Old run: January 28, 2026 (Top of Main scaled GEMM before swizzle)
New run: February 19, 2026 (Top of Main scaled GEMM after swizzle)
Hardware: MI355, profiled with rocprofv3
Muzammiluddin-Syed-ECE
/ MXFP4_GEMM_Perf_Comparison_after_XOR.md
Last active
February 19, 2026 21:15
2026-02-19-SCALED_GEMM
Muzammiluddin-Syed-ECE
/ reproducer.ll
Created
February 5, 2026 21:48
Reproducing the multi mem operand failure
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| ; Reproducer for MemSDNode::getMemOperand() assertion (multi-MMO). See AMDGPU-multi-MMO-fix.md. | |
| ; To trigger: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx950 -O3 repro-amdgcn-multi-mmo-e2e-matmul.ll -o - | |
| ; | |
| ; To reproduce the .optimized.ll from the .linked.ll, run: | |
| ; opt -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx950 --passes='verify,memprof-remove-attributes,annotation2metadata,forceattrs,inferattrs,coro-early,function<eager-inv>(ee-instrument<>,lower-expect,simplifycfg<bonus-inst-threshold=1;no-forward-switch-cond;no-switch-range-to-icmp;no-switch-to-arithmetic;no-switch-to-lookup;keep-loops;no-hoist-common-insts;no-hoist-loads-stores-with-cond-faulting;no-sink-common-insts;speculate-blocks;simplify-cond-branch;no-speculate-unpredictables>,sroa<modify-cfg>,early-cse<>),openmp-opt,amdgpu-printf-runtime-binding,ipsccp,called-value-propagation,globalopt,function<eager-inv>(mem2reg,instcombine<max-iterations=1;no-verify-fixpoint>,amdgpu-usenative,amdgpu-simplifylib,amdgpu-uniform-intrinsic-combine,simplifycfg<bonus-inst |
Muzammiluddin-Syed-ECE
/ ALL_DATA_FOR_FEATURE.csv
Last active
January 29, 2026 21:29
Units for SQ_LDS_BANK_CONFLICT is cycles lost for bank conflicts
We can make this file beautiful and searchable if this error is corrected: It looks like row 9 should actually have 11 columns, instead of 3 in line 8.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| shape,name,throughput (TFLOPS),stddev,AI,IREE-AI,created_date,benchmark,instrument,version,SQ_LDS_BANK_CONFLICT | |
| 500_512_256_16,scaled_matmul,15.134461058830322,0,201.76345992808237,331.6867762571868,2026-01-28,mxfp4 gemm,rocprofv3,,0.0 | |
| 512_512_256_16,scaled_matmul,22.748767457627117,0,202.2716049382716,334.3673469387755,2026-01-28,mxfp4 gemm,rocprofv3,,0.0 | |
| 1000_512_256_16,scaled_matmul,30.624299065420562,0,213.18898662364023,400.1172218423366,2026-01-28,mxfp4 gemm,rocprofv3,,0.0 | |
| 1024_512_256_16,scaled_matmul,45.497534915254235,0,213.47231270358307,402.06134969325154,2026-01-28,mxfp4 gemm,rocprofv3,,0.0 | |
| 8100_512_256_16,scaled_matmul,227.3289866709491,0,224.32378524751437,488.4482035200848,2026-01-28,mxfp4 gemm,rocprofv3,,0.0 | |
| 8192_512_256_16,scaled_matmul,305.0402909090909,0,224.3423192126658,488.618825722274,2026-01-28,mxfp4 gemm,rocprofv3,,0.0 | |
| 16300_512_256_16,scaled_matmul,434.24260162601627,0,225.1570269673266,496.20994503881474,2026-01-28,mxfp4 gemm,rocprofv3,,0.0 | |
| 16384_512_256_16,scaled_matmul,502.8941275 |
Muzammiluddin-Syed-ECE
/ SwizzleImprovementOverMain.csv
Last active
January 29, 2026 21:09
Perf gains from swizzling
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| shape | Feature_Throughput (TFLOps) | Top_of_main_Throughput(TFLOps) | improvement_ratio | improvement_percent | |
|---|---|---|---|---|---|
| 1000_1024_8192_512 | 96.044057 | 86.408117 | 1.111517 | 11.151661 | |
| 1000_16384_26624_1664 | 1409.736175 | 1366.244197 | 1.031833 | 3.183324 | |
| 1000_16384_8192_512 | 1377.701308 | 1247.127461 | 1.104700 | 10.469968 | |
| 1000_512_256_16 | 30.624299 | 28.066809 | 1.091121 | 9.112150 | |
| 1000_53248_8192_512 | 1043.042983 | 984.096396 | 1.059899 | 5.989920 | |
| 1024_1024_8192_512 | 167.149430 | 160.287635 | 1.042809 | 4.280926 | |
| 1024_16384_26624_1664 | 2414.689992 | 2327.495265 | 1.037463 | 3.746290 | |
| 1024_16384_8192_512 | 2381.916318 | 2284.148917 | 1.042803 | 4.280255 | |
| 1024_512_256_16 | 45.497535 | 41.943040 | 1.084746 | 8.474576 |
Muzammiluddin-Syed-ECE
/ Issues.md
Last active
October 24, 2025 23:24
Installing windows difficulties
Muzammiluddin-Syed-ECE
/ 8192_8192_256_32_sample_scaled_MFMA.mlir
Created
October 14, 2025 16:28
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| //<dim> 8192 </dim> | |
| //<dim> 8192 </dim> | |
| //<dim> 256 </dim> | |
| //<dim> 32 </dim> | |
| //<input>tensor<8192x256x32xi8></input> | |
| //<input>tensor<8192x256xi8></input> | |
| //<input>tensor<8192x256x32xi8></input> | |
| //<input>tensor<8192x256xi8></input> | |
| !lhs = f4E2M1FN |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| /home/kdrewnia/fp4-benchmark/output/1024_1024_32_32_sample_scaled_MFMA.mlir:43:8: error: failed to legalize operation 'linalg.generic' that was explicitly marked illegal | |
| %D = linalg.generic { | |
| ^ | |
| /home/kdrewnia/fp4-benchmark/output/1024_1024_32_32_sample_scaled_MFMA.mlir:31:1: note: called from | |
| func.func @scaled_matmul(%lhs : !A, %lhs_scales : !A_s, %rhs : !B, %rhs_scales : !B_s) -> !C_size { | |
| ^ | |
| /home/kdrewnia/fp4-benchmark/output/1024_1024_32_32_sample_scaled_MFMA.mlir:43:8: note: see current operation: | |
| %19 = "linalg.generic"(%11, %12, %13, %14, %18) <{indexing_maps = [affine_map<(d0, d1, d2, d3) -> (d0, d2, d3)>, affine_map<(d0, d1, d2, d3) -> (d1, d2, d3)>, affine_map<(d0, d1, d2, d3) -> (d0, d2)>, affine_map<(d0, d1, d2, d3) -> (d1, d2)>, affine_map<(d0, d1, d2, d3) -> (d0, d1)>], iterator_types = [#linalg.iterator_type<parallel>, #linalg.iterator_type<parallel>, #linalg.iterator_type<reduction>, #linalg.iterator_type<reduction>], operandSegmentSizes = array<i32: 4, 1>}> ({ | |
| ^bb0(%arg4: f4E2M1FN, % |
Muzammiluddin-Syed-ECE
/ mxfp_gemm_asm
Created
September 12, 2025 15:48
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| // <input> tensor<1024x1024xi8> </input> | |
| // <input> tensor<1024x1024xi8> </input> | |
| // <input> tensor<1024x32xi8> </input> | |
| // <input> tensor<1024x32xi8> </input> | |
| #executable_target_rocm_hsaco_fb = #hal.executable.target<"rocm", "rocm-hsaco-fb", {target_arch = "gfx950", ukernels = "none"}> | |
| #map = affine_map<(d0) -> (d0)> | |
| #map1 = affine_map<(d0, d1) -> (d0, d1)> | |
| #map2 = affine_map<(d0, d1) -> (d0)> | |
| #map3 = affine_map<()[s0] -> ((s0 ceildiv 256) * 256)> | |
| #map4 = affine_map<()[s0] -> ((s0 ceildiv 32) * 32)> |
Muzammiluddin-Syed-ECE
/ ScaledGemm.mlir
Created
August 25, 2025 20:08
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| func.func @scaled_matmul_1x1() -> tensor<1x1xf32> { | |
| %lhs = util.unfoldable_constant dense<1.0> : tensor<1x1x32xf32> | |
| %rhs = util.unfoldable_constant dense<1.0> : tensor<1x1x32xf32> | |
| %lhs_scales = util.unfoldable_constant dense<126> : tensor<1x1xi8> | |
| %rhs_scales = util.unfoldable_constant dense<126> : tensor<1x1xi8> | |
| %A_scales = arith.bitcast %lhs_scales : tensor<1x1xi8> to tensor<1x1xf8E8M0FNU> | |
| %B_scales = arith.bitcast %rhs_scales : tensor<1x1xi8> to tensor<1x1xf8E8M0FNU> | |
| %lhs_splat = util.unfoldable_constant dense<1.0> : tensor<1x1x32xf8E8M0FNU> | |
| %rhs_splat = util.unfoldable_constant dense<1.0> : tensor<1x1x32xf8E8M0FNU> | |
| %A = arith.scaling_truncf %lhs, %lhs_splat : tensor<1x1x32xf32>, tensor<1x1x32xf8E8M0FNU> to tensor<1x1x32xf4E2M1FN> |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| func.func @scaled_matmul_1x1() { | |
| %lhs = util.unfoldable_constant dense<1.0> : tensor<1x1x32xf32> | |
| %rhs = util.unfoldable_constant dense<1.0> : tensor<1x1x32xf32> | |
| %lhs_scales = util.unfoldable_constant dense<126> : tensor<1x1xi8> | |
| %rhs_scales = util.unfoldable_constant dense<126> : tensor<1x1xi8> | |
| %A_scales = arith.bitcast %lhs_scales : tensor<1x1xi8> to tensor<1x1xf8E8M0FNU> | |
| %B_scales = arith.bitcast %rhs_scales : tensor<1x1xi8> to tensor<1x1xf8E8M0FNU> | |
| %lhs_splat = util.unfoldable_constant dense<1.0> : tensor<1x1x32xf8E8M0FNU> | |
| %rhs_splat = util.unfoldable_constant dense<1.0> : tensor<1x1x32xf8E8M0FNU> | |
| %A = arith.scaling_truncf %lhs, %lhs_splat : tensor<1x1x32xf32>, tensor<1x1x32xf8E8M0FNU> to tensor<1x1x32xf4E2M1FN> |
NewerOlder