void
fd3_program_emit(struct fd_ringbuffer *ring, struct fd3_emit *emit,
int nr, struct pipe_surface **bufs)
{
const struct ir3_shader_variant *vp, *fp;
const struct ir3_info *vsi, *fsi;
enum a3xx_instrbuffermode fpbuffer, vpbuffer;
uint32_t fpbuffersz, vpbuffersz, fsoff;
uint32_t pos_regid, posz_regid, psize_regid, color_regid[4] = {0};
int constmode;
int i, j, k;
debug_assert(nr <= ARRAY_SIZE(color_regid));
vp = fd3_emit_get_vp(emit);
if (emit->key.binning_pass) {
/* use dummy stateobj to simplify binning vs non-binning: */
static const struct ir3_shader_variant binning_fp = {};
fp = &binning_fp;
} else {
fp = fd3_emit_get_fp(emit);
}
vsi = &vp->info;
fsi = &fp->info;
fpbuffer = BUFFER;
vpbuffer = BUFFER;
fpbuffersz = fp->instrlen;
vpbuffersz = vp->instrlen;
/*
* Decide whether to use BUFFER or CACHE mode for VS and FS. It
* appears like 256 is the hard limit, but when the combined size
* exceeds 128 then blob will try to keep FS in BUFFER mode and
* switch to CACHE for VS until VS is too large. The blob seems
* to switch FS out of BUFFER mode at slightly under 128. But
* a bit fuzzy on the decision tree, so use slightly conservative
* limits.
*
* TODO check if these thresholds for BUFFER vs CACHE mode are the
* same for all a3xx or whether we need to consider the gpuid
*/
if ((fpbuffersz + vpbuffersz) > 128) {
if (fpbuffersz < 112) {
/* FP:BUFFER VP:CACHE */
vpbuffer = CACHE;
vpbuffersz = 256 - fpbuffersz;
} else if (vpbuffersz < 112) {
/* FP:CACHE VP:BUFFER */
fpbuffer = CACHE;
fpbuffersz = 256 - vpbuffersz;
} else {
/* FP:CACHE VP:CACHE */
vpbuffer = fpbuffer = CACHE;
vpbuffersz = fpbuffersz = 192;
}
}
if (fpbuffer == BUFFER) {
fsoff = 128 - fpbuffersz;
} else {
fsoff = 256 - fpbuffersz;
}
/* seems like vs->constlen + fs->constlen > 256, then CONSTMODE=1 */
constmode = ((vp->constlen + fp->constlen) > 256) ? 1 : 0;
pos_regid = ir3_find_output_regid(vp,
ir3_semantic_name(TGSI_SEMANTIC_POSITION, 0));
posz_regid = ir3_find_output_regid(fp,
ir3_semantic_name(TGSI_SEMANTIC_POSITION, 0));
psize_regid = ir3_find_output_regid(vp,
ir3_semantic_name(TGSI_SEMANTIC_PSIZE, 0));
if (fp->color0_mrt) {
color_regid[0] = color_regid[1] = color_regid[2] = color_regid[3] =
ir3_find_output_regid(fp, ir3_semantic_name(TGSI_SEMANTIC_COLOR, 0));
} else {
for (i = 0; i < fp->outputs_count; i++) {
ir3_semantic sem = fp->outputs[i].semantic;
unsigned idx = sem2idx(sem);
if (sem2name(sem) != TGSI_SEMANTIC_COLOR)
continue;
debug_assert(idx < ARRAY_SIZE(color_regid));
color_regid[idx] = fp->outputs[i].regid;
}
}
/* adjust regids for alpha output formats. there is no alpha render
* format, so it's just treated like red
*/
for (i = 0; i < nr; i++)
if (util_format_is_alpha(pipe_surface_format(bufs[i])))
color_regid[i] += 3;
/* we could probably divide this up into things that need to be
* emitted if frag-prog is dirty vs if vert-prog is dirty..
*/
OUT_PKT0(ring, REG_A3XX_HLSQ_CONTROL_0_REG, 6);
OUT_RING(ring, A3XX_HLSQ_CONTROL_0_REG_FSTHREADSIZE(FOUR_QUADS) |
A3XX_HLSQ_CONTROL_0_REG_CONSTMODE(constmode) |
/* NOTE: I guess SHADERRESTART and CONSTFULLUPDATE maybe
* flush some caches? I think we only need to set those
* bits if we have updated const or shader..
*/
A3XX_HLSQ_CONTROL_0_REG_SPSHADERRESTART |
A3XX_HLSQ_CONTROL_0_REG_SPCONSTFULLUPDATE);
OUT_RING(ring, A3XX_HLSQ_CONTROL_1_REG_VSTHREADSIZE(TWO_QUADS) |
A3XX_HLSQ_CONTROL_1_REG_VSSUPERTHREADENABLE |
COND(fp->frag_coord, A3XX_HLSQ_CONTROL_1_REG_ZWCOORD));
OUT_RING(ring, A3XX_HLSQ_CONTROL_2_REG_PRIMALLOCTHRESHOLD(31));
OUT_RING(ring, A3XX_HLSQ_CONTROL_3_REG_REGID(fp->pos_regid));
OUT_RING(ring, A3XX_HLSQ_VS_CONTROL_REG_CONSTLENGTH(vp->constlen) |
A3XX_HLSQ_VS_CONTROL_REG_CONSTSTARTOFFSET(0) |
A3XX_HLSQ_VS_CONTROL_REG_INSTRLENGTH(vpbuffersz));
OUT_RING(ring, A3XX_HLSQ_FS_CONTROL_REG_CONSTLENGTH(fp->constlen) |
A3XX_HLSQ_FS_CONTROL_REG_CONSTSTARTOFFSET(128) |
A3XX_HLSQ_FS_CONTROL_REG_INSTRLENGTH(fpbuffersz));
OUT_PKT0(ring, REG_A3XX_SP_SP_CTRL_REG, 1);
OUT_RING(ring, A3XX_SP_SP_CTRL_REG_CONSTMODE(constmode) |
COND(emit->key.binning_pass, A3XX_SP_SP_CTRL_REG_BINNING) |
A3XX_SP_SP_CTRL_REG_SLEEPMODE(1) |
A3XX_SP_SP_CTRL_REG_L0MODE(0));
OUT_PKT0(ring, REG_A3XX_SP_VS_LENGTH_REG, 1);
OUT_RING(ring, A3XX_SP_VS_LENGTH_REG_SHADERLENGTH(vp->instrlen));
OUT_PKT0(ring, REG_A3XX_SP_VS_CTRL_REG0, 3);
OUT_RING(ring, A3XX_SP_VS_CTRL_REG0_THREADMODE(MULTI) |
A3XX_SP_VS_CTRL_REG0_INSTRBUFFERMODE(vpbuffer) |
COND(vpbuffer == CACHE, A3XX_SP_VS_CTRL_REG0_CACHEINVALID) |
A3XX_SP_VS_CTRL_REG0_HALFREGFOOTPRINT(vsi->max_half_reg + 1) |
A3XX_SP_VS_CTRL_REG0_FULLREGFOOTPRINT(vsi->max_reg + 1) |
A3XX_SP_VS_CTRL_REG0_INOUTREGOVERLAP(0) |
A3XX_SP_VS_CTRL_REG0_THREADSIZE(TWO_QUADS) |
A3XX_SP_VS_CTRL_REG0_SUPERTHREADMODE |
COND(vp->has_samp, A3XX_SP_VS_CTRL_REG0_PIXLODENABLE) |
A3XX_SP_VS_CTRL_REG0_LENGTH(vpbuffersz));
OUT_RING(ring, A3XX_SP_VS_CTRL_REG1_CONSTLENGTH(vp->constlen) |
A3XX_SP_VS_CTRL_REG1_INITIALOUTSTANDING(vp->total_in) |
A3XX_SP_VS_CTRL_REG1_CONSTFOOTPRINT(MAX2(vp->constlen + 1, 0)));
OUT_RING(ring, A3XX_SP_VS_PARAM_REG_POSREGID(pos_regid) |
A3XX_SP_VS_PARAM_REG_PSIZEREGID(psize_regid) |
A3XX_SP_VS_PARAM_REG_TOTALVSOUTVAR(align(fp->total_in, 4) / 4));
for (i = 0, j = -1; (i < 8) && (j < (int)fp->inputs_count); i++) {
uint32_t reg = 0;
OUT_PKT0(ring, REG_A3XX_SP_VS_OUT_REG(i), 1);
j = ir3_next_varying(fp, j);
if (j < fp->inputs_count) {
k = ir3_find_output(vp, fp->inputs[j].semantic);
reg |= A3XX_SP_VS_OUT_REG_A_REGID(vp->outputs[k].regid);
reg |= A3XX_SP_VS_OUT_REG_A_COMPMASK(fp->inputs[j].compmask);
}
j = ir3_next_varying(fp, j);
if (j < fp->inputs_count) {
k = ir3_find_output(vp, fp->inputs[j].semantic);
reg |= A3XX_SP_VS_OUT_REG_B_REGID(vp->outputs[k].regid);
reg |= A3XX_SP_VS_OUT_REG_B_COMPMASK(fp->inputs[j].compmask);
}
OUT_RING(ring, reg);
}
for (i = 0, j = -1; (i < 4) && (j < (int)fp->inputs_count); i++) {
uint32_t reg = 0;
OUT_PKT0(ring, REG_A3XX_SP_VS_VPC_DST_REG(i), 1);
j = ir3_next_varying(fp, j);
if (j < fp->inputs_count)
reg |= A3XX_SP_VS_VPC_DST_REG_OUTLOC0(fp->inputs[j].inloc);
j = ir3_next_varying(fp, j);
if (j < fp->inputs_count)
reg |= A3XX_SP_VS_VPC_DST_REG_OUTLOC1(fp->inputs[j].inloc);
j = ir3_next_varying(fp, j);
if (j < fp->inputs_count)
reg |= A3XX_SP_VS_VPC_DST_REG_OUTLOC2(fp->inputs[j].inloc);
j = ir3_next_varying(fp, j);
if (j < fp->inputs_count)
reg |= A3XX_SP_VS_VPC_DST_REG_OUTLOC3(fp->inputs[j].inloc);
OUT_RING(ring, reg);
}
OUT_PKT0(ring, REG_A3XX_SP_VS_OBJ_OFFSET_REG, 2);
OUT_RING(ring, A3XX_SP_VS_OBJ_OFFSET_REG_CONSTOBJECTOFFSET(0) |
A3XX_SP_VS_OBJ_OFFSET_REG_SHADEROBJOFFSET(0));
OUT_RELOC(ring, vp->bo, 0, 0, 0); /* SP_VS_OBJ_START_REG */
if (emit->key.binning_pass) {
OUT_PKT0(ring, REG_A3XX_SP_FS_LENGTH_REG, 1);
OUT_RING(ring, 0x00000000);
OUT_PKT0(ring, REG_A3XX_SP_FS_CTRL_REG0, 2);
OUT_RING(ring, A3XX_SP_FS_CTRL_REG0_THREADMODE(MULTI) |
A3XX_SP_FS_CTRL_REG0_INSTRBUFFERMODE(BUFFER));
OUT_RING(ring, 0x00000000);
OUT_PKT0(ring, REG_A3XX_SP_FS_OBJ_OFFSET_REG, 1);
OUT_RING(ring, A3XX_SP_FS_OBJ_OFFSET_REG_CONSTOBJECTOFFSET(128) |
A3XX_SP_FS_OBJ_OFFSET_REG_SHADEROBJOFFSET(0));
} else {
OUT_PKT0(ring, REG_A3XX_SP_FS_LENGTH_REG, 1);
OUT_RING(ring, A3XX_SP_FS_LENGTH_REG_SHADERLENGTH(fp->instrlen));
OUT_PKT0(ring, REG_A3XX_SP_FS_CTRL_REG0, 2);
OUT_RING(ring, A3XX_SP_FS_CTRL_REG0_THREADMODE(MULTI) |
A3XX_SP_FS_CTRL_REG0_INSTRBUFFERMODE(fpbuffer) |
COND(fpbuffer == CACHE, A3XX_SP_FS_CTRL_REG0_CACHEINVALID) |
A3XX_SP_FS_CTRL_REG0_HALFREGFOOTPRINT(fsi->max_half_reg + 1) |
A3XX_SP_FS_CTRL_REG0_FULLREGFOOTPRINT(fsi->max_reg + 1) |
A3XX_SP_FS_CTRL_REG0_INOUTREGOVERLAP(1) |
A3XX_SP_FS_CTRL_REG0_THREADSIZE(FOUR_QUADS) |
A3XX_SP_FS_CTRL_REG0_SUPERTHREADMODE |
COND(fp->has_samp > 0, A3XX_SP_FS_CTRL_REG0_PIXLODENABLE) |
A3XX_SP_FS_CTRL_REG0_LENGTH(fpbuffersz));
OUT_RING(ring, A3XX_SP_FS_CTRL_REG1_CONSTLENGTH(fp->constlen) |
A3XX_SP_FS_CTRL_REG1_INITIALOUTSTANDING(fp->total_in) |
A3XX_SP_FS_CTRL_REG1_CONSTFOOTPRINT(MAX2(fp->constlen + 1, 0)) |
A3XX_SP_FS_CTRL_REG1_HALFPRECVAROFFSET(63));
OUT_PKT0(ring, REG_A3XX_SP_FS_OBJ_OFFSET_REG, 2);
OUT_RING(ring, A3XX_SP_FS_OBJ_OFFSET_REG_CONSTOBJECTOFFSET(
MAX2(128, vp->constlen)) |
A3XX_SP_FS_OBJ_OFFSET_REG_SHADEROBJOFFSET(fsoff));
OUT_RELOC(ring, fp->bo, 0, 0, 0); /* SP_FS_OBJ_START_REG */
}
OUT_PKT0(ring, REG_A3XX_SP_FS_OUTPUT_REG, 1);
OUT_RING(ring,
COND(fp->writes_pos, A3XX_SP_FS_OUTPUT_REG_DEPTH_ENABLE) |
A3XX_SP_FS_OUTPUT_REG_DEPTH_REGID(posz_regid) |
A3XX_SP_FS_OUTPUT_REG_MRT(MAX2(1, nr) - 1));
OUT_PKT0(ring, REG_A3XX_SP_FS_MRT_REG(0), 4);
for (i = 0; i < 4; i++) {
uint32_t mrt_reg = A3XX_SP_FS_MRT_REG_REGID(color_regid[i]) |
COND(fp->key.half_precision, A3XX_SP_FS_MRT_REG_HALF_PRECISION);
if (i < nr) {
enum pipe_format fmt = pipe_surface_format(bufs[i]);
mrt_reg |= COND(util_format_is_pure_uint(fmt), A3XX_SP_FS_MRT_REG_UINT) |
COND(util_format_is_pure_sint(fmt), A3XX_SP_FS_MRT_REG_SINT);
}
OUT_RING(ring, mrt_reg);
}
if (emit->key.binning_pass) {
OUT_PKT0(ring, REG_A3XX_VPC_ATTR, 2);
OUT_RING(ring, A3XX_VPC_ATTR_THRDASSIGN(1) |
A3XX_VPC_ATTR_LMSIZE(1) |
COND(vp->writes_psize, A3XX_VPC_ATTR_PSIZE));
OUT_RING(ring, 0x00000000);
} else {
uint32_t vinterp[4], flatshade[2], vpsrepl[4];
memset(vinterp, 0, sizeof(vinterp));
memset(flatshade, 0, sizeof(flatshade));
memset(vpsrepl, 0, sizeof(vpsrepl));
/* figure out VARYING_INTERP / FLAT_SHAD register values: */
for (j = -1; (j = ir3_next_varying(fp, j)) < (int)fp->inputs_count; ) {
uint32_t interp = fp->inputs[j].interpolate;
/* TODO might be cleaner to just +8 in SP_VS_VPC_DST_REG
* instead.. rather than -8 everywhere else..
*/
uint32_t inloc = fp->inputs[j].inloc - 8;
/* currently assuming varyings aligned to 4 (not
* packed):
*/
debug_assert((inloc % 4) == 0);
if ((interp == TGSI_INTERPOLATE_CONSTANT) ||
((interp == TGSI_INTERPOLATE_COLOR) && emit->rasterflat)) {
uint32_t loc = inloc;
for (i = 0; i < 4; i++, loc++) {
vinterp[loc / 16] |= FLAT << ((loc % 16) * 2);
flatshade[loc / 32] |= 1 << (loc % 32);
}
}
/* Replace the .xy coordinates with S/T from the point sprite. Set
* interpolation bits for .zw such that they become .01
*/
if (emit->sprite_coord_enable & (1 << sem2idx(fp->inputs[j].semantic))) {
vpsrepl[inloc / 16] |= (emit->sprite_coord_mode ? 0x0d : 0x09)
<< ((inloc % 16) * 2);
vinterp[(inloc + 2) / 16] |= 2 << (((inloc + 2) % 16) * 2);
vinterp[(inloc + 3) / 16] |= 3 << (((inloc + 3) % 16) * 2);
}
}
OUT_PKT0(ring, REG_A3XX_VPC_ATTR, 2);
OUT_RING(ring, A3XX_VPC_ATTR_TOTALATTR(fp->total_in) |
A3XX_VPC_ATTR_THRDASSIGN(1) |
A3XX_VPC_ATTR_LMSIZE(1) |
COND(vp->writes_psize, A3XX_VPC_ATTR_PSIZE));
OUT_RING(ring, A3XX_VPC_PACK_NUMFPNONPOSVAR(fp->total_in) |
A3XX_VPC_PACK_NUMNONPOSVSVAR(fp->total_in));
OUT_PKT0(ring, REG_A3XX_VPC_VARYING_INTERP_MODE(0), 4);
OUT_RING(ring, vinterp[0]); /* VPC_VARYING_INTERP[0].MODE */
OUT_RING(ring, vinterp[1]); /* VPC_VARYING_INTERP[1].MODE */
OUT_RING(ring, vinterp[2]); /* VPC_VARYING_INTERP[2].MODE */
OUT_RING(ring, vinterp[3]); /* VPC_VARYING_INTERP[3].MODE */
OUT_PKT0(ring, REG_A3XX_VPC_VARYING_PS_REPL_MODE(0), 4);
OUT_RING(ring, vpsrepl[0]); /* VPC_VARYING_PS_REPL[0].MODE */
OUT_RING(ring, vpsrepl[1]); /* VPC_VARYING_PS_REPL[1].MODE */
OUT_RING(ring, vpsrepl[2]); /* VPC_VARYING_PS_REPL[2].MODE */
OUT_RING(ring, vpsrepl[3]); /* VPC_VARYING_PS_REPL[3].MODE */
OUT_PKT0(ring, REG_A3XX_SP_FS_FLAT_SHAD_MODE_REG_0, 2);
OUT_RING(ring, flatshade[0]); /* SP_FS_FLAT_SHAD_MODE_REG_0 */
OUT_RING(ring, flatshade[1]); /* SP_FS_FLAT_SHAD_MODE_REG_1 */
}
if (vpbuffer == BUFFER)
emit_shader(ring, vp);
OUT_PKT0(ring, REG_A3XX_VFD_PERFCOUNTER0_SELECT, 1);
OUT_RING(ring, 0x00000000); /* VFD_PERFCOUNTER0_SELECT */
if (!emit->key.binning_pass) {
if (fpbuffer == BUFFER)
emit_shader(ring, fp);
OUT_PKT0(ring, REG_A3XX_VFD_PERFCOUNTER0_SELECT, 1);
OUT_RING(ring, 0x00000000); /* VFD_PERFCOUNTER0_SELECT */
}
}
void
fd3_program_emit(struct fd_ringbuffer *ring,
struct fd_program_stateobj *prog)
{
struct fd3_shader_stateobj *vp = prog->vp;
struct fd3_shader_stateobj *fp = prog->fp;
struct ir3_shader_info *vsi = &vp->info;
struct ir3_shader_info *fsi = &fp->info;
int i;
/* we could probably divide this up into things that need to be
* emitted if frag-prog is dirty vs if vert-prog is dirty..
*/
OUT_PKT0(ring, REG_A3XX_HLSQ_CONTROL_0_REG, 6);
OUT_RING(ring, A3XX_HLSQ_CONTROL_0_REG_FSTHREADSIZE(FOUR_QUADS) |
A3XX_HLSQ_CONTROL_0_REG_SPSHADERRESTART |
A3XX_HLSQ_CONTROL_0_REG_SPCONSTFULLUPDATE);
OUT_RING(ring, A3XX_HLSQ_CONTROL_1_REG_VSTHREADSIZE(TWO_QUADS) |
A3XX_HLSQ_CONTROL_1_REG_VSSUPERTHREADENABLE);
OUT_RING(ring, A3XX_HLSQ_CONTROL_2_REG_PRIMALLOCTHRESHOLD(31));
OUT_RING(ring, 0x00000000); /* HLSQ_CONTROL_3_REG */
OUT_RING(ring, A3XX_HLSQ_VS_CONTROL_REG_CONSTLENGTH(vp->constlen) |
A3XX_HLSQ_VS_CONTROL_REG_CONSTSTARTOFFSET(0) |
A3XX_HLSQ_VS_CONTROL_REG_INSTRLENGTH(vp->instrlen));
OUT_RING(ring, A3XX_HLSQ_FS_CONTROL_REG_CONSTLENGTH(fp->constlen) |
A3XX_HLSQ_FS_CONTROL_REG_CONSTSTARTOFFSET(128) |
A3XX_HLSQ_FS_CONTROL_REG_INSTRLENGTH(fp->instrlen));
OUT_PKT0(ring, REG_A3XX_SP_SP_CTRL_REG, 1);
OUT_RING(ring, A3XX_SP_SP_CTRL_REG_CONSTMODE(0) |
A3XX_SP_SP_CTRL_REG_SLEEPMODE(1) |
// XXX "resolve" (?) bit set on gmem->mem pass..
// COND(!uniforms, A3XX_SP_SP_CTRL_REG_RESOLVE) |
// XXX sometimes 0, sometimes 1:
A3XX_SP_SP_CTRL_REG_LOMODE(1));
/* emit unknown sequence of perfcounter disables that the blob
* emits as part of the program state..
*/
for (i = 0; i < 6; i++) {
OUT_PKT0(ring, REG_A3XX_SP_PERFCOUNTER0_SELECT, 1);
OUT_RING(ring, 0x00000000); /* SP_PERFCOUNTER4_SELECT */
OUT_PKT0(ring, REG_A3XX_SP_PERFCOUNTER4_SELECT, 1);
OUT_RING(ring, 0x00000000); /* SP_PERFCOUNTER4_SELECT */
}
OUT_PKT0(ring, REG_A3XX_SP_VS_LENGTH_REG, 1);
OUT_RING(ring, A3XX_SP_VS_LENGTH_REG_SHADERLENGTH(vp->instrlen));
OUT_PKT0(ring, REG_A3XX_SP_VS_CTRL_REG0, 3);
OUT_RING(ring, A3XX_SP_VS_CTRL_REG0_THREADMODE(MULTI) |
A3XX_SP_VS_CTRL_REG0_INSTRBUFFERMODE(BUFFER) |
A3XX_SP_VS_CTRL_REG0_HALFREGFOOTPRINT(vsi->max_half_reg + 1) |
A3XX_SP_VS_CTRL_REG0_FULLREGFOOTPRINT(vsi->max_reg + 1) |
A3XX_SP_VS_CTRL_REG0_INOUTREGOVERLAP(0) |
A3XX_SP_VS_CTRL_REG0_THREADSIZE(TWO_QUADS) |
A3XX_SP_VS_CTRL_REG0_SUPERTHREADMODE |
COND(vp->samplers_count > 0, A3XX_SP_VS_CTRL_REG0_PIXLODENABLE) |
A3XX_SP_VS_CTRL_REG0_LENGTH(vp->instrlen));
OUT_RING(ring, A3XX_SP_VS_CTRL_REG1_CONSTLENGTH(vp->constlen) |
A3XX_SP_VS_CTRL_REG1_INITIALOUTSTANDING(vp->total_in) |
A3XX_SP_VS_CTRL_REG1_CONSTFOOTPRINT(MAX2(vsi->max_const, 0)));
OUT_RING(ring, A3XX_SP_VS_PARAM_REG_POSREGID(vp->pos_regid) |
A3XX_SP_VS_PARAM_REG_PSIZEREGID(vp->psize_regid) |
A3XX_SP_VS_PARAM_REG_TOTALVSOUTVAR(vp->outputs_count));
assert(vp->outputs_count >= fp->inputs_count);
for (i = 0; i < fp->inputs_count; ) {
uint32_t reg = 0;
OUT_PKT0(ring, REG_A3XX_SP_VS_OUT_REG(i/2), 1);
reg |= A3XX_SP_VS_OUT_REG_A_REGID(vp->outputs[i].regid);
reg |= A3XX_SP_VS_OUT_REG_A_COMPMASK(fp->inputs[i].compmask);
i++;
reg |= A3XX_SP_VS_OUT_REG_B_REGID(vp->outputs[i].regid);
reg |= A3XX_SP_VS_OUT_REG_B_COMPMASK(fp->inputs[i].compmask);
i++;
OUT_RING(ring, reg);
}
for (i = 0; i < fp->inputs_count; ) {
uint32_t reg = 0;
OUT_PKT0(ring, REG_A3XX_SP_VS_VPC_DST_REG(i/4), 1);
reg |= A3XX_SP_VS_VPC_DST_REG_OUTLOC0(fp->inputs[i++].inloc);
reg |= A3XX_SP_VS_VPC_DST_REG_OUTLOC1(fp->inputs[i++].inloc);
reg |= A3XX_SP_VS_VPC_DST_REG_OUTLOC2(fp->inputs[i++].inloc);
reg |= A3XX_SP_VS_VPC_DST_REG_OUTLOC3(fp->inputs[i++].inloc);
OUT_RING(ring, reg);
}
#if 0
/* for some reason, when I write SP_{VS,FS}_OBJ_START_REG I get:
[ 666.663665] kgsl kgsl-3d0: |a3xx_err_callback| RBBM | AHB bus error | READ | addr=201 | ports=1:3
[ 666.664001] kgsl kgsl-3d0: |a3xx_err_callback| ringbuffer AHB error interrupt
[ 670.680909] kgsl kgsl-3d0: |adreno_idle| spun too long waiting for RB to idle
[ 670.681062] kgsl kgsl-3d0: |kgsl-3d0| Dump Started
[ 670.681123] kgsl kgsl-3d0: POWER: FLAGS = 00000007 | ACTIVE POWERLEVEL = 00000001
[ 670.681214] kgsl kgsl-3d0: POWER: INTERVAL TIMEOUT = 0000000A
[ 670.681367] kgsl kgsl-3d0: GRP_CLK = 325000000
[ 670.681489] kgsl kgsl-3d0: BUS CLK = 0
*/
OUT_PKT0(ring, REG_A3XX_SP_VS_OBJ_OFFSET_REG, 2);
OUT_RING(ring, A3XX_SP_VS_OBJ_OFFSET_REG_CONSTOBJECTOFFSET(0) |
A3XX_SP_VS_OBJ_OFFSET_REG_SHADEROBJOFFSET(0));
OUT_RELOC(ring, vp->bo, 0, 0); /* SP_VS_OBJ_START_REG */
#endif
OUT_PKT0(ring, REG_A3XX_SP_FS_LENGTH_REG, 1);
OUT_RING(ring, A3XX_SP_FS_LENGTH_REG_SHADERLENGTH(fp->instrlen));
OUT_PKT0(ring, REG_A3XX_SP_FS_CTRL_REG0, 2);
OUT_RING(ring, A3XX_SP_FS_CTRL_REG0_THREADMODE(MULTI) |
A3XX_SP_FS_CTRL_REG0_INSTRBUFFERMODE(BUFFER) |
A3XX_SP_FS_CTRL_REG0_HALFREGFOOTPRINT(fsi->max_half_reg + 1) |
A3XX_SP_FS_CTRL_REG0_FULLREGFOOTPRINT(fsi->max_reg + 1) |
A3XX_SP_FS_CTRL_REG0_INOUTREGOVERLAP(1) |
A3XX_SP_FS_CTRL_REG0_THREADSIZE(FOUR_QUADS) |
A3XX_SP_FS_CTRL_REG0_SUPERTHREADMODE |
COND(fp->samplers_count > 0, A3XX_SP_FS_CTRL_REG0_PIXLODENABLE) |
A3XX_SP_FS_CTRL_REG0_LENGTH(fp->instrlen));
OUT_RING(ring, A3XX_SP_FS_CTRL_REG1_CONSTLENGTH(fp->constlen) |
A3XX_SP_FS_CTRL_REG1_INITIALOUTSTANDING(fp->total_in) |
A3XX_SP_FS_CTRL_REG1_CONSTFOOTPRINT(MAX2(fsi->max_const, 0)) |
A3XX_SP_FS_CTRL_REG1_HALFPRECVAROFFSET(63));
#if 0
OUT_PKT0(ring, REG_A3XX_SP_FS_OBJ_OFFSET_REG, 2);
OUT_RING(ring, A3XX_SP_FS_OBJ_OFFSET_REG_CONSTOBJECTOFFSET(128) |
A3XX_SP_FS_OBJ_OFFSET_REG_SHADEROBJOFFSET(128 - fp->instrlen));
OUT_RELOC(ring, fp->bo, 0, 0); /* SP_FS_OBJ_START_REG */
#endif
OUT_PKT0(ring, REG_A3XX_SP_FS_FLAT_SHAD_MODE_REG_0, 2);
OUT_RING(ring, 0x00000000); /* SP_FS_FLAT_SHAD_MODE_REG_0 */
OUT_RING(ring, 0x00000000); /* SP_FS_FLAT_SHAD_MODE_REG_1 */
OUT_PKT0(ring, REG_A3XX_SP_FS_OUTPUT_REG, 1);
OUT_RING(ring, 0x00000000); /* SP_FS_OUTPUT_REG */
OUT_PKT0(ring, REG_A3XX_SP_FS_MRT_REG(0), 4);
OUT_RING(ring, A3XX_SP_FS_MRT_REG_REGID(fp->color_regid) |
COND(fp->half_precision, A3XX_SP_FS_MRT_REG_HALF_PRECISION));
OUT_RING(ring, A3XX_SP_FS_MRT_REG_REGID(0));
OUT_RING(ring, A3XX_SP_FS_MRT_REG_REGID(0));
OUT_RING(ring, A3XX_SP_FS_MRT_REG_REGID(0));
OUT_PKT0(ring, REG_A3XX_VPC_ATTR, 2);
OUT_RING(ring, A3XX_VPC_ATTR_TOTALATTR(fp->total_in) |
A3XX_VPC_ATTR_THRDASSIGN(1) |
A3XX_VPC_ATTR_LMSIZE(1));
OUT_RING(ring, A3XX_VPC_PACK_NUMFPNONPOSVAR(fp->total_in) |
A3XX_VPC_PACK_NUMNONPOSVSVAR(fp->total_in));
OUT_PKT0(ring, REG_A3XX_VPC_VARYING_INTERP_MODE(0), 4);
OUT_RING(ring, fp->vinterp[0]); /* VPC_VARYING_INTERP[0].MODE */
OUT_RING(ring, fp->vinterp[1]); /* VPC_VARYING_INTERP[1].MODE */
OUT_RING(ring, fp->vinterp[2]); /* VPC_VARYING_INTERP[2].MODE */
OUT_RING(ring, fp->vinterp[3]); /* VPC_VARYING_INTERP[3].MODE */
OUT_PKT0(ring, REG_A3XX_VPC_VARYING_PS_REPL_MODE(0), 4);
OUT_RING(ring, fp->vpsrepl[0]); /* VPC_VARYING_PS_REPL[0].MODE */
OUT_RING(ring, fp->vpsrepl[1]); /* VPC_VARYING_PS_REPL[1].MODE */
OUT_RING(ring, fp->vpsrepl[2]); /* VPC_VARYING_PS_REPL[2].MODE */
OUT_RING(ring, fp->vpsrepl[3]); /* VPC_VARYING_PS_REPL[3].MODE */
OUT_PKT0(ring, REG_A3XX_VFD_VS_THREADING_THRESHOLD, 1);
OUT_RING(ring, A3XX_VFD_VS_THREADING_THRESHOLD_REGID_THRESHOLD(15) |
A3XX_VFD_VS_THREADING_THRESHOLD_REGID_VTXCNT(252));
emit_shader(ring, vp);
OUT_PKT0(ring, REG_A3XX_VFD_PERFCOUNTER0_SELECT, 1);
OUT_RING(ring, 0x00000000); /* VFD_PERFCOUNTER0_SELECT */
emit_shader(ring, fp);
OUT_PKT0(ring, REG_A3XX_VFD_PERFCOUNTER0_SELECT, 1);
OUT_RING(ring, 0x00000000); /* VFD_PERFCOUNTER0_SELECT */
OUT_PKT0(ring, REG_A3XX_VFD_CONTROL_0, 2);
OUT_RING(ring, A3XX_VFD_CONTROL_0_TOTALATTRTOVS(vp->total_in) |
A3XX_VFD_CONTROL_0_PACKETSIZE(2) |
A3XX_VFD_CONTROL_0_STRMDECINSTRCNT(vp->inputs_count) |
A3XX_VFD_CONTROL_0_STRMFETCHINSTRCNT(vp->inputs_count));
OUT_RING(ring, A3XX_VFD_CONTROL_1_MAXSTORAGE(1) | // XXX
A3XX_VFD_CONTROL_1_REGID4VTX(regid(63,0)) |
A3XX_VFD_CONTROL_1_REGID4INST(regid(63,0)));
}
void fd_program_emit_state(struct fd_program *program, uint32_t first,
struct fd_parameters *uniforms, struct fd_parameters *attr,
struct fd_ringbuffer *ring)
{
struct fd_shader *vs = get_shader(program, FD_SHADER_VERTEX);
struct fd_shader *fs = get_shader(program, FD_SHADER_FRAGMENT);
struct ir3_shader_info *vsi = &vs->info;
struct ir3_shader_info *fsi = &fs->info;
uint32_t vsconstlen = constlen(vs);
uint32_t fsconstlen = constlen(fs);
uint32_t i, outloc;
uint32_t posregid = getpos(vs, "gl_Position", 0);
uint32_t psizeregid = getpos(vs, "gl_PointSize", (63 << 2));
uint32_t colorregid = getpos(fs, "gl_FragColor", 0);
uint32_t numvar = totalvar(fs);
assert (vs->ir->varyings_count == fs->ir->varyings_count);
OUT_PKT0(ring, REG_A3XX_HLSQ_CONTROL_0_REG, 6);
OUT_RING(ring, A3XX_HLSQ_CONTROL_0_REG_FSTHREADSIZE(FOUR_QUADS) |
A3XX_HLSQ_CONTROL_0_REG_SPSHADERRESTART |
A3XX_HLSQ_CONTROL_0_REG_SPCONSTFULLUPDATE);
OUT_RING(ring, A3XX_HLSQ_CONTROL_1_REG_VSTHREADSIZE(TWO_QUADS) |
A3XX_HLSQ_CONTROL_1_REG_VSSUPERTHREADENABLE);
OUT_RING(ring, A3XX_HLSQ_CONTROL_2_REG_PRIMALLOCTHRESHOLD(31));
OUT_RING(ring, 0x00000000); /* HLSQ_CONTROL_3_REG */
OUT_RING(ring, A3XX_HLSQ_VS_CONTROL_REG_CONSTLENGTH(vsconstlen) |
A3XX_HLSQ_VS_CONTROL_REG_CONSTSTARTOFFSET(0) |
A3XX_HLSQ_VS_CONTROL_REG_INSTRLENGTH(instrlen(vs)));
OUT_RING(ring, A3XX_HLSQ_FS_CONTROL_REG_CONSTLENGTH(fsconstlen) |
A3XX_HLSQ_FS_CONTROL_REG_CONSTSTARTOFFSET(128) |
A3XX_HLSQ_FS_CONTROL_REG_INSTRLENGTH(instrlen(fs)));
OUT_PKT0(ring, REG_A3XX_SP_SP_CTRL_REG, 1);
OUT_RING(ring, A3XX_SP_SP_CTRL_REG_CONSTMODE(0) |
A3XX_SP_SP_CTRL_REG_SLEEPMODE(1) |
// XXX "resolve" (?) bit set on gmem->mem pass..
COND(!uniforms, A3XX_SP_SP_CTRL_REG_RESOLVE) |
// XXX sometimes 0, sometimes 1:
A3XX_SP_SP_CTRL_REG_LOMODE(1));
/* emit unknown sequence of writes to 0x0ec4/0x0ec8 that the blob
* emits as part of the program state (it seems)..
*/
for (i = 0; i < 6; i++) {
OUT_PKT0(ring, REG_A3XX_SP_PERFCOUNTER0_SELECT, 1);
OUT_RING(ring, 0x00000000); /* SP_PERFCOUNTER0_SELECT */
OUT_PKT0(ring, REG_A3XX_SP_PERFCOUNTER3_SELECT, 1);
OUT_RING(ring, 0x00000000); /* SP_PERFCOUNTER3_SELECT */
}
OUT_PKT0(ring, REG_A3XX_SP_VS_LENGTH_REG, 1);
OUT_RING(ring, A3XX_SP_VS_LENGTH_REG_SHADERLENGTH(instrlen(vs)));
OUT_PKT0(ring, REG_A3XX_SP_VS_CTRL_REG0, 3);
OUT_RING(ring, A3XX_SP_VS_CTRL_REG0_THREADMODE(MULTI) |
A3XX_SP_VS_CTRL_REG0_INSTRBUFFERMODE(BUFFER) |
A3XX_SP_VS_CTRL_REG0_HALFREGFOOTPRINT(vsi->max_half_reg + 1) |
A3XX_SP_VS_CTRL_REG0_FULLREGFOOTPRINT(vsi->max_reg + 1) |
A3XX_SP_VS_CTRL_REG0_INOUTREGOVERLAP(0) |
A3XX_SP_VS_CTRL_REG0_THREADSIZE(TWO_QUADS) |
A3XX_SP_VS_CTRL_REG0_SUPERTHREADMODE |
A3XX_SP_VS_CTRL_REG0_LENGTH(instrlen(vs)));
OUT_RING(ring, A3XX_SP_VS_CTRL_REG1_CONSTLENGTH(vsconstlen) |
A3XX_SP_VS_CTRL_REG1_INITIALOUTSTANDING(totalattr(vs)) |
A3XX_SP_VS_CTRL_REG1_CONSTFOOTPRINT(max(vsi->max_const, 0)));
OUT_RING(ring, A3XX_SP_VS_PARAM_REG_POSREGID(posregid) |
A3XX_SP_VS_PARAM_REG_PSIZEREGID(psizeregid) |
A3XX_SP_VS_PARAM_REG_TOTALVSOUTVAR(fs->ir->varyings_count));
for (i = 0; i < vs->ir->varyings_count; ) {
struct ir3_varying *v;
uint32_t reg = 0;
OUT_PKT0(ring, REG_A3XX_SP_VS_OUT_REG(i/2), 1);
v = vs->ir->varyings[i++];
if (v) {
reg |= A3XX_SP_VS_OUT_REG_A_REGID(v->rstart->num);
reg |= A3XX_SP_VS_OUT_REG_A_COMPMASK(regmask(v->num));
}
v = vs->ir->varyings[i++];
if (v) {
reg |= A3XX_SP_VS_OUT_REG_B_REGID(v->rstart->num);
reg |= A3XX_SP_VS_OUT_REG_B_COMPMASK(regmask(v->num));
}
OUT_RING(ring, reg);
}
outloc = 8; /* I assume 0 and 4 are gl_Position/gl_PointSize? */
for (i = 0; i < vs->ir->varyings_count; ) {
struct ir3_varying *v;
uint32_t reg = 0;
OUT_PKT0(ring, REG_A3XX_SP_VS_VPC_DST_REG(i/4), 1);
/* note: if we supported anything other than vec4 varyings, we'd
* actually be incrementing outloc by the actual varying size in
* units of scalar registers (ie. vec3 -> 3)
*/
v = vs->ir->varyings[i++];
if (v) {
reg |= A3XX_SP_VS_VPC_DST_REG_OUTLOC0(outloc);
outloc += v->num;
}
v = vs->ir->varyings[i++];
if (v) {
reg |= A3XX_SP_VS_VPC_DST_REG_OUTLOC1(outloc);
outloc += v->num;
}
v = vs->ir->varyings[i++];
if (v) {
reg |= A3XX_SP_VS_VPC_DST_REG_OUTLOC2(outloc);
outloc += v->num;
}
v = vs->ir->varyings[i++];
if (v) {
reg |= A3XX_SP_VS_VPC_DST_REG_OUTLOC3(outloc);
outloc += v->num;
}
OUT_RING(ring, reg);
}
// TODO SP_VS_OBJ_OFFSET_REG / SP_VS_OBJ_START_REG
OUT_PKT0(ring, REG_A3XX_SP_FS_LENGTH_REG, 1);
OUT_RING(ring, A3XX_SP_FS_LENGTH_REG_SHADERLENGTH(instrlen(fs)));
OUT_PKT0(ring, REG_A3XX_SP_FS_CTRL_REG0, 2);
OUT_RING(ring, A3XX_SP_FS_CTRL_REG0_THREADMODE(MULTI) |
A3XX_SP_FS_CTRL_REG0_INSTRBUFFERMODE(BUFFER) |
A3XX_SP_FS_CTRL_REG0_HALFREGFOOTPRINT(fsi->max_half_reg + 1) |
A3XX_SP_FS_CTRL_REG0_FULLREGFOOTPRINT(fsi->max_reg + 1) |
A3XX_SP_FS_CTRL_REG0_INOUTREGOVERLAP(1) |
A3XX_SP_FS_CTRL_REG0_THREADSIZE(FOUR_QUADS) |
A3XX_SP_FS_CTRL_REG0_SUPERTHREADMODE |
COND(fs->ir->samplers_count > 0, A3XX_SP_FS_CTRL_REG0_PIXLODENABLE) |
A3XX_SP_FS_CTRL_REG0_LENGTH(instrlen(fs)));
OUT_RING(ring, A3XX_SP_FS_CTRL_REG1_CONSTLENGTH(fsconstlen) |
A3XX_SP_FS_CTRL_REG1_INITIALOUTSTANDING(0) |
A3XX_SP_FS_CTRL_REG1_CONSTFOOTPRINT(max(fsi->max_const, 0)) |
A3XX_SP_FS_CTRL_REG1_HALFPRECVAROFFSET(63));
// TODO SP_FS_OBJ_OFFSET_REG / SP_FS_OBJ_START_REG
OUT_PKT0(ring, REG_A3XX_SP_FS_FLAT_SHAD_MODE_REG_0, 2);
OUT_RING(ring, 0x00000000); /* SP_FS_FLAT_SHAD_MODE_REG_0 */
OUT_RING(ring, 0x00000000); /* SP_FS_FLAT_SHAD_MODE_REG_1 */
OUT_PKT0(ring, REG_A3XX_SP_FS_OUTPUT_REG, 1);
OUT_RING(ring, 0x00000000); /* SP_FS_OUTPUT_REG */
OUT_PKT0(ring, REG_A3XX_SP_FS_MRT_REG(0), 4);
OUT_RING(ring, A3XX_SP_FS_MRT_REG_REGID(colorregid) | /* SP_FS_MRT[0].REG */
A3XX_SP_FS_MRT_REG_HALF_PRECISION);
OUT_RING(ring, A3XX_SP_FS_MRT_REG_REGID(0)); /* SP_FS_MRT[1].REG */
OUT_RING(ring, A3XX_SP_FS_MRT_REG_REGID(0)); /* SP_FS_MRT[2].REG */
OUT_RING(ring, A3XX_SP_FS_MRT_REG_REGID(0)); /* SP_FS_MRT[3].REG */
OUT_PKT0(ring, REG_A3XX_VPC_ATTR, 2);
OUT_RING(ring, A3XX_VPC_ATTR_TOTALATTR(numvar) |
A3XX_VPC_ATTR_THRDASSIGN(1) |
A3XX_VPC_ATTR_LMSIZE(1));
OUT_RING(ring, A3XX_VPC_PACK_NUMFPNONPOSVAR(numvar) |
A3XX_VPC_PACK_NUMNONPOSVSVAR(numvar));
OUT_PKT0(ring, REG_A3XX_VPC_VARYING_INTERP_MODE(0), 4);
OUT_RING(ring, 0x00000000); /* VPC_VARYING_INTERP[0].MODE */
OUT_RING(ring, 0x00000000); /* VPC_VARYING_INTERP[1].MODE */
OUT_RING(ring, 0x00000000); /* VPC_VARYING_INTERP[2].MODE */
OUT_RING(ring, 0x00000000); /* VPC_VARYING_INTERP[3].MODE */
OUT_PKT0(ring, REG_A3XX_VPC_VARYING_PS_REPL_MODE(0), 4);
OUT_RING(ring, 0x00000000); /* VPC_VARYING_PS_REPL[0].MODE */
OUT_RING(ring, 0x00000000); /* VPC_VARYING_PS_REPL[1].MODE */
OUT_RING(ring, 0x00000000); /* VPC_VARYING_PS_REPL[2].MODE */
OUT_RING(ring, 0x00000000); /* VPC_VARYING_PS_REPL[3].MODE */
OUT_PKT0(ring, REG_A3XX_VFD_VS_THREADING_THRESHOLD, 1);
OUT_RING(ring, A3XX_VFD_VS_THREADING_THRESHOLD_REGID_THRESHOLD(15) |
A3XX_VFD_VS_THREADING_THRESHOLD_REGID_VTXCNT(252));
emit_shader(ring, vs, SB_VERT_SHADER);
OUT_PKT0(ring, REG_A3XX_VFD_PERFCOUNTER0_SELECT, 1);
OUT_RING(ring, 0x00000000); /* VFD_PERFCOUNTER0_SELECT */
emit_shader(ring, fs, SB_FRAG_SHADER);
OUT_PKT0(ring, REG_A3XX_VFD_PERFCOUNTER0_SELECT, 1);
OUT_RING(ring, 0x00000000); /* VFD_PERFCOUNTER0_SELECT */
OUT_PKT0(ring, REG_A3XX_VFD_CONTROL_0, 2);
OUT_RING(ring, A3XX_VFD_CONTROL_0_TOTALATTRTOVS(totalattr(vs)) |
A3XX_VFD_CONTROL_0_PACKETSIZE(2) |
A3XX_VFD_CONTROL_0_STRMDECINSTRCNT(vs->ir->attributes_count) |
A3XX_VFD_CONTROL_0_STRMFETCHINSTRCNT(vs->ir->attributes_count));
OUT_RING(ring, A3XX_VFD_CONTROL_1_MAXSTORAGE(1) | // XXX
A3XX_VFD_CONTROL_1_REGID4VTX(63 << 2) |
A3XX_VFD_CONTROL_1_REGID4INST(63 << 2));
emit_vtx_fetch(ring, vs, attr, first);
/* we have this sometimes, not others.. perhaps we could be clever
* and figure out actually when we need to invalidate cache:
*/
OUT_PKT0(ring, REG_A3XX_UCHE_CACHE_INVALIDATE0_REG, 2);
OUT_RING(ring, A3XX_UCHE_CACHE_INVALIDATE0_REG_ADDR(0));
OUT_RING(ring, A3XX_UCHE_CACHE_INVALIDATE1_REG_ADDR(0) |
A3XX_UCHE_CACHE_INVALIDATE1_REG_OPCODE(INVALIDATE) |
A3XX_UCHE_CACHE_INVALIDATE1_REG_ENTIRE_CACHE);
/* for RB_RESOLVE_PASS, I think the consts are not needed: */
if (uniforms) {
emit_uniconst(ring, vs, uniforms, SB_VERT_SHADER);
emit_uniconst(ring, fs, uniforms, SB_FRAG_SHADER);
}
}