mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-05 06:01:23 +00:00
7330ec37ee
* Create new copies of existing device struct and gridwise struct for batched_gemm_softmax_gemm and disable the softmax part. Still based on old wmma pipelines. Also copy the example and remove the softmax part from the reference calculation. Works and results match reference except for tiny float errors in problem 2. * Turn DeviceBatchedGemmGemm_Wmma_CShuffleV3 into a proper DeviceBatchedGemmGemm derived class, with the right argument and invoker functions. Update example to use new definitions. * Remove unused cross-attention and self-attention kernels, arguments, and invokers. Also remove other unused Argument types. * Remove masking related code, test unusual sizes in example. * Remove remaining softmax related code from GridwiseBatchedGemmGemm_wmma_cshuffle_v3 and example. * Remove code related to numDims, bias, and TensorSpec from Device struct and example. * Add layout template parameters to device struct * Move (NPerBlock, LTilePerBlock) device struct template arguments up by two places to match XDL template argument ordering. * Merge accumulation data types into one type to match XDL device struct. * Remove NPerWmma template parameter from device struct and just set it equal to LPerWmma. Now device struct template params exactly match those for XDL batched gemm gemm. * Add support for RCCR layout and test this in example * Add batched_gemm_gemm_wmma to instance library + profiler, and add gtest just like for xdl. * Add RCCR instance and additional RCRR instance to library. * Remove unused permute and alpha related code. Time all tests. Fix B1 strides in argument verification. * Remove references to G0, G1 in favor of batch, reduce dimensionality of length and stride arrays. * Managed to replace old wmma gridwise pipeline and blockwise struct with new wmma blockwise pipeline. Some cleanup required but all tests pass. * Make TransposeC a proper template parameter that gets passed all the way from BlockGemmPipeline_Selector to WmmaGemm so we can use the correct settings for bacthed gemm gemm as well as regular gemm. Gemm universal tests now pass again. * Replace old LoopSched and PipelineVer params with BlockwiseGemm pipeline equivalents, and use these in instance factory. The v3 pipeline does not work yet, but v1 works for intrawave and interwave. * Adapt the A wave descriptor to deal with RDNA4 wmma. This fixes batched gemm gemm functionality on RDNA4. * Fixed two aspects of the v3 pipeline that were incorrect: First of all the blockwise copy operator was invoked once too many in all cases (RunRead and move window), which broke batched gemm gemm when the blockwise pipeline was used multiple times. Furthermore we should be using the mainloop (hotloop) for num_k_loop >=2 instead of num_k_loop >=3. Now we can use support any K dimension. * Remove num prefetch parameter from gridwise struct since we don't use it and it doesn't do anything, * Remove unused non-lds paths. * Test and update the IsSupportedArgument() and CheckValidity() functions for all layouts + padding modes and various problem sizes. * Add a lot of instances to the profiler with various blocksizes and pipelines, all verified. * Add support for BF16: instance library, tests, and examples. * Add examples for int8 and fp8, had to add type_convert_sp template specializations for the latter. * Template the library instance lists and add default padding instances. * Move memory calculations from the kernel to the Argument contructor. Also actually parse and use the user-provided batch strides. * Actually parse and use user-provided regular strides. * More refactor: remove references to multiple dims per dims, and g0 / g1. Also move xdl specific test utils out of generic test util header. * Small post-rebase-on-develop fix due to bscale-related pipeline changes. All tests rerun + tested bscale and regular gemm. * Introduce the correct GetCThreadDescriptor function in the blockwise gemm pipelines for the TransposeC=true case. It turns out to be identical for our batched gemm gemm (gemm0) usecases, but could theoretically be different for wmma_gemm instances with smaller-than-4-byte output data size. * Remove unused NumPrefetch template parameter, we don't need to match the XDL template params one-to-one. * Implement proper TailNum and HasMainLoop template parameters for the v3 pipeline. Now the Run() function knows at compile time whether there are 1, 2, or more loops in total, and adds or removes sections accordingly. It still uses the blockwise copy operators the correct amount of times. * Add print lambda with env check and file and func to device and gridwise level compatibility error messages. Also respect compatibility in example script. * RDNA3 does not support fp8
Composable Kernel profiler
Profiler GEMM UNIVERSAL kernels
# arg1: tensor operation (gemm_universal: Universal GEMM)
# arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8; 4: f8@f16; 5: f16@f8; 6: f16->f8; 7: f8->bf16, comp f8; 8: f16@i4; 9: bf16@i4
# arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];
# 1: A[m, k] * B[n, k] = C[m, n];
# 2: A[k, m] * B[k, n] = C[m, n];
# 3: A[k, m] * B[n, k] = C[m, n])
# arg4: verification (0: no; 1: yes)
# arg5: initialization (0: no init; 1: integer value; 2: decimal value)
# arg6: print tensor value (0: no; 1: yes)
# arg7: time kernel (0=no, 1=yes)
# arg8 to 13: M, N, K, StrideA, StrideB, StrideC
# arg14: split k into mulitiple batch
# optional:
# arg15: number of warm-up cycles (default 1)
# arg16: number of iterations (default 10)
# arg17: memory for rotating buffer (default 0, size in MB)
################ op datatype layout verify init print time M N K StrideA StrideB StrideC SplitK WarmupCycles Iterations MemoryBuffer
./bin/ckProfiler gemm_universal 1 0 1 1 0 1 4096 4096 4096 4096 4096 4096 1 1 10 0
Profile GEMM kernels
#arg1: tensor operation (gemm=GEMM)
#arg2: data type (0=fp32, 1=fp16)
#arg3: matrix layout (0=NN, 1=NT, 2=TN, 3=TT)
#arg4: verification (0=no, 1=yes)
#arg5: initialization (0=no init, 1=integer value, 2=decimal value)
#arg6: print matrix value (0=no, 1=yes)
#arg7: run kernel # of times (>1)
#arg8 to 13: M, N, K, StrideA, StrideB, StrideC
################ op datatype layout verify init log repeat M___ N___ K___ StrideA StrideB StrideC
./bin/ckProfiler gemm 1 1 1 1 0 5 3840 4096 4096 4096 4096 4096
Profile 2D forward convolution kernels
#arg1: tensor operation (conv=Convolution)
#arg2: data type (0=fp32, 1=fp16)
#arg3: input tensor layout (0=NCHW, 1=NHWC)
#arg4: weight tensor layout (0=KCYX, 1=KYXC)
#arg5: output tensor layout (0=NKHW, 1=NHWK)
#arg6: verification (0=no, 1=yes)
#arg7: initialization (0=no init, 1=integer value, 2=decimal value)
#arg8: print matrix value (0=no, 1=yes)
#arg9: run kernel # of times (>1)
#arg10 to 24: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, RightPx
################ op datatype in_layout wei_layout out_layout verify init log repeat N__ K___ C___ Y X Hi__ Wi__ Strides Dilations LeftPads RightPads
./bin/ckProfiler conv2d_fwd 1 1 1 1 1 1 0 5 128 256 192 3 3 71 71 2 2 1 1 1 1 1 1
Profile contraction kernels
#arg1: tensor operation (contraction_bilinear=CONTRACTION+Bilinear)
#arg2: data type (0: fp32; 1: f64; 2: f16; 3: bf16)
#arg3: compute data type (0: fp32; 1: f64; 2: f16; 3: bf16)
#arg4: Number of dimension for M, N and K (one for all)
#arg5: matrix layout (0: A[m0, m1, k0, k1] * B[k0, k1, n0, n1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1];
# 1: A[m0, m1, k0, k1] * B[n0, n1, k0, k1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1];
# 2: A[k0, k1, m0, m1] * B[k0, k1, n0, n1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1];
# 3: A[k0, k1, m0, m1] * B[n0, n1, k0, k1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1])
#arg6: verification (0: no; 1: yes)
#arg7: initialization (0: no init; 1: integer value; 2: decimal
# value)
#arg8: print tensor value (0: no; 1: yes)
#arg9: time kernel (0: no, 1: yes)
#arg10: alpha
#arg11: beta
#arg12 to 17/29: M0, M1, N0, N1, K0, K1
#arg18/30 to 33/77: Strides for A, B, D and E (skip for default)
################ op datatype compute_datatype num_dim layout verify init log time alpha beta M0 M1 N0 N1 K0 K1
./bin/ckProfiler contraction_bilinear 0 0 2 1 0 0 0 1 1.0 1.0 128 128 128 128 128 128
Profile batched gemm multiple D kernels
#arg1: tensor operation (batched_gemm_multi_d=Batched GEMM multi D);
#arg2: data type (0: fp16; 1: int8)
#arg3: matrix layout (0: A[g, m, k] * B[g, k, n] = C[g, m, n];
# 1: A[g, m, k] * B[g, n, k] = C[g, m, n];
# 2: A[g, k, m] * B[g, k, n] = C[g, m, n];
# 3: A[g, k, m] * B[g, n, k] = C[g, m, n])
#arg4: verification (0: no; 1: yes)
#arg5: initialization (0: no init; 1: integer value; 2: decimal value)
#arg6: print tensor value (0: no; 1: yes)
#arg7: time kernel (0=n0, 1=yes)
#arg8 to 17: M, N, K, StrideA, StrideB, StrideC, BatchStrideA, BatchStrideB, BatchStrideC, BatchCount
################ op datatype layout verify init log time M N K StrideA StrideB StrideC BatchStrideA BatchStrideB BatchStrideC BatchCount
./bin/ckProfiler batched_gemm_multi_d 0 1 0 0 0 1 4096 4096 4096 4096 4096 4096 16777216 16777216 16777216 16
Profile grouped convolution backward data kernels
# arg1: tensor operation (grouped_conv_bwd_data: Grouped Convolution Backward Data)
# arg2: data type (0: Output fp32, Weight fp32, Input fp32
# 1: Output fp16, Weight fp16, Input fp16
# 2: Output bf16, Weight bf16, Input bf16
# arg3: tensor layout (0: Output[G, N, Hi, Wi, C], Weight[G, K, Y, X, C], Input[G, N, Ho, Wo, K]
# 1: Output[N, Hi, Wi, G, C], Weight[G, K, Y, X, C], Input[N, Ho, Wo, G, K])
# arg4: verification (0: no, 1: yes)
# arg5: initialization (0: no init, 1: integer value, 2: decimal value)
# arg6: print tensor value (0: no; 1: yes)
# arg7: time kernel (0: no, 1: yes)
# Following arguments (depending on number of spatial dims):
# Number of spatial dimensions (1=Conv1D, 2=Conv2D, 3=Conv3D)
# G, N, K, C,
# <filter spatial dimensions>, (ie Y, X for 2D)
# <input image spatial dimensions>, (ie Hi, Wi for 2D)
# <strides>, (ie Sy, Sx for 2D)
# <dilations>, (ie Dy, Dx for 2D)
# <left padding>, (ie LeftPy, LeftPx for 2D)
# <right padding>, (ie RightPy, RightPx for 2D)
################ op datatype layout verify init log time Ndims G N K C Y X Hi Wi Sy Sx Dy Dx LeftPy LeftPx RightPy RightPx
./bin/ckProfiler grouped_conv_bwd_data 1 0 1 1 0 1 2 32 4 192 192 3 3 28 28 1 1 1 1 1 1 1 1
Profile grouped convolution backward weight kernels
# arg1: tensor operation (grouped_conv_bwd_weight: Grouped Convolution Backward Weight)
# arg2: data type (0: Input fp32, Weight fp32, Output fp32
# 1: Input fp16, Weight fp16, Output fp16
# 2: Input bf16, Weight fp32, Output bf16
# 3: Input fp16, Weight fp16, Output fp16, Gemm bf8@fp8
# 4: Input int8, Weight int8, Output int8)
# arg3: tensor layout (0: Input[G, N, C, Hi, Wi], Weight[G, K, C, Y, X], Output[G, N, K, Ho, Wo]
# 1: Input[G, N, Hi, Wi, C], Weight[G, K, Y, X, C], Output[G, N, Ho, Wo, K]
# 2: Input[N, Hi, Wi, G, C], Weight[G, K, Y, X, C], Output[N, Ho, Wo, G, K]
# arg4: verification (0: no, 1: yes)
# arg5: initialization (0: no init, 1: integer value, 2: decimal value)
# arg6: print tensor value (0: no; 1: yes)
# arg7: time kernel (0: no, 1: yes)
# Following arguments (depending on number of spatial dims):
# Number of spatial dimensions (1=Conv1D, 2=Conv2D, 3=Conv3D)
# G, N, K, C,
# <filter spatial dimensions>, (ie Y, X for 2D)
# <input image spatial dimensions>, (ie Hi, Wi for 2D)
# <strides>, (ie Sy, Sx for 2D)
# <dilations>, (ie Dy, Dx for 2D)
# <left padding>, (ie LeftPy, LeftPx for 2D)
# <right padding>, (ie RightPy, RightPx for 2D)
# SplitK (-1 for internally computed split-K value, positive value to set k batches explicitly, or 'all' to test all internal split-K values)
################ op datatype layout verify init log time Ndims G N K C Y X Hi Wi Sy Sx Dy Dx LeftPy LeftPx RightPy RightPx SplitK
./bin/ckProfiler grouped_conv_bwd_weight 1 1 0 1 0 1 2 32 256 256 512 3 3 28 28 1 1 1 1 1 0 0 0 1
Note: This kernel use atomic add, this will cause output buffer to be accumulated multiple times, causing verification failure. To work around it, do not use CK's own timer and do verification at the same time.
Profile image to column/column to image kernels
# arg1: tensor operation ( conv_tensor_rearrange : Conv Tensor Rearrange )
# arg2: data type (0: Input fp32, Weight fp32, Output fp32
# 1: Input fp16, Weight fp16, Output fp16
# 2: Input bf16, Weight bf16, Output bf16
# 3: Input int8, Weight int8, Output int8)
# arg3: tensor layout (0: Input[G, N, Hi, Wi, C], Output[G * N * Ho * Wo, Y * X * C],
# 1: Input[N, Hi, Wi, G, C], Output[N * Ho * Wo * G, Y * X * C])
# arg4: verification (0: no, 1: yes)
# arg5: initialization (0: no init, 1: integer value, 2: decimal value)
# arg6: print tensor value (0: no; 1: yes)
# arg7: time kernel (0: no, 1: yes)
# arg8: operation type (0: ImageToColumn, 1: ColumnToImage)
# Following arguments (depending on number of spatial dims):
# Number of spatial dimensions (1=Conv1D, 2=Conv2D, 3=Conv3D)
# G, N, K, C,
# <filter spatial dimensions>, (ie Y, X for 2D)
# <input image spatial dimensions>, (ie Hi, Wi for 2D)
# <strides>, (ie Sy, Sx for 2D)
# <dilations>, (ie Dy, Dx for 2D)
# <left padding>, (ie LeftPy, LeftPx for 2D)
# <right padding>, (ie RightPy, RightPx for 2D)
################ op datatype layout verify init log time opType Ndims G N K C Y X Hi Wi Sy Sx Dy Dx LeftPy LeftPx RightPy RightPx
./bin/ckProfiler conv_tensor_rearrange 0 0 0 1 0 1 0 2 1 256 1 512 3 3 28 28 1 1 1 1 0 0 0 0
Note: Column to image kernel adds to the output memory, this will cause output buffer to be accumulated multiple times, causing verification failure. To work around it, do not use CK's own timer and do verification at the same time.
Profile Permute scale kernels
# arg1: tensor operation ( permute_scale : Permute Scale )
# arg2: data type (0: Input fp32, Output fp32
# 1: Input fp16, Output fp16
# arg4: verification (0: no, 1: yes)
# arg5: initialization (0: no init, 1: integer value, 2: decimal value)
# arg6: print tensor value (0: no; 1: yes)
# arg7: time kernel (0: no, 1: yes)
# from arg8: tensor lengths
# input strides
# output strides
################ op datatype verify init log time dim0 dim1 dim2 in_stride0 in_stride1 in_stride2 out_stride0 out_stride1 out_stride2
./bin/ckProfiler permute_scale 0 1 1 0 1 64 64 64 4096 64 1 1 64 4096
Convert MIOpen driver command to CKProfiler
python3 ../script/convert_miopen_driver_to_profiler.py
/opt/rocm/bin/MIOpenDriver conv -n 32 -c 64 -H 28 -W 28 -k 64 -y 3 -x 3
-p 1 -q 1 -u 2 -v 2 -l 1 -j 1 -m conv -g 32 -F 1 -t 1
Only convolution driver is supported.