Reborn in 2010, the AI ​​in my mind is too powerful.

AMD’s headquarters are located in Sunnyvale and consist of just five gray office buildings.

In reality, this is the typical form of a Silicon Valley company; not every company can build a headquarters as grand as Oracle's.

In particular, AMD and Intel have a fierce rivalry, and AMD has never been able to beat Nvidia in the graphics card market. AMD has always been short of funds and has no right to pursue superficial grandeur.

Chen Yiran had recently visited two impressive headquarters, and seeing AMD's, he even felt it was a bit shabby. But once he actually went inside, it wasn't bad at all.

The only somewhat unique feature is the several glass display cases in the lobby, which display chips ranging from the K6 to the latest Bulldozer architecture, arranged in a row according to their age.

Unlike the quiet atmosphere of Oracle, this place is more lively. Engineers carrying motherboards, graphics cards, and various components often walk around in the corridors, and sometimes there is even a strange mixture of solder and coffee smells in the air.

Harry, who had been working on the project, and acting CEO Thomas Seifert greeted Chen Yiran. They showed Chen Yiran around and then led him to the fourth floor.

Although Chen Yiran doesn't need to come to work every day, he only needs to align his work process with AMD's staff every week, AMD still provided him with a very spacious office of over 40 square meters.

The office window faces a low mountain range, and there's a large L-shaped corner desk. Next to it is a small workstation, complete with a separate coffee machine and a small refrigerator stocked with Coca-Cola.

It's clear that a lot of thought went into the arrangement.

Chen Yiran nodded: "Okay, you guys go ahead and get busy. Have the liaison team come over, and we'll get started right away."

Harry and Thomas nodded and went out. A moment later, four engineers knocked on the door and came in. The one leading them was a Chinese man in his early forties, wearing metal-rimmed glasses.

"Mr. Chen, I am Marvin, the technical lead for the OpenCL compiler team. These three are Lin, Sunil, and Amy, all from our compiler backend and memory management group."

We have a few engineers in our team who can work with us. Shall we get started?

Chen Yiran nodded: "Okay, you guys tell me your thoughts first."

The group didn't go to the conference room. Marvin connected a laptop to the projector on the conference table and projected an architecture diagram onto the white wall next to it.

The above is the complete compilation pipeline for the AMD OpenCL runtime. The OpenCL C source files are parsed by the front end to generate an intermediate representation, then converted into AMD IL intermediate language by the unified optimizer, and finally the GPU back-end compiler maps the IL to specific VLSI slots.

"This is our current compilation pipeline, which has four layers and about 60,000 lines of C++. The main bottleneck is in the backend, in the process of packaging very long instructions."

Our GPUs use a very long instruction word (VLSI) architecture. The compiler needs to bundle multiple independent instructions into a single execution bundle to fully utilize the stream processing units, but this instruction bundling is highly dependent on the structure of the input code.

Marvin displayed another slide on the projector, which contained several lines of code highlighted in red.

"For a relatively well-organized, computationally intensive kernel, our backend can fill 70% to 80% of the instruction slots. But once there are complex branches or non-contiguous global memory accesses in the kernel, the fill rate will drop to below 40%."

Sunil added, "Our current approach is two-step. First, we'll add a pre-analysis to global memory access pattern recognition. Second, we'll add a heuristic algorithm to the VLSI scheduler."

Chen Yiran looked at the screen without speaking immediately, actually discussing with Jasmine.

【( > ρ < 」) No wonder AMD is no good, what is all this nonsense?】

Seeing that Chen Yiran didn't interrupt, the four people discussed it amongst themselves for a long time, looking like they had a lot of ideas.

Chen Yiran couldn't hold back any longer and covered her face, asking, "Did you write your own AMD IL intermediate language layer, or did you reuse the one from ATI?"

Marvin glanced at Lin, the only one in the team who came from ATI, having joined AMD after ATI was acquired.

"It's inherited from the ATI era. We've done several rounds of restructuring, but the core structure hasn't changed."

"The middle indicates the number of opcodes."

"About four hundred or so."

"How many of these are actually being used by you right now?"

Lin thought for a moment and said, "A little over half."

Chen Yiran took a sip of cola and said, "More than four hundred opcodes, but only two hundred are used. I remember ATI was founded in 1985, right? And then it was acquired by AMD in 2006. Who knows how many years these opcodes have been sitting in there?"

Your current compilation pipeline has four layers. Each layer has its own data structure, each layer has its own optimization rules, and each layer needs to perform format conversion when passing data to the next layer.

What you're talking about are classic methods, but the problem now isn't with the methods themselves.

The problem is that an overly bloated intermediate representation can lead to information loss during inter-layer transformations, and the packaging stage may not have enough dependency information for optimal scheduling.

This is rubbish, I'll rewrite it from scratch.

The four engineers were somewhat unconvinced, but Chen Yiran's technical authority was too strong, and it was widely rumored that he had an exceptionally wide range of technical expertise, so they dared not refute him.

Chen Yiran ignored them and said to himself, "Very Long Instruction (VLI) instruction packing doesn't necessarily have to be placed at the end. If the compiler already knows the dependencies and data paths of each instruction when it generates the code, it can mark which instructions can be parallelized during the generation process."

Marvin adjusted his glasses: "It's possible, but it means the AMD IL intermediate language layer will have to be scrapped."

"Is there a problem?"

Marvin hesitated to speak rashly, and after thinking for a moment, said, "I need to report this. First, if we really want to do this, our project team will need more people to support us. Second, the related adaptation work could be enormous. I need to report this; this is a very difficult matter."

Chen Yiran nodded: "You can go and report it. I'm not going to worry about you guys. I'm going home first. My dog ​​isn't around, which will affect my work efficiency. I'm going home to start working."

【(๑•̀ㅂ•́)و✧ Let's go back and get started. It's time to show them what real skill is! Give me a week, and I'll scare them to death.】