ROCm/composable_kernel

mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-14 10:09:41 +00:00

Files

History

juuso-oskari f5beedb2e9 Add CK-UA decode_d128_mha_m32 / _m16 small-Q tiers

For pure-decode workloads (sq=1) at d=128 the m128 tile wastes most of
its 128 query rows, capping CK below Triton on every batch size in our
sweep (4..256). Add two small-Q tiers that mirror the d=64 GQA-8 ladder:

  * decode_d128_mha_m16 : kBlockM=16, 1 warp, 16x16 MFMA  (tiny-decode)
  * decode_d128_mha_m32 : kBlockM=32, 1 warp, 32x32 MFMA  (tiny-decode)

select_config now ladders by (avg_q, max_q): m16 -> m32 -> m128 -> prefill.

d=128 MHA, hq=16/hk=16, sq=1, sk=120k, num_blocks=60k:
  batch  before    after    CK BW
      4  ~0.95x   0.98x   4.76 TB/s
      8  ~0.85x   1.29x   5.00 TB/s
     32  ~0.85x   1.14x   5.29 TB/s
     64  ~0.75x   0.93x   5.35 TB/s
    128  ~1.00x   1.09x   5.39 TB/s
    256  ~1.03x   1.02x   5.41 TB/s

Correctness suite stays at 241/245 (same 4 known int32-overflow
failures in the prefill path).

Co-authored-by: Cursor <cursoragent@cursor.com>

2026-05-12 11:48:19 +00:00

..

Padding support for wave transfer (#3537 )

2026-01-26 12:57:09 -08:00

02_gemm_bilinear

[CI, CK examples] Disable time_kernel for CI tests and examples (#3464 )

2026-01-07 16:30:57 +01:00

03_gemm_bias_relu

chore(copyright) update library wide CMakeLists.txt copyright header template (#3313 )

2025-11-28 13:49:54 -08:00

04_gemm_add_add_fastgelu

chore(copyright) update library wide CMakeLists.txt copyright header template (#3313 )

2025-11-28 13:49:54 -08:00

[CK] Integrate GPU reference into ckProfiler for convolutions (#3379 )

2025-12-18 07:59:45 +01:00

10_convnd_fwd_multiple_d_multiple_reduce

chore(copyright) update library wide CMakeLists.txt copyright header template (#3313 )

2025-11-28 13:49:54 -08:00

11_convnd_fwd_bias

[DOCS] Documentation Addition (Readme updates) (#2495 )

2025-10-16 03:10:57 -07:00

[rocm-libraries] ROCm/rocm-libraries#5030 (commit 8e02a26)

2026-03-06 16:28:22 +00:00

[CI, CK examples] Disable time_kernel for CI tests and examples (#3464 )

2026-01-07 16:30:57 +01:00

14_gemm_quantization

[rocm-libraries] ROCm/rocm-libraries#5348 (commit 7b18234)

2026-03-12 08:48:36 +00:00

15_grouped_gemm

[rocm-libraries] ROCm/rocm-libraries#4340 (commit 70a312f)

2026-02-26 00:28:58 +00:00

16_gemm_multi_d_multi_reduces

[rocm-libraries] ROCm/rocm-libraries#5348 (commit 7b18234)

2026-03-12 08:48:36 +00:00

17_convnd_bwd_data

[CK] Integrate GPU reference into ckProfiler for convolutions (#3379 )

2025-12-18 07:59:45 +01:00

18_batched_gemm_reduce

chore(copyright) update library wide CMakeLists.txt copyright header template (#3313 )

2025-11-28 13:49:54 -08:00

19_binary_elementwise

chore(copyright) update library wide CMakeLists.txt copyright header template (#3313 )

2025-11-28 13:49:54 -08:00

20_grouped_conv_bwd_weight

[rocm-libraries] ROCm/rocm-libraries#4872 (commit ca623f7)

2026-02-25 20:11:01 +00:00

21_gemm_layernorm

chore(copyright) update library wide CMakeLists.txt copyright header template (#3313 )

2025-11-28 13:49:54 -08:00

[CI, CK examples] Disable time_kernel for CI tests and examples (#3464 )

2026-01-07 16:30:57 +01:00

[CI, CK examples] Disable time_kernel for CI tests and examples (#3464 )

2026-01-07 16:30:57 +01:00

24_batched_gemm

[CI, CK examples] Disable time_kernel for CI tests and examples (#3464 )

2026-01-07 16:30:57 +01:00

25_gemm_bias_e_permute

Implement batched gemm bias permute for RDNA4 (#3534 )

2026-01-17 08:30:27 +01:00

Add support to fp16 + compute fp16 and bf16 + compute bf16 contractions (#3598 )

2026-01-20 09:39:57 -08:00

27_layernorm2d_fwd

chore(copyright) update library wide CMakeLists.txt copyright header template (#3313 )

2025-11-28 13:49:54 -08:00

28_grouped_gemm_bias_e_permute

chore(copyright) update library wide CMakeLists.txt copyright header template (#3313 )

2025-11-28 13:49:54 -08:00

29_batched_gemm_bias_e_permute

Implement batched gemm bias permute for RDNA4 (#3534 )

2026-01-17 08:30:27 +01:00

30_grouped_conv_fwd_multiple_d

[CI, CK examples] Disable time_kernel for CI tests and examples (#3464 )

2026-01-07 16:30:57 +01:00

31_batched_gemm_gemm

chore(copyright) update library wide CMakeLists.txt copyright header template (#3313 )

2025-11-28 13:49:54 -08:00

32_batched_gemm_scale_softmax_gemm

chore(copyright) update library wide CMakeLists.txt copyright header template (#3313 )

2025-11-28 13:49:54 -08:00

33_multiple_reduce

[CI, CK examples] Disable time_kernel for CI tests and examples (#3464 )

2026-01-07 16:30:57 +01:00

chore(copyright) update library wide CMakeLists.txt copyright header template (#3313 )

2025-11-28 13:49:54 -08:00

[rocm-libraries] ROCm/rocm-libraries#4762 (commit 5598eb5)

2026-02-20 22:41:34 +00:00

36_sparse_embedding

[CI, CK examples] Disable time_kernel for CI tests and examples (#3464 )

2026-01-07 16:30:57 +01:00

37_batched_gemm_add_add_relu_gemm_add

Implement batched gemm add relu gemm add for rdna4 (#3391 )

2026-01-20 13:06:59 -08:00

38_grouped_conv_bwd_data_multiple_d

Grouped convolution backward data WMMA v3 implementation (#3460 )

2025-12-30 16:25:08 +01:00

chore(copyright) update library wide CMakeLists.txt copyright header template (#3313 )

2025-11-28 13:49:54 -08:00

40_conv2d_fwd_quantization

[CI, CK examples] Disable time_kernel for CI tests and examples (#3464 )

2026-01-07 16:30:57 +01:00

41_grouped_conv_conv_fwd

chore(copyright) update library wide CMakeLists.txt copyright header template (#3313 )

2025-11-28 13:49:54 -08:00

42_groupnorm_fwd

[CI, CK examples] Disable time_kernel for CI tests and examples (#3464 )

2026-01-07 16:30:57 +01:00

43_splitk_gemm_bias_e_permute

chore(copyright) update library wide CMakeLists.txt copyright header template (#3313 )

2025-11-28 13:49:54 -08:00

44_elementwise_permute

[CI, CK examples] Disable time_kernel for CI tests and examples (#3464 )

2026-01-07 16:30:57 +01:00

45_elementwise_normalization

[CI, CK examples] Disable time_kernel for CI tests and examples (#3464 )

2026-01-07 16:30:57 +01:00

46_gemm_add_multiply

chore(copyright) update library wide CMakeLists.txt copyright header template (#3313 )

2025-11-28 13:49:54 -08:00

47_gemm_bias_softmax_gemm_permute

chore(copyright) update library wide CMakeLists.txt copyright header template (#3313 )

2025-11-28 13:49:54 -08:00

chore(copyright) update library wide CMakeLists.txt copyright header template (#3313 )

2025-11-28 13:49:54 -08:00

49_maxpool2d_bwd

chore(copyright) update library wide CMakeLists.txt copyright header template (#3313 )

2025-11-28 13:49:54 -08:00

chore(copyright) update library wide CMakeLists.txt copyright header template (#3313 )

2025-11-28 13:49:54 -08:00

51_avgpool3d_bwd

chore(copyright) update library wide CMakeLists.txt copyright header template (#3313 )

2025-11-28 13:49:54 -08:00

52_im2col_col2im

chore(copyright) update library wide CMakeLists.txt copyright header template (#3313 )

2025-11-28 13:49:54 -08:00

53_layernorm2d_bwd

chore(copyright) update library wide CMakeLists.txt copyright header template (#3313 )

2025-11-28 13:49:54 -08:00

54_groupnorm_bwd

chore(copyright) update library wide CMakeLists.txt copyright header template (#3313 )

2025-11-28 13:49:54 -08:00

59_grouped_gemm_multi_ABD

[rocm-libraries] ROCm/rocm-libraries#4425 (commit 513cf9f)

2026-02-25 05:17:08 +00:00

60_gemm_multi_ABD

Multi AB support for wave transfer (#3578 )

2026-01-29 10:29:40 -08:00

61_contraction_multi_ABD

chore(copyright) update library wide CMakeLists.txt copyright header template (#3313 )

2025-11-28 13:49:54 -08:00

62_convnd_activ

Adding remaining conv, dynamic_op, and scaleadd_scaleadd_relu flavors for grouped conv fwd (#3529 )

2026-01-30 17:02:14 +01:00

63_layernorm4d_fwd

chore(copyright) update library wide CMakeLists.txt copyright header template (#3313 )

2025-11-28 13:49:54 -08:00

64_fpAintB_gemm

chore(copyright) update library wide CMakeLists.txt copyright header template (#3313 )

2025-11-28 13:49:54 -08:00

65_gemm_multiply_multiply

[rocm-libraries] ROCm/rocm-libraries#4762 (commit 5598eb5)

2026-02-20 22:41:34 +00:00

66_complex_contraction_bilinear

chore(copyright) update library wide CMakeLists.txt copyright header template (#3313 )

2025-11-28 13:49:54 -08:00

67_gemm_microscaling

[CI, CK examples] Disable time_kernel for CI tests and examples (#3464 )

2026-01-07 16:30:57 +01:00

[CK][Examples] Fixing stride issues in ck examples 14/65/68/69 by workaround - Bypassing hostTensor validation

2026-01-15 16:43:02 +01:00

69_gemm_add_relu

[CK][Examples] Fixing stride issues in ck examples 14/65/68/69 by workaround - Bypassing hostTensor validation

2026-01-15 16:43:02 +01:00

Add CK-UA decode_d128_mha_m32 / _m16 small-Q tiers

2026-05-12 11:48:19 +00:00

CMakeLists.txt

Build CK on Windows (#3458 )

2026-01-14 07:31:45 -08:00

README.md

Add basic documentation structure (#1715 )

2024-12-04 00:46:47 +01:00

README.md

Back to the main page

Composable Kernel examples