mirror of
https://github.com/libretro/parallel-ps2.git
synced 2026-11-10 11:21:36 -10:00
LibretroAdmin
c0c82fec27
Seven ways the GS gets values in and out of a vector, each written both ways and counted: loading and storing a packed pair, building from four ints or a splat, reading one lane, writing one lane, and the rect round trip that reads several lanes by name, does scalar arithmetic and writes them back. Six compile identically on both compilers. The round trip is four instructions shorter through the union under clang, which vectorises it where the class's members stay scalar -- the shape GSDirtyRect uses, and so far the only one where the difference is in the reading rather than the operation.
40 lines
2.4 KiB
C++
40 lines
2.4 KiB
C++
/* How a value is read and written around the operations.
|
|
*
|
|
* The operations themselves compile identically either way -- the codegen
|
|
* lane records that. What can differ is the shape around them, so these are
|
|
* the seven ways the GS actually gets values in and out of a vector, each
|
|
* written both ways and counted.
|
|
*
|
|
* The one that moves is the last: several lanes read by name, scalar
|
|
* arithmetic on each, written back. clang vectorises that through the union
|
|
* and leaves the class's members scalar.
|
|
*/
|
|
#include "GS/GSVector.h"
|
|
extern "C" {
|
|
#include "GS/gs_vector.h"
|
|
}
|
|
/* 1. read a 2-int pair from memory into a vector (GSVector2i -> vector) */
|
|
extern "C" __m128i cpp_pair(const void*p){ return GSVector4i(*(const GSVector2i*)p); }
|
|
extern "C" __m128i c89_pair(const void*p){ return gs_v4i_loadl(p); }
|
|
/* 2. write two lanes back out as a pair */
|
|
extern "C" void cpp_wpair(__m128i v, void*o){ GSVector4i::storel(o, GSVector4i(v)); }
|
|
extern "C" void c89_wpair(__m128i v, void*o){ gs_v4i_storel(o, v); }
|
|
/* 3. build a vector from four runtime ints */
|
|
extern "C" __m128i cpp_four(int a,int b,int c,int d){ return GSVector4i(a,b,c,d); }
|
|
extern "C" __m128i c89_four(int a,int b,int c,int d){ return gs_v4i_set4(a,b,c,d); }
|
|
/* 4. splat one runtime int */
|
|
extern "C" __m128i cpp_splat(int a){ return GSVector4i(a); }
|
|
extern "C" __m128i c89_splat(int a){ return gs_v4i_set32(a); }
|
|
/* 5. read one lane and use it as a scalar */
|
|
extern "C" int cpp_lane(const void*p){ GSVector4i A=GSVector4i::load<true>(p); return A.z; }
|
|
extern "C" int c89_lane(const void*p){ union gs_v4i_view w; w.v=gs_v4i_load(p); return w.lane.z; }
|
|
/* 6. write one lane then use the whole vector */
|
|
extern "C" __m128i cpp_setl(const void*p,int x){ GSVector4i A=GSVector4i::load<true>(p); A.y=x; return A; }
|
|
extern "C" __m128i c89_setl(const void*p,int x){ union gs_v4i_view w; w.v=gs_v4i_load(p); w.lane.y=x; return w.v; }
|
|
/* 7. the full rect->scalar->rect round trip GSDirtyRect does */
|
|
extern "C" __m128i cpp_rt(const void*p,int dx,int sx){ GSVector4i A=GSVector4i::load<true>(p); GSVector4i R;
|
|
R.left=(A.left*dx)/sx; R.top=(A.top*dx)/sx; R.right=(A.right*dx)/sx; R.bottom=(A.bottom*dx)/sx; return R; }
|
|
extern "C" __m128i c89_rt(const void*p,int dx,int sx){ union gs_v4i_view a,r; a.v=gs_v4i_load(p);
|
|
r.rect.left=(a.rect.left*dx)/sx; r.rect.top=(a.rect.top*dx)/sx;
|
|
r.rect.right=(a.rect.right*dx)/sx; r.rect.bottom=(a.rect.bottom*dx)/sx; return r.v; }
|