ed/pikuma_ps1
1
0
Fork 0
mirror of https://github.com/Ed94/pikuma_ps1.git synced 2026-08-25 02:20:33 +00:00
Code Issues Packages Projects Releases Wiki Activity

reducing cross product atom procs to a single one in gte for once.

Browse Source
This commit is contained in:
ed
2026-08-17 01:05:58 -04:00
parent 18b1d5a04b
commit e79c364b40
8 changed files with 128 additions and 186 deletions
Download Patch File Download Diff File Expand all files Collapse all files
code/hello_camera/hello_camera.atom.c
+3 -138
View File
@@ -104,9 +104,9 @@ MipsAtomComp_Proc_(ab, {
typedef AtomBundle_(resolve_look_at) { MipsAtom*
input_and_sub,
normalize_fwd_uz,
cross_uz_up_into_right,
cross_to_right,
normalize_right_ux,
cross_uz_ux_to_up,
cross_to_up,
normalize_up_uy,
populate,
set_gte_mt3s2s4,
@@ -178,141 +178,6 @@ atom_info(atom_bind(Binds_ResolveLookAtSub)) MipsAtom_Proc_(aa, {
mac_yield()
})
typedef Struct_(RegUse_resolve_look_at_cross_uz_up_into_right) {
Reg scratch;
Reg a; Reg b; Reg c; /* load a.x/y/z; result out.x/y/z */
Reg d; /* load b.x */
Reg f; /* r_f = &right (out ptr), r_g = &uz, r_h = &up_in */
union { Reg t1, g, target0; };
union { Reg t2, h, target1; };
Reg t0;
};
/* Atom 2: cross uz × up_in → right. */
// internal MipsAtom* AtomBundleEntry_(resolve_look_at, cross_uz_up_to_right)(AtomArena_R aa,
internal MipsAtom* resolve_look_at_cross_uz_up_into_right(AtomArena_R aa,
RegUse_resolve_look_at_cross_uz_up_into_right r
) MipsAtom_Proc_(aa, {
/* FIX: build packed RT22+RT33 with proper sign extension. */
add_si(r.g, r.scratch, O_(ResolveLookAtScratch,uz)), /* r_g = &uz */
add_si(r.h, r.scratch, O_(ResolveLookAtScratch,up_in)), /* r_h = &up_in */
add_si(r.f, r.scratch, O_(ResolveLookAtScratch,right)), /* r_f = &right (out) */
nop,
/* Load a (uz).x/y/z into r_a/r_b/r_c. */
mac_load_word_v3(r.a, r.b, r.c, r.g, 0), LdSlot_
/* Load b (up_in).x/y/z into r_d + R_AT/R_V0 (R_AT/R_V0 are hardcoded scratch). */
mac_load_word_v3(r.d, R_AT, r.t0, r.h, 0), LdSlot_ // (taken by gte_mv_from_ctrl_r)
/* Save the two RT control-register slots OP will clobber. We reuse r_g/r_h (scratch pointers, no longer needed) as the save targets. */
gte_mv_from_ctrl_r(r.target0, gte_cr_RT11), /* r_g = C2 r0 (RT11|RT12) */
gte_mv_from_ctrl_r(r.target1, gte_cr_RT22), /* r_h = C2 r4 (RT22|RT33) */
/* Load uz.x/uz.y/uz.z into COP2 control registers.
* OP reads D1 = RT11 from $0.low, D2 = RT22 from $2.high, D3 = RT33 from $4.high.
* RT22 is in BOTH $2.high AND $4.low (shared bit position). OP reads from $2.high.
* So set RT22 via ctc2 r_b, $2 (sets $2.high = a.y.high = RT22, $2.low = a.y.low = RT13).
* Then set RT33 via ctc2 r_c, $4 (sets $4.high = a.z.high = RT33, $4.low = a.z.low).
* The $2 and $4 writes don't clobber each other (separate registers).
* The 2nd ctc2 DOES clobber $4.low (becomes a.z.low, NOT a.y.high), but since OP
* reads RT22 from $2.high (which the 2nd ctc2 doesn't touch), D2 is still a.y.high.
* This is libpsyx's OuterProduct12 convention. */
gte_mv_to_ctrl_r(r.b, gte_cr_RT13), /* $2 = r_b = a.y. RT13=a.y.low, RT22=a.y.high. */
gte_mv_to_ctrl_r(r.c, gte_cr_RT22), /* $4 = r_c = a.z. RT22=a.z.low, RT33=a.z.high. */
/* Load uz into the RT diagonal. */
gte_mv_to_ctrl_r(r.a, gte_cr_RT11), /* D1 = RT11 = uz.x. */
GteDelay_ nop2, /* CTC2 retirement */
/* Load up_in into IR (the second operand for OP). */
gte_mv_to_data_r(r.d, C2_IR1), /* IR1 = up_in.x */
gte_mv_to_data_r(R_AT, C2_IR2), /* IR2 = up_in.y */
gte_mv_to_data_r(r.t0, C2_IR3), /* IR3 = up_in.z */
GteDelay_ nop2, /* MTC2 retirement */
gte_cmdw_cross, /* OP: MAC1/2/3 = uz × up_in
* MAC1 = IR3*D2 - IR2*D3 = up_in.z*uz.y.high - up_in.y*uz.z.high
* MAC2 = IR1*D3 - IR3*D1 = up_in.x*uz.z.high - up_in.z*uz.x
* MAC3 = IR2*D1 - IR1*D2 = up_in.y*uz.x - up_in.x*uz.y.high
* For up_in = (0, -fp_one, 0):
* MAC1 = 0 - (-fp_one)*uz.z.high = fp_one*uz.z.high
* MAC2 = 0 - 0 = 0
* MAC3 = (-fp_one)*uz.x - 0 = -fp_one*uz.x */
/* Restore the RT slots we clobbered. */
gte_mv_to_ctrl_r(r.target0, gte_cr_RT11), /* restore C2 r0 (RT11|RT12) */
gte_mv_to_ctrl_r(r.target1, gte_cr_RT22), /* restore C2 r4 (RT22|RT33) */
/* mfc2 MAC1/2/3 → r_a/r_b/r_c (out.x/y/z). */
mac_gte_mv_from_data_r_mac123(r.a, r.b, r.c),
GteDelay_ nop, /* MFC2 retirement */
/* Right-shift MAC by 12 to convert from GTE's S12.20 fixed-point scale back to libpsyx OuterProduct12 convention (S12.0, fp_one=4096=1<<12).
* Without this, MAC values (~16M for unit-vector cross products) overflow the GTE's 16-bit IR registers when atom 3 normalizes via mtc2. */
mac_shift_aright_v3_self(r.a, r.b, r.c, 12),
/* Store out.x/y/z to r_f (out ptr = scratch+32). */
mac_store_word_v3(r.a, r.b, r.c, r.f, 0),
mac_yield()
})
typedef Struct_(RegUse_resolve_look_at__cross_uz_ux_to_up_proc) {
Reg const scratch; /* pinned T4 */
Reg_(V3_S4) a; /* uz components, then MAC / out */
Reg_(V3_S4) b; /* ux components */
union { Reg t0, up; }; /* &up, dedicated */
union { Reg t1, uz, rt11; }; /* &uz, then RT11 save */
union { Reg t2, ux, rt22; }; /* &ux, then RT22 save */
};
/* Atom 4: cross uz × ux → up. */
internal MipsAtom* resolve_look_at__cross_uz_ux_to_up_proc(AtomArena_R aa,
RegUse_resolve_look_at__cross_uz_ux_to_up_proc r
) MipsAtom_Proc_(aa, {
add_si(r.uz, r.scratch, O_(ResolveLookAtScratch,uz)), /* r.uz = &uz */
add_si(r.ux, r.scratch, O_(ResolveLookAtScratch,ux)), /* r.ux = &ux */
add_si(r.up, r.scratch, O_(ResolveLookAtScratch,up)), /* r.up = &up (out) */
mac_load_v3s4(r.a, r.uz, 0), LdSlot_
mac_load_v3s4(r.b, r.ux, 0), LdSlot_ /* taken by gte_mv_from_ctrl_r */
/* OP reads D1/D2/D3 from RT11/RT22/RT33 ($0/$2/$4), not V0/V1/V2.
* Mirror atom 1: cfc2 RT save, ctc2 RT diagonal from uz, mtc2 IR from ux, ctc2 RT restore. */
/* Save the two RT control-register slots OP will clobber (reusing r.uz/r.ux — they're no longer needed as scratch pointers). */
gte_mv_from_ctrl_r(r.rt11, gte_cr_RT11), /* r.rt11 = C2 $0 (RT11|RT12) */
gte_mv_from_ctrl_r(r.rt22, gte_cr_RT22), /* r.rt22 = C2 $4 (RT22|RT33) */
/* Load uz into the RT diagonal — same packing as atom 1.
* OP reads D1 = RT11 from $0.low, D2 = RT22 from $2.high, D3 = RT33 from $4.high.
* RT22 is shared between $2.high and $4.low — the ctc2 sequence to $2 then $4
* sets RT22 to uz.y.high (via $2), then to uz.z.low (via $4).
* OP reads RT22 from $2.high which the second ctc2 doesn't touch, so D2 stays uz.y.high. */
gte_mv_to_ctrl_r(r.a.y, gte_cr_RT13), /* $2 = uz.y. RT13=uz.y.low, RT22=uz.y.high. */
gte_mv_to_ctrl_r(r.a.z, gte_cr_RT22), /* $4 = uz.z. RT22=uz.z.low, RT33=uz.z.high. */
gte_mv_to_ctrl_r(r.a.x, gte_cr_RT11), /* $0 = uz.x. RT11=uz.x. */
GteDelay_ nop2, /* CTC2 retirement (CPU→COP2 2-slot delay) */
/* Load ux into the IR registers (the second operand for OP). */
gte_mv_to_data_r(r.b.x, C2_IR1), /* IR1 = ux.x */
gte_mv_to_data_r(r.b.y, C2_IR2), /* IR2 = ux.y */
gte_mv_to_data_r(r.b.z, C2_IR3), /* IR3 = ux.z */
GteDelay_ nop2, /* MTC2 retirement (CPU→COP2 2-slot delay) */
gte_cmdw_cross,
/* Restore the RT slots we clobbered. */
gte_mv_to_ctrl_r(r.rt11, gte_cr_RT11), /* restore C2 $0 (RT11|RT12) */
gte_mv_to_ctrl_r(r.rt22, gte_cr_RT22), /* restore C2 $4 (RT22|RT33) */
mac_gte_mv_from_data_r_mac123(r.a.x, r.a.y, r.a.z),
GteDelay_ nop,
/* Right-shift MAC by 12 to convert from GTE's S12.20 scale back to libpsyx
* OuterProduct12 convention (S12.0, fp_one=4096). See atom 1 for rationale. */
mac_shift_aright_v3_self(r.a.x, r.a.y, r.a.z, 12),
mac_store_v3s4(r.a, r.up, 0),
mac_yield()
})
typedef Struct_(Binds_ResolveLookAtPopAndTrans) {
U4 look_at; /* U4 (MT3_S2S4* — destination matrix address) */
@@ -407,7 +272,7 @@ internal MipsAtom* resolve_look_at__matrix_vector_proc(AtomArena_R aa,
*
* Pool cost: r_look_at (1) + r_scratch (carrier) + r_off_ptr + 1 clobber = 4 GPRs.
*/
I_ MipsAtom* resolve_look_at__trans_matrix_proc(AtomArena_R aa
I_ MipsAtom* AtomBundleEntry_(resolve_look_at,trans_matrix)(AtomArena_R aa
, U4 r_look_at, U4 r_scratch, U4 r_off_ptr
, U4 r_tmp0, U4 r_tmp1, U4 r_tmp2
) MipsAtom_Proc_(aa, {
code/hello_camera/hello_camera.c
+41 -29
View File
@@ -186,21 +186,16 @@ internal void compile_resolve_look_at(void) {
});
regfile_reset_to_mask(& rf, pin_mask);
/* === ATOM 2: cross uz×up_in→right === */
// smem.resolve_look_at_bundle[2] = AtomBundleEntry_(resolve_look_at,cross_uz_up_to_right)(& ab,
smem.resolve_look_at_bundle[2] = resolve_look_at_cross_uz_up_into_right(& ab,
RegUse_(resolve_look_at_cross_uz_up_into_right) {
.scratch = R_ResolveScratch,
.a = ralloc(),
.b = ralloc(),
.c = ralloc(),
.d = ralloc(),
.f = ralloc(),
/* === ATOM 2: cross uz×up_in→right (Binds_gte_cross_v3s4) === */
smem.resolve_look_at_bundle[2] = gte_cross_v3s4(& ab,
RegUse_(gte_cross_v3s4) {
.a = ralloc_v3(),
.b = ralloc_v3(),
.t0 = ralloc(),
.t1 = ralloc(),
.t2 = ralloc(),
.t0 = ralloc(),
});
regfile_reset_to_mask(& rf, pin_mask);
regfile_reset_to_mask(& rf, pin_mask);
/* === ATOM 3: normalize right→ux === */
src_offset = O_(ResolveLookAtScratch, right);
@@ -220,17 +215,16 @@ internal void compile_resolve_look_at(void) {
});
regfile_reset_to_mask(& rf, pin_mask);
/* === ATOM 4: cross uz×ux→up === */
smem.resolve_look_at_bundle[4] = resolve_look_at__cross_uz_ux_to_up_proc(& ab,
RegUse_(resolve_look_at__cross_uz_ux_to_up_proc){
.scratch = R_ResolveScratch,
.a = ralloc_v3(), /* T0 T1 T2 */
.b = ralloc_v3(), /* T3 T5 T6 */
.t0 = ralloc(), /* T7 = up */
.t1 = ralloc(), /* V0 = uz / rt11 */
.t2 = ralloc(), /* V1 = ux / rt22 */
/* === ATOM 4: cross uz×ux→up (Binds_gte_cross_v3s4) === */
smem.resolve_look_at_bundle[4] = gte_cross_v3s4(& ab,
RegUse_(gte_cross_v3s4) {
.a = ralloc_v3(),
.b = ralloc_v3(),
.t0 = ralloc(),
.t1 = ralloc(),
.t2 = ralloc(),
});
regfile_reset_to_mask(& rf, pin_mask);
regfile_reset_to_mask(& rf, pin_mask);
/* === ATOM 5: normalize up→uy === */
src_offset = O_(ResolveLookAtScratch, up);
@@ -280,8 +274,13 @@ internal void compile_resolve_look_at(void) {
U4 r_scratch_6c = R_ResolveScratch;
U4 r_off_ptr_6c = R_T1; /* &scratch.eye (= off dst) */
U4 r_tmp0_6c = R_T2;
smem.resolve_look_at_bundle[9] = resolve_look_at__trans_matrix_proc(& ab,
r_look_at_6c, r_scratch_6c, r_off_ptr_6c, r_tmp0_6c, R_T3, R_T4);
smem.resolve_look_at_bundle[9] = AtomBundleEntry_(resolve_look_at,trans_matrix)(& ab,
r_look_at_6c,
r_scratch_6c,
r_off_ptr_6c,
r_tmp0_6c,
R_T3,
R_T4);
/* Sanity check: arena didn't overflow. */
assert(ab.used <= ResolveLookAtArena_Size);
@@ -295,6 +294,9 @@ I_ void resolve_look_at(TapeBuilder_R tb
, P3_S4* target
, V3_S4* up_in
){
/* Typed view of the scratchpad for field-address arithmetic. */
ResolveLookAtScratch* sp = C_scratch(ResolveLookAtScratch*);
tb_emit(tb, smem.resolve_look_at_bundle[0]); {
tb_data(tb, u4_(target));
tb_data(tb, u4_(eye));
@@ -302,15 +304,25 @@ I_ void resolve_look_at(TapeBuilder_R tb
tb_data(tb, u4_(smem.scratchpad));
}
tb_emit(tb, smem.resolve_look_at_bundle[1]); {
tb_emit(tb, smem.resolve_look_at_bundle[1]); {
// tb_data(tb, u4_(Scratchpad_Loc));
}
tb_emit(tb, smem.resolve_look_at_bundle[2]); { }
tb_emit(tb, smem.resolve_look_at_bundle[3]); {
tb_emit(tb, smem.resolve_look_at_bundle[2]); {
/* Binds_gte_cross_v3s4: src_a, src_b, out (3 pointers). */
tb_data(tb, u4_(& sp->uz)); /* src_a */
tb_data(tb, u4_(& sp->up_in)); /* src_b */
tb_data(tb, u4_(& sp->right)); /* out */
}
tb_emit(tb, smem.resolve_look_at_bundle[3]); {
// tb_data(tb, u4_(Scratchpad_Loc));
}
tb_emit(tb, smem.resolve_look_at_bundle[4]); { }
tb_emit(tb, smem.resolve_look_at_bundle[5]); {
tb_emit(tb, smem.resolve_look_at_bundle[4]); {
/* Binds_gte_cross_v3s4: src_a, src_b, out (3 pointers). */
tb_data(tb, u4_(& sp->uz)); /* src_a */
tb_data(tb, u4_(& sp->ux)); /* src_b */
tb_data(tb, u4_(& sp->up)); /* out */
}
tb_emit(tb, smem.resolve_look_at_bundle[5]); {
// tb_data(tb, u4_(Scratchpad_Loc));
}