From a5be0a517771814e8c115d905529e299439054f3 Mon Sep 17 00:00:00 2001 From: Matt Fishman Date: Fri, 14 Jun 2024 15:07:16 -0400 Subject: [PATCH 1/2] [Docs] Update RunningOnGPUs.md --- docs/src/RunningOnGPUs.md | 17 +++++++++++------ 1 file changed, 11 insertions(+), 6 deletions(-) diff --git a/docs/src/RunningOnGPUs.md b/docs/src/RunningOnGPUs.md index 45b9d26bed..25a1df4cfc 100644 --- a/docs/src/RunningOnGPUs.md +++ b/docs/src/RunningOnGPUs.md @@ -62,9 +62,14 @@ The table below summarizes each backend's current capabilities. | | CUDA | cuTENSOR | ROCm | Metal | oneAPI | |------------------------------|------|------------|--------|--------|--------| -| Contractions (dense) | ✓ (cuBLAS) | ✓ | ✓ | ✓ | N/A | -| QR (dense) | ✓ (cuSOLVER) | ✓ (cuSOLVER) | On CPU | On CPU | N/A | -| SVD (dense) | ✓ (cuSOLVER) | ✓ (cuSOLVER) | On CPU | On CPU | N/A | -| Eigendecomposition (dense) | ✓ (cuSOLVER) | ✓ (cuSOLVER) | On CPU | On CPU | N/A | -| Double precision (`Float64`) | ✓ (cuSOLVER) | ✓ | ✓ | N/A | N/A | -| Block sparse | In progress | In progress | In progress | In progress | N/A | +| Contractions (dense) | ✓ (cuBLAS) | ✓ | ✓ | ✓ | N/A[^oneapi] | +| QR (dense) | ✓ (cuSOLVER) | ✓ (cuSOLVER) | On CPU[^linalg] | On CPU[^linalg] | N/A[^oneapi] | +| SVD (dense) | ✓ (cuSOLVER) | ✓ (cuSOLVER) | On CPU[^linalg] | On CPU[^linalg] | N/A[^oneapi] | +| Eigendecomposition (dense) | ✓ (cuSOLVER) | ✓ (cuSOLVER) | On CPU[^linalg] | On CPU[^linalg] | N/A[^oneapi] | +| Double precision (`Float64`) | ✓ | ✓ | ✓ | N/A[^metal] | N/A[^oneapi] | +| Block sparse | ✓[^blocksparse] | ✓[^blocksparse] | ✓[^blocksparse] | ✓[^blocksparse] | N/A[^oneapi] | + +[^linalg]: Some GPU vendors have not implemented certain matrix factorizations, or the ones they have implemented are not efficient compared to running on CPU, so as a workaround we perform those operations on CPU by transferring the data back and forth from GPU to CPU. We will add support for running those operations on GPU as they become available. If your algorithm's cost is dominated by those operations you won't see any speedup by trying to run it on those kinds of GPUs. +[^blocksparse]: Support is experimental. Operations may not be fully optimized and could have bugs. +[^oneapi]: We plan to add Intel GPU support through Julia's oneAPI.jl interface but don't have any Intel GPUs to test on right now. +[^metal]: Apple doesn't support double precision floating point operations on their GPUs, see Section 2.1 of the [Metal Shading Language Specification](https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf). Until it does, we can't support double precision operations on Apple GPUs. From eaf8cb4d3e6f2dac19af249d8f21ec651b9274de Mon Sep 17 00:00:00 2001 From: Matt Fishman Date: Fri, 14 Jun 2024 15:14:34 -0400 Subject: [PATCH 2/2] [ITensors] Bump to v0.6.13 --- Project.toml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Project.toml b/Project.toml index 8ca62a6d3f..3b05f24e1e 100644 --- a/Project.toml +++ b/Project.toml @@ -1,7 +1,7 @@ name = "ITensors" uuid = "9136182c-28ba-11e9-034c-db9fb085ebd5" authors = ["Matthew Fishman ", "Miles Stoudenmire "] -version = "0.6.12" +version = "0.6.13" [deps] Adapt = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"