A small parallel-compute chip for the ECOS 2610 shuttle: tiny-gpu with four cores of four threads, its program and data memories on chip, loaded and started by a host over SPI.
Sixteen threads run the same kernel at once, each finding its own data from its thread number: the computing model of a GPU at its smallest. The same hardware shades an image, one pixel a thread, and runs a small neural network over a batch, one sample a thread. There is no processor on the chip. A host writes the kernel and the data through an SPI port that speaks the protocol of spis, starts it, waits for the done pin and reads the data memory back.
tiny-gpu is taken unmodified through sv2v, as upstream builds it, and htest/tieoff.py ties to zero the four inputs that upstream leaves unconnected on the program memory controller. This repository adds the two memories and the management port in hwsrc/.
The board wiring, the register map, the chip tests and the limits are in docs/流片说明.md; the tape-out report is generated from that file.
$ python3 sw/gpu.py emu sw/kernels/shade.asm --show 8 # the software model alone
$ python3 sw/gpu.py --ftdi ftdi://ftdi:232h/1 ident # 4 cores, 4 threads a block
$ python3 sw/gpu.py --ftdi ftdi://ftdi:232h/1 run sw/kernels/shade.asm --show 8sw/gpu.py holds an assembler for tiny-gpu kernels, a software model that follows what the hardware actually does, and the SPI operations. run compares what the chip returns with the model.
| Kernel | Threads | Computes |
|---|---|---|
saxpy.asm |
16 | y = 3x + y |
matmul.asm |
16 | a 4 × 4 matrix product, one element a thread, with a loop |
shade.asm |
64 | an 8 × 8 greyscale image, one pixel a thread |
mlp.asm |
16 | batch inference: a 4-4-1 perceptron with ReLU tells the parity of a 2 × 2 pattern, one pattern a thread, weights in the data memory |
One thing the model has to copy: the comparison upstream implements only tells equal from unequal, so BRn branches when the operands differ.
sw/pocl/ adds a tgpu device to PoCL. Host programs call the plain OpenCL 1.2 API.
$ bash sw/pocl/build.sh build/pocl ~/.local/pocl-tgpu # PoCL v7.2 without LLVM, plus the clrun example
$ python3 sw/gpu.py --ftdi ftdi://ftdi:232h/1 serve 2610 & # or --model, with no board at all
$ export POCL_DEVICES=tgpu POCL_TGPU0_PARAMETERS=tcp:127.0.0.1:2610
$ clrun builtin 100 # pocl.add.i8, pocl.mul.i8, pocl.copy.i8
$ python3 sw/gpu.py build sw/kernels/shade.asm -o shade.tgk && clrun tgk shade.tgk 8Built-in kernels keep PoCL's names and semantics, and the device splits the data into tiles that fit the 256-byte data memory. A .tgk file goes through clCreateProgramWithBinary: one kernel, main, whose only argument is the whole data memory. There is no OpenCL C compiler, as that would need an LLVM backend for tiny-gpu.
The same computations and the same triangles, written against the standard APIs and run on Mesa's software drivers, give byte-identical results to the chip. sw/vk/ holds five GLSL compute shaders, one per assembly kernel, with vkrun to dispatch one; vkdraw sends triangles through the Vulkan graphics pipeline into an 8 × 8 framebuffer. sw/gl/gldraw does the same through OpenGL 4.5 on a surfaceless EGL display. Vertices from gpu.py are pixel centres, shifted by 0.5 on the way in, and edges follow the top-left rule.
$ bash sw/vk/build.sh build/vk && bash sw/gl/build.sh build/gl
$ echo "1.5 1.5 6.5 2.5 3.5 6.5 200" | build/vk/vkdraw build/vk/tri_vert.spv build/vk/tri_frag.spv | xxd
$ echo "1.5 1.5 6.5 2.5 3.5 6.5 200" | EGL_PLATFORM=surfaceless build/gl/gldraw | xxd$ ran test to2610-gpu # the chip, OpenCL, Vulkan and OpenGL tests, on the Verilog file that goes to the shuttle
$ ran asic to2610-gpu # to2610_gpu.v, ecc at 50 MHz, report.json
$ ran asic to2610-gpu --no-run # only the Verilog file and ecc.tomlThis repository: 任选其一 MIT · Apache 2.0 · 木兰宽松许可证 第2版.
SPDX-License-Identifier: MIT OR Apache-2.0 OR MulanPSL-2.0
tiny-gpu itself has no licence file upstream. It is fetched through the tiny-gpu repository as a submodule and neither changed nor copied here.
除非另行说明，你提交的贡献按上述三者同时授权，不附加其他条件。