iree
iree copied to clipboard
[Codegen] Add support for vectorizing tensor.pack ops with masking.
It introduces inferVectorSizesFromIR(Value val) method, so every op can infer the input vector size through the chain. This is important for fusion cases (e.g., generic + pack) because it assumed extract_slice op being a producer op. It improves the input vector size inference generally.
Abbreviated Benchmark Summary
@ commit 14047dde9b63497c3bb46bf599a7596f9945efbd (vs. base 573ff1ff02347266ed747dd316cefaeb4c710396)
Data-Tiling Comparison Table
Click to show
| Name | No-DT (baseline) | DT-Only | DT-UK |
|---|---|---|---|
| BertForMaskedLMTF(stablehlo) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_task(embedded_elf)[30-thread,full-inference,default-flags] with default @ c2-standard-60[cpu] | 216.140 (1.0X) | N/A | 106.034 (2.0X) |
| BertLargePTBatch1(linalg) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_task(embedded_elf)[30-thread,full-inference,default-flags] with default @ c2-standard-60[cpu] | 655.015 (1.0X) | N/A | 225.109 (2.9X) |
| BertLargeTF(stablehlo) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_task(embedded_elf)[30-thread,full-inference,default-flags] with default @ c2-standard-60[cpu] | 701.476 (1.0X) | N/A | 222.570 (3.2X) |
| DeepLabV3_fp32(tflite) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_task(embedded_elf)[1-thread,full-inference,default-flags] with default @ c2-standard-60[cpu] | 31.596 (1.0X) | N/A | 33.329 (0.9X) |
| DeepLabV3_fp32(tflite) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_task(embedded_elf)[8-thread,full-inference,default-flags] with default @ c2-standard-60[cpu] | 6.953 (1.0X) | N/A | 8.572 (0.8X) |
| EfficientNetV2STF(stablehlo) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_task(embedded_elf)[1-thread,full-inference,default-flags] with default @ c2-standard-60[cpu] | 272.646 (1.0X) | N/A | 227.781 (1.2X) |
| EfficientNetV2STF(stablehlo) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_task(embedded_elf)[15-thread,full-inference,default-flags] with default @ c2-standard-60[cpu] | 35.064 (1.0X) | N/A | 33.573 (1.0X) |
| EfficientNet_int8(tflite) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_task(embedded_elf)[1-thread,full-inference,default-flags] with default @ c2-standard-60[cpu] | 26.003 (1.0X) | N/A | 15.566 (1.7X) |
| EfficientNet_int8(tflite) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_task(embedded_elf)[8-thread,full-inference,default-flags] with default @ c2-standard-60[cpu] | 5.568 (1.0X) | N/A | 5.280 (1.1X) |
| Falcon7bGptqPT(linalg) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_task(embedded_elf)[30-thread,full-inference,default-flags] with default @ c2-standard-60[cpu] | 20400.041 (1.0X) | N/A | 4141.295 (4.9X) |
| Falcon7bInt4GptqPT(linalg) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_task(embedded_elf)[30-thread,full-inference,default-flags] with default @ c2-standard-60[cpu] | 20410.537 (1.0X) | N/A | 4053.921 (5.0X) |
| GPT2_117M_TF_1X1XI32(stablehlo) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_task(embedded_elf)[1-thread,full-inference,default-flags] with default @ c2-standard-60[cpu] | 69.791 (1.0X) | N/A | 39.266 (1.8X) |
| GPT2_117M_TF_1X1XI32(stablehlo) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_task(embedded_elf)[15-thread,full-inference,default-flags] with default @ c2-standard-60[cpu] | 9.109 (1.0X) | N/A | 8.545 (1.1X) |
| GPT2_117M_TF_1X4XI32(stablehlo) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_task(embedded_elf)[1-thread,full-inference,default-flags] with default @ c2-standard-60[cpu] | 88.519 (1.0X) | N/A | 41.098 (2.2X) |
| GPT2_117M_TF_1X4XI32(stablehlo) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_task(embedded_elf)[15-thread,full-inference,default-flags] with default @ c2-standard-60[cpu] | 10.725 (1.0X) | N/A | 8.020 (1.3X) |
| MiniLML12H384Uncased(stablehlo) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_task(embedded_elf)[1-thread,full-inference,default-flags] with default @ c2-standard-60[cpu] | 81.515 (1.0X) | N/A | 58.656 (1.4X) |
| MiniLML12H384Uncased(stablehlo) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_task(embedded_elf)[15-thread,full-inference,default-flags] with default @ c2-standard-60[cpu] | 12.266 (1.0X) | N/A | 12.898 (1.0X) |
| MobileBertSquad_fp16(tflite) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_task(embedded_elf)[1-thread,full-inference,default-flags] with default @ c2-standard-60[cpu] | 178.796 (1.0X) | N/A | 186.773 (1.0X) |
| MobileBertSquad_fp16(tflite) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_task(embedded_elf)[15-thread,full-inference,default-flags] with default @ c2-standard-60[cpu] | 33.197 (1.0X) | N/A | 58.774 (0.6X) |
| MobileBertSquad_fp32(tflite) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_task(embedded_elf)[1-thread,full-inference,default-flags] with default @ c2-standard-60[cpu] | 181.224 (1.0X) | N/A | 190.552 (1.0X) |
| MobileBertSquad_fp32(tflite) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_task(embedded_elf)[15-thread,full-inference,default-flags] with default @ c2-standard-60[cpu] | 33.541 (1.0X) | N/A | 59.552 (0.6X) |
| MobileBertSquad_int8(tflite) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_task(embedded_elf)[1-thread,full-inference,default-flags] with default @ c2-standard-60[cpu] | 481.539 (1.0X) | N/A | 215.223 (2.2X) |
| MobileBertSquad_int8(tflite) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_task(embedded_elf)[15-thread,full-inference,default-flags] with default @ c2-standard-60[cpu] | 60.670 (1.0X) | N/A | 65.658 (0.9X) |
| MobileNetV1_fp32(tflite) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_task(embedded_elf)[1-thread,full-inference,default-flags] with default @ c2-standard-60[cpu] | 27.486 (1.0X) | N/A | 19.845 (1.4X) |
| MobileNetV1_fp32(tflite) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_task(embedded_elf)[8-thread,full-inference,default-flags] with default @ c2-standard-60[cpu] | 5.482 (1.0X) | N/A | 5.048 (1.1X) |
| MobileNetV2_fp32(tflite) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_task(embedded_elf)[1-thread,full-inference,default-flags] with default @ c2-standard-60[cpu] | 11.036 (1.0X) | N/A | 12.695 (0.9X) |
| MobileNetV2_fp32(tflite) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_task(embedded_elf)[8-thread,full-inference,default-flags] with default @ c2-standard-60[cpu] | 3.661 (1.0X) | N/A | 4.932 (0.7X) |
| MobileNetV2_int8(tflite) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_task(embedded_elf)[1-thread,full-inference,default-flags] with default @ c2-standard-60[cpu] | 21.098 (1.0X) | N/A | 14.202 (1.5X) |
| MobileNetV2_int8(tflite) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_task(embedded_elf)[8-thread,full-inference,default-flags] with default @ c2-standard-60[cpu] | 5.702 (1.0X) | N/A | 5.578 (1.0X) |
| MobileNetV3Small_fp32(tflite) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_sync(embedded_elf)[full-inference,default-flags] with default @ c2-standard-60[cpu] | 2.852 (1.0X) | N/A | 3.217 (0.9X) |
| MobileNetV3Small_fp32(tflite) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_task(embedded_elf)[1-thread,full-inference,default-flags] with default @ c2-standard-60[cpu] | 2.836 (1.0X) | N/A | 3.274 (0.9X) |
| MobileSSD_fp32(tflite) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_task(embedded_elf)[1-thread,full-inference,default-flags] with default @ c2-standard-60[cpu] | 34.354 (1.0X) | N/A | 32.138 (1.1X) |
| MobileSSD_fp32(tflite) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_task(embedded_elf)[8-thread,full-inference,default-flags] with default @ c2-standard-60[cpu] | 8.311 (1.0X) | N/A | 9.468 (0.9X) |
| PersonDetect_int8(tflite) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_sync(embedded_elf)[full-inference,default-flags] with default @ c2-standard-60[cpu] | 0.734 (1.0X) | N/A | 0.614 (1.2X) |
| PersonDetect_int8(tflite) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_task(embedded_elf)[1-thread,full-inference,default-flags] with default @ c2-standard-60[cpu] | 0.799 (1.0X) | N/A | 0.674 (1.2X) |
| PoseNet_fp32(tflite) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_task(embedded_elf)[1-thread,full-inference,default-flags] with default @ c2-standard-60[cpu] | 18.038 (1.0X) | N/A | 20.885 (0.9X) |
| PoseNet_fp32(tflite) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_task(embedded_elf)[8-thread,full-inference,default-flags] with default @ c2-standard-60[cpu] | 4.161 (1.0X) | N/A | 5.290 (0.8X) |
| matmul_1x256x2048_i8_i4_i32_tile_config_default(linalg) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_sync(embedded_elf)[full-inference,default-flags] with default @ c2-standard-60[cpu] | 0.065 (1.0X) | N/A | 0.065 (1.0X) |
| matmul_1x256x2048_i8_i8_i32_tile_config_default(linalg) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_sync(embedded_elf)[full-inference,default-flags] with default @ c2-standard-60[cpu] | 0.042 (1.0X) | N/A | 0.021 (2.0X) |
| matmul_256x256x2048_i8_i4_i32_tile_config_default(linalg) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_sync(embedded_elf)[full-inference,default-flags] with default @ c2-standard-60[cpu] | 7.435 (1.0X) | N/A | 7.441 (1.0X) |
| matmul_256x256x2048_i8_i8_i32_tile_config_default(linalg) [x86_64-cascadelake-linux_gnu-llvm_cpu] local_sync(embedded_elf)[full-inference,default-flags] with default @ c2-standard-60[cpu] | 6.742 (1.0X) | N/A | 1.977 (3.4X) |
| DeepLabV3_fp32(tflite) [armv8.2-a-generic-linux_android29-llvm_cpu] local_sync(embedded_elf)[full-inference,default-flags] with default @ pixel-6-pro[big-cores] | 51.440 (1.0X) | N/A | 86.020 (0.6X) |
| DeepLabV3_fp32(tflite) [armv8.2-a-generic-linux_android29-llvm_cpu] local_task(embedded_elf)[1-thread,full-inference,system-scheduling] with default @ pixel-6-pro[big-cores] | 54.051 (1.0X) | N/A | 85.419 (0.6X) |
| DeepLabV3_fp32(tflite) [armv8.2-a-generic-linux_android29-llvm_cpu] local_task(embedded_elf)[2-thread,full-inference,system-scheduling] with default @ pixel-6-pro[big-cores] | 33.669 (1.0X) | N/A | 53.429 (0.6X) |
| GPT2_117M_TF_1X1XI32(stablehlo) [armv8.2-a-generic-linux_android29-llvm_cpu] local_sync(embedded_elf)[full-inference,default-flags] with default @ pixel-6-pro[big-cores] | 92.225 (1.0X) | N/A | 22.239 (4.1X) |
| GPT2_117M_TF_1X1XI32(stablehlo) [armv8.2-a-generic-linux_android29-llvm_cpu] local_task(embedded_elf)[1-thread,full-inference,system-scheduling] with default @ pixel-6-pro[big-cores] | 92.689 (1.0X) | N/A | 22.517 (4.1X) |
| GPT2_117M_TF_1X1XI32(stablehlo) [armv8.2-a-generic-linux_android29-llvm_cpu] local_task(embedded_elf)[2-thread,full-inference,system-scheduling] with default @ pixel-6-pro[big-cores] | 53.116 (1.0X) | N/A | 22.035 (2.4X) |
| GPT2_117M_TF_1X4XI32(stablehlo) [armv8.2-a-generic-linux_android29-llvm_cpu] local_sync(embedded_elf)[full-inference,default-flags] with default @ pixel-6-pro[big-cores] | 139.417 (1.0X) | N/A | 28.329 (4.9X) |
| GPT2_117M_TF_1X4XI32(stablehlo) [armv8.2-a-generic-linux_android29-llvm_cpu] local_task(embedded_elf)[1-thread,full-inference,system-scheduling] with default @ pixel-6-pro[big-cores] | 133.055 (1.0X) | N/A | 30.905 (4.3X) |
| GPT2_117M_TF_1X4XI32(stablehlo) [armv8.2-a-generic-linux_android29-llvm_cpu] local_task(embedded_elf)[2-thread,full-inference,system-scheduling] with default @ pixel-6-pro[big-cores] | 74.650 (1.0X) | N/A | 26.637 (2.8X) |
| MobileBertSquad_fp32(tflite) [armv8.2-a-generic-linux_android29-llvm_cpu] local_sync(embedded_elf)[full-inference,default-flags] with default @ pixel-6-pro[big-cores] | 758.776 (1.0X) | N/A | 452.874 (1.7X) |
| MobileBertSquad_fp32(tflite) [armv8.2-a-generic-linux_android29-llvm_cpu] local_task(embedded_elf)[1-thread,full-inference,system-scheduling] with default @ pixel-6-pro[big-cores] | 756.923 (1.0X) | N/A | 405.381 (1.9X) |
| MobileBertSquad_fp32(tflite) [armv8.2-a-generic-linux_android29-llvm_cpu] local_task(embedded_elf)[2-thread,full-inference,system-scheduling] with default @ pixel-6-pro[big-cores] | 446.125 (1.0X) | N/A | 246.400 (1.8X) |
| MobileBertSquad_int8(tflite) [armv8.2-a-generic-linux_android29-llvm_cpu] local_sync(embedded_elf)[full-inference,default-flags] with default @ pixel-6-pro[big-cores] | 1098.169 (1.0X) | N/A | 257.919 (4.3X) |
| MobileBertSquad_int8(tflite) [armv8.2-a-generic-linux_android29-llvm_cpu] local_task(embedded_elf)[1-thread,full-inference,system-scheduling] with default @ pixel-6-pro[big-cores] | 1102.532 (1.0X) | N/A | 260.858 (4.2X) |
| MobileBertSquad_int8(tflite) [armv8.2-a-generic-linux_android29-llvm_cpu] local_task(embedded_elf)[2-thread,full-inference,system-scheduling] with default @ pixel-6-pro[big-cores] | 598.704 (1.0X) | N/A | 153.164 (3.9X) |
| Vit_int8(tflite) [armv8.2-a-generic-linux_android29-llvm_cpu] local_sync(embedded_elf)[full-inference,default-flags] with default @ pixel-6-pro[big-cores] | 2177.568 (1.0X) | N/A | 309.368 (7.0X) |
| Vit_int8(tflite) [armv8.2-a-generic-linux_android29-llvm_cpu] local_task(embedded_elf)[1-thread,full-inference,system-scheduling] with default @ pixel-6-pro[big-cores] | 2174.275 (1.0X) | N/A | 313.893 (6.9X) |
| Vit_int8(tflite) [armv8.2-a-generic-linux_android29-llvm_cpu] local_task(embedded_elf)[2-thread,full-inference,system-scheduling] with default @ pixel-6-pro[big-cores] | 1227.048 (1.0X) | N/A | 189.735 (6.5X) |
| matmul_1x256x2048_i8_i4_i32_tile_config_default(linalg) [armv8.2-a-generic-linux_android29-llvm_cpu] local_sync(embedded_elf)[full-inference,default-flags] with default @ pixel-6-pro[big-cores] | 0.081 (1.0X) | N/A | 0.016 (5.0X) |
| matmul_1x256x2048_i8_i8_i32_tile_config_default(linalg) [armv8.2-a-generic-linux_android29-llvm_cpu] local_sync(embedded_elf)[full-inference,default-flags] with default @ pixel-6-pro[big-cores] | 0.073 (1.0X) | N/A | 0.017 (4.3X) |
| matmul_256x256x2048_i8_i4_i32_tile_config_default(linalg) [armv8.2-a-generic-linux_android29-llvm_cpu] local_sync(embedded_elf)[full-inference,default-flags] with default @ pixel-6-pro[big-cores] | 12.014 (1.0X) | N/A | 1.429 (8.4X) |
| matmul_256x256x2048_i8_i8_i32_tile_config_default(linalg) [armv8.2-a-generic-linux_android29-llvm_cpu] local_sync(embedded_elf)[full-inference,default-flags] with default @ pixel-6-pro[big-cores] | 16.546 (1.0X) | N/A | 1.245 (13.3X) |
Regressed Latencies 🚩
| Benchmark Name | Average Latency (ms) | Median Latency (ms) | Latency Standard Deviation (ms) |
|---|---|---|---|
| MobileBertSquad\_fp32(tflite) [armv8.2-a-generic-linux\_android29-llvm\_cpu][default-flags,dt-uk] local\_sync(embedded\_elf)[full-inference,default-flags] with default @ pixel-6-pro[big-cores] | 452.874 (vs. 374.014, 21.08%↑) | 451.174 | 4.913 |
| DeepLabV3\_fp32(tflite) [armv8.2-a-generic-linux\_android29-llvm\_cpu][default-flags,dt-uk] local\_task(embedded\_elf)[2-thread,full-inference,system-scheduling] with default @ pixel-6-pro[big-cores] | 53.429 (vs. 49.226, 8.54%↑) | 53.979 | 2.584 |
| MobileBertSquad\_fp32(tflite) [armv8.2-a-generic-linux\_android29-llvm\_cpu][experimental-flags,no-dt] local\_sync(embedded\_elf)[full-inference,default-flags] with default @ pixel-6-pro[big-cores] | 758.776 (vs. 714.709, 6.17%↑) | 764.421 | 15.542 |
[Top 3 out of 5 results showed]
Improved Latencies 🎉
| Benchmark Name | Average Latency (ms) | Median Latency (ms) | Latency Standard Deviation (ms) |
|---|---|---|---|
| MobileBertSquad\_int8(tflite) [armv8.2-a-generic-linux\_android29-llvm\_cpu][default-flags,dt-uk] local\_task(embedded\_elf)[2-thread,full-inference,system-scheduling] with default @ pixel-6-pro[big-cores] | 153.164 (vs. 162.861, 5.95%↓) | 154.220 | 4.399 |
No improved or regressed compilation metrics 🏖️
For more information:
This is an NFC for e2e tests because all the pack ops are decomposed before vectorization.