[trimul] add CUDA, HIP, and SYCL benchmarks

README.md

@@ -85,7 +85,7 @@ Each benchmark falls into a single category. While such classification is not ac
     adjacent, aligned-types, asta, asyncAllocation, awbarrier, blockAccess, blockexchange, blockScan, collision, concurrentKernels, conversion, dispatch, dp4a, graphExecution, ert, interleave, intrinsics-cast, kernelLaunch, layout, mallocFree, maxFlops, mixbench, nosync, openmp, overlap, pad, pitch, popcount, pointerchase, prefetch, reverse, ring, saxpy-ompt, shuffle, simpleMultiDevice, streamCreateCopyDestroy, streamOrderedAllocation, streamPriority, streamUM, tensorAccessor, threadfence, warpexchange, vote, wmma, wordcount, zerocopy 
 
 ### Machine learning  
-    accuracy, adam, adamw, addBiasQKV, addBiasResidualLayerNorm, attention, attention-paged, attentionMergeState, attentionMultiHead, attentionMultiHeadKVCache, backprop, bincount, bn, channelShuffle, channelSum, clink, concat, crossEntropy, dense-embedding, dropout, dwconv, dwconv1d, expdist, flip, gd, gelu, ge-spmm, geglu, glu, gmm, gru, gru2, kalman, kmc, kmeans, knn, layernorm, lda, lif, logprob, lr, lrn, mask, matern, maxpool3d, mcpr, meanshift, megablocks, mergeVS, mf-sgd, mlaDecode, mlp, mmcsf, mnist, moe, moe-align, moe-sum, mrc, multinomial, muon, nlll, nonzero, overlay, p4, page-rank, permute, perplexity, pointwise, pool, qkv, qtclustering, remap, relu, resnet-kernels, rmsnorm, rowwiseMoments, rotary, sampling, scel, silu, snicit, softmax, softmax-fused, softmax-online, ssm, stddev, streamcluster, tsne, twell, unfold, vol2col, wedford, winograd, word2vec
+    accuracy, adam, adamw, addBiasQKV, addBiasResidualLayerNorm, attention, attention-paged, attentionMergeState, attentionMultiHead, attentionMultiHeadKVCache, backprop, bincount, bn, channelShuffle, channelSum, clink, concat, crossEntropy, dense-embedding, dropout, dwconv, dwconv1d, expdist, flip, gd, gelu, ge-spmm, geglu, glu, gmm, gru, gru2, kalman, kmc, kmeans, knn, layernorm, lda, lif, logprob, lr, lrn, mask, matern, maxpool3d, mcpr, meanshift, megablocks, mergeVS, mf-sgd, mlaDecode, mlp, mmcsf, mnist, moe, moe-align, moe-sum, mrc, multinomial, muon, nlll, nonzero, overlay, p4, page-rank, permute, perplexity, pointwise, pool, qkv, qtclustering, remap, relu, resnet-kernels, rmsnorm, rowwiseMoments, rotary, sampling, scel, silu, snicit, softmax, softmax-fused, softmax-online, ssm, stddev, streamcluster, trimul, tsne, twell, unfold, vol2col, wedford, winograd, word2vec
 
 ### Math
     atan2, axpby, bgmv, blas-dot, blas-fp4gemm, blas-fp8gemm, blas-gemm, blas-gemmBatched, blas-gemmStridedBatched, blas-gemmEx, blas-gemmEx2, blas-groupgemm, blas-mxfp6gemm, blas-mxfp8gemm, braycurtis, complex, cross, determinant, divergence, dp, eigenvalue, f16max, f16sp, f8cast, fresnel, fwt, gaussian, geam, gels, gemv, hadamard, hellinger, hmm, idivide, interval, jaccard, jacobi, jacobian, kurtosis, lanczos, langford, lci, lebesgue, leukocyte, lfib4, log2, lud, ludb, lut-gemm, michalewicz, matrix-rotate, matrixT, minkowski, mr, mrg32k3a, norm2, nqueen, ntt, oziMMU, phmm, pnpoly, quant3MatMul, reverse2D, rfs, romberg, rsc, schur-complement, sddmm-batch, secp256k1, simpleSpmv, slu, spd2s, spgeam, spgemm, spmm, spmv, spnnz, sps2d, spsort, sptrsv, thomas, wyllie, zeropoint
@@ -1755,6 +1755,9 @@ Early results are shown [here](results/README.md)
 ### tridiagonal (opencl)
   Matrix solvers for large number of small independent tridiagonal linear systems(http://developer.download.nvidia.com/compute/cuda/3_0/sdk/website/OpenCL/website/samples.html)
 
+### trimul (cuda)
+  Triangle multiplicative update in the AlphaFold Evoformer (https://github.com/NVIDIA/cuEquivariance)
+
 ### tsa (cuda)
   Trotter-Suzuki approximation (https://bitbucket.org/zzzoom/trottersuzuki/src/master/)
 

src/CMakeLists.txt

@@ -431,6 +431,7 @@ set(HECBENCH_POC_BENCHMARKS
     tonemapping
     tpacf
     tqs
+    trimul
     tsne
     tsp
     unfold

src/trimul-cuda/CMakeLists.txt

@@ -0,0 +1,11 @@
+# trimul-cuda/CMakeLists.txt
+
+# -fopenmp parallelizes the CPU reference in reference.h
+add_hecbench_benchmark(
+    NAME trimul
+    MODEL cuda
+    SOURCES main.cu
+    COMPILE_OPTIONS -Xcompiler -fopenmp
+    LINK_LIBRARIES cublas -fopenmp
+    CATEGORIES ml
+)

src/trimul-cuda/LICENSE

@@ -0,0 +1,20 @@
+The triangle multiplicative update algorithm benchmarked here is derived from
+NVIDIA cuEquivariance:
+  https://github.com/NVIDIA/cuEquivariance
+  cuequivariance_jax/cuequivariance_jax/triangle/_triangle_multiplicative_update.py
+  cuequivariance_jax/cuequivariance_jax/triangle/_sigmoid_gated_dual_gemm.py
+
+SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+SPDX-License-Identifier: Apache-2.0
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.

src/trimul-cuda/Makefile

@@ -0,0 +1,67 @@
+#===============================================================================
+# User Options
+#===============================================================================
+
+# Compiler can be set below, or via environment variable
+CC        = nvcc
+OPTIMIZE  = yes
+DEBUG     = no
+# BF16 WMMA requires Ampere or newer.
+ARCH      = sm_80
+LAUNCHER  ?=
+
+#===============================================================================
+# Program name & source code list
+#===============================================================================
+
+program = main
+
+source = main.cu
+
+obj = $(source:.cu=.o)
+
+#===============================================================================
+# Sets Flags
+#===============================================================================
+
+# Standard Flags
+# The CPU reference is OpenMP-parallel; without -fopenmp it still builds and
+# runs, just serially, which is slow at the larger sequence lengths.
+CFLAGS := $(EXTRA_CFLAGS) -std=c++17 -Xcompiler -Wall -Xcompiler -fopenmp \
+          -arch=$(ARCH)
+
+# Linker Flags
+LDFLAGS = -lcublas -Xcompiler -fopenmp
+
+# Debug Flags
+ifeq ($(DEBUG),yes)
+  CFLAGS += -g -DDEBUG
+  LDFLAGS  += -g
+endif
+
+# Optimization Flags
+ifeq ($(OPTIMIZE),yes)
+  CFLAGS += -O3
+endif
+
+#===============================================================================
+# Targets to Build
+#===============================================================================
+
+$(program): $(obj) Makefile
+	$(CC) $(CFLAGS) $(obj) -o $@ $(LDFLAGS)
+
+%.o: %.cu reference.h Makefile
+	$(CC) $(CFLAGS) -c $< -o $@
+
+clean:
+	rm -rf $(program) $(obj)
+
+# Sequence lengths AlphaFold actually runs at, with the pair-representation
+# width of 128: the AF2 training crops (256, 384) and the AF3 ones (384, 640,
+# 768).  Batch is 1 per device in both.
+run: $(program)
+	$(LAUNCHER) ./$(program) 1 256 128 100
+	$(LAUNCHER) ./$(program) 1 384 128 100
+	$(LAUNCHER) ./$(program) 1 640 128 100
+	$(LAUNCHER) ./$(program) 1 768 128 100