blob: d0599074d5fa4e0652a4ccd3414769549b7f8c48 [file] [log] [blame]
* Copyright 2015 Google Inc.
* Use of this source code is governed by a BSD-style license that can be
* found in the LICENSE file.
#ifndef SkOpts_DEFINED
#define SkOpts_DEFINED
#include "include/private/base/SkSpan_impl.h"
#include "src/core/SkRasterPipelineOpContexts.h"
#include "src/core/SkRasterPipelineOpList.h"
#include <cstddef>
#include <cstdint>
* SkOpts (short for SkOptimizations) is a mechanism where we can ship with multiple implementations
* of a set of functions and dynamically choose the best one at runtime (e.g. the call to
* SkGraphics::Init(), which calls SkOpts::Init()) depending on the detected CPU features. This is
* also referred to as having "specializations" of a given function.
* For example, Skia might be compiled to support CPUs that only have the sse2 instruction set
* (
* but may be run on a more modern CPU that supports AVX2 instructions.
* (
* SkOpts allow Skia to have two versions of a row-blitting function, one that uses normal C++
* code (e.g. loops, scalar integer math) and one that makes use of the AVX2 vector types and
* intrinsic functions. This function is declared here in the SkOpts namespace, and then the
* implementation (see SkOpts.cpp) is deferred to a function of the same name in the sse2::
* namespace (the minimum Skia is compiled with) using DEFINE_DEFAULT.
* All implementations of this blit function are done in a header file file in //src/opts
* (e.g. //src/opts/SkBlitRow_opts.h). ifdefs guard each of the implementations, such that only
* one implementation is possible for a given SK_CPU_SSE_LEVEL. This header will be compiled
* *multiple* times with a different SK_CPU_SSE_LEVEL each compilation.
* Each CPU instruction set that we want specializations for has a .cpp file in //src/opts which
* defines an Init() function that replaces the function pointers in the SkOpts namespace with the
* ones from the specialized namespace (e.g. hsw::). These .cpp files don't implement the
* specializations, they just refer to the specialization created in the header files (e.g.
* SkBlitRow_opts.h).
* At compile time:
* - SkOpts.cpp is compiled with the minimum CPU level (e.g. SSE2). Because this
* file includes all the headers in //src/opts/, those headers add "the default implementation"
* of all their functions to the SK_OPTS_NS namespace (e.g. sse2::blit_row_color32).
* - Each of the specialized .cpp files in //src/opts/ are compiled with their respective
* compiler flags. Because the specialized .cpp file includes the headers that implement the
* functions using intrinsics or other CPU-specific code, those specialized functions end up
* in the specialized namespace, e.g. (hsw::blit_row_color32).
* At link time, the default implementations and all specializations of all SkOpts functions are
* included in the resulting library/binary file.
* At runtime, SkOpts::Init() will run the appropriate Init functions that the current CPU level
* supports specializations for (e.g. Init_hsw, Init_ssse3). Note multiple Init functions can
* be called as CPU instruction sets are typically super sets of older instruction sets
struct SkRasterPipelineStage;
namespace SkOpts {
// Call to replace pointers to portable functions with pointers to CPU-specific functions.
// Thread-safe and idempotent.
// Called by SkGraphics::Init().
void Init();
// We can't necessarily express the type of SkRasterPipeline stage functions here,
// so we just use this void(*)(void) as a stand-in.
using StageFn = void(*)(void);
extern StageFn ops_highp[kNumRasterPipelineHighpOps], just_return_highp;
extern StageFn ops_lowp [kNumRasterPipelineLowpOps ], just_return_lowp;
extern void (*start_pipeline_highp)(size_t,size_t,size_t,size_t, SkRasterPipelineStage*,
extern void (*start_pipeline_lowp )(size_t,size_t,size_t,size_t, SkRasterPipelineStage*,
extern size_t raster_pipeline_lowp_stride;
extern size_t raster_pipeline_highp_stride;
} // namespace SkOpts
#endif // SkOpts_DEFINED