Remove GEMMLOWP_ALLOW_SLOW_SCALAR_FALLBACK

Fixes #117

README.md

@@ -54,13 +54,10 @@ Current optimized code paths:
 *   ARM with NEON (both 32bit and 64bit).
 *   Intel x86 with SSE 4.1 (both 32bit and 64bit).
 
-If you are building for x86, it's important that you pass in the `-msse4.1`
-compiler flag when building, or you'll end up using slow reference code. If
-you're building with Bazel, you can do this by running `bazel build gemmlowp:all
---copt=-msse4.1`. If you're building for a machine with no SIMD support in
-gemmlowp then by default you'll see an error. If you want to run with the
-reference implementations anyway, you can override the error by specifying
-`GEMMLOWP_ALLOW_SLOW_SCALAR_FALLBACK` as a build define.
+When building for x86, it's important to pass `-msse4.1` or `-march=native` to
+the compiler, otherwise gemmlowp will use slow reference code. Bazel users can
+compile by running `bazel build --config=opt //gemmlowp:all` or alternatively
+pass the `--copt=-msse4.1` flag.
 
 Details of what it takes to make an efficient port of gemmlowp, namely writing a
 suitable GEMM kernel and accompanying packing code, are explained in this file:

contrib/CMakeLists.txt

@@ -12,11 +12,6 @@ set(CMAKE_CXX_STANDARD 11)
 
 get_filename_component(gemmlowp_src ${gemmlowp_SOURCE_DIR} PATH)
 
-# Enabling SIMD is recommended for modern x86 machines
-#add_definitions(-msse4)
-# However here we take the best compatibility
-add_definitions(-DGEMMLOWP_ALLOW_SLOW_SCALAR_FALLBACK)
-
 if(WIN32)
   # one can enable simd from the cmake command line, ie -DCMAKE_CXX_FLAGS="/arch:AVX2
   add_definitions(-DNOMINMAX -DWIN64 -DWIN32_LEAN_AND_MEAN -DNOGDI)

eight_bit_int_gemm/eight_bit_int_gemm.cc

@@ -12,9 +12,6 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#ifndef GEMMLOWP_ALLOW_SLOW_SCALAR_FALLBACK
-#define GEMMLOWP_ALLOW_SLOW_SCALAR_FALLBACK
-#endif
 #include "eight_bit_int_gemm.h"
 
 #include <memory>

internal/kernel_default.h

@@ -79,23 +79,6 @@ GEMMLOWP_SET_DEFAULT_KERNEL(false, false, SSE4_32_Kernel4x4Depth2)
 #include "kernel_sse.h"
 GEMMLOWP_SET_DEFAULT_KERNEL(false, false, SSE4_64_Kernel12x4Depth2)
 #else
-#ifndef GEMMLOWP_ALLOW_SLOW_SCALAR_FALLBACK
-#if defined __ARM_ARCH_5TE__
-// SIMD is not available on this platform. The slow fallback will be used.
-// Don't require GEMMLOWP_ALLOW_SLOW_SCALAR_FALLBACK because there's nothing
-// the user can do about it.
-#elif defined __powerpc__
-// There is currently no fast kernel using SIMD instructions on POWER. Don't
-// require GEMMLOWP_ALLOW_SLOW_SCALAR_FALLBACK because there's nothing the user
-// can do about it.
-#else
-#error \
-    "SIMD not enabled, you'd be getting a slow software fallback. Consider \
-enabling SIMD extensions (for example using -msse4 if you're on modern x86). \
-If that's not an option, and you would like to continue with the \
-slow fallback, define GEMMLOWP_ALLOW_SLOW_SCALAR_FALLBACK."
-#endif
-#endif
 #include "kernel_reference.h"
 namespace gemmlowp {
 typedef ReferenceKernel<KernelFormat<