Resolve "Optimize SSSE3, AVX2, AVX512 Implementations of GF(2^8) SIMD Arithmetic"

Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "rlnc"
-version = "0.8.6"
+version = "0.8.7"
 edition = "2024"
 resolver = "3"
 rust-version = "1.89.0"

README.md

@@ -157,27 +157,17 @@ For visualizing benchmark results, run following command, which will produce PNG
 make bench_then_plot # Only runs with `default` features enabled
 ```
 
-### On 12th Gen Intel(R) Core(TM) i7-1260P
-
-Running benchmarks on `Linux 6.14.0-27-generic x86_64`, compiled with `rustc 1.88.0 (6b00bc388 2025-06-23)`.
-
-Component | Peak Median Throughput (`default` feature) | Peak Median Throughput (`parallel` feature) | Impact of number of pieces on performance
---- | --- | --- | ---
-Full RLNC Encoder | **30.14 GiB/s** | **23.39 GiB/s** | The number of pieces original data got split into has a **minimal** impact on the encoding speed.
-Full RLNC Recoder | **27.26 GiB/s** | **12.63 GiB/s** | Similar to the encoder, the recoder's performance remains largely consistent regardless of how many pieces the original data is split into.
-Full RLNC Decoder | **1.59 GiB/s** | **Doesn't yet implement a parallel decoding mode** | As the number of pieces increases, the decoding time increases substantially, leading to a considerable drop in throughput. This indicates that decoding is the most computationally intensive part of the full RLNC scheme, and its performance is inversely proportional to the number of pieces.
-
-In summary, the full RLNC implementation demonstrates excellent encoding and recoding speeds, consistently achieving GiB/s throughputs with minimal sensitivity to the number of data pieces. The `parallel` feature, leveraging Rust `rayon` data-parallelism framework, also provides good performance for both encoding and recoding. Whether you want to use that feature, completely depends on your usecase. However, decoding remains a much slower operation, with its performance significantly diminishing as the data is split into a greater number of pieces, and currently does **not** implement a parallel decoding algorithm.
+More performance benchmarking results are displayed on README inside [./plots](./plots) directory.
 
 ## Usage
 
 To use `rlnc` library crate in your Rust project, add it as a dependency in your `Cargo.toml`:
 
 ```toml
 [dependencies]
-rlnc = "=0.8.6"                                      # On x86_64 and aarch64 targets, it offers fast encoding, recoding and decoding, using SIMD intrinsics.
+rlnc = "=0.8.7"                                      # On x86_64 and aarch64 targets, it offers fast encoding, recoding and decoding, using SIMD intrinsics.
 # or
-rlnc = { version = "=0.8.6", features = "parallel" } # Uses `rayon`-based data-parallelism for fast encoding and recoding. Note, this feature, doesn't yet parallelize RLNC decoding.
+rlnc = { version = "=0.8.7", features = "parallel" } # Uses `rayon`-based data-parallelism for fast encoding and recoding. Note, this feature, doesn't yet parallelize RLNC decoding.
 
 rand = { version = "=0.9.2" } # Required for random number generation
 ```

plots/README.md

@@ -0,0 +1,98 @@
+# Plotted Performance Benchmark Results
+
+Following plots are generated by running this make recipe, from root of the repository, on machines with specified configuration.
+
+> [!NOTE]
+> These benchmark results don't capture the performance of running RLNC encoder, recoder and decoder with `parallel` feature.
+
+```bash
+make bench_then_plot
+```
+
+## Performance Characteristics
+
+Algorithm | Characteristics
+--- | ---
+Encoding | The number of pieces original data got split into has a **minimal** impact on the encoding speed.
+Recoding | Recoding is a wrapper over encoding, with an additional matrix-vector multiplication. If number of pieces increases, the dimension of matrix-vector multiplication also increases, resulting in higher computational complexity, during Recoding.
+Decoding | As the number of pieces increases, the decoding time increases substantially, leading to a considerable drop in throughput. This indicates that decoding is the most computationally intensive part of the full RLNC scheme, and its performance is inversely proportional to the number of pieces.
+
+In summary, this RLNC implementation demonstrates excellent encoding and recoding speeds, consistently achieving GiB/s throughputs with relatively minimal sensitivity to the number of data pieces. The `parallel` feature, leveraging Rust `rayon` data-parallelism framework, also provides good performance for both encoding and recoding. Whether you want to use that feature, completely depends on your use case. However, decoding remains a much slower operation, with its performance significantly diminishing as the data is split into a greater number of pieces, and currently does **not** implement a parallel decoding algorithm.
+
+## On 12th Gen Intel(R) Core(TM) i7-1260P
+
+Running Linux kernel
+
+```bash
+$ uname -srm
+Linux 6.17.0-5-generic x86_64
+```
+
+with Rust compiler
+
+```bash
+$ rustc --version
+rustc 1.90.0 (1159e78c4 2025-09-14)
+```
+
+and following CPU feature flags
+
+```bash
+$ lscpu | awk -F': *' '/Flags/{gsub(" ", ", ", $2);print $2}'
+fpu, vme, de, pse, tsc, msr, pae, mce, cx8, apic, sep, mtrr, pge, mca, cmov, pat, pse36, clflush, dts, acpi, mmx, fxsr, sse, sse2, ss, ht, tm, pbe, syscall, nx, pdpe1gb, rdtscp, lm, constant_tsc, art, arch_perfmon, pebs, bts, rep_good, nopl, xtopology, nonstop_tsc, cpuid, aperfmperf, tsc_known_freq, pni, pclmulqdq, dtes64, monitor, ds_cpl, vmx, smx, est, tm2, ssse3, sdbg, fma, cx16, xtpr, pdcm, pcid, sse4_1, sse4_2, x2apic, movbe, popcnt, tsc_deadline_timer, aes, xsave, avx, f16c, rdrand, lahf_lm, abm, 3dnowprefetch, cpuid_fault, epb, ssbd, ibrs, ibpb, stibp, ibrs_enhanced, tpr_shadow, flexpriority, ept, vpid, ept_ad, fsgsbase, tsc_adjust, bmi1, avx2, smep, bmi2, erms, invpcid, rdseed, adx, smap, clflushopt, clwb, intel_pt, sha_ni, xsaveopt, xsavec, xgetbv1, xsaves, split_lock_detect, user_shstk, avx_vnni, dtherm, ida, arat, pln, pts, hwp, hwp_notify, hwp_act_window, hwp_epp, hwp_pkg_req, hfi, vnmi, umip, pku, ospke, waitpkg, gfni, vaes, vpclmulqdq, rdpid, movdiri, movdir64b, fsrm, md_clear, serialize, arch_lbr, ibt, flush_l1d, arch_capabilities
+```
+
+### Encoder
+
+![benchmark_encode_on_12th_Gen_IntelR_CoreTM_i7-1260P_with_rustc_1.90.0_1159e78c4_2025-09-14](./benchmark_encode_on_12th_Gen_IntelR_CoreTM_i7-1260P_with_rustc_1.90.0_1159e78c4_2025-09-14.png)
+
+![benchmark_encode_zero_alloc_on_12th_Gen_IntelR_CoreTM_i7-1260P_with_rustc_1.90.0_1159e78c4_2025-09-14](./benchmark_encode_zero_alloc_on_12th_Gen_IntelR_CoreTM_i7-1260P_with_rustc_1.90.0_1159e78c4_2025-09-14.png)
+
+### Recoder
+
+![benchmark_recode_on_12th_Gen_IntelR_CoreTM_i7-1260P_with_rustc_1.90.0_1159e78c4_2025-09-14](./benchmark_recode_on_12th_Gen_IntelR_CoreTM_i7-1260P_with_rustc_1.90.0_1159e78c4_2025-09-14.png)
+
+![benchmark_recode_zero_alloc_on_12th_Gen_IntelR_CoreTM_i7-1260P_with_rustc_1.90.0_1159e78c4_2025-09-14](./benchmark_recode_zero_alloc_on_12th_Gen_IntelR_CoreTM_i7-1260P_with_rustc_1.90.0_1159e78c4_2025-09-14.png)
+
+### Decoder
+
+![benchmark_decode_on_12th_Gen_IntelR_CoreTM_i7-1260P_with_rustc_1.90.0_1159e78c4_2025-09-14](./benchmark_decode_on_12th_Gen_IntelR_CoreTM_i7-1260P_with_rustc_1.90.0_1159e78c4_2025-09-14.png)
+
+## On AMD EPYC 9R14 (AWS EC2 `m7a.large`)
+
+Running Linux kernel
+
+```bash
+$ uname -srm
+Linux 6.14.0-1011-aws x86_64
+```
+
+with Rust compiler
+
+```bash
+$ rustc --version
+rustc 1.90.0 (1159e78c4 2025-09-14)
+```
+
+and following CPU feature flags
+
+```bash
+$ lscpu | awk -F': *' '/Flags/{gsub(" ", ", ", $2);print $2}'
+fpu, vme, de, pse, tsc, msr, pae, mce, cx8, apic, sep, mtrr, pge, mca, cmov, pat, pse36, clflush, dts, acpi, mmx, fxsr, sse, sse2, ss, ht, tm, pbe, syscall, nx, pdpe1gb, rdtscp, lm, constant_tsc, art, arch_perfmon, pebs, bts, rep_good, nopl, xtopology, nonstop_tsc, cpuid, aperfmperf, tsc_known_freq, pni, pclmulqdq, dtes64, monitor, ds_cpl, vmx, smx, est, tm2, ssse3, sdbg, fma, cx16, xtpr, pdcm, pcid, sse4_1, sse4_2, x2apic, movbe, popcnt, tsc_deadline_timer, aes, xsave, avx, f16c, rdrand, lahf_lm, abm, 3dnowprefetch, cpuid_fault, epb, ssbd, ibrs, ibpb, stibp, ibrs_enhanced, tpr_shadow, flexpriority, ept, vpid, ept_ad, fsgsbase, tsc_adjust, bmi1, avx2, smep, bmi2, erms, invpcid, rdseed, adx, smap, clflushopt, clwb, intel_pt, sha_ni, xsaveopt, xsavec, xgetbv1, xsaves, split_lock_detect, user_shstk, avx_vnni, dtherm, ida, arat, pln, pts, hwp, hwp_notify, hwp_act_window, hwp_epp, hwp_pkg_req, hfi, vnmi, umip, pku, ospke, waitpkg, gfni, vaes, vpclmulqdq, rdpid, movdiri, movdir64b, fsrm, md_clear, serialize, arch_lbr, ibt, flush_l1d, arch_capabilities
+```
+
+### Encoder
+
+![benchmark_encode_on_AMD_EPYC_9R14_with_rustc_1.90.0_1159e78c4_2025-09-14](./benchmark_encode_on_AMD_EPYC_9R14_with_rustc_1.90.0_1159e78c4_2025-09-14.png)
+
+![benchmark_encode_zero_alloc_on_AMD_EPYC_9R14_with_rustc_1.90.0_1159e78c4_2025-09-14](./benchmark_encode_zero_alloc_on_AMD_EPYC_9R14_with_rustc_1.90.0_1159e78c4_2025-09-14.png)
+
+### Recoder
+
+![benchmark_recode_on_AMD_EPYC_9R14_with_rustc_1.90.0_1159e78c4_2025-09-14](./benchmark_recode_on_AMD_EPYC_9R14_with_rustc_1.90.0_1159e78c4_2025-09-14.png)
+
+![benchmark_recode_zero_alloc_on_AMD_EPYC_9R14_with_rustc_1.90.0_1159e78c4_2025-09-14](./benchmark_recode_zero_alloc_on_AMD_EPYC_9R14_with_rustc_1.90.0_1159e78c4_2025-09-14.png)
+
+### Decoder
+
+![benchmark_decode_on_AMD_EPYC_9R14_with_rustc_1.90.0_1159e78c4_2025-09-14](./benchmark_decode_on_AMD_EPYC_9R14_with_rustc_1.90.0_1159e78c4_2025-09-14.png)