --every-group runs every group and compares each stored image whole. A texel the device changed that no host store wrote also counts as a difference.

A global access's offset is signed. The OFFSET field of a global or scratch access is a signed twelve-bit number (RDNA 2 instruction set, FLAT). The reference assembler encodes offset:-2048 as 0x800. The decoder hands the field as it lies. The translator, the host evaluator and native lane runs read it through model::flat_offset, which sign-extends it for those two segments. Before this, a probe's store at offset:-2048 landed 2048 bytes above its address on the device. The test is a_global_access_below_its_address_reaches_below_it.

Measured on GTA V's ray-traversal pass at 0x74005b7e8600. Its 2,171 instructions run as [240, 135, 1] groups of one wave of sixty-four, device-addressed (D964), two lanes an invocation (D973). The capture is from run 347's in-world frames. The pass writes a one-channel shadow mask at half resolution. It casts one shadow ray a pixel, wave-coherently: the wave walks the hierarchy together on a stack held in lanes, and box and triangle nodes go to image_bvh_intersect_ray.

• The pass is deterministic on the device: four replays left the same output hash.
• With --every-group, all 32,400 groups ran to their end on the host, in 12 minutes. Traversal is deepest over the characters, up to 31,790 instructions a wave. All 152,045 texels the host stored are the device's, bit for bit, and the device changed no texel the host did not store. The mask is the same whole.
• The first comparison, over 360 groups, disagreed on 22 texels of 8,932. A probe stored registers to guest memory from patched code (--patch, --jump). It found the host reading a normal map's channels reordered twice. The capture already holds an 8_8_8_8_UNORM image's texels in the order its selects give (channel_order, as the binding hands them to the device), and the host applied the selects again. With that fixed, nothing differs.
• The hierarchy's node format, read from the capture: the TLAS and the BLASes use only boxes of floats, apart from five type-6 nodes, which the guest sends to the instruction too. Leaf lists name triangle nodes. Almost all (82,185) are type 0, a single triangle with its word 15 (9) naming the barycentric order the test already returns. A node whose type selects another of several triangles, or other id words, is not exercised by this pass. Amended by D992, which completes the node tests for the rest. D993 makes the fused multiply-adds round once on the device and in the host trace alike.

So the pass's output is correct as translated. The coloured static on in-world frames is not this pass's: it writes one channel, and its replayed mask is plain lit or shadowed, noisy only on the characters, as a one-ray-a-pixel shadow is before its denoiser.

Limits. The device and the host approximate the transcendentals and v_rcp_f32 differently in their last places, as the guest's hardware does too. For this pass, one-ulp changes to each of them on the host changed no texel. A sampled image is filtered at the eight sub-texel bits a device uses, so a bilinear sample can differ in its last bits. Atomics are not modelled. The tile flags the pass ORs into its buffer are compared only through the stored image.

Rejected:

• Comparing only the replay's output hash between builds: two builds agreeing says nothing about either being right.
• Matching a descriptor to a captured image by the order the program loads descriptors: the binding's order is the translation's, which the host does not run.