Book contents +

XIV · AMD acceleration (welvet.rocm) / CHAPTER 13

AMD ROCm: results and suite pointers

Samuel Watson · Welvet feature bookOriginal source ↗

host-only pointers


Why it exists

Find the AMD sibling repo and live ROCm parity/timing matrices after the small usage chapters.

What it is

Print Go v1.5.0 / welvet.rocm 0.79 pins and commands for w2a go_v_rocm and rocm_smoke. No GPU required.

Where live results live

This chapter is host-only. It does not claim timings from loading the page. Sibling AMD engine: github.com/openfluke/welvet.rocm. Run the matrices on a machine with a ROCm-capable AMD GPU:

cd ../../w2a && go test ./tests/go_v_rocm -count=1 -timeout 180m
cd ../../welvet.rocm/examples/rocm_smoke && go run .
cd welvet/examples && ./run_amd.sh

Engine pin: welvet.rocm 0.79, model ABI 1, Go binding nativerocm (Welvet v1.5.0). Platforms: Linux x86_64 + Windows amd64 for Mount; Linux amd64 for EnableRouting. GPU compute untested until a ROCm-capable HSA device is available.

Requirements overview: chapter 96.

Go example

examples/102-amd-rocm-results/main.go

Run:cd welvet/examples/102-amd-rocm-results && source ../env.sh && go run .
// Pointers to ROCm validation matrices; no GPU run is implied.
package main

import "fmt"

func main() {
	fmt.Println("AMD ROCm results pointers (not a live benchmark)")
	fmt.Println()
	fmt.Println("Engine pin: welvet.rocm 0.79 · model ABI 1 · Go binding: nativerocm (Welvet v1.5.0)")
	fmt.Println("Usage tour: 96 mount · 97 Mamba dtypes · 98 train modes · 99 SyncToGo/ENTITY · 100 JSON · 101 mixed")
	fmt.Println()
	fmt.Println("Live matrices (sibling checkouts):")
	fmt.Println("  cd ../../w2a && go test ./tests/go_v_rocm -count=1 -timeout 180m")
	fmt.Println("  cd ../../welvet.rocm/examples/rocm_smoke && go run .")
	fmt.Println()
	fmt.Println("Install shared library into Go:")
	fmt.Println("  cd ../../welvet.rocm && ./scripts/linux_install_go.sh")
	fmt.Println()
	fmt.Println("Small usage tour without the large matrix:")
	fmt.Println("  ../run_amd.sh")
}

Output

exit 0 · last run via go run .

AMD ROCm results pointers (not a live benchmark)

Engine pin: welvet.rocm 0.79 · model ABI 1 · Go binding: nativerocm (Welvet v1.5.0)
Usage tour: 96 mount · 97 Mamba dtypes · 98 train modes · 99 SyncToGo/ENTITY · 100 JSON · 101 mixed

Live matrices (sibling checkouts):
  cd ../../w2a && go test ./tests/go_v_rocm -count=1 -timeout 180m
  cd ../../welvet.rocm/examples/rocm_smoke && go run .

Install shared library into Go:
  cd ../../welvet.rocm && ./scripts/linux_install_go.sh

Small usage tour without the large matrix:
  ../run_amd.sh

Validate (harness)

cd welvet/examples/102-amd-rocm-results && source ../env.sh && go run .