Remove platform-specific compiler options from the main CMakeLists.txt

CMakeLists.txt

@@ -40,10 +40,12 @@ target_link_libraries(blast
 )
 
 target_compile_options(blast
-    INTERFACE "-Wno-ignored-attributes" "-fno-math-errno" "-ftemplate-backtrace-limit=0"
-    # Enable SIMD instruction sets, otherwise it does not compile.
-    # This will change when we support multiple architectures.
-    INTERFACE "-march=native" "-mfma" "-mavx" "-mavx2" "-msse4"
+    INTERFACE "-Wno-ignored-attributes" "-fno-math-errno"
+)
+
+target_compile_options(blast
+    INTERFACE
+        $<$<COMPILE_LANGUAGE:CXX>:-ftemplate-backtrace-limit=0>
 )
 
 # BLAST_WITH_BLASFEO

Makefile

@@ -3,8 +3,10 @@ BENCH_BLASFEO = build/bin/bench-blasfeo
 BENCH_BLAZE = build/bin/bench-blaze
 BENCH_EIGEN = build/bin/bench-eigen
 BENCH_BLAST = build/bin/bench-blast
+BENCH_BLAST_OUTPUT_DIR = $(shell git rev-parse --short HEAD)
+BENCH_BLAST_CPU_CORE = 11
 BENCH_LIBXSMM = build/bin/bench-libxsmm
-BENCHMARK_OPTIONS = --benchmark_repetitions=5 --benchmark_counters_tabular=true --benchmark_out_format=json --benchmark_report_aggregates_only=true
+BENCHMARK_OPTIONS = --benchmark_repetitions=30 --benchmark_counters_tabular=true --benchmark_out_format=json --benchmark_enable_random_interleaving=true --benchmark_min_warmup_time=10 --benchmark_min_time=1000000x
 RUN_MATLAB = matlab -nodisplay -nosplash -nodesktop -r
 BENCH_DATA = bench_result/data
 BENCH_IMAGE = bench_result/image
@@ -61,29 +63,29 @@ ${BENCH_DATA}/sgemm-blaze-static.json: $(BENCH_BLAZE)
 	$(BENCH_BLAZE) --benchmark_filter="BM_gemm_static<float>*" $(BENCHMARK_OPTIONS) \
 		--benchmark_out=${BENCH_DATA}/sgemm-blaze-static.json
 
-${BENCH_DATA}/dgemm-blast-static-panel.json: $(BENCH_BLAST)
-	$(BENCH_BLAST) --benchmark_filter="BM_gemm_static_panel<double, .+>" $(BENCHMARK_OPTIONS) \
-		--benchmark_out=${BENCH_DATA}/dgemm-blast-static-panel.json
+${BENCH_DATA}/${BENCH_BLAST_OUTPUT_DIR}/dgemm-blast-static-panel.json: $(BENCH_BLAST)
+	taskset -c ${BENCH_BLAST_CPU_CORE} $(BENCH_BLAST) --benchmark_filter="BM_gemm_static_panel<double, .+>" $(BENCHMARK_OPTIONS) \
+		--benchmark_out=${BENCH_DATA}/${BENCH_BLAST_OUTPUT_DIR}/dgemm-blast-static-panel.json
 
-${BENCH_DATA}/dgemm-blast-dynamic-panel.json: $(BENCH_BLAST)
-	$(BENCH_BLAST) --benchmark_filter="BM_gemm_dynamic_panel<double>" $(BENCHMARK_OPTIONS) \
-		--benchmark_out=${BENCH_DATA}/dgemm-blast-dynamic-panel.json
+${BENCH_DATA}/${BENCH_BLAST_OUTPUT_DIR}/dgemm-blast-dynamic-panel.json: $(BENCH_BLAST)
+	taskset -c ${BENCH_BLAST_CPU_CORE} $(BENCH_BLAST) --benchmark_filter="BM_gemm_dynamic_panel<double>" $(BENCHMARK_OPTIONS) \
+		--benchmark_out=${BENCH_DATA}/${BENCH_BLAST_OUTPUT_DIR}/dgemm-blast-dynamic-panel.json
 
-${BENCH_DATA}/dgemm-blast-static-plain.json: $(BENCH_BLAST)
-	$(BENCH_BLAST) --benchmark_filter="BM_gemm_static_plain<double, .+>" $(BENCHMARK_OPTIONS) \
-		--benchmark_out=${BENCH_DATA}/dgemm-blast-static-plain.json
+${BENCH_DATA}/${BENCH_BLAST_OUTPUT_DIR}/dgemm-blast-static-plain.json: $(BENCH_BLAST)
+	taskset -c ${BENCH_BLAST_CPU_CORE} $(BENCH_BLAST) --benchmark_filter="BM_gemm_static_plain<double, .+>" $(BENCHMARK_OPTIONS) \
+		--benchmark_out=${BENCH_DATA}/${BENCH_BLAST_OUTPUT_DIR}/dgemm-blast-static-plain.json
 
-${BENCH_DATA}/dgemm-blast-dynamic-plain.json: $(BENCH_BLAST)
-	$(BENCH_BLAST) --benchmark_filter="BM_gemm_dynamic_plain<double>" $(BENCHMARK_OPTIONS) \
-		--benchmark_out=${BENCH_DATA}/dgemm-blast-dynamic-plain.json
+${BENCH_DATA}/${BENCH_BLAST_OUTPUT_DIR}/dgemm-blast-dynamic-plain.json: $(BENCH_BLAST)
+	taskset -c ${BENCH_BLAST_CPU_CORE} $(BENCH_BLAST) --benchmark_filter="BM_gemm_dynamic_plain<double>" $(BENCHMARK_OPTIONS) \
+		--benchmark_out=${BENCH_DATA}/${BENCH_BLAST_OUTPUT_DIR}/dgemm-blast-dynamic-plain.json
 
 ${BENCH_DATA}/sgemm-blast-static-panel.json: $(BENCH_BLAST)
-	$(BENCH_BLAST) --benchmark_filter="BM_gemm_static_panel<float, .+>" $(BENCHMARK_OPTIONS) \
-		--benchmark_out=${BENCH_DATA}/sgemm-blast-static-panel.json
+	taskset -c ${BENCH_BLAST_CPU_CORE} $(BENCH_BLAST) --benchmark_filter="BM_gemm_static_panel<float, .+>" $(BENCHMARK_OPTIONS) \
+		--benchmark_out=${BENCH_DATA}/${BENCH_BLAST_OUTPUT_DIR}/sgemm-blast-static-panel.json
 
 ${BENCH_DATA}/sgemm-blast-dynamic-panel.json: $(BENCH_BLAST)
-	$(BENCH_BLAST) --benchmark_filter="BM_gemm_dynamic_panel<float>" $(BENCHMARK_OPTIONS) \
-		--benchmark_out=${BENCH_DATA}/sgemm-blast-dynamic-panel.json
+	taskset -c ${BENCH_BLAST_CPU_CORE} $(BENCH_BLAST) --benchmark_filter="BM_gemm_dynamic_panel<float>" $(BENCHMARK_OPTIONS) \
+		--benchmark_out=${BENCH_DATA}/${BENCH_BLAST_OUTPUT_DIR}/sgemm-blast-dynamic-panel.json
 
 ${BENCH_DATA}/dgemm-libxsmm.json: $(BENCH_LIBXSMM)
 	$(BENCH_LIBXSMM) --benchmark_filter="BM_gemm_nt<double>" $(BENCHMARK_OPTIONS) \
@@ -94,16 +96,17 @@ ${BENCH_DATA}/sgemm-libxsmm.json: $(BENCH_LIBXSMM)
 		--benchmark_out=${BENCH_DATA}/sgemm-libxsmm.json
 
 dgemm-benchmarks: \
+	$(shell mkdir -p ${BENCH_DATA}/${BENCH_BLAST_OUTPUT_DIR}) \
 	${BENCH_DATA}/dgemm-openblas.json \
 	${BENCH_DATA}/dgemm-mkl.json \
 	${BENCH_DATA}/dgemm-libxsmm.json \
 	${BENCH_DATA}/dgemm-blasfeo.json \
 	${BENCH_DATA}/dgemm-blaze-static.json \
 	${BENCH_DATA}/dgemm-eigen-static.json \
-	${BENCH_DATA}/dgemm-blast-static-panel.json \
-	${BENCH_DATA}/dgemm-blast-dynamic-panel.json \
-	${BENCH_DATA}/dgemm-blast-static-plain.json \
-	${BENCH_DATA}/dgemm-blast-dynamic-plain.json
+	${BENCH_DATA}/${BENCH_BLAST_OUTPUT_DIR}/dgemm-blast-static-panel.json \
+	${BENCH_DATA}/${BENCH_BLAST_OUTPUT_DIR}/dgemm-blast-dynamic-panel.json \
+	${BENCH_DATA}/${BENCH_BLAST_OUTPUT_DIR}/dgemm-blast-static-plain.json \
+	${BENCH_DATA}/${BENCH_BLAST_OUTPUT_DIR}/dgemm-blast-dynamic-plain.json
 
 
 #
@@ -238,4 +241,4 @@ ${BENCH_IMAGE}/mpc_software.pdf_tex: ${BENCH_IMAGE}/mpc_software.svg
 	/usr/bin/inkscape --without-gui --file=${BENCH_IMAGE}/mpc_software.svg --export-pdf=${BENCH_IMAGE}/mpc_software.pdf --export-latex --export-area-drawing
 
 ${BENCH_IMAGE}/mpc_software.pdf: ${BENCH_IMAGE}/mpc_software.svg
-	/usr/bin/inkscape --without-gui --file=${BENCH_IMAGE}/mpc_software.svg --export-pdf=${BENCH_IMAGE}/mpc_software.pdf --export-area-drawing
+	/usr/bin/inkscape --without-gui --file=${BENCH_IMAGE}/mpc_software.svg --export-pdf=${BENCH_IMAGE}/mpc_software.pdf --export-area-drawing

bench/analysis/dgemm_performance.py

@@ -2,10 +2,18 @@
 matplotlib.use("Agg")
 import matplotlib.pyplot as plt
 import json
-
+import glob
+import pathlib
 
 def filter_aggregate(benchmarks, name):
-    return [b for b in benchmarks if b['aggregate_name'] == name]
+    result = []
+    for b in benchmarks:
+        try:
+            if b['aggregate_name'] == name:
+                result.append(b)
+        except KeyError:
+            continue
+    return result
 
 
 factor = 1e+9 # Giga
@@ -17,20 +25,21 @@ def filter_aggregate(benchmarks, name):
 
 plots = [
     # {'data_file': 'dgemm-openblas.json', 'label': 'OpenBLAS'},
-    {'data_file': 'dgemm-mkl.json', 'label': 'MKL'},
+    # {'data_file': 'dgemm-mkl.json', 'label': 'MKL'},
     {'data_file': 'dgemm-blasfeo.json', 'label': 'BLASFEO'},
     # {'data_file': 'dgemm-blasfeo-blas.json', 'label': 'BLASFEO*'},
-    {'data_file': 'dgemm-libxsmm.json', 'label': 'LIBXSMM'},
+    # {'data_file': 'dgemm-libxsmm.json', 'label': 'LIBXSMM'},
     # {'data_file': 'dgemm-eigen-dynamic.json', 'label': 'Eigen (D)'},
     # {'data_file': 'dgemm-eigen-static.json', 'label': 'Eigen (S)'},
     # {'data_file': 'dgemm-blaze-dynamic.json', 'label': 'Blaze (D)'},
-    {'data_file': 'dgemm-blaze-static.json', 'label': 'Blaze (S)'},
-    {'data_file': 'dgemm-blast-static-panel.json', 'label': 'BLAST (SP)'},
-    {'data_file': 'dgemm-blast-static-plain.json', 'label': 'BLAST (SD)'},
-    {'data_file': 'dgemm-blast-dynamic-panel.json', 'label': 'BLAST (DP)'},
-    {'data_file': 'dgemm-blast-dynamic-plain.json', 'label': 'BLAST (DD)'},
+    # {'data_file': 'dgemm-blaze-static.json', 'label': 'Blaze (S)'},
 ]
 
+for benchmark_file, benchmark_label in [('dgemm-blast-static-panel.json', 'SP'), ('dgemm-blast-static-plain.json', 'SD'), ('dgemm-blast-dynamic-panel.json', 'DP'), ('dgemm-blast-dynamic-plain.json', 'DD')]:
+    files = glob.glob('./**/' + benchmark_file, recursive=True, root_dir='bench_result/data')
+    for file in files:
+        plots.append({'data_file': file, 'label': f'BLAST ({benchmark_label}) {pathlib.Path(file).parent.stem}'})
+
 fig = plt.figure(figsize=[10, 6])
 ax = fig.subplots()
 

bench/analysis/dgemm_performance_ratio.py

@@ -5,8 +5,14 @@
 
 
 def filter_aggregate(benchmarks, name):
-    return [b for b in benchmarks if b['aggregate_name'] == name]
-
+    result = []
+    for b in benchmarks:
+        try:
+            if b['aggregate_name'] == name:
+                result.append(b)
+        except KeyError:
+            continue
+    return result
 
 def load_benchmark(file_name):
     with open(file_name) as f:

bench/blast/math/dense/DynamicGemm.cpp

@@ -2,12 +2,12 @@
 // Use of this source code is governed by a BSD-style
 // license that can be found in the LICENSE file.
 
-#include <blast/math/dense/Gemm.hpp>
+#include <blast/math/algorithm/Gemm.hpp>
+#include <blast/math/Matrix.hpp>
+#include <blast/blaze/Math.hpp>
 
 #include <bench/Gemm.hpp>
 
-#include <blaze/math/DynamicMatrix.h>
-
 #include <test/Randomize.hpp>
 
 

bench/blast/math/dense/StaticGemm.cpp

@@ -2,12 +2,12 @@
 // Use of this source code is governed by a BSD-style
 // license that can be found in the LICENSE file.
 
-#include <blast/math/dense/Gemm.hpp>
+#include <blast/math/algorithm/Gemm.hpp>
+#include <blast/math/Matrix.hpp>
+#include <blast/blaze/Math.hpp>
 
 #include <bench/Gemm.hpp>
 
-#include <blaze/math/StaticMatrix.h>
-
 #include <test/Randomize.hpp>
 
 

bench/blast/math/dense/StaticIamax.cpp

@@ -14,7 +14,7 @@
 
 #include <blast/math/dense/Iamax.hpp>
 
-#include <blaze/math/DynamicVector.h>
+#include <blast/blaze/Math.hpp>
 
 #include <bench/Iamax.hpp>
 #include <bench/Complexity.hpp>

bench/blast/math/dense/StaticTrmm.cpp

@@ -3,15 +3,12 @@
 // license that can be found in the LICENSE file.
 
 #include <blast/math/dense/Trmm.hpp>
-
-#include <blaze/math/StaticMatrix.h>
+#include <blast/math/Matrix.hpp>
+#include <blast/blaze/Math.hpp>
 
 #include <bench/Gemm.hpp>
 #include <test/Randomize.hpp>
 
-#include <random>
-#include <memory>
-
 
 namespace blast :: benchmark
 {

bench/blast/math/panel/DynamicGemm.cpp

@@ -3,7 +3,7 @@
 // license that can be found in the LICENSE file.
 
 #include <blast/math/DynamicPanelMatrix.hpp>
-#include <blast/math/panel/Gemm.hpp>
+#include <blast/math/algorithm/Gemm.hpp>
 
 #include <bench/Gemm.hpp>
 

bench/blast/math/panel/StaticGemm.cpp

@@ -3,7 +3,9 @@
 // license that can be found in the LICENSE file.
 
 #include <blast/math/StaticPanelMatrix.hpp>
-#include <blast/math/panel/Gemm.hpp>
+#include <blast/math/Matrix.hpp>
+#include <blast/math/algorithm/Gemm.hpp>
+#include <blast/blaze/Math.hpp>
 
 #include <bench/Gemm.hpp>
 

bench/blast/math/simd/Trmm.cpp

@@ -2,14 +2,14 @@
 // Use of this source code is governed by a BSD-style
 // license that can be found in the LICENSE file.
 
-#include <blast/math/dense/StaticMatrixPointer.hpp>
+#include <blast/math/Matrix.hpp>
 #include <blast/math/RegisterMatrix.hpp>
 
 #include <bench/Benchmark.hpp>
 
 #include <test/Randomize.hpp>
 
-#include <blaze/math/StaticMatrix.h>
+#include <blast/blaze/Math.hpp>
 
 
 namespace blast :: benchmark
@@ -81,4 +81,4 @@ namespace blast :: benchmark
     BENCHMARK_TEMPLATE(BM_RegisterMatrix_trmmRightLower, float, 24, 4, columnMajor);
     BENCHMARK_TEMPLATE(BM_RegisterMatrix_trmmRightLower, float, 16, 5, columnMajor);
     BENCHMARK_TEMPLATE(BM_RegisterMatrix_trmmRightLower, float, 16, 6, columnMajor);
-}
+}

include/blast/blaze/Math.hpp

@@ -0,0 +1,10 @@
+// Copyright 2024 Mikhail Katliar. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+#pragma once
+
+#include <blast/blaze/math/Vector.hpp>
+#include <blast/blaze/math/TypeTraits.hpp>
+
+#include <blaze/Math.h>

include/blast/blaze/math/TypeTraits.hpp

@@ -0,0 +1,12 @@
+// Copyright 2024 Mikhail Katliar. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+#pragma once
+
+#include <blast/blaze/math/typetraits/IsContiguous.hpp>
+#include <blast/blaze/math/typetraits/IsStatic.hpp>
+#include <blast/blaze/math/typetraits/IsStaticallySpaced.hpp>
+#include <blast/blaze/math/typetraits/IsDenseVector.hpp>
+#include <blast/blaze/math/typetraits/IsDenseMatrix.hpp>
+#include <blast/blaze/math/typetraits/Spacing.hpp>

include/blast/blaze/math/Vector.hpp

@@ -0,0 +1,38 @@
+// Copyright 2024 Mikhail Katliar. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+#pragma once
+
+#include <blast/blaze/math/typetraits/IsContiguous.hpp>
+
+#include <blaze/math/typetraits/IsDenseVector.h>
+#include <blaze/math/typetraits/IsView.h>
+
+
+namespace blaze
+{
+    /**
+     * @brief Memory distance between consecutive elements of a dense Blaze vector
+     *
+     * NOTE: The function is declared in blaze namespace s.t. it can be found by ADL.
+     *
+     * @tparam VT vector type
+     *
+     * @param v vector
+     *
+     * @return memory distance between consecutive elements of @a v
+     */
+    template <typename VT>
+    requires blaze::IsDenseVector_v<VT>
+    inline size_t spacing(VT const& v) noexcept
+    {
+        if constexpr (IsContiguous_v<VT>)
+            return 1;
+        else
+            if constexpr (blaze::IsView_v<VT>)
+                return spacing(v.operand());
+            else
+                static_assert(false, "Spacing is not defined for a type which is not a view and is not contiguous");
+    }
+}

include/blast/blaze/math/typetraits/IsContiguous.hpp

@@ -0,0 +1,41 @@
+// Copyright 2024 Mikhail Katliar. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+#pragma once
+
+#include <blast/math/typetraits/IsContiguous.hpp>
+
+#include <blaze/math/Aliases.h>
+#include <blaze/math/typetraits/IsContiguous.h>
+#include <blaze/math/typetraits/IsVector.h>
+#include <blaze/math/typetraits/IsTransExpr.h>
+
+#include <type_traits>
+
+
+namespace blast
+{
+    /**
+     * @brief Specialization for Blaze vectors which are not transpose expressions
+     *
+     * @tparam T type
+     */
+    template <typename T>
+    requires blaze::IsVector_v<T> && (!blaze::IsTransExpr_v<T>)
+    struct IsContiguous<T> : blaze::IsContiguous<T> {};
+
+
+    /**
+     * @brief Specialization for Blaze vector transpose expressions
+     *
+     * The trasnposed vector expression is contiguous iff its operand is contiguous.
+     *
+     * This specialization is required to fix this Blaze bug: https://bitbucket.org/blaze-lib/blaze/issues/474
+     *
+     * @tparam T type
+     */
+    template <typename T>
+    requires blaze::IsVector_v<T> && blaze::IsTransExpr_v<T>
+    struct IsContiguous<T> : IsContiguous<std::remove_reference_t<blaze::Operand_t<T>>> {};
+}