Change to INTGEMM_COMPILER_SUPPORTS_AVX512BW and update test

34 files changed, 73 insertions, 73 deletions

diff --git a/CMakeLists.txt b/CMakeLists.txt
index f29ccff..bbb9b83 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -16,15 +16,13 @@ else()
   add_compile_options(-Wall -Wextra -pedantic -Werror -Wno-unknown-pragmas)
 endif()
 
-# Check if compiler supports AVX512
-try_compile(INTGEMM_COMPILER_SUPPORTS_AVX512
+# Check if compiler supports AVX512BW
+try_compile(INTGEMM_COMPILER_SUPPORTS_AVX512BW
   ${CMAKE_CURRENT_BINARY_DIR}/compile_tests
-  ${CMAKE_CURRENT_SOURCE_DIR}/compile_test_avx512.cc
-  #Hack: pass compiler arguments as definitions because the test code overrides CXX_FLAGS :'(
-  COMPILE_DEFINITIONS -mavx512f -mavx512bw -mavx512dq)
+  ${CMAKE_CURRENT_SOURCE_DIR}/compile_test_avx512bw.cc)
 
-if(NOT INTGEMM_COMPILER_SUPPORTS_AVX512)
-  message(WARNING "${Orange}Not building AVX512-based multiplication because your compiler is too old.\nFor details rerun cmake with --debug-trycompile then try to build in compile_tests/CMakeFiles/CMakeTmp.${ColourReset}")
+if(NOT INTGEMM_COMPILER_SUPPORTS_AVX512BW)
+  message(WARNING "${Orange}Not building AVX512BW-based multiplication because your compiler is too old.\nFor details rerun cmake with --debug-trycompile then try to build in compile_tests/CMakeFiles/CMakeTmp.${ColourReset}")
 endif()
 
 try_compile(INTGEMM_COMPILER_SUPPORTS_AVX512VNNI
diff --git a/avx512_gemm.h b/avx512_gemm.h
index 267dc6d..cdbfff5 100644
--- a/avx512_gemm.h
+++ b/avx512_gemm.h
@@ -2,7 +2,7 @@
 
 #include "intgemm_config.h"
 
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 
 #include "interleave.h"
 #include "kernels.h"
diff --git a/benchmarks/benchmark.cc b/benchmarks/benchmark.cc
index 26b0ac5..6063d5c 100644
--- a/benchmarks/benchmark.cc
+++ b/benchmarks/benchmark.cc
@@ -196,7 +196,7 @@ int main(int, char ** argv) {
     RunAll<AVX2_16bit>(matrices, end, stats.avx2_16bit);
   }
 
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
   std::cerr << "AVX512 8bit, 100 samples..." << std::endl;
   for (int samples = 0; samples < kSamples; ++samples) {
     RandomMatrices *end = (samples < 4) ? matrices_end : full_sample;
@@ -225,7 +225,7 @@ int main(int, char ** argv) {
     std::cout << "Multiply\t" << matrices[i].A_rows << '\t' << matrices[i].width << '\t' << matrices[i].B_cols << '\t' << "Samples=" << (kOutlierThreshold * stats.sse2_16bit[i].size()) << '\n';
     Print<SSSE3_8bit>(stats.ssse3_8bit, i);
     Print<AVX2_8bit>(stats.avx2_8bit, i);
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
     Print<AVX512_8bit>(stats.avx512_8bit, i);
 #endif
 #ifdef INTGEMM_COMPILER_SUPPORTS_AVX512VNNI
@@ -233,7 +233,7 @@ int main(int, char ** argv) {
 #endif
     Print<SSE2_16bit>(stats.sse2_16bit, i);
     Print<AVX2_16bit>(stats.avx2_16bit, i);
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
     Print<AVX512_16bit>(stats.avx512_16bit, i);
 #endif
   }
diff --git a/benchmarks/benchmark_quantizer.cc b/benchmarks/benchmark_quantizer.cc
index b9b0782..eb96499 100644
--- a/benchmarks/benchmark_quantizer.cc
+++ b/benchmarks/benchmark_quantizer.cc
@@ -36,7 +36,7 @@ int main() {
     }
     QuantizerBench<intgemm::SSSE3_8bit>(in.begin(), out.begin(), count);
     QuantizerBench<intgemm::AVX2_8bit>(in.begin(), out.begin(), count);
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
     QuantizerBench<intgemm::AVX512_8bit>(in.begin(), out.begin(), count);
 #endif
   }
diff --git a/benchmarks/biasmultiply.cc b/benchmarks/biasmultiply.cc
index 220422a..515536e 100644
--- a/benchmarks/biasmultiply.cc
+++ b/benchmarks/biasmultiply.cc
@@ -197,7 +197,7 @@ int main(int argc, char ** argv) {
 	}
 
 	std::cout << repeat << " iterations of Shifted AVX2 took: " << newTimeAVX2.count() << " seconds." << std::endl;
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 	if (kCPU < CPUType::AVX512BW) return 0;
 	std::chrono::duration<double> oldAVX512_nobias = testOld_nobias<AVX512_8bit>(1, 64, 8);
 	for (int i = 0; i<repeat; i++) {
diff --git a/callbacks.h b/callbacks.h
index c8a29df..24f9009 100644
--- a/callbacks.h
+++ b/callbacks.h
@@ -18,7 +18,7 @@
 #include "callbacks/implementations.inl"
 #undef CALLBACKS_THIS_IS_AVX2
 
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 #define CALLBACKS_THIS_IS_AVX512BW
 #include "callbacks/implementations.inl"
 #undef CALLBACKS_THIS_IS_AVX512BW
diff --git a/compile_test_avx512.cc b/compile_test_avx512.cc
deleted file mode 100644
index f56cc12..0000000
--- a/compile_test_avx512.cc
+++ /dev/null
@@ -1,16 +0,0 @@
-// Some compilers don't have AVX512BW support.  Test for them.
-#include <immintrin.h>
-
-#include <iostream>
-
-int main() {
-  // AVX512F
-  __m512i value = _mm512_set1_epi32(1);
-  // AVX512BW
-  value = _mm512_maddubs_epi16(value, value);
-
-  __m256i value2 = _mm256_set1_epi8(1);
-  // AVX512DQ
-  value = _mm512_inserti32x8(value, value2, 1);
-  return *(int*)&value && __builtin_cpu_supports("avx512f");
-}
diff --git a/compile_test_avx512bw.cc b/compile_test_avx512bw.cc
new file mode 100644
index 0000000..8d07551
--- /dev/null
+++ b/compile_test_avx512bw.cc
@@ -0,0 +1,18 @@
+// Some compilers don't have AVX512BW support.  Test for them.
+#include <immintrin.h>
+
+#if defined __INTEL_COMPILER
+#define INTGEMM_AVX512BW __attribute__ ((target ("avx512f")))
+#else
+#define INTGEMM_AVX512BW __attribute__ ((target ("avx512bw")))
+#endif
+
+INTGEMM_AVX512BW int Test() {
+  // AVX512BW
+  __m512i value = _mm512_set1_epi32(1);
+  value = _mm512_maddubs_epi16(value, value);
+  return *(int*)&value;
+}
+
+int main() {
+}
diff --git a/interleave.h b/interleave.h
index 9bd4fdd..231be46 100644
--- a/interleave.h
+++ b/interleave.h
@@ -28,7 +28,7 @@ INTGEMM_INTERLEAVE_N(target, type, 64)
 
 INTGEMM_INTERLEAVE(INTGEMM_SSE2, __m128i)
 INTGEMM_INTERLEAVE(INTGEMM_AVX2, __m256i)
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 INTGEMM_INTERLEAVE(INTGEMM_AVX512BW, __m512i)
 #endif
 
@@ -44,7 +44,7 @@ target static inline void Swap(Register &a, Register &b) { \
 
 INTGEMM_SWAP(INTGEMM_SSE2, __m128i)
 INTGEMM_SWAP(INTGEMM_AVX2, __m256i)
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 /* Only INTGEMM_AVX512F is necessary but due to GCC 5.4 bug we have to set INTGEMM_AVX512BW */
 INTGEMM_SWAP(INTGEMM_AVX512BW, __m512i)
 #endif
@@ -97,7 +97,7 @@ target static inline void Transpose16InLane(Register &r0, Register &r1, Register
 
 INTGEMM_TRANSPOSE16(INTGEMM_SSE2, __m128i)
 INTGEMM_TRANSPOSE16(INTGEMM_AVX2, __m256i)
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 /* Only INTGEMM_AVX512F is necessary but due to GCC 5.4 bug we have to set INTGEMM_AVX512BW */
 INTGEMM_TRANSPOSE16(INTGEMM_AVX512BW, __m512i)
 #endif
diff --git a/intgemm.h b/intgemm.h
index ba86e53..6d4d95f 100644
--- a/intgemm.h
+++ b/intgemm.h
@@ -51,7 +51,7 @@
 #include "avx512_gemm.h"
 #include "avx512vnni_gemm.h"
 
-#if defined(__GNUC__) && defined(INTGEMM_COMPILER_SUPPORTS_AVX512)
+#if defined(__GNUC__) && defined(INTGEMM_COMPILER_SUPPORTS_AVX512BW)
 #include "cpuid.h"
 #endif
 
@@ -119,7 +119,7 @@ struct Unsupported_8bit {
   constexpr static const char *const kName = "8-bit Unsupported";
 };
 
-#ifndef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifndef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 // These won't ever be called in this capacity, but it does let the code below compile.
 typedef Unsupported_16bit AVX512_16bit;
 typedef Unsupported_8bit AVX512_8bit;
@@ -136,7 +136,7 @@ typedef Unsupported_8bit AVX512VNNI_8bit;
 #endif
 
 
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 // gcc 5.4.0 bizarrely supports avx512bw targets but not __builtin_cpu_supports("avx512bw").  So implement it manually.
 inline bool CheckAVX512BW() {
   __builtin_cpu_init ();
@@ -185,7 +185,7 @@ template <class T> T ChooseCPU(T
     return avx512vnni;
   }
 #endif
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
   if (CheckAVX512BW()) {
     return avx512bw;
   }
diff --git a/intgemm_config.h.in b/intgemm_config.h.in
index 11be91c..920e9ae 100644
--- a/intgemm_config.h.in
+++ b/intgemm_config.h.in
@@ -1,4 +1,4 @@
 #pragma once
 
-#cmakedefine INTGEMM_COMPILER_SUPPORTS_AVX512
+#cmakedefine INTGEMM_COMPILER_SUPPORTS_AVX512BW
 #cmakedefine INTGEMM_COMPILER_SUPPORTS_AVX512VNNI
diff --git a/intrinsics.h b/intrinsics.h
index 5fe3159..d662472 100644
--- a/intrinsics.h
+++ b/intrinsics.h
@@ -381,7 +381,7 @@ INTGEMM_AVX2 static inline __m256i xor_si(__m256i a, __m256i b) {
  * AVX512
  *
  */
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 
 INTGEMM_AVX512BW static inline __m512i abs_epi8(__m512i arg) {
   return _mm512_abs_epi8(arg);
diff --git a/kernels.h b/kernels.h
index ef63fec..84631b5 100644
--- a/kernels.h
+++ b/kernels.h
@@ -16,7 +16,7 @@
 #include "kernels/implementations.inl"
 #undef KERNELS_THIS_IS_AVX2
 
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 #define KERNELS_THIS_IS_AVX512BW
 #include "kernels/implementations.inl"
 #undef KERNELS_THIS_IS_AVX512BW
diff --git a/multiply.h b/multiply.h
index a9766d3..a313c16 100644
--- a/multiply.h
+++ b/multiply.h
@@ -36,7 +36,7 @@ INTGEMM_AVX2 static inline __m256i PermuteSummer(__m256i pack0123, __m256i pack4
   return _mm256_add_epi32(rev, blended);
 }
 
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 /* Only INTGEMM_AVX512F is necessary but due to GCC 5.4 bug we have to set INTGEMM_AVX512BW */
 INTGEMM_AVX512BW static inline __m256i PermuteSummer(__m512i pack0123, __m512i pack4567) {
   // Form [0th 128-bit register of pack0123, 0st 128-bit register of pack4567, 2nd 128-bit register of pack0123, 2nd 128-bit register of pack4567]
@@ -104,7 +104,7 @@ target inline Register Pack0123(Register sum0, Register sum1, Register sum2, Reg
 
 INTGEMM_PACK0123(INTGEMM_SSE2, __m128i)
 INTGEMM_PACK0123(INTGEMM_AVX2, __m256i)
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 /* Only INTGEMM_AVX512F is necessary but due to GCC 5.4 bug we have to set INTGEMM_AVX512BW */
 INTGEMM_PACK0123(INTGEMM_AVX512BW, __m512i)
 #endif
diff --git a/test/add127_test.cc b/test/add127_test.cc
index c271a0c..d959b14 100644
--- a/test/add127_test.cc
+++ b/test/add127_test.cc
@@ -295,7 +295,7 @@ TEST_CASE("PrepareBias AVX2", "[Add127]") {
 
 TEST_CASE("PrepareBias AVX512F", "[Add127]") {
 	if (kCPU < CPUType::AVX512BW) return;
-	#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+	#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 	TestPrepareBias<AVX512_8bit>(256,256);
 	TestPrepareBias<AVX512_8bit>(2048,256);
 	TestPrepareBias<AVX512_8bit>(512,512);
@@ -321,7 +321,7 @@ TEST_CASE("PrepareA AVX2", "[Add127]") {
 
 TEST_CASE("PrepareA AVX512F", "[Add127]") {
 	if (kCPU < CPUType::AVX512BW) return;
-	#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+	#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 	TestPrepareA<AVX512_8bit>(64,64);
 	TestPrepareA<AVX512_8bit>(256,256);
 	TestPrepareA<AVX512_8bit>(512,512);
@@ -352,7 +352,7 @@ TEST_CASE ("Multiply AVX2 8bit Shift with bias", "[Add127]") {
   TestMultiplyBiasNew<AVX2_8bit>(248, 256, 256, 0.48, 0.64, 0.16, 0.15);
   TestMultiplyBiasNew<AVX2_8bit>(200, 256, 256, 0.55, 0.74, 0.17, 0.16);
 }
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 TEST_CASE ("Multiply AVX512F 8bit Shift with bias", "[Add127]") {
   if (kCPU < CPUType::AVX512BW) return;
   TestMultiplyBiasNew<AVX512_8bit>(1, 64, 8, 0.0001, 0.05, 0.03, 0.001);
@@ -400,7 +400,7 @@ TEST_CASE ("Multiply AVX2 8bit Shift vs nonshift", "[Add127]") {
   TestMultiplyShiftNonShift<AVX2_8bit>(248, 256, 256, 0.0001, 0.64, 0.16, 0.0001);
   TestMultiplyShiftNonShift<AVX2_8bit>(200, 256, 256, 0.0001, 0.74, 0.17, 0.0001);
 }
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 TEST_CASE ("Multiply AVX512F 8bit Shift vs nonshift", "[Add127]") {
   if (kCPU < CPUType::AVX512BW) return;
   TestMultiplyShiftNonShift<AVX512_8bit>(1, 64, 8, 0.0001, 0.05, 0.03, 0.001);
@@ -448,7 +448,7 @@ TEST_CASE ("Multiply AVX2 8bit Shift vs Int", "[Add127]") {
   TestMultiplyShiftInt<AVX2_8bit>(248, 256, 256, 0.0001f, 0.64, 0.16, 0.0001f);
   TestMultiplyShiftInt<AVX2_8bit>(200, 256, 256, 0.0001f, 0.74, 0.17, 0.0001f);
 }
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 TEST_CASE ("Multiply AVX512F 8bit Shift vs Int", "[Add127]") {
   if (kCPU < CPUType::AVX512BW) return;
   TestMultiplyShiftInt<AVX512_8bit>(1, 64, 8, 0.0001f, 0.05, 0.03, 0.0001f);
diff --git a/test/kernels/add_bias_test.cc b/test/kernels/add_bias_test.cc
index 7b10b56..2dd4e3d 100644
--- a/test/kernels/add_bias_test.cc
+++ b/test/kernels/add_bias_test.cc
@@ -48,7 +48,7 @@ KERNEL_TEST_CASE("add_bias/int AVX2") { return kernel_add_bias_test<CPUType::AVX
 KERNEL_TEST_CASE("add_bias/float AVX2") { return kernel_add_bias_test<CPUType::AVX2, float>(); }
 KERNEL_TEST_CASE("add_bias/double AVX2") { return kernel_add_bias_test<CPUType::AVX2, double>(); }
 
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 template INTGEMM_AVX512BW void kernel_add_bias_test<CPUType::AVX512BW, int8_t>();
 template INTGEMM_AVX512BW void kernel_add_bias_test<CPUType::AVX512BW, int16_t>();
 template INTGEMM_AVX512BW void kernel_add_bias_test<CPUType::AVX512BW, int>();
diff --git a/test/kernels/bitwise_not_test.cc b/test/kernels/bitwise_not_test.cc
index 02b700b..309be7e 100644
--- a/test/kernels/bitwise_not_test.cc
+++ b/test/kernels/bitwise_not_test.cc
@@ -30,7 +30,7 @@ KERNEL_TEST_CASE("bitwise_not SSE2") { return kernel_bitwise_not_test<CPUType::S
 template INTGEMM_AVX2 void kernel_bitwise_not_test<CPUType::AVX2>();
 KERNEL_TEST_CASE("bitwise_not AVX2") { return kernel_bitwise_not_test<CPUType::AVX2>(); }
 
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 template INTGEMM_AVX512BW void kernel_bitwise_not_test<CPUType::AVX512BW>();
 KERNEL_TEST_CASE("bitwise_not AVX512BW") { return kernel_bitwise_not_test<CPUType::AVX512BW>(); }
 #endif
diff --git a/test/kernels/downcast_test.cc b/test/kernels/downcast_test.cc
index 5ecc084..d3261c7 100644
--- a/test/kernels/downcast_test.cc
+++ b/test/kernels/downcast_test.cc
@@ -32,7 +32,7 @@ KERNEL_TEST_CASE("downcast32to8 SSE2") { return kernel_downcast32to8_test<CPUTyp
 template INTGEMM_AVX2 void kernel_downcast32to8_test<CPUType::AVX2>();
 KERNEL_TEST_CASE("downcast32to8 AVX2") { return kernel_downcast32to8_test<CPUType::AVX2>(); }
 
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 template INTGEMM_AVX512BW void kernel_downcast32to8_test<CPUType::AVX512BW>();
 KERNEL_TEST_CASE("downcast32to8 AVX512BW") { return kernel_downcast32to8_test<CPUType::AVX512BW>(); }
 #endif
@@ -62,7 +62,7 @@ KERNEL_TEST_CASE("downcast32to16 SSE2") { return kernel_downcast32to16_test<CPUT
 template INTGEMM_AVX2 void kernel_downcast32to16_test<CPUType::AVX2>();
 KERNEL_TEST_CASE("downcast32to16 AVX2") { return kernel_downcast32to16_test<CPUType::AVX2>(); }
 
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 template INTGEMM_AVX512BW void kernel_downcast32to16_test<CPUType::AVX512BW>();
 KERNEL_TEST_CASE("downcast32to16 AVX512BW") { return kernel_downcast32to16_test<CPUType::AVX512BW>(); }
 #endif
@@ -92,7 +92,7 @@ KERNEL_TEST_CASE("downcast16to8 SSE2") { return kernel_downcast16to8_test<CPUTyp
 template INTGEMM_AVX2 void kernel_downcast16to8_test<CPUType::AVX2>();
 KERNEL_TEST_CASE("downcast16to8 AVX2") { return kernel_downcast16to8_test<CPUType::AVX2>(); }
 
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 template INTGEMM_AVX512BW void kernel_downcast16to8_test<CPUType::AVX512BW>();
 KERNEL_TEST_CASE("downcast16to8 AVX512BW") { return kernel_downcast16to8_test<CPUType::AVX512BW>(); }
 #endif
diff --git a/test/kernels/exp_test.cc b/test/kernels/exp_test.cc
index d54f2ca..cf85562 100644
--- a/test/kernels/exp_test.cc
+++ b/test/kernels/exp_test.cc
@@ -27,7 +27,7 @@ void kernel_exp_approx_taylor_test() {
 template INTGEMM_AVX2 void kernel_exp_approx_taylor_test<CPUType::AVX2>();
 KERNEL_TEST_CASE("exp_approx_taylor AVX2") { return kernel_exp_approx_taylor_test<CPUType::AVX2>(); }
 
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 template INTGEMM_AVX512BW void kernel_exp_approx_taylor_test<CPUType::AVX512BW>();
 KERNEL_TEST_CASE("exp_approx_taylor AVX512BW") { return kernel_exp_approx_taylor_test<CPUType::AVX512BW>(); }
 #endif
diff --git a/test/kernels/floor_test.cc b/test/kernels/floor_test.cc
index 10914a3..365d16d 100644
--- a/test/kernels/floor_test.cc
+++ b/test/kernels/floor_test.cc
@@ -30,7 +30,7 @@ KERNEL_TEST_CASE("floor SSE2") { return kernel_floor_test<CPUType::SSE2>(); }
 template INTGEMM_AVX2 void kernel_floor_test<CPUType::AVX2>();
 KERNEL_TEST_CASE("floor AVX2") { return kernel_floor_test<CPUType::AVX2>(); }
 
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 template INTGEMM_AVX512BW void kernel_floor_test<CPUType::AVX512BW>();
 KERNEL_TEST_CASE("floor AVX512BW") { return kernel_floor_test<CPUType::AVX512BW>(); }
 #endif
diff --git a/test/kernels/multiply_sat_test.cc b/test/kernels/multiply_sat_test.cc
index 87fc09a..36a9b9f 100644
--- a/test/kernels/multiply_sat_test.cc
+++ b/test/kernels/multiply_sat_test.cc
@@ -41,7 +41,7 @@ template INTGEMM_AVX2 void kernel_multiply_sat_test<CPUType::AVX2, int16_t>();
 KERNEL_TEST_CASE("multiply_sat/int8 AVX2") { return kernel_multiply_sat_test<CPUType::AVX2, int8_t>(); }
 KERNEL_TEST_CASE("multiply_sat/int16 AVX2") { return kernel_multiply_sat_test<CPUType::AVX2, int16_t>(); }
 
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 template INTGEMM_AVX512BW void kernel_multiply_sat_test<CPUType::AVX512BW, int8_t>();
 template INTGEMM_AVX512BW void kernel_multiply_sat_test<CPUType::AVX512BW, int16_t>();
 KERNEL_TEST_CASE("multiply_sat/int8 AVX512BW") { return kernel_multiply_sat_test<CPUType::AVX512BW, int8_t>(); }
diff --git a/test/kernels/multiply_test.cc b/test/kernels/multiply_test.cc
index 0eea965..30f1640 100644
--- a/test/kernels/multiply_test.cc
+++ b/test/kernels/multiply_test.cc
@@ -48,7 +48,7 @@ KERNEL_TEST_CASE("multiply/int AVX2") { return kernel_multiply_test<CPUType::AVX
 KERNEL_TEST_CASE("multiply/float AVX2") { return kernel_multiply_test<CPUType::AVX2, float>(); }
 KERNEL_TEST_CASE("multiply/double AVX2") { return kernel_multiply_test<CPUType::AVX2, double>(); }
 
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 template INTGEMM_AVX512BW void kernel_multiply_test<CPUType::AVX512BW, int8_t>();
 template INTGEMM_AVX512BW void kernel_multiply_test<CPUType::AVX512BW, int16_t>();
 template INTGEMM_AVX512BW void kernel_multiply_test<CPUType::AVX512BW, int>();
diff --git a/test/kernels/quantize_test.cc b/test/kernels/quantize_test.cc
index c9eae0a..07bfe21 100644
--- a/test/kernels/quantize_test.cc
+++ b/test/kernels/quantize_test.cc
@@ -31,7 +31,7 @@ KERNEL_TEST_CASE("quantize SSE2") { return kernel_quantize_test<CPUType::SSE2>()
 template INTGEMM_AVX2 void kernel_quantize_test<CPUType::AVX2>();
 KERNEL_TEST_CASE("quantize AVX2") { return kernel_quantize_test<CPUType::AVX2>(); }
 
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 template INTGEMM_AVX512BW void kernel_quantize_test<CPUType::AVX512BW>();
 KERNEL_TEST_CASE("quantize AVX512BW") { return kernel_quantize_test<CPUType::AVX512BW>(); }
 #endif
diff --git a/test/kernels/relu_test.cc b/test/kernels/relu_test.cc
index 25a212a..c291dea 100644
--- a/test/kernels/relu_test.cc
+++ b/test/kernels/relu_test.cc
@@ -46,7 +46,7 @@ KERNEL_TEST_CASE("relu/int AVX2") { return kernel_relu_test<CPUType::AVX2, int>(
 KERNEL_TEST_CASE("relu/float AVX2") { return kernel_relu_test<CPUType::AVX2, float>(); }
 KERNEL_TEST_CASE("relu/double AVX2") { return kernel_relu_test<CPUType::AVX2, double>(); }
 
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 template INTGEMM_AVX512BW void kernel_relu_test<CPUType::AVX512BW, int8_t>();
 template INTGEMM_AVX512BW void kernel_relu_test<CPUType::AVX512BW, int16_t>();
 template INTGEMM_AVX512BW void kernel_relu_test<CPUType::AVX512BW, int>();
diff --git a/test/kernels/rescale_test.cc b/test/kernels/rescale_test.cc
index d380c8d..2f79d39 100644
--- a/test/kernels/rescale_test.cc
+++ b/test/kernels/rescale_test.cc
@@ -32,7 +32,7 @@ KERNEL_TEST_CASE("rescale SSE2") { return kernel_rescale_test<CPUType::SSE2>();
 template INTGEMM_AVX2 void kernel_rescale_test<CPUType::AVX2>();
 KERNEL_TEST_CASE("rescale AVX2") { return kernel_rescale_test<CPUType::AVX2>(); }
 
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 template INTGEMM_AVX512BW void kernel_rescale_test<CPUType::AVX512BW>();
 KERNEL_TEST_CASE("rescale AVX512BW") { return kernel_rescale_test<CPUType::AVX512BW>(); }
 #endif
diff --git a/test/kernels/sigmoid_test.cc b/test/kernels/sigmoid_test.cc
index e4743e2..a8b0b3c 100644
--- a/test/kernels/sigmoid_test.cc
+++ b/test/kernels/sigmoid_test.cc
@@ -34,7 +34,7 @@ void kernel_sigmoid_test() {
 template INTGEMM_AVX2 void kernel_sigmoid_test<CPUType::AVX2>();
 KERNEL_TEST_CASE("sigmoid AVX2") { return kernel_sigmoid_test<CPUType::AVX2>(); }
 
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 template INTGEMM_AVX512BW void kernel_sigmoid_test<CPUType::AVX512BW>();
 KERNEL_TEST_CASE("sigmoid AVX512BW") { return kernel_sigmoid_test<CPUType::AVX512BW>(); }
 #endif
diff --git a/test/kernels/tanh_test.cc b/test/kernels/tanh_test.cc
index 737ac9b..2d688ea 100644
--- a/test/kernels/tanh_test.cc
+++ b/test/kernels/tanh_test.cc
@@ -27,7 +27,7 @@ void kernel_tanh_test() {
 template INTGEMM_AVX2 void kernel_tanh_test<CPUType::AVX2>();
 KERNEL_TEST_CASE("tanh AVX2") { return kernel_tanh_test<CPUType::AVX2>(); }
 
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 template INTGEMM_AVX512BW void kernel_tanh_test<CPUType::AVX512BW>();
 KERNEL_TEST_CASE("tanh AVX512BW") { return kernel_tanh_test<CPUType::AVX512BW>(); }
 #endif
diff --git a/test/kernels/unquantize_test.cc b/test/kernels/unquantize_test.cc
index 6f40da6..20c3d6a 100644
--- a/test/kernels/unquantize_test.cc
+++ b/test/kernels/unquantize_test.cc
@@ -31,7 +31,7 @@ KERNEL_TEST_CASE("unquantize SSE2") { return kernel_unquantize_test<CPUType::SSE
 template INTGEMM_AVX2 void kernel_unquantize_test<CPUType::AVX2>();
 KERNEL_TEST_CASE("unquantize AVX2") { return kernel_unquantize_test<CPUType::AVX2>(); }
 
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 template INTGEMM_AVX512BW void kernel_unquantize_test<CPUType::AVX512BW>();
 KERNEL_TEST_CASE("unquantize AVX512BW") { return kernel_unquantize_test<CPUType::AVX512BW>(); }
 #endif
diff --git a/test/kernels/upcast_test.cc b/test/kernels/upcast_test.cc
index df6a62e..497c734 100644
--- a/test/kernels/upcast_test.cc
+++ b/test/kernels/upcast_test.cc
@@ -33,7 +33,7 @@ KERNEL_TEST_CASE("upcast8to16 SSE2") { return kernel_upcast8to16_test<CPUType::S
 template INTGEMM_AVX2 void kernel_upcast8to16_test<CPUType::AVX2>();
 KERNEL_TEST_CASE("upcast8to16 AVX2") { return kernel_upcast8to16_test<CPUType::AVX2>(); }
 
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 template INTGEMM_AVX512BW void kernel_upcast8to16_test<CPUType::AVX512BW>();
 KERNEL_TEST_CASE("upcast8to16 AVX512BW") { return kernel_upcast8to16_test<CPUType::AVX512BW>(); }
 #endif
@@ -65,7 +65,7 @@ KERNEL_TEST_CASE("upcast16to32 SSE2") { return kernel_upcast16to32_test<CPUType:
 template INTGEMM_AVX2 void kernel_upcast16to32_test<CPUType::AVX2>();
 KERNEL_TEST_CASE("upcast16to32 AVX2") { return kernel_upcast16to32_test<CPUType::AVX2>(); }
 
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 template INTGEMM_AVX512BW void kernel_upcast16to32_test<CPUType::AVX512BW>();
 KERNEL_TEST_CASE("upcast16to32 AVX512BW") { return kernel_upcast16to32_test<CPUType::AVX512BW>(); }
 #endif
@@ -99,7 +99,7 @@ KERNEL_TEST_CASE("upcast8to32 SSE2") { return kernel_upcast8to32_test<CPUType::S
 template INTGEMM_AVX2 void kernel_upcast8to32_test<CPUType::AVX2>();
 KERNEL_TEST_CASE("upcast8to32 AVX2") { return kernel_upcast8to32_test<CPUType::AVX2>(); }
 
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 template INTGEMM_AVX512BW void kernel_upcast8to32_test<CPUType::AVX512BW>();
 KERNEL_TEST_CASE("upcast8to32 AVX512BW") { return kernel_upcast8to32_test<CPUType::AVX512BW>(); }
 #endif
diff --git a/test/kernels/write_test.cc b/test/kernels/write_test.cc
index aeaafcb..f2e3d35 100644
--- a/test/kernels/write_test.cc
+++ b/test/kernels/write_test.cc
@@ -46,7 +46,7 @@ KERNEL_TEST_CASE("write/int AVX2") { return kernel_write_test<CPUType::AVX2, int
 KERNEL_TEST_CASE("write/float AVX2") { return kernel_write_test<CPUType::AVX2, float>(); }
 KERNEL_TEST_CASE("write/double AVX2") { return kernel_write_test<CPUType::AVX2, double>(); }
 
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 template INTGEMM_AVX512BW void kernel_write_test<CPUType::AVX512BW, int8_t>();
 template INTGEMM_AVX512BW void kernel_write_test<CPUType::AVX512BW, int16_t>();
 template INTGEMM_AVX512BW void kernel_write_test<CPUType::AVX512BW, int>();
diff --git a/test/multiply_test.cc b/test/multiply_test.cc
index 7316bb0..260dd76 100644
--- a/test/multiply_test.cc
+++ b/test/multiply_test.cc
@@ -82,7 +82,7 @@ template <class Routine> void TestPrepare(Index rows = 32, Index cols = 16) {
 
 TEST_CASE("Prepare AVX512", "[prepare]") {
   if (kCPU < CPUType::AVX512BW) return;
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
 	TestPrepare<AVX512_8bit>(64, 8);
 	TestPrepare<AVX512_8bit>(256, 32);
     TestPrepare<AVX512_16bit>(64, 8);
@@ -147,7 +147,7 @@ template <class Routine> void TestSelectColumnsB(Index rows = 64, Index cols = 1
 
 TEST_CASE("SelectColumnsB AVX512", "[select]") {
   if (kCPU < CPUType::AVX512BW) return;
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
     TestSelectColumnsB<AVX512_8bit>();
     TestSelectColumnsB<AVX512_16bit>(256, 256);
 #endif
@@ -223,7 +223,7 @@ TEST_CASE("MaxAbsolute AVX2", "[max]") {
 
 TEST_CASE("MaxAbsolute AVX512F", "[max]") {
   if (kCPU < CPUType::AVX512BW) return;
-  #ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+  #ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
   TestMaxAbsolute<avx512f::MaxAbsolute>();
   #endif
 }
@@ -434,7 +434,7 @@ TEST_CASE ("Multiply AVX2 16bit with bias", "[biased_multiply]") {
   TestMultiplyBias<AVX2_16bit>(200, 256, 256, .1, 1, 0.01);
 }
 
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
   TEST_CASE ("Multiply AVX512 8bit", "[multiply]") {
     if (kCPU < CPUType::AVX512BW) return;
     TestMultiply<AVX512_8bit>(8, 256, 256, 0, 0.25, 0.062);
diff --git a/test/prepare_b_quantized_transposed.cc b/test/prepare_b_quantized_transposed.cc
index 68c0a29..3a5faaf 100644
--- a/test/prepare_b_quantized_transposed.cc
+++ b/test/prepare_b_quantized_transposed.cc
@@ -80,7 +80,7 @@ TEST_CASE("PrepareBQuantizedTransposed AVX2", "") {
   CHECK(TestMany<AVX2_16bit>(32, 128));
 }
 
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
   TEST_CASE("PrepareBQuantizedTransposed AVX512", "") {
     if (kCPU < CPUType::AVX512BW)
       return;
diff --git a/test/prepare_b_transposed.cc b/test/prepare_b_transposed.cc
index 219e56a..1a4ed88 100644
--- a/test/prepare_b_transposed.cc
+++ b/test/prepare_b_transposed.cc
@@ -81,7 +81,7 @@ TEST_CASE("PrepareBTransposed AVX2", "") {
   CHECK(TestMany<AVX2_16bit>(8, 128, 2.0f));
 }
 
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
   TEST_CASE("PrepareBTransposed AVX512", "") {
     if (kCPU < CPUType::AVX512BW)
       return;
diff --git a/test/quantize_test.cc b/test/quantize_test.cc
index 3263812..ee27261 100644
--- a/test/quantize_test.cc
+++ b/test/quantize_test.cc
@@ -86,7 +86,7 @@ TEST_CASE ("Quantize AVX2", "[quantize]") {
   TestMany<AVX2_8bit>(1);
   TestMany<AVX2_16bit>(16);
 }
-#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
+#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
   TEST_CASE ("Quantize AVX512", "[quantize]") {
     if (kCPU < CPUType::AVX512BW) return;
     TestMany<AVX512_8bit>(1);