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

Moving definitions to use dedicated scratch register.

Browse Source
This commit is contained in:
ed
2026-08-18 09:26:49 -04:00
parent da007d342e
commit 86fe189b4e
5 changed files with 108 additions and 67 deletions
Download Patch File Download Diff File Expand all files Collapse all files
code/duffle/gte.atom.c
+2 -2
View File
@@ -275,7 +275,7 @@ typedef Struct_(RegUse_build_normalize_v3s4) {
Reg scratch; /* scratchpad base; loaded via load_word_imm below. */
Reg src_ptr;
Reg dst_ptr;
Reg recip_est; /* |v|² sum + shift-input + sqrtbl[index] */
Reg recip_est; /* |v|² sum + shift-input + sqrtbl[index] */
Reg norm; Reg shift;
Reg src_x;
union { Reg mac1_scratch, dst_offset; } t3;
@@ -372,7 +372,7 @@ atom_info(atom_bind(Binds_gte_cross_v3s4)) MipsAtom_Proc_(aa, {
load_word(r.y.src_a, R_TapePtr, O_(Binds_gte_cross_v3s4,src_a)),
load_word(r.z.src_b, R_TapePtr, O_(Binds_gte_cross_v3s4,src_b)),
load_word(r.x.out, R_TapePtr, O_(Binds_gte_cross_v3s4,out)),
LdSlot_ add_ui_self(R_TapePtr, S_(Binds_gte_cross_v3s4)),
LdSlot_ add_ui_self( R_TapePtr, S_(Binds_gte_cross_v3s4)),
mac_load_v3s4(r.a, r.y.src_a, 0), LdSlot_
mac_load_v3s4(r.b, r.z.src_b, 0), LdSlot_
code/duffle/lottes_tape.h
+4 -4
View File
@@ -97,8 +97,8 @@ enum {
R_Atom5 = R_T5,
R_Atom6 = R_T6,
R_Atom7 = R_T7,
R_Atom8 = R_T8,
R_Atom9 = R_T9,
R_Atom8 = R_T8,
R_Atom9 = R_T9,
R_Atom10 = R_V0, // Tend to be used with gte DMAs
R_Atom11 = R_V1, // Tend to be used with gte DMAs
R_Atom12 = R_A0,
@@ -383,8 +383,8 @@ FI_ Reg regfile__alloc_helper(A2_U2 file, Reg r_id) {
* any of them while R0, R1, R26-R31 remain reserved. */
I_ Reg regfile_alloc(RegFile_R rf) {
Reg allocated = 0;
for index_iter(U4, idx, 0, <, Array_len(regfile_alloc_order)) {
allocated = regfile__alloc_helper(rf->GPR, idx);
for index_iter(U4, r_id, R_V0, <, R_T9) {
allocated = regfile__alloc_helper(rf->GPR, r_id);
Jmp_nZero_(allocated,resolved);
}
assert(allocated != 0);
code/hello_camera/hello_camera.atom.c
+32 -45
View File
@@ -111,66 +111,59 @@ typedef AtomBundle_(resolve_look_at) { MipsAtom
*pop_mv_trans;
};
enum {
// TODO(Ed): We can resolve scratch at anytime its fixed to a specific address.
R_ResolveScratch = R_T4 atom_reg atom_type(U4*),
#define R_ResolveScratch_Code R_T4_Code
typedef Struct_(ResolveLookAtScratch) {
V3_S4 fwd;
V3_S4 uz;
V3_S4 right;
V3_S4 ux;
V3_S4 up;
V3_S4 uy;
P3_S4 eye;
P3_S4 target;
V3_S4 up_in;
};
/* Binds_ResolveLookAtSub — what the C side pushes onto the tape before input_and_sub.
* The scratchpad base is no longer pushed because R_ScratchBase (= R_SP) is a tape carrier
* preserved across atoms; the atom body reads 0x1F800000 directly from R_SP. */
typedef Struct_(Binds_ResolveLookAt) {
MT3_S2S4* look_at;
P3_S4* eye;
P3_S4* target;
V3_S4* up_in;
};
/* ─── ResolveLookAtScratch — offset schema for the resolve_look_at bundle's */
typedef Struct_(ResolveLookAtScratch) {
V3_S4 fwd; /* offset +0 (16 bytes — 4 S4 fields incl. internal pad) */
V3_S4 uz; /* offset +16 (16 bytes) */
V3_S4 right; /* offset +32 (16 bytes) */
V3_S4 ux; /* offset +48 (16 bytes) */
V3_S4 up; /* offset +64 (16 bytes) */
V3_S4 uy; /* offset +80 (16 bytes) */
P3_S4 eye; /* offset +96 (16 bytes; storage alias of V3_S4) */
P3_S4 target; /* offset +112 (16 bytes; storage alias of V3_S4) */
V3_S4 up_in; /* offset +128 (16 bytes) */
};
typedef Struct_(Binds_ResolveLookAtSub) {
P3_S4* target;
P3_S4* eye;
P3_S4* eye;
V3_S4* up_in;
ResolveLookAtScratch* scratchpad;
};
typedef Struct_(RegUse_resolve_look_at_input_and_sub) {
Reg scratch;
Reg target; Reg eye; Reg up_in;
Reg t0; Reg t1; Reg t2; Reg t3; Reg t4;
};
/* Atom 0 in the bundle: input_and_sub. Stages C-side inputs into the scratchpad and computes fwd = target - eye. */
internal MipsAtom* AtomBundleEntry_(resolve_look_at,input_and_sub)(AtomArena_R aa, RegUse_resolve_look_at_input_and_sub r)
internal MipsAtom* AtomBundleEntry_(resolve_look_at,input_and_sub)(AtomArena_R aa, RegUse_resolve_look_at_input_and_sub r)
atom_info(atom_bind(Binds_ResolveLookAtSub)) MipsAtom_Proc_(aa, {
load_word(r.target, R_TapePtr, O_(Binds_ResolveLookAtSub,target)),
load_word(r.eye, R_TapePtr, O_(Binds_ResolveLookAtSub,eye)),
load_word(r.up_in, R_TapePtr, O_(Binds_ResolveLookAtSub,up_in)),
load_word(r.scratch, R_TapePtr, O_(Binds_ResolveLookAtSub,scratchpad)),
LdSlot_ add_ui_self( R_TapePtr, S_(Binds_ResolveLookAtSub)),
load_word(r.target, R_TapePtr, O_(Binds_ResolveLookAtSub,target)),
load_word(r.eye, R_TapePtr, O_(Binds_ResolveLookAtSub,eye)),
load_word(r.up_in, R_TapePtr, O_(Binds_ResolveLookAtSub,up_in)),
LdSlot_ add_ui_self(R_TapePtr, S_(Binds_ResolveLookAtSub)),
/* Stage up_in.x/y/z into the scratchpad. */
/* Stage up_in.x/y/z into the scratchpad. R_ScratchBase = R_SP = 0x1F800000. */
mac_load_word_v3( r.t0, r.t1, r.t2, r.up_in, 0), LdSlot_
mac_store_word_v3(r.t0, r.t1, r.t2, r.scratch, O_(ResolveLookAtScratch,up_in)),
mac_store_word_v3(r.t0, r.t1, r.t2, R_ScratchBase, O_(ResolveLookAtScratch,up_in)),
// Stage eye.x/y/z into the scratchpad (atom 6 reads these for the translation column).
mac_load_word_v3( r.t0, r.t1, r.t2, r.eye, 0), LdSlot_
mac_store_word_v3(r.t0, r.t1, r.t2, r.scratch, O_(ResolveLookAtScratch,eye)),
mac_store_word_v3(r.t0, r.t1, r.t2, R_ScratchBase, O_(ResolveLookAtScratch,eye)),
/* Compute fwd = target - eye. */
// mac_load_p3s4(t3, R_AT, t4, r.eye, 0),
mac_load_word_v3(r.t3, R_AT, r.t4, r.target, 0), LdSlot_
mac_load_word_v3(r.t3, R_AT, r.t4, r.target, 0), LdSlot_
mac_sub_s_v3_self(
r.t3, R_AT, r.t4,
r.t0, r.t1, r.t2),
mac_store_word_v3(r.t3, R_AT, r.t4, r.scratch, O_(ResolveLookAtScratch,fwd)),
mac_store_word_v3(r.t3, R_AT, r.t4, R_ScratchBase, O_(ResolveLookAtScratch,fwd)),
mac_yield()
})
@@ -180,8 +173,6 @@ typedef Struct_(Binds_ResolveLookAtPopMvTrans) {
U4 look_at; /* MT3_S2S4* — destination matrix address */
};
typedef Struct_(RegUse_resolve_look_at__pop_mv_trans) {
Reg scratch; /* loaded via load_word_imm below — can't rely on
* R_T4 surviving across the tape_run boundary */
Reg look_at;
Reg_(V3_S4) row; /* populate phase: load ux/uy/uz */
union { Reg ux, v_x; } t6; /* populate addr (canonical) → matrix_vector v_x */
@@ -236,17 +227,13 @@ internal MipsAtom* resolve_look_at__pop_mv_trans(AtomArena_R aa,
load_word(r.look_at, R_TapePtr, O_(Binds_ResolveLookAtPopMvTrans,look_at)),
LdSlot_ add_ui_self(R_TapePtr, S_(Binds_ResolveLookAtPopMvTrans)),
/* Load scratch base via immediate (Scratchpad_Loc = 0x1F800000). We can't rely on
* R_T4 (= R_ResolveScratch) surviving across the tape_run boundary — the compiler
* treats it as clobberable per the tape_run asm_clobber list. Baking the scratch
* address via load_word_imm is robust. */
mac_load_word_imm(r.scratch, Scratchpad_Loc),
/* --- Scratch addresses for ux/uy/uz/eye (populate phase; t6/t7/t8 alias ux/uy/uz) --- */
add_si(r.t6.ux, r.scratch, O_(ResolveLookAtScratch, ux)), LdSlot_
add_si(r.t7.uy, r.scratch, O_(ResolveLookAtScratch, uy)),
add_si(r.t8.uz, r.scratch, O_(ResolveLookAtScratch, uz)),
add_si(r.eye, r.scratch, O_(ResolveLookAtScratch, eye)),
/* --- Scratch addresses for ux/uy/uz/eye (populate phase; t6/t7/t8 alias ux/uy/uz).
* R_ScratchBase (= R_SP) holds 0x1F800000; no per-atom bake is required because
* R_SP is a tape carrier preserved across atoms. --- */
add_si(r.t6.ux, R_ScratchBase, O_(ResolveLookAtScratch, ux)), LdSlot_
add_si(r.t7.uy, R_ScratchBase, O_(ResolveLookAtScratch, uy)),
add_si(r.t8.uz, R_ScratchBase, O_(ResolveLookAtScratch, uz)),
add_si(r.eye, R_ScratchBase, O_(ResolveLookAtScratch, eye)),
/* --- POPULATE phase: write look_at->m[][] from ux/uy/uz as packed S2 --- */
mac_load_v3s4(r.row, r.t6.ux, 0), LdSlot_ mac_store_v3s2(r.row, r.look_at, O_(MT3_S2S4, m[0])),
code/hello_camera/hello_camera.c
+15 -16
View File
@@ -151,16 +151,17 @@ internal void compile_init_atoms(void) {
smem.normalize_v3s4 = build_normalize_v3s4(& ab,
RegUse_(build_normalize_v3s4) {
.scratch = ralloc(),
.src_ptr = ralloc(),
.dst_ptr = ralloc(),
.recip_est = ralloc(),
.norm = ralloc(),
.shift = ralloc(),
.src_x = ralloc(),
.t3 = ralloc(),
.t4 = ralloc(),
.t5 = ralloc(),
.scratch = ralloc(),
.src_ptr = ralloc(),
.dst_ptr = ralloc(),
.recip_est = ralloc(),
.norm = ralloc(),
.shift = ralloc(),
.src_x = ralloc(),
// .shift_count = ralloc(), /* dedicated slot for stage-3 → stage-4 shift count */
.t3 = ralloc(),
.t4 = ralloc(),
.t5 = ralloc(),
});
regfile_reset(& rf);
@@ -173,14 +174,14 @@ internal void compile_resolve_look_at(void) {
AtomArena ab = atomarena_make(slice_ut_arr(smem.resolve_look_at_mem));
AtomBundle_resolve_look_at_R bundle = C_(void*, smem.resolve_look_at_bundle);
U4 pin_mask = regfile_abi_mask | (1 << R_ResolveScratch);
RegFile rf = regfile(pin_mask);
/* R_ScratchBase (= R_SP) is a tape carrier preserved across atoms; no carrier
* pin is needed in the regfile. The standard 24-register pool is sufficient. */
RegFile rf = regfile(regfile_abi_mask);
#define ralloc() regfile_alloc(& rf)
#define ralloc_v3() { ralloc(), ralloc(), ralloc() }
bundle->input_and_sub = AtomBundleEntry_(resolve_look_at, input_and_sub)(& ab,
RegUse_(resolve_look_at_input_and_sub) {
.scratch = R_ResolveScratch,
.target = ralloc(),
.eye = ralloc(),
.up_in = ralloc(),
@@ -190,7 +191,7 @@ internal void compile_resolve_look_at(void) {
.t3 = ralloc(),
.t4 = ralloc(),
});
regfile_reset_to_mask(& rf, pin_mask);
regfile_reset(& rf);
bundle->normalize_fwd_uz = smem.normalize_v3s4;
bundle->cross_to_right = smem.gte_cross_v3s4;
@@ -200,7 +201,6 @@ internal void compile_resolve_look_at(void) {
bundle->pop_mv_trans = resolve_look_at__pop_mv_trans(& ab,
RegUse_(resolve_look_at__pop_mv_trans){
.scratch = R_ResolveScratch,
.look_at = ralloc(),
.eye = ralloc(),
.row = ralloc_v3(),
@@ -229,7 +229,6 @@ I_ void resolve_look_at(TapeBuilder_R tb
tb_data(tb, u4_(target));
tb_data(tb, u4_(eye));
tb_data(tb, u4_(up_in));
tb_data(tb, u4_(smem.scratchpad));
}
tb_emit(tb, bundle->normalize_fwd_uz); {
tb_data(tb, u4_(O_(ResolveLookAtScratch, fwd) | (O_(ResolveLookAtScratch, uz) << 16)));