/* Creates shader:
* (sy)(ss)(rpt3)mov.f16f16 hr0.x, (r)hc0.x
* end
*/
static struct fd3_shader_stateobj *
create_solid_fp(struct pipe_context *pctx)
{
struct fd3_shader_stateobj *so;
struct ir3_shader *ir = ir3_shader_create();
struct ir3_instruction *instr;
/* (sy)(ss)(rpt3)mov.f16f16 hr0.x, (r)hc0.x */
instr = ir3_instr_create(ir, 1, 0); /* mov/cov instructions have no opc */
instr->flags = IR3_INSTR_SY | IR3_INSTR_SS;
instr->repeat = 3;
instr->cat1.src_type = TYPE_F16;
instr->cat1.dst_type = TYPE_F16;
ir3_reg_create(instr, regid(0,0), IR3_REG_HALF); /* hr0.x */
ir3_reg_create(instr, regid(0,0), IR3_REG_HALF | /* (r)hc0.x */
IR3_REG_CONST | IR3_REG_R);
/* end */
instr = ir3_instr_create(ir, 0, OPC_END);
so = create_internal_shader(pctx, SHADER_FRAGMENT, ir);
if (!so)
return NULL;
so->color_regid = regid(0,0);
so->half_precision = true;
so->inputs_count = 0;
so->total_in = 0;
return so;
}
/* Creates shader:
* (sy)(ss)end
*/
static struct fd3_shader_stateobj *
create_solid_vp(struct pipe_context *pctx)
{
struct fd3_shader_stateobj *so;
struct ir3_shader *ir = ir3_shader_create();
struct ir3_instruction *instr;
/* (sy)(ss)end */
instr = ir3_instr_create(ir, 0, OPC_END);
instr->flags = IR3_INSTR_SY | IR3_INSTR_SS;
so = create_internal_shader(pctx, SHADER_VERTEX, ir);
if (!so)
return NULL;
so->pos_regid = regid(0,0);
so->psize_regid = regid(63,0);
so->inputs_count = 1;
so->inputs[0].regid = regid(0,0);
so->inputs[0].compmask = 0xf;
so->total_in = 4;
so->outputs_count = 0;
fixup_vp_regfootprint(so);
return so;
}
/* for vertex shader, the inputs are loaded into registers before the shader
* is executed, so max_regs from the shader instructions might not properly
* reflect the # of registers actually used, especially in case passthrough
* varyings.
*
* Likewise, for fragment shader, we can have some regs which are passed
* input values but never touched by the resulting shader (ie. as result
* of dead code elimination or simply because we don't know how to turn
* the reg off.
*/
static void
fixup_regfootprint(struct ir3_shader_variant *v, uint32_t gpu_id)
{
unsigned i;
for (i = 0; i < v->inputs_count; i++) {
/* skip frag inputs fetch via bary.f since their reg's are
* not written by gpu before shader starts (and in fact the
* regid's might not even be valid)
*/
if (v->inputs[i].bary)
continue;
/* ignore high regs that are global to all threads in a warp
* (they exist by default) (a5xx+)
*/
if (v->inputs[i].regid >= regid(48,0))
continue;
if (v->inputs[i].compmask) {
unsigned n = util_last_bit(v->inputs[i].compmask) - 1;
int32_t regid = v->inputs[i].regid + n;
if (v->inputs[i].half) {
if (gpu_id < 500) {
v->info.max_half_reg = MAX2(v->info.max_half_reg, regid >> 2);
} else {
v->info.max_reg = MAX2(v->info.max_reg, regid >> 3);
}
} else {
static uint32_t
find_output_regid(const struct ir3_shader_variant *so, ir3_semantic semantic)
{
int j;
for (j = 0; j < so->outputs_count; j++)
if (so->outputs[j].semantic == semantic)
return so->outputs[j].regid;
return regid(63, 0);
}
Region
makeregion()
{
Region reg;
if( (reg = XCreateRegion()) == (Region) NULL)
delete("XCreateRegion() returned NULL.");
else
regid(0, (union regtypes *)®, REG_REGION);
return(reg);
}
/* Creates shader:
* (sy)(ss)(rpt1)bary.f (ei)r0.z, (r)0, r0.x
* (rpt5)nop
* sam (f32)(xyzw)r0.x, r0.z, s#0, t#0
* (sy)(rpt3)cov.f32f16 hr0.x, (r)r0.x
* end
*/
static struct fd3_shader_stateobj *
create_blit_fp(struct pipe_context *pctx)
{
struct fd3_shader_stateobj *so;
struct ir3_shader *ir = ir3_shader_create();
struct ir3_instruction *instr;
/* (sy)(ss)(rpt1)bary.f (ei)r0.z, (r)0, r0.x */
instr = ir3_instr_create(ir, 2, OPC_BARY_F);
instr->flags = IR3_INSTR_SY | IR3_INSTR_SS;
instr->repeat = 1;
ir3_reg_create(instr, regid(0,2), IR3_REG_EI); /* (ei)r0.z */
ir3_reg_create(instr, 0, IR3_REG_R | /* (r)0 */
IR3_REG_IMMED)->iim_val = 0;
ir3_reg_create(instr, regid(0,0), 0); /* r0.x */
/* (rpt5)nop */
instr = ir3_instr_create(ir, 0, OPC_NOP);
instr->repeat = 5;
/* sam (f32)(xyzw)r0.x, r0.z, s#0, t#0 */
instr = ir3_instr_create(ir, 5, OPC_SAM);
instr->cat5.samp = 0;
instr->cat5.tex = 0;
instr->cat5.type = TYPE_F32;
ir3_reg_create(instr, regid(0,0), /* (xyzw)r0.x */
0)->wrmask = 0xf;
ir3_reg_create(instr, regid(0,2), 0); /* r0.z */
/* (sy)(rpt3)cov.f32f16 hr0.x, (r)r0.x */
instr = ir3_instr_create(ir, 1, 0); /* mov/cov instructions have no opc */
instr->flags = IR3_INSTR_SY;
instr->repeat = 3;
instr->cat1.src_type = TYPE_F32;
instr->cat1.dst_type = TYPE_F16;
ir3_reg_create(instr, regid(0,0), IR3_REG_HALF); /* hr0.x */
ir3_reg_create(instr, regid(0,0), IR3_REG_R); /* (r)r0.x */
/* end */
instr = ir3_instr_create(ir, 0, OPC_END);
so = create_internal_shader(pctx, SHADER_FRAGMENT, ir);
if (!so)
return NULL;
so->color_regid = regid(0,0);
so->half_precision = true;
so->inputs_count = 1;
so->inputs[0].inloc = 8;
so->inputs[0].compmask = 0x3;
so->total_in = 2;
so->samplers_count = 1;
so->vpsrepl[0] = 0x99999999;
so->vpsrepl[1] = 0x99999999;
so->vpsrepl[2] = 0x99999999;
so->vpsrepl[3] = 0x99999999;
return so;
}
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
fd5_program_emit(struct fd_context *ctx, struct fd_ringbuffer *ring,
struct fd5_emit *emit)
{
struct stage s[MAX_STAGES];
uint32_t pos_regid, psize_regid, color_regid[8];
uint32_t face_regid, coord_regid, zwcoord_regid;
uint32_t vcoord_regid, vertex_regid, instance_regid;
enum a3xx_threadsize fssz;
uint8_t psize_loc = ~0;
int i, j;
setup_stages(emit, s);
fssz = (s[FS].i->max_reg >= 24) ? TWO_QUADS : FOUR_QUADS;
pos_regid = ir3_find_output_regid(s[VS].v, VARYING_SLOT_POS);
psize_regid = ir3_find_output_regid(s[VS].v, VARYING_SLOT_PSIZ);
vertex_regid = ir3_find_sysval_regid(s[VS].v, SYSTEM_VALUE_VERTEX_ID_ZERO_BASE);
instance_regid = ir3_find_sysval_regid(s[VS].v, SYSTEM_VALUE_INSTANCE_ID);
if (s[FS].v->color0_mrt) {
color_regid[0] = color_regid[1] = color_regid[2] = color_regid[3] =
color_regid[4] = color_regid[5] = color_regid[6] = color_regid[7] =
ir3_find_output_regid(s[FS].v, FRAG_RESULT_COLOR);
} else {
color_regid[0] = ir3_find_output_regid(s[FS].v, FRAG_RESULT_DATA0);
color_regid[1] = ir3_find_output_regid(s[FS].v, FRAG_RESULT_DATA1);
color_regid[2] = ir3_find_output_regid(s[FS].v, FRAG_RESULT_DATA2);
color_regid[3] = ir3_find_output_regid(s[FS].v, FRAG_RESULT_DATA3);
color_regid[4] = ir3_find_output_regid(s[FS].v, FRAG_RESULT_DATA4);
color_regid[5] = ir3_find_output_regid(s[FS].v, FRAG_RESULT_DATA5);
color_regid[6] = ir3_find_output_regid(s[FS].v, FRAG_RESULT_DATA6);
color_regid[7] = ir3_find_output_regid(s[FS].v, FRAG_RESULT_DATA7);
}
/* TODO get these dynamically: */
face_regid = s[FS].v->frag_face ? regid(0,0) : regid(63,0);
coord_regid = s[FS].v->frag_coord ? regid(0,0) : regid(63,0);
zwcoord_regid = s[FS].v->frag_coord ? regid(0,2) : regid(63,0);
vcoord_regid = (s[FS].v->total_in > 0) ? s[FS].v->pos_regid : regid(63,0);
/* 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_PKT4(ring, REG_A5XX_HLSQ_VS_CONFIG, 5);
OUT_RING(ring, A5XX_HLSQ_VS_CONFIG_CONSTOBJECTOFFSET(s[VS].constoff) |
A5XX_HLSQ_VS_CONFIG_SHADEROBJOFFSET(s[VS].instroff) |
COND(s[VS].v, A5XX_HLSQ_VS_CONFIG_ENABLED));
OUT_RING(ring, A5XX_HLSQ_FS_CONFIG_CONSTOBJECTOFFSET(s[FS].constoff) |
A5XX_HLSQ_FS_CONFIG_SHADEROBJOFFSET(s[FS].instroff) |
COND(s[FS].v, A5XX_HLSQ_FS_CONFIG_ENABLED));
OUT_RING(ring, A5XX_HLSQ_HS_CONFIG_CONSTOBJECTOFFSET(s[HS].constoff) |
A5XX_HLSQ_HS_CONFIG_SHADEROBJOFFSET(s[HS].instroff) |
COND(s[HS].v, A5XX_HLSQ_HS_CONFIG_ENABLED));
OUT_RING(ring, A5XX_HLSQ_DS_CONFIG_CONSTOBJECTOFFSET(s[DS].constoff) |
A5XX_HLSQ_DS_CONFIG_SHADEROBJOFFSET(s[DS].instroff) |
COND(s[DS].v, A5XX_HLSQ_DS_CONFIG_ENABLED));
OUT_RING(ring, A5XX_HLSQ_GS_CONFIG_CONSTOBJECTOFFSET(s[GS].constoff) |
A5XX_HLSQ_GS_CONFIG_SHADEROBJOFFSET(s[GS].instroff) |
COND(s[GS].v, A5XX_HLSQ_GS_CONFIG_ENABLED));
OUT_PKT4(ring, REG_A5XX_HLSQ_CS_CONFIG, 1);
OUT_RING(ring, 0x00000000);
OUT_PKT4(ring, REG_A5XX_HLSQ_VS_CNTL, 5);
OUT_RING(ring, A5XX_HLSQ_VS_CNTL_INSTRLEN(s[VS].instrlen) |
COND(s[VS].v && s[VS].v->has_ssbo, A5XX_HLSQ_VS_CNTL_SSBO_ENABLE));
OUT_RING(ring, A5XX_HLSQ_FS_CNTL_INSTRLEN(s[FS].instrlen) |
COND(s[FS].v && s[FS].v->has_ssbo, A5XX_HLSQ_FS_CNTL_SSBO_ENABLE));
OUT_RING(ring, A5XX_HLSQ_HS_CNTL_INSTRLEN(s[HS].instrlen) |
COND(s[HS].v && s[HS].v->has_ssbo, A5XX_HLSQ_HS_CNTL_SSBO_ENABLE));
OUT_RING(ring, A5XX_HLSQ_DS_CNTL_INSTRLEN(s[DS].instrlen) |
COND(s[DS].v && s[DS].v->has_ssbo, A5XX_HLSQ_DS_CNTL_SSBO_ENABLE));
OUT_RING(ring, A5XX_HLSQ_GS_CNTL_INSTRLEN(s[GS].instrlen) |
COND(s[GS].v && s[GS].v->has_ssbo, A5XX_HLSQ_GS_CNTL_SSBO_ENABLE));
OUT_PKT4(ring, REG_A5XX_SP_VS_CONFIG, 5);
OUT_RING(ring, A5XX_SP_VS_CONFIG_CONSTOBJECTOFFSET(s[VS].constoff) |
A5XX_SP_VS_CONFIG_SHADEROBJOFFSET(s[VS].instroff) |
COND(s[VS].v, A5XX_SP_VS_CONFIG_ENABLED));
OUT_RING(ring, A5XX_SP_FS_CONFIG_CONSTOBJECTOFFSET(s[FS].constoff) |
A5XX_SP_FS_CONFIG_SHADEROBJOFFSET(s[FS].instroff) |
COND(s[FS].v, A5XX_SP_FS_CONFIG_ENABLED));
OUT_RING(ring, A5XX_SP_HS_CONFIG_CONSTOBJECTOFFSET(s[HS].constoff) |
A5XX_SP_HS_CONFIG_SHADEROBJOFFSET(s[HS].instroff) |
COND(s[HS].v, A5XX_SP_HS_CONFIG_ENABLED));
OUT_RING(ring, A5XX_SP_DS_CONFIG_CONSTOBJECTOFFSET(s[DS].constoff) |
A5XX_SP_DS_CONFIG_SHADEROBJOFFSET(s[DS].instroff) |
COND(s[DS].v, A5XX_SP_DS_CONFIG_ENABLED));
OUT_RING(ring, A5XX_SP_GS_CONFIG_CONSTOBJECTOFFSET(s[GS].constoff) |
A5XX_SP_GS_CONFIG_SHADEROBJOFFSET(s[GS].instroff) |
COND(s[GS].v, A5XX_SP_GS_CONFIG_ENABLED));
OUT_PKT4(ring, REG_A5XX_SP_CS_CONFIG, 1);
OUT_RING(ring, 0x00000000);
OUT_PKT4(ring, REG_A5XX_HLSQ_VS_CONSTLEN, 2);
OUT_RING(ring, s[VS].constlen); /* HLSQ_VS_CONSTLEN */
OUT_RING(ring, s[VS].instrlen); /* HLSQ_VS_INSTRLEN */
OUT_PKT4(ring, REG_A5XX_HLSQ_FS_CONSTLEN, 2);
OUT_RING(ring, s[FS].constlen); /* HLSQ_FS_CONSTLEN */
OUT_RING(ring, s[FS].instrlen); /* HLSQ_FS_INSTRLEN */
OUT_PKT4(ring, REG_A5XX_HLSQ_HS_CONSTLEN, 2);
OUT_RING(ring, s[HS].constlen); /* HLSQ_HS_CONSTLEN */
OUT_RING(ring, s[HS].instrlen); /* HLSQ_HS_INSTRLEN */
OUT_PKT4(ring, REG_A5XX_HLSQ_DS_CONSTLEN, 2);
OUT_RING(ring, s[DS].constlen); /* HLSQ_DS_CONSTLEN */
OUT_RING(ring, s[DS].instrlen); /* HLSQ_DS_INSTRLEN */
OUT_PKT4(ring, REG_A5XX_HLSQ_GS_CONSTLEN, 2);
OUT_RING(ring, s[GS].constlen); /* HLSQ_GS_CONSTLEN */
OUT_RING(ring, s[GS].instrlen); /* HLSQ_GS_INSTRLEN */
OUT_PKT4(ring, REG_A5XX_HLSQ_CS_CONSTLEN, 2);
OUT_RING(ring, 0x00000000); /* HLSQ_CS_CONSTLEN */
OUT_RING(ring, 0x00000000); /* HLSQ_CS_INSTRLEN */
OUT_PKT4(ring, REG_A5XX_SP_VS_CTRL_REG0, 1);
OUT_RING(ring, A5XX_SP_VS_CTRL_REG0_HALFREGFOOTPRINT(s[VS].i->max_half_reg + 1) |
A5XX_SP_VS_CTRL_REG0_FULLREGFOOTPRINT(s[VS].i->max_reg + 1) |
0x6 | /* XXX seems to be always set? */
A5XX_SP_VS_CTRL_REG0_BRANCHSTACK(0x3) | // XXX need to figure this out somehow..
COND(s[VS].v->has_samp, A5XX_SP_VS_CTRL_REG0_PIXLODENABLE));
struct ir3_shader_linkage l = {0};
ir3_link_shaders(&l, s[VS].v, s[FS].v);
if ((s[VS].v->shader->stream_output.num_outputs > 0) &&
!emit->key.binning_pass)
link_stream_out(&l, s[VS].v);
BITSET_DECLARE(varbs, 128) = {0};
uint32_t *varmask = (uint32_t *)varbs;
for (i = 0; i < l.cnt; i++)
for (j = 0; j < util_last_bit(l.var[i].compmask); j++)
BITSET_SET(varbs, l.var[i].loc + j);
OUT_PKT4(ring, REG_A5XX_VPC_VAR_DISABLE(0), 4);
OUT_RING(ring, ~varmask[0]); /* VPC_VAR[0].DISABLE */
OUT_RING(ring, ~varmask[1]); /* VPC_VAR[1].DISABLE */
OUT_RING(ring, ~varmask[2]); /* VPC_VAR[2].DISABLE */
OUT_RING(ring, ~varmask[3]); /* VPC_VAR[3].DISABLE */
/* a5xx appends pos/psize to end of the linkage map: */
if (pos_regid != regid(63,0))
ir3_link_add(&l, pos_regid, 0xf, l.max_loc);
if (psize_regid != regid(63,0)) {
psize_loc = l.max_loc;
ir3_link_add(&l, psize_regid, 0x1, l.max_loc);
}
if ((s[VS].v->shader->stream_output.num_outputs > 0) &&
!emit->key.binning_pass) {
emit_stream_out(ring, s[VS].v, &l);
OUT_PKT4(ring, REG_A5XX_VPC_SO_OVERRIDE, 1);
OUT_RING(ring, 0x00000000);
} else {
OUT_PKT4(ring, REG_A5XX_VPC_SO_OVERRIDE, 1);
OUT_RING(ring, A5XX_VPC_SO_OVERRIDE_SO_DISABLE);
}
for (i = 0, j = 0; (i < 16) && (j < l.cnt); i++) {
uint32_t reg = 0;
OUT_PKT4(ring, REG_A5XX_SP_VS_OUT_REG(i), 1);
reg |= A5XX_SP_VS_OUT_REG_A_REGID(l.var[j].regid);
reg |= A5XX_SP_VS_OUT_REG_A_COMPMASK(l.var[j].compmask);
j++;
reg |= A5XX_SP_VS_OUT_REG_B_REGID(l.var[j].regid);
reg |= A5XX_SP_VS_OUT_REG_B_COMPMASK(l.var[j].compmask);
j++;
OUT_RING(ring, reg);
}
for (i = 0, j = 0; (i < 8) && (j < l.cnt); i++) {
uint32_t reg = 0;
OUT_PKT4(ring, REG_A5XX_SP_VS_VPC_DST_REG(i), 1);
reg |= A5XX_SP_VS_VPC_DST_REG_OUTLOC0(l.var[j++].loc);
reg |= A5XX_SP_VS_VPC_DST_REG_OUTLOC1(l.var[j++].loc);
reg |= A5XX_SP_VS_VPC_DST_REG_OUTLOC2(l.var[j++].loc);
reg |= A5XX_SP_VS_VPC_DST_REG_OUTLOC3(l.var[j++].loc);
OUT_RING(ring, reg);
}
OUT_PKT4(ring, REG_A5XX_SP_VS_OBJ_START_LO, 2);
OUT_RELOC(ring, s[VS].v->bo, 0, 0, 0); /* SP_VS_OBJ_START_LO/HI */
if (s[VS].instrlen)
fd5_emit_shader(ring, s[VS].v);
// TODO depending on other bits in this reg (if any) set somewhere else?
OUT_PKT4(ring, REG_A5XX_PC_PRIM_VTX_CNTL, 1);
OUT_RING(ring, COND(s[VS].v->writes_psize, A5XX_PC_PRIM_VTX_CNTL_PSIZE));
OUT_PKT4(ring, REG_A5XX_SP_PRIMITIVE_CNTL, 1);
OUT_RING(ring, A5XX_SP_PRIMITIVE_CNTL_VSOUT(l.cnt));
OUT_PKT4(ring, REG_A5XX_VPC_CNTL_0, 1);
OUT_RING(ring, A5XX_VPC_CNTL_0_STRIDE_IN_VPC(l.max_loc) |
COND(s[FS].v->total_in > 0, A5XX_VPC_CNTL_0_VARYING) |
COND(s[FS].v->frag_coord, A5XX_VPC_CNTL_0_VARYING) |
0x10000); // XXX
fd5_context(ctx)->max_loc = l.max_loc;
if (emit->key.binning_pass) {
OUT_PKT4(ring, REG_A5XX_SP_FS_OBJ_START_LO, 2);
OUT_RING(ring, 0x00000000); /* SP_FS_OBJ_START_LO */
OUT_RING(ring, 0x00000000); /* SP_FS_OBJ_START_HI */
} else {
OUT_PKT4(ring, REG_A5XX_SP_FS_OBJ_START_LO, 2);
OUT_RELOC(ring, s[FS].v->bo, 0, 0, 0); /* SP_FS_OBJ_START_LO/HI */
}
OUT_PKT4(ring, REG_A5XX_HLSQ_CONTROL_0_REG, 5);
OUT_RING(ring, A5XX_HLSQ_CONTROL_0_REG_FSTHREADSIZE(fssz) |
A5XX_HLSQ_CONTROL_0_REG_CSTHREADSIZE(TWO_QUADS) |
0x00000880); /* XXX HLSQ_CONTROL_0 */
OUT_RING(ring, A5XX_HLSQ_CONTROL_1_REG_PRIMALLOCTHRESHOLD(63));
OUT_RING(ring, A5XX_HLSQ_CONTROL_2_REG_FACEREGID(face_regid) |
0xfcfcfc00); /* XXX */
OUT_RING(ring, A5XX_HLSQ_CONTROL_3_REG_FRAGCOORDXYREGID(vcoord_regid) |
0xfcfcfc00); /* XXX */
OUT_RING(ring, A5XX_HLSQ_CONTROL_4_REG_XYCOORDREGID(coord_regid) |
A5XX_HLSQ_CONTROL_4_REG_ZWCOORDREGID(zwcoord_regid) |
0x0000fcfc); /* XXX */
OUT_PKT4(ring, REG_A5XX_SP_FS_CTRL_REG0, 1);
OUT_RING(ring, COND(s[FS].v->total_in > 0, A5XX_SP_FS_CTRL_REG0_VARYING) |
COND(s[FS].v->frag_coord, A5XX_SP_FS_CTRL_REG0_VARYING) |
0x40006 | /* XXX set pretty much everywhere */
A5XX_SP_FS_CTRL_REG0_THREADSIZE(fssz) |
A5XX_SP_FS_CTRL_REG0_HALFREGFOOTPRINT(s[FS].i->max_half_reg + 1) |
A5XX_SP_FS_CTRL_REG0_FULLREGFOOTPRINT(s[FS].i->max_reg + 1) |
A5XX_SP_FS_CTRL_REG0_BRANCHSTACK(0x3) | // XXX need to figure this out somehow..
COND(s[FS].v->has_samp, A5XX_SP_FS_CTRL_REG0_PIXLODENABLE));
OUT_PKT4(ring, REG_A5XX_HLSQ_UPDATE_CNTL, 1);
OUT_RING(ring, 0x020fffff); /* XXX */
OUT_PKT4(ring, REG_A5XX_VPC_GS_SIV_CNTL, 1);
OUT_RING(ring, 0x0000ffff); /* XXX */
OUT_PKT4(ring, REG_A5XX_SP_SP_CNTL, 1);
OUT_RING(ring, 0x00000010); /* XXX */
OUT_PKT4(ring, REG_A5XX_GRAS_CNTL, 1);
OUT_RING(ring, COND(s[FS].v->total_in > 0, A5XX_GRAS_CNTL_VARYING) |
COND(s[FS].v->frag_coord, A5XX_GRAS_CNTL_XCOORD |
A5XX_GRAS_CNTL_YCOORD |
A5XX_GRAS_CNTL_ZCOORD |
A5XX_GRAS_CNTL_WCOORD |
A5XX_GRAS_CNTL_UNK3) |
COND(s[FS].v->frag_face, A5XX_GRAS_CNTL_UNK3));
OUT_PKT4(ring, REG_A5XX_RB_RENDER_CONTROL0, 2);
OUT_RING(ring, COND(s[FS].v->total_in > 0, A5XX_RB_RENDER_CONTROL0_VARYING) |
COND(s[FS].v->frag_coord, A5XX_RB_RENDER_CONTROL0_XCOORD |
A5XX_RB_RENDER_CONTROL0_YCOORD |
A5XX_RB_RENDER_CONTROL0_ZCOORD |
A5XX_RB_RENDER_CONTROL0_WCOORD |
A5XX_RB_RENDER_CONTROL0_UNK3) |
COND(s[FS].v->frag_face, A5XX_RB_RENDER_CONTROL0_UNK3));
OUT_RING(ring, COND(s[FS].v->frag_face, A5XX_RB_RENDER_CONTROL1_FACENESS));
OUT_PKT4(ring, REG_A5XX_SP_FS_OUTPUT_REG(0), 8);
for (i = 0; i < 8; i++) {
OUT_RING(ring, A5XX_SP_FS_OUTPUT_REG_REGID(color_regid[i]) |
COND(emit->key.half_precision,
A5XX_SP_FS_OUTPUT_REG_HALF_PRECISION));
}
OUT_PKT4(ring, REG_A5XX_VPC_PACK, 1);
OUT_RING(ring, A5XX_VPC_PACK_NUMNONPOSVAR(s[FS].v->total_in) |
A5XX_VPC_PACK_PSIZELOC(psize_loc));
if (!emit->key.binning_pass) {
uint32_t vinterp[8], vpsrepl[8];
memset(vinterp, 0, sizeof(vinterp));
memset(vpsrepl, 0, sizeof(vpsrepl));
/* looks like we need to do int varyings in the frag
* shader on a5xx (no flatshad reg? or a420.0 bug?):
*
* (sy)(ss)nop
* (sy)ldlv.u32 r0.x,l[r0.x], 1
* ldlv.u32 r0.y,l[r0.x+1], 1
* (ss)bary.f (ei)r63.x, 0, r0.x
* (ss)(rpt1)cov.s32f16 hr0.x, (r)r0.x
* (rpt5)nop
* sam (f16)(xyzw)hr0.x, hr0.x, s#0, t#0
*
* Possibly on later a5xx variants we'll be able to use
* something like the code below instead of workaround
* in the shader:
*/
/* figure out VARYING_INTERP / VARYING_PS_REPL register values: */
for (j = -1; (j = ir3_next_varying(s[FS].v, j)) < (int)s[FS].v->inputs_count; ) {
/* NOTE: varyings are packed, so if compmask is 0xb
* then first, third, and fourth component occupy
* three consecutive varying slots:
*/
unsigned compmask = s[FS].v->inputs[j].compmask;
uint32_t inloc = s[FS].v->inputs[j].inloc;
if ((s[FS].v->inputs[j].interpolate == INTERP_MODE_FLAT) ||
(s[FS].v->inputs[j].rasterflat && emit->rasterflat)) {
uint32_t loc = inloc;
for (i = 0; i < 4; i++) {
if (compmask & (1 << i)) {
vinterp[loc / 16] |= 1 << ((loc % 16) * 2);
//flatshade[loc / 32] |= 1 << (loc % 32);
loc++;
}
}
}
gl_varying_slot slot = s[FS].v->inputs[j].slot;
/* since we don't enable PIPE_CAP_TGSI_TEXCOORD: */
if (slot >= VARYING_SLOT_VAR0) {
unsigned texmask = 1 << (slot - VARYING_SLOT_VAR0);
/* 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 & texmask) {
/* mask is two 2-bit fields, where:
* '01' -> S
* '10' -> T
* '11' -> 1 - T (flip mode)
*/
unsigned mask = emit->sprite_coord_mode ? 0b1101 : 0b1001;
uint32_t loc = inloc;
if (compmask & 0x1) {
vpsrepl[loc / 16] |= ((mask >> 0) & 0x3) << ((loc % 16) * 2);
loc++;
}
if (compmask & 0x2) {
vpsrepl[loc / 16] |= ((mask >> 2) & 0x3) << ((loc % 16) * 2);
loc++;
}
void
fd4_emit_vertex_bufs(struct fd_ringbuffer *ring, struct fd4_emit *emit)
{
int32_t i, j, last = -1;
uint32_t total_in = 0;
const struct fd_vertex_state *vtx = emit->vtx;
struct ir3_shader_variant *vp = fd4_emit_get_vp(emit);
unsigned vertex_regid = regid(63, 0), instance_regid = regid(63, 0);
for (i = 0; i < vp->inputs_count; i++) {
uint8_t semantic = sem2name(vp->inputs[i].semantic);
if (semantic == TGSI_SEMANTIC_VERTEXID_NOBASE)
vertex_regid = vp->inputs[i].regid;
else if (semantic == TGSI_SEMANTIC_INSTANCEID)
instance_regid = vp->inputs[i].regid;
else if ((i < vtx->vtx->num_elements) && vp->inputs[i].compmask)
last = i;
}
/* hw doesn't like to be configured for zero vbo's, it seems: */
if ((vtx->vtx->num_elements == 0) &&
(vertex_regid == regid(63, 0)) &&
(instance_regid == regid(63, 0)))
return;
for (i = 0, j = 0; i <= last; i++) {
assert(sem2name(vp->inputs[i].semantic) == 0);
if (vp->inputs[i].compmask) {
struct pipe_vertex_element *elem = &vtx->vtx->pipe[i];
const struct pipe_vertex_buffer *vb =
&vtx->vertexbuf.vb[elem->vertex_buffer_index];
struct fd_resource *rsc = fd_resource(vb->buffer);
enum pipe_format pfmt = elem->src_format;
enum a4xx_vtx_fmt fmt = fd4_pipe2vtx(pfmt);
bool switchnext = (i != last) ||
(vertex_regid != regid(63, 0)) ||
(instance_regid != regid(63, 0));
bool isint = util_format_is_pure_integer(pfmt);
uint32_t fs = util_format_get_blocksize(pfmt);
uint32_t off = vb->buffer_offset + elem->src_offset;
uint32_t size = fd_bo_size(rsc->bo) - off;
debug_assert(fmt != ~0);
OUT_PKT0(ring, REG_A4XX_VFD_FETCH(j), 4);
OUT_RING(ring, A4XX_VFD_FETCH_INSTR_0_FETCHSIZE(fs - 1) |
A4XX_VFD_FETCH_INSTR_0_BUFSTRIDE(vb->stride) |
COND(elem->instance_divisor, A4XX_VFD_FETCH_INSTR_0_INSTANCED) |
COND(switchnext, A4XX_VFD_FETCH_INSTR_0_SWITCHNEXT));
OUT_RELOC(ring, rsc->bo, off, 0, 0);
OUT_RING(ring, A4XX_VFD_FETCH_INSTR_2_SIZE(size));
OUT_RING(ring, A4XX_VFD_FETCH_INSTR_3_STEPRATE(MAX2(1, elem->instance_divisor)));
OUT_PKT0(ring, REG_A4XX_VFD_DECODE_INSTR(j), 1);
OUT_RING(ring, A4XX_VFD_DECODE_INSTR_CONSTFILL |
A4XX_VFD_DECODE_INSTR_WRITEMASK(vp->inputs[i].compmask) |
A4XX_VFD_DECODE_INSTR_FORMAT(fmt) |
A4XX_VFD_DECODE_INSTR_SWAP(fd4_pipe2swap(pfmt)) |
A4XX_VFD_DECODE_INSTR_REGID(vp->inputs[i].regid) |
A4XX_VFD_DECODE_INSTR_SHIFTCNT(fs) |
A4XX_VFD_DECODE_INSTR_LASTCOMPVALID |
COND(isint, A4XX_VFD_DECODE_INSTR_INT) |
COND(switchnext, A4XX_VFD_DECODE_INSTR_SWITCHNEXT));
total_in += vp->inputs[i].ncomp;
j++;
}
}
OUT_PKT0(ring, REG_A4XX_VFD_CONTROL_0, 5);
OUT_RING(ring, A4XX_VFD_CONTROL_0_TOTALATTRTOVS(total_in) |
0xa0000 | /* XXX */
A4XX_VFD_CONTROL_0_STRMDECINSTRCNT(j) |
A4XX_VFD_CONTROL_0_STRMFETCHINSTRCNT(j));
OUT_RING(ring, A4XX_VFD_CONTROL_1_MAXSTORAGE(129) | // XXX
A4XX_VFD_CONTROL_1_REGID4VTX(vertex_regid) |
A4XX_VFD_CONTROL_1_REGID4INST(instance_regid));
OUT_RING(ring, 0x00000000); /* XXX VFD_CONTROL_2 */
OUT_RING(ring, A4XX_VFD_CONTROL_3_REGID_VTXCNT(regid(63, 0)));
OUT_RING(ring, 0x00000000); /* XXX VFD_CONTROL_4 */
/* cache invalidate, otherwise vertex fetch could see
* stale vbo contents:
*/
OUT_PKT0(ring, REG_A4XX_UCHE_INVALIDATE0, 2);
OUT_RING(ring, 0x00000000);
OUT_RING(ring, 0x00000012);
}
/**
* Find instruction src's which are mov's that can be collapsed, replacing
* the mov dst with the mov src
*/
static void
instr_cp(struct ir3_cp_ctx *ctx, struct ir3_instruction *instr)
{
struct ir3_register *reg;
if (instr->regs_count == 0)
return;
if (ir3_instr_check_mark(instr))
return;
/* walk down the graph from each src: */
foreach_src_n(reg, n, instr) {
struct ir3_instruction *src = ssa(reg);
if (!src)
continue;
instr_cp(ctx, src);
/* TODO non-indirect access we could figure out which register
* we actually want and allow cp..
*/
if (reg->flags & IR3_REG_ARRAY)
continue;
reg_cp(ctx, instr, reg, n);
}
if (instr->regs[0]->flags & IR3_REG_ARRAY) {
struct ir3_instruction *src = ssa(instr->regs[0]);
if (src)
instr_cp(ctx, src);
}
if (instr->address) {
instr_cp(ctx, instr->address);
ir3_instr_set_address(instr, eliminate_output_mov(instr->address));
}
/* we can end up with extra cmps.s from frontend, which uses a
*
* cmps.s p0.x, cond, 0
*
* as a way to mov into the predicate register. But frequently 'cond'
* is itself a cmps.s/cmps.f/cmps.u. So detect this special case and
* just re-write the instruction writing predicate register to get rid
* of the double cmps.
*/
if ((instr->opc == OPC_CMPS_S) &&
(instr->regs[0]->num == regid(REG_P0, 0)) &&
ssa(instr->regs[1]) &&
(instr->regs[2]->flags & IR3_REG_IMMED) &&
(instr->regs[2]->iim_val == 0)) {
struct ir3_instruction *cond = ssa(instr->regs[1]);
switch (cond->opc) {
case OPC_CMPS_S:
case OPC_CMPS_F:
case OPC_CMPS_U:
instr->opc = cond->opc;
instr->flags = cond->flags;
instr->cat2 = cond->cat2;
instr->address = cond->address;
instr->regs[1] = cond->regs[1];
instr->regs[2] = cond->regs[2];
break;
default:
break;
}
}
}
void
ir3_emit_vs_consts(const struct ir3_shader_variant *v, struct fd_ringbuffer *ring,
struct fd_context *ctx, const struct pipe_draw_info *info)
{
debug_assert(v->type == MESA_SHADER_VERTEX);
emit_common_consts(v, ring, ctx, PIPE_SHADER_VERTEX);
/* emit driver params every time: */
/* TODO skip emit if shader doesn't use driver params to avoid WFI.. */
if (info) {
const struct ir3_const_state *const_state = &v->shader->const_state;
uint32_t offset = const_state->offsets.driver_param;
if (v->constlen > offset) {
uint32_t vertex_params[IR3_DP_VS_COUNT] = {
[IR3_DP_VTXID_BASE] = info->index_size ?
info->index_bias : info->start,
[IR3_DP_VTXCNT_MAX] = max_tf_vtx(ctx, v),
};
/* if no user-clip-planes, we don't need to emit the
* entire thing:
*/
uint32_t vertex_params_size = 4;
if (v->key.ucp_enables) {
struct pipe_clip_state *ucp = &ctx->ucp;
unsigned pos = IR3_DP_UCP0_X;
for (unsigned i = 0; pos <= IR3_DP_UCP7_W; i++) {
for (unsigned j = 0; j < 4; j++) {
vertex_params[pos] = fui(ucp->ucp[i][j]);
pos++;
}
}
vertex_params_size = ARRAY_SIZE(vertex_params);
}
ring_wfi(ctx->batch, ring);
bool needs_vtxid_base =
ir3_find_sysval_regid(v, SYSTEM_VALUE_VERTEX_ID_ZERO_BASE) != regid(63, 0);
/* for indirect draw, we need to copy VTXID_BASE from
* indirect-draw parameters buffer.. which is annoying
* and means we can't easily emit these consts in cmd
* stream so need to copy them to bo.
*/
if (info->indirect && needs_vtxid_base) {
struct pipe_draw_indirect_info *indirect = info->indirect;
struct pipe_resource *vertex_params_rsc =
pipe_buffer_create(&ctx->screen->base,
PIPE_BIND_CONSTANT_BUFFER, PIPE_USAGE_STREAM,
vertex_params_size * 4);
unsigned src_off = info->indirect->offset;;
void *ptr;
ptr = fd_bo_map(fd_resource(vertex_params_rsc)->bo);
memcpy(ptr, vertex_params, vertex_params_size * 4);
if (info->index_size) {
/* indexed draw, index_bias is 4th field: */
src_off += 3 * 4;
} else {
/* non-indexed draw, start is 3rd field: */
src_off += 2 * 4;
}
/* copy index_bias or start from draw params: */
ctx->mem_to_mem(ring, vertex_params_rsc, 0,
indirect->buffer, src_off, 1);
ctx->emit_const(ring, MESA_SHADER_VERTEX, offset * 4, 0,
vertex_params_size, NULL, vertex_params_rsc);
pipe_resource_reference(&vertex_params_rsc, NULL);
} else {
ctx->emit_const(ring, MESA_SHADER_VERTEX, offset * 4, 0,
vertex_params_size, vertex_params, NULL);
}
/* if needed, emit stream-out buffer addresses: */
if (vertex_params[IR3_DP_VTXCNT_MAX] > 0) {
emit_tfbos(ctx, v, ring);
}
}
}
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);
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, VARYING_SLOT_POS);
posz_regid = ir3_find_output_regid(fp, FRAG_RESULT_DEPTH);
psize_regid = ir3_find_output_regid(vp, VARYING_SLOT_PSIZ);
if (fp->color0_mrt) {
color_regid[0] = color_regid[1] = color_regid[2] = color_regid[3] =
ir3_find_output_regid(fp, FRAG_RESULT_COLOR);
} else {
color_regid[0] = ir3_find_output_regid(fp, FRAG_RESULT_DATA0);
color_regid[1] = ir3_find_output_regid(fp, FRAG_RESULT_DATA1);
color_regid[2] = ir3_find_output_regid(fp, FRAG_RESULT_DATA2);
color_regid[3] = ir3_find_output_regid(fp, FRAG_RESULT_DATA3);
}
/* 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_FRAGCOORDXYREGID(regid(0,0)) |
A3XX_HLSQ_CONTROL_1_REG_FRAGCOORDZWREGID(regid(0,2))));
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_THREADSIZE(TWO_QUADS) |
A3XX_SP_VS_CTRL_REG0_SUPERTHREADMODE |
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(fp->varying_in));
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].slot);
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].slot);
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 |
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; ) {
/* NOTE: varyings are packed, so if compmask is 0xb
* then first, third, and fourth component occupy
* three consecutive varying slots:
*/
unsigned compmask = fp->inputs[j].compmask;
/* 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;
if ((fp->inputs[j].interpolate == INTERP_QUALIFIER_FLAT) ||
(fp->inputs[j].rasterflat && emit->rasterflat)) {
uint32_t loc = inloc;
for (i = 0; i < 4; i++) {
if (compmask & (1 << i)) {
vinterp[loc / 16] |= FLAT << ((loc % 16) * 2);
flatshade[loc / 32] |= 1 << (loc % 32);
loc++;
}
}
}
gl_varying_slot slot = fp->inputs[j].slot;
/* since we don't enable PIPE_CAP_TGSI_TEXCOORD: */
if (slot >= VARYING_SLOT_VAR0) {
unsigned texmask = 1 << (slot - VARYING_SLOT_VAR0);
/* 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 & texmask) {
/* mask is two 2-bit fields, where:
* '01' -> S
* '10' -> T
* '11' -> 1 - T (flip mode)
*/
unsigned mask = emit->sprite_coord_mode ? 0b1101 : 0b1001;
uint32_t loc = inloc;
if (compmask & 0x1) {
vpsrepl[loc / 16] |= ((mask >> 0) & 0x3) << ((loc % 16) * 2);
loc++;
}
if (compmask & 0x2) {
vpsrepl[loc / 16] |= ((mask >> 2) & 0x3) << ((loc % 16) * 2);
loc++;
}
if (compmask & 0x4) {
/* .z <- 0.0f */
vinterp[loc / 16] |= 0b10 << ((loc % 16) * 2);
loc++;
}
if (compmask & 0x8) {
/* .w <- 1.0f */
vinterp[loc / 16] |= 0b11 << ((loc % 16) * 2);
loc++;
}
}
void
fd4_program_emit(struct fd_ringbuffer *ring, struct fd4_emit *emit)
{
struct stage s[MAX_STAGES];
uint32_t pos_regid, posz_regid, psize_regid, color_regid;
uint32_t face_regid, coord_regid, zwcoord_regid;
int constmode;
int i, j, k;
setup_stages(emit, s);
/* blob seems to always use constmode currently: */
constmode = 1;
pos_regid = ir3_find_output_regid(s[VS].v,
ir3_semantic_name(TGSI_SEMANTIC_POSITION, 0));
posz_regid = ir3_find_output_regid(s[FS].v,
ir3_semantic_name(TGSI_SEMANTIC_POSITION, 0));
psize_regid = ir3_find_output_regid(s[VS].v,
ir3_semantic_name(TGSI_SEMANTIC_PSIZE, 0));
color_regid = ir3_find_output_regid(s[FS].v,
ir3_semantic_name(TGSI_SEMANTIC_COLOR, 0));
if (util_format_is_alpha(emit->pformat))
color_regid += 3;
/* TODO get these dynamically: */
face_regid = s[FS].v->frag_face ? regid(0,0) : regid(63,0);
coord_regid = s[FS].v->frag_coord ? regid(0,0) : regid(63,0);
zwcoord_regid = s[FS].v->frag_coord ? regid(0,2) : regid(63,0);
/* 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_A4XX_HLSQ_UPDATE_CONTROL, 1);
OUT_RING(ring, 0x00000003);
OUT_PKT0(ring, REG_A4XX_HLSQ_CONTROL_0_REG, 5);
OUT_RING(ring, A4XX_HLSQ_CONTROL_0_REG_FSTHREADSIZE(FOUR_QUADS) |
A4XX_HLSQ_CONTROL_0_REG_CONSTMODE(constmode) |
A4XX_HLSQ_CONTROL_0_REG_FSSUPERTHREADENABLE |
/* 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..
*/
A4XX_HLSQ_CONTROL_0_REG_SPSHADERRESTART |
A4XX_HLSQ_CONTROL_0_REG_SPCONSTFULLUPDATE);
OUT_RING(ring, A4XX_HLSQ_CONTROL_1_REG_VSTHREADSIZE(TWO_QUADS) |
A4XX_HLSQ_CONTROL_1_REG_VSSUPERTHREADENABLE |
A4XX_HLSQ_CONTROL_1_REG_COORDREGID(coord_regid) |
A4XX_HLSQ_CONTROL_1_REG_ZWCOORDREGID(zwcoord_regid));
OUT_RING(ring, A4XX_HLSQ_CONTROL_2_REG_PRIMALLOCTHRESHOLD(63) |
0x3f3f000 | /* XXX */
A4XX_HLSQ_CONTROL_2_REG_FACEREGID(face_regid));
OUT_RING(ring, A4XX_HLSQ_CONTROL_3_REG_REGID(s[FS].v->pos_regid) |
0xfcfcfc00);
OUT_RING(ring, 0x00fcfcfc); /* XXX HLSQ_CONTROL_4 */
OUT_PKT0(ring, REG_A4XX_HLSQ_VS_CONTROL_REG, 5);
OUT_RING(ring, A4XX_HLSQ_VS_CONTROL_REG_CONSTLENGTH(s[VS].constlen) |
A4XX_HLSQ_VS_CONTROL_REG_CONSTOBJECTOFFSET(s[VS].constoff) |
A4XX_HLSQ_VS_CONTROL_REG_INSTRLENGTH(s[VS].instrlen) |
A4XX_HLSQ_VS_CONTROL_REG_SHADEROBJOFFSET(s[VS].instroff));
OUT_RING(ring, A4XX_HLSQ_FS_CONTROL_REG_CONSTLENGTH(s[FS].constlen) |
A4XX_HLSQ_FS_CONTROL_REG_CONSTOBJECTOFFSET(s[FS].constoff) |
A4XX_HLSQ_FS_CONTROL_REG_INSTRLENGTH(s[FS].instrlen) |
A4XX_HLSQ_FS_CONTROL_REG_SHADEROBJOFFSET(s[FS].instroff));
OUT_RING(ring, A4XX_HLSQ_HS_CONTROL_REG_CONSTLENGTH(s[HS].constlen) |
A4XX_HLSQ_HS_CONTROL_REG_CONSTOBJECTOFFSET(s[HS].constoff) |
A4XX_HLSQ_HS_CONTROL_REG_INSTRLENGTH(s[HS].instrlen) |
A4XX_HLSQ_HS_CONTROL_REG_SHADEROBJOFFSET(s[HS].instroff));
OUT_RING(ring, A4XX_HLSQ_DS_CONTROL_REG_CONSTLENGTH(s[DS].constlen) |
A4XX_HLSQ_DS_CONTROL_REG_CONSTOBJECTOFFSET(s[DS].constoff) |
A4XX_HLSQ_DS_CONTROL_REG_INSTRLENGTH(s[DS].instrlen) |
A4XX_HLSQ_DS_CONTROL_REG_SHADEROBJOFFSET(s[DS].instroff));
OUT_RING(ring, A4XX_HLSQ_GS_CONTROL_REG_CONSTLENGTH(s[GS].constlen) |
A4XX_HLSQ_GS_CONTROL_REG_CONSTOBJECTOFFSET(s[GS].constoff) |
A4XX_HLSQ_GS_CONTROL_REG_INSTRLENGTH(s[GS].instrlen) |
A4XX_HLSQ_GS_CONTROL_REG_SHADEROBJOFFSET(s[GS].instroff));
OUT_PKT0(ring, REG_A4XX_SP_SP_CTRL_REG, 1);
OUT_RING(ring, 0x140010 | /* XXX */
COND(emit->key.binning_pass, A4XX_SP_SP_CTRL_REG_BINNING_PASS));
OUT_PKT0(ring, REG_A4XX_SP_INSTR_CACHE_CTRL, 1);
OUT_RING(ring, 0x7f | /* XXX */
COND(s[VS].instrlen, A4XX_SP_INSTR_CACHE_CTRL_VS_BUFFER) |
COND(s[FS].instrlen, A4XX_SP_INSTR_CACHE_CTRL_FS_BUFFER) |
COND(s[VS].instrlen && s[FS].instrlen,
A4XX_SP_INSTR_CACHE_CTRL_INSTR_BUFFER));
OUT_PKT0(ring, REG_A4XX_SP_VS_LENGTH_REG, 1);
OUT_RING(ring, s[VS].v->instrlen); /* SP_VS_LENGTH_REG */
OUT_PKT0(ring, REG_A4XX_SP_VS_CTRL_REG0, 3);
OUT_RING(ring, A4XX_SP_VS_CTRL_REG0_THREADMODE(MULTI) |
A4XX_SP_VS_CTRL_REG0_HALFREGFOOTPRINT(s[VS].i->max_half_reg + 1) |
A4XX_SP_VS_CTRL_REG0_FULLREGFOOTPRINT(s[VS].i->max_reg + 1) |
A4XX_SP_VS_CTRL_REG0_INOUTREGOVERLAP(0) |
A4XX_SP_VS_CTRL_REG0_THREADSIZE(TWO_QUADS) |
A4XX_SP_VS_CTRL_REG0_SUPERTHREADMODE |
COND(s[VS].v->has_samp, A4XX_SP_VS_CTRL_REG0_PIXLODENABLE));
OUT_RING(ring, A4XX_SP_VS_CTRL_REG1_CONSTLENGTH(s[VS].constlen) |
A4XX_SP_VS_CTRL_REG1_INITIALOUTSTANDING(s[VS].v->total_in));
OUT_RING(ring, A4XX_SP_VS_PARAM_REG_POSREGID(pos_regid) |
A4XX_SP_VS_PARAM_REG_PSIZEREGID(psize_regid) |
A4XX_SP_VS_PARAM_REG_TOTALVSOUTVAR(align(s[FS].v->total_in, 4) / 4));
for (i = 0, j = -1; (i < 16) && (j < (int)s[FS].v->inputs_count); i++) {
uint32_t reg = 0;
OUT_PKT0(ring, REG_A4XX_SP_VS_OUT_REG(i), 1);
j = ir3_next_varying(s[FS].v, j);
if (j < s[FS].v->inputs_count) {
k = ir3_find_output(s[VS].v, s[FS].v->inputs[j].semantic);
reg |= A4XX_SP_VS_OUT_REG_A_REGID(s[VS].v->outputs[k].regid);
reg |= A4XX_SP_VS_OUT_REG_A_COMPMASK(s[FS].v->inputs[j].compmask);
}
j = ir3_next_varying(s[FS].v, j);
if (j < s[FS].v->inputs_count) {
k = ir3_find_output(s[VS].v, s[FS].v->inputs[j].semantic);
reg |= A4XX_SP_VS_OUT_REG_B_REGID(s[VS].v->outputs[k].regid);
reg |= A4XX_SP_VS_OUT_REG_B_COMPMASK(s[FS].v->inputs[j].compmask);
}
OUT_RING(ring, reg);
}
for (i = 0, j = -1; (i < 8) && (j < (int)s[FS].v->inputs_count); i++) {
uint32_t reg = 0;
OUT_PKT0(ring, REG_A4XX_SP_VS_VPC_DST_REG(i), 1);
j = ir3_next_varying(s[FS].v, j);
if (j < s[FS].v->inputs_count)
reg |= A4XX_SP_VS_VPC_DST_REG_OUTLOC0(s[FS].v->inputs[j].inloc);
j = ir3_next_varying(s[FS].v, j);
if (j < s[FS].v->inputs_count)
reg |= A4XX_SP_VS_VPC_DST_REG_OUTLOC1(s[FS].v->inputs[j].inloc);
j = ir3_next_varying(s[FS].v, j);
if (j < s[FS].v->inputs_count)
reg |= A4XX_SP_VS_VPC_DST_REG_OUTLOC2(s[FS].v->inputs[j].inloc);
j = ir3_next_varying(s[FS].v, j);
if (j < s[FS].v->inputs_count)
reg |= A4XX_SP_VS_VPC_DST_REG_OUTLOC3(s[FS].v->inputs[j].inloc);
OUT_RING(ring, reg);
}
OUT_PKT0(ring, REG_A4XX_SP_VS_OBJ_OFFSET_REG, 2);
OUT_RING(ring, A4XX_SP_VS_OBJ_OFFSET_REG_CONSTOBJECTOFFSET(s[VS].constoff) |
A4XX_SP_VS_OBJ_OFFSET_REG_SHADEROBJOFFSET(s[VS].instroff));
OUT_RELOC(ring, s[VS].v->bo, 0, 0, 0); /* SP_VS_OBJ_START_REG */
OUT_PKT0(ring, REG_A4XX_SP_FS_LENGTH_REG, 1);
OUT_RING(ring, s[FS].v->instrlen); /* SP_FS_LENGTH_REG */
OUT_PKT0(ring, REG_A4XX_SP_FS_CTRL_REG0, 2);
OUT_RING(ring, A4XX_SP_FS_CTRL_REG0_THREADMODE(MULTI) |
COND(s[FS].v->total_in > 0, A4XX_SP_FS_CTRL_REG0_VARYING) |
A4XX_SP_FS_CTRL_REG0_HALFREGFOOTPRINT(s[FS].i->max_half_reg + 1) |
A4XX_SP_FS_CTRL_REG0_FULLREGFOOTPRINT(s[FS].i->max_reg + 1) |
A4XX_SP_FS_CTRL_REG0_INOUTREGOVERLAP(1) |
A4XX_SP_FS_CTRL_REG0_THREADSIZE(FOUR_QUADS) |
A4XX_SP_FS_CTRL_REG0_SUPERTHREADMODE |
COND(s[FS].v->has_samp, A4XX_SP_FS_CTRL_REG0_PIXLODENABLE));
OUT_RING(ring, A4XX_SP_FS_CTRL_REG1_CONSTLENGTH(s[FS].constlen) |
0x80000000 | /* XXX */
COND(s[FS].v->frag_face, A4XX_SP_FS_CTRL_REG1_FACENESS) |
COND(s[FS].v->total_in > 0, A4XX_SP_FS_CTRL_REG1_VARYING) |
COND(s[FS].v->frag_coord, A4XX_SP_FS_CTRL_REG1_FRAGCOORD));
OUT_PKT0(ring, REG_A4XX_SP_FS_OBJ_OFFSET_REG, 2);
OUT_RING(ring, A4XX_SP_FS_OBJ_OFFSET_REG_CONSTOBJECTOFFSET(s[FS].constoff) |
A4XX_SP_FS_OBJ_OFFSET_REG_SHADEROBJOFFSET(s[FS].instroff));
if (emit->key.binning_pass)
OUT_RING(ring, 0x00000000);
else
OUT_RELOC(ring, s[FS].v->bo, 0, 0, 0); /* SP_FS_OBJ_START_REG */
OUT_PKT0(ring, REG_A4XX_SP_HS_OBJ_OFFSET_REG, 1);
OUT_RING(ring, A4XX_SP_HS_OBJ_OFFSET_REG_CONSTOBJECTOFFSET(s[HS].constoff) |
A4XX_SP_HS_OBJ_OFFSET_REG_SHADEROBJOFFSET(s[HS].instroff));
OUT_PKT0(ring, REG_A4XX_SP_DS_OBJ_OFFSET_REG, 1);
OUT_RING(ring, A4XX_SP_DS_OBJ_OFFSET_REG_CONSTOBJECTOFFSET(s[DS].constoff) |
A4XX_SP_DS_OBJ_OFFSET_REG_SHADEROBJOFFSET(s[DS].instroff));
OUT_PKT0(ring, REG_A4XX_SP_GS_OBJ_OFFSET_REG, 1);
OUT_RING(ring, A4XX_SP_GS_OBJ_OFFSET_REG_CONSTOBJECTOFFSET(s[GS].constoff) |
A4XX_SP_GS_OBJ_OFFSET_REG_SHADEROBJOFFSET(s[GS].instroff));
OUT_PKT0(ring, REG_A4XX_RB_RENDER_CONTROL2, 1);
OUT_RING(ring, A4XX_RB_RENDER_CONTROL2_MSAA_SAMPLES(0) |
COND(s[FS].v->total_in > 0, A4XX_RB_RENDER_CONTROL2_VARYING) |
COND(s[FS].v->frag_face, A4XX_RB_RENDER_CONTROL2_FACENESS) |
COND(s[FS].v->frag_coord, A4XX_RB_RENDER_CONTROL2_XCOORD |
A4XX_RB_RENDER_CONTROL2_YCOORD |
// TODO enabling gl_FragCoord.z is causing lockups on 0ad (but seems
// to work everywhere else).
// A4XX_RB_RENDER_CONTROL2_ZCOORD |
A4XX_RB_RENDER_CONTROL2_WCOORD));
OUT_PKT0(ring, REG_A4XX_RB_FS_OUTPUT_REG, 1);
OUT_RING(ring, A4XX_RB_FS_OUTPUT_REG_MRT(1) |
COND(s[FS].v->writes_pos, A4XX_RB_FS_OUTPUT_REG_FRAG_WRITES_Z));
OUT_PKT0(ring, REG_A4XX_SP_FS_OUTPUT_REG, 1);
if (s[FS].v->writes_pos) {
OUT_RING(ring, 0x00000001 |
A4XX_SP_FS_OUTPUT_REG_DEPTH_ENABLE |
A4XX_SP_FS_OUTPUT_REG_DEPTH_REGID(posz_regid));
} else {
OUT_RING(ring, 0x00000001);
}
OUT_PKT0(ring, REG_A4XX_SP_FS_MRT_REG(0), 8);
OUT_RING(ring, A4XX_SP_FS_MRT_REG_REGID(color_regid) |
A4XX_SP_FS_MRT_REG_MRTFORMAT(emit->format) |
COND(emit->key.half_precision, A4XX_SP_FS_MRT_REG_HALF_PRECISION));
OUT_RING(ring, A4XX_SP_FS_MRT_REG_REGID(0));
OUT_RING(ring, A4XX_SP_FS_MRT_REG_REGID(0));
OUT_RING(ring, A4XX_SP_FS_MRT_REG_REGID(0));
OUT_RING(ring, A4XX_SP_FS_MRT_REG_REGID(0));
OUT_RING(ring, A4XX_SP_FS_MRT_REG_REGID(0));
OUT_RING(ring, A4XX_SP_FS_MRT_REG_REGID(0));
OUT_RING(ring, A4XX_SP_FS_MRT_REG_REGID(0));
if (emit->key.binning_pass) {
OUT_PKT0(ring, REG_A4XX_VPC_ATTR, 2);
OUT_RING(ring, A4XX_VPC_ATTR_THRDASSIGN(1) |
0x40000000 | /* XXX */
COND(s[VS].v->writes_psize, A4XX_VPC_ATTR_PSIZE));
OUT_RING(ring, 0x00000000);
} else {
uint32_t vinterp[8], flatshade[2];
memset(vinterp, 0, sizeof(vinterp));
memset(flatshade, 0, sizeof(flatshade));
/* looks like we need to do int varyings in the frag
* shader on a4xx (no flatshad reg? or a420.0 bug?):
*
* (sy)(ss)nop
* (sy)ldlv.u32 r0.x,l[r0.x], 1
* ldlv.u32 r0.y,l[r0.x+1], 1
* (ss)bary.f (ei)r63.x, 0, r0.x
* (ss)(rpt1)cov.s32f16 hr0.x, (r)r0.x
* (rpt5)nop
* sam (f16)(xyzw)hr0.x, hr0.x, s#0, t#0
*
* Possibly on later a4xx variants we'll be able to use
* something like the code below instead of workaround
* in the shader:
*/
#if 0
/* figure out VARYING_INTERP / FLAT_SHAD register values: */
for (j = -1; (j = ir3_next_varying(s[FS].v, j)) < (int)s[FS].v->inputs_count; ) {
uint32_t interp = s[FS].v->inputs[j].interpolate;
if ((interp == TGSI_INTERPOLATE_CONSTANT) ||
((interp == TGSI_INTERPOLATE_COLOR) && emit->rasterflat)) {
/* TODO might be cleaner to just +8 in SP_VS_VPC_DST_REG
* instead.. rather than -8 everywhere else..
*/
uint32_t loc = s[FS].v->inputs[j].inloc - 8;
/* currently assuming varyings aligned to 4 (not
* packed):
*/
debug_assert((loc % 4) == 0);
for (i = 0; i < 4; i++, loc++) {
vinterp[loc / 16] |= 1 << ((loc % 16) * 2);
flatshade[loc / 32] |= 1 << (loc % 32);
}
}
}
#endif
OUT_PKT0(ring, REG_A4XX_VPC_ATTR, 2);
OUT_RING(ring, A4XX_VPC_ATTR_TOTALATTR(s[FS].v->total_in) |
A4XX_VPC_ATTR_THRDASSIGN(1) |
COND(s[FS].v->total_in > 0, A4XX_VPC_ATTR_ENABLE) |
0x40000000 | /* XXX */
COND(s[VS].v->writes_psize, A4XX_VPC_ATTR_PSIZE));
OUT_RING(ring, A4XX_VPC_PACK_NUMFPNONPOSVAR(s[FS].v->total_in) |
A4XX_VPC_PACK_NUMNONPOSVSVAR(s[FS].v->total_in));
OUT_PKT0(ring, REG_A4XX_VPC_VARYING_INTERP_MODE(0), 8);
for (i = 0; i < 8; i++)
OUT_RING(ring, vinterp[i]); /* VPC_VARYING_INTERP[i].MODE */
OUT_PKT0(ring, REG_A4XX_VPC_VARYING_PS_REPL_MODE(0), 8);
for (i = 0; i < 8; i++)
OUT_RING(ring, s[FS].v->shader->vpsrepl[i]); /* VPC_VARYING_PS_REPL[i] */
}
if (s[VS].instrlen)
emit_shader(ring, s[VS].v);
if (!emit->key.binning_pass)
if (s[FS].instrlen)
emit_shader(ring, s[FS].v);
}
/**
* Find instruction src's which are mov's that can be collapsed, replacing
* the mov dst with the mov src
*/
static void
instr_cp(struct ir3_cp_ctx *ctx, struct ir3_instruction *instr)
{
struct ir3_register *reg;
if (instr->regs_count == 0)
return;
if (ir3_instr_check_mark(instr))
return;
/* walk down the graph from each src: */
foreach_src_n(reg, n, instr) {
struct ir3_instruction *src = ssa(reg);
if (!src)
continue;
instr_cp(ctx, src);
/* TODO non-indirect access we could figure out which register
* we actually want and allow cp..
*/
if (reg->flags & IR3_REG_ARRAY)
continue;
/* Don't CP absneg into meta instructions, that won't end well: */
if (is_meta(instr) && (src->opc != OPC_MOV))
continue;
reg_cp(ctx, instr, reg, n);
}
if (instr->regs[0]->flags & IR3_REG_ARRAY) {
struct ir3_instruction *src = ssa(instr->regs[0]);
if (src)
instr_cp(ctx, src);
}
if (instr->address) {
instr_cp(ctx, instr->address);
ir3_instr_set_address(instr, eliminate_output_mov(instr->address));
}
/* we can end up with extra cmps.s from frontend, which uses a
*
* cmps.s p0.x, cond, 0
*
* as a way to mov into the predicate register. But frequently 'cond'
* is itself a cmps.s/cmps.f/cmps.u. So detect this special case and
* just re-write the instruction writing predicate register to get rid
* of the double cmps.
*/
if ((instr->opc == OPC_CMPS_S) &&
(instr->regs[0]->num == regid(REG_P0, 0)) &&
ssa(instr->regs[1]) &&
(instr->regs[2]->flags & IR3_REG_IMMED) &&
(instr->regs[2]->iim_val == 0)) {
struct ir3_instruction *cond = ssa(instr->regs[1]);
switch (cond->opc) {
case OPC_CMPS_S:
case OPC_CMPS_F:
case OPC_CMPS_U:
instr->opc = cond->opc;
instr->flags = cond->flags;
instr->cat2 = cond->cat2;
instr->address = cond->address;
instr->regs[1] = cond->regs[1];
instr->regs[2] = cond->regs[2];
instr->barrier_class |= cond->barrier_class;
instr->barrier_conflict |= cond->barrier_conflict;
unuse(cond);
break;
default:
break;
}
}
/* Handle converting a sam.s2en (taking samp/tex idx params via
* register) into a normal sam (encoding immediate samp/tex idx)
* if they are immediate. This saves some instructions and regs
* in the common case where we know samp/tex at compile time:
*/
if (is_tex(instr) && (instr->flags & IR3_INSTR_S2EN) &&
!(ir3_shader_debug & IR3_DBG_FORCES2EN)) {
/* The first src will be a fan-in (collect), if both of it's
* two sources are mov from imm, then we can
*/
struct ir3_instruction *samp_tex = ssa(instr->regs[1]);
debug_assert(samp_tex->opc == OPC_META_FI);
struct ir3_instruction *samp = ssa(samp_tex->regs[1]);
struct ir3_instruction *tex = ssa(samp_tex->regs[2]);
if ((samp->opc == OPC_MOV) &&
(samp->regs[1]->flags & IR3_REG_IMMED) &&
(tex->opc == OPC_MOV) &&
(tex->regs[1]->flags & IR3_REG_IMMED)) {
instr->flags &= ~IR3_INSTR_S2EN;
instr->cat5.samp = samp->regs[1]->iim_val;
instr->cat5.tex = tex->regs[1]->iim_val;
instr->regs[1]->instr = NULL;
}
}
}
void
fd4_program_emit(struct fd_ringbuffer *ring, struct fd4_emit *emit,
int nr, struct pipe_surface **bufs)
{
struct stage s[MAX_STAGES];
uint32_t pos_regid, posz_regid, psize_regid, color_regid[8];
uint32_t face_regid, coord_regid, zwcoord_regid;
enum a3xx_threadsize fssz;
int constmode;
int i, j, k;
debug_assert(nr <= ARRAY_SIZE(color_regid));
if (emit->key.binning_pass)
nr = 0;
setup_stages(emit, s);
fssz = (s[FS].i->max_reg >= 24) ? TWO_QUADS : FOUR_QUADS;
/* blob seems to always use constmode currently: */
constmode = 1;
pos_regid = ir3_find_output_regid(s[VS].v, VARYING_SLOT_POS);
if (pos_regid == regid(63, 0)) {
/* hw dislikes when there is no position output, which can
* happen for transform-feedback vertex shaders. Just tell
* the hw to use r0.x, with whatever random value is there:
*/
pos_regid = regid(0, 0);
}
posz_regid = ir3_find_output_regid(s[FS].v, FRAG_RESULT_DEPTH);
psize_regid = ir3_find_output_regid(s[VS].v, VARYING_SLOT_PSIZ);
if (s[FS].v->color0_mrt) {
color_regid[0] = color_regid[1] = color_regid[2] = color_regid[3] =
color_regid[4] = color_regid[5] = color_regid[6] = color_regid[7] =
ir3_find_output_regid(s[FS].v, FRAG_RESULT_COLOR);
} else {
color_regid[0] = ir3_find_output_regid(s[FS].v, FRAG_RESULT_DATA0);
color_regid[1] = ir3_find_output_regid(s[FS].v, FRAG_RESULT_DATA1);
color_regid[2] = ir3_find_output_regid(s[FS].v, FRAG_RESULT_DATA2);
color_regid[3] = ir3_find_output_regid(s[FS].v, FRAG_RESULT_DATA3);
color_regid[4] = ir3_find_output_regid(s[FS].v, FRAG_RESULT_DATA4);
color_regid[5] = ir3_find_output_regid(s[FS].v, FRAG_RESULT_DATA5);
color_regid[6] = ir3_find_output_regid(s[FS].v, FRAG_RESULT_DATA6);
color_regid[7] = ir3_find_output_regid(s[FS].v, FRAG_RESULT_DATA7);
}
/* TODO get these dynamically: */
face_regid = s[FS].v->frag_face ? regid(0,0) : regid(63,0);
coord_regid = s[FS].v->frag_coord ? regid(0,0) : regid(63,0);
zwcoord_regid = s[FS].v->frag_coord ? regid(0,2) : regid(63,0);
/* 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_A4XX_HLSQ_UPDATE_CONTROL, 1);
OUT_RING(ring, 0x00000003);
OUT_PKT0(ring, REG_A4XX_HLSQ_CONTROL_0_REG, 5);
OUT_RING(ring, A4XX_HLSQ_CONTROL_0_REG_FSTHREADSIZE(fssz) |
A4XX_HLSQ_CONTROL_0_REG_CONSTMODE(constmode) |
A4XX_HLSQ_CONTROL_0_REG_FSSUPERTHREADENABLE |
/* 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..
*/
A4XX_HLSQ_CONTROL_0_REG_SPSHADERRESTART |
A4XX_HLSQ_CONTROL_0_REG_SPCONSTFULLUPDATE);
OUT_RING(ring, A4XX_HLSQ_CONTROL_1_REG_VSTHREADSIZE(TWO_QUADS) |
A4XX_HLSQ_CONTROL_1_REG_VSSUPERTHREADENABLE |
A4XX_HLSQ_CONTROL_1_REG_COORDREGID(coord_regid) |
A4XX_HLSQ_CONTROL_1_REG_ZWCOORDREGID(zwcoord_regid));
OUT_RING(ring, A4XX_HLSQ_CONTROL_2_REG_PRIMALLOCTHRESHOLD(63) |
0x3f3f000 | /* XXX */
A4XX_HLSQ_CONTROL_2_REG_FACEREGID(face_regid));
OUT_RING(ring, A4XX_HLSQ_CONTROL_3_REG_REGID(s[FS].v->pos_regid) |
0xfcfcfc00);
OUT_RING(ring, 0x00fcfcfc); /* XXX HLSQ_CONTROL_4 */
OUT_PKT0(ring, REG_A4XX_HLSQ_VS_CONTROL_REG, 5);
OUT_RING(ring, A4XX_HLSQ_VS_CONTROL_REG_CONSTLENGTH(s[VS].constlen) |
A4XX_HLSQ_VS_CONTROL_REG_CONSTOBJECTOFFSET(s[VS].constoff) |
A4XX_HLSQ_VS_CONTROL_REG_INSTRLENGTH(s[VS].instrlen) |
A4XX_HLSQ_VS_CONTROL_REG_SHADEROBJOFFSET(s[VS].instroff));
OUT_RING(ring, A4XX_HLSQ_FS_CONTROL_REG_CONSTLENGTH(s[FS].constlen) |
A4XX_HLSQ_FS_CONTROL_REG_CONSTOBJECTOFFSET(s[FS].constoff) |
A4XX_HLSQ_FS_CONTROL_REG_INSTRLENGTH(s[FS].instrlen) |
A4XX_HLSQ_FS_CONTROL_REG_SHADEROBJOFFSET(s[FS].instroff));
OUT_RING(ring, A4XX_HLSQ_HS_CONTROL_REG_CONSTLENGTH(s[HS].constlen) |
A4XX_HLSQ_HS_CONTROL_REG_CONSTOBJECTOFFSET(s[HS].constoff) |
A4XX_HLSQ_HS_CONTROL_REG_INSTRLENGTH(s[HS].instrlen) |
A4XX_HLSQ_HS_CONTROL_REG_SHADEROBJOFFSET(s[HS].instroff));
OUT_RING(ring, A4XX_HLSQ_DS_CONTROL_REG_CONSTLENGTH(s[DS].constlen) |
A4XX_HLSQ_DS_CONTROL_REG_CONSTOBJECTOFFSET(s[DS].constoff) |
A4XX_HLSQ_DS_CONTROL_REG_INSTRLENGTH(s[DS].instrlen) |
A4XX_HLSQ_DS_CONTROL_REG_SHADEROBJOFFSET(s[DS].instroff));
OUT_RING(ring, A4XX_HLSQ_GS_CONTROL_REG_CONSTLENGTH(s[GS].constlen) |
A4XX_HLSQ_GS_CONTROL_REG_CONSTOBJECTOFFSET(s[GS].constoff) |
A4XX_HLSQ_GS_CONTROL_REG_INSTRLENGTH(s[GS].instrlen) |
A4XX_HLSQ_GS_CONTROL_REG_SHADEROBJOFFSET(s[GS].instroff));
OUT_PKT0(ring, REG_A4XX_SP_SP_CTRL_REG, 1);
OUT_RING(ring, 0x140010 | /* XXX */
COND(emit->key.binning_pass, A4XX_SP_SP_CTRL_REG_BINNING_PASS));
OUT_PKT0(ring, REG_A4XX_SP_INSTR_CACHE_CTRL, 1);
OUT_RING(ring, 0x7f | /* XXX */
COND(s[VS].instrlen, A4XX_SP_INSTR_CACHE_CTRL_VS_BUFFER) |
COND(s[FS].instrlen, A4XX_SP_INSTR_CACHE_CTRL_FS_BUFFER) |
COND(s[VS].instrlen && s[FS].instrlen,
A4XX_SP_INSTR_CACHE_CTRL_INSTR_BUFFER));
OUT_PKT0(ring, REG_A4XX_SP_VS_LENGTH_REG, 1);
OUT_RING(ring, s[VS].v->instrlen); /* SP_VS_LENGTH_REG */
OUT_PKT0(ring, REG_A4XX_SP_VS_CTRL_REG0, 3);
OUT_RING(ring, A4XX_SP_VS_CTRL_REG0_THREADMODE(MULTI) |
A4XX_SP_VS_CTRL_REG0_HALFREGFOOTPRINT(s[VS].i->max_half_reg + 1) |
A4XX_SP_VS_CTRL_REG0_FULLREGFOOTPRINT(s[VS].i->max_reg + 1) |
A4XX_SP_VS_CTRL_REG0_INOUTREGOVERLAP(0) |
A4XX_SP_VS_CTRL_REG0_THREADSIZE(TWO_QUADS) |
A4XX_SP_VS_CTRL_REG0_SUPERTHREADMODE |
COND(s[VS].v->has_samp, A4XX_SP_VS_CTRL_REG0_PIXLODENABLE));
OUT_RING(ring, A4XX_SP_VS_CTRL_REG1_CONSTLENGTH(s[VS].constlen) |
A4XX_SP_VS_CTRL_REG1_INITIALOUTSTANDING(s[VS].v->total_in));
OUT_RING(ring, A4XX_SP_VS_PARAM_REG_POSREGID(pos_regid) |
A4XX_SP_VS_PARAM_REG_PSIZEREGID(psize_regid) |
A4XX_SP_VS_PARAM_REG_TOTALVSOUTVAR(s[FS].v->varying_in));
for (i = 0, j = -1; (i < 16) && (j < (int)s[FS].v->inputs_count); i++) {
uint32_t reg = 0;
OUT_PKT0(ring, REG_A4XX_SP_VS_OUT_REG(i), 1);
j = ir3_next_varying(s[FS].v, j);
if (j < s[FS].v->inputs_count) {
k = ir3_find_output(s[VS].v, s[FS].v->inputs[j].slot);
reg |= A4XX_SP_VS_OUT_REG_A_REGID(s[VS].v->outputs[k].regid);
reg |= A4XX_SP_VS_OUT_REG_A_COMPMASK(s[FS].v->inputs[j].compmask);
}
j = ir3_next_varying(s[FS].v, j);
if (j < s[FS].v->inputs_count) {
k = ir3_find_output(s[VS].v, s[FS].v->inputs[j].slot);
reg |= A4XX_SP_VS_OUT_REG_B_REGID(s[VS].v->outputs[k].regid);
reg |= A4XX_SP_VS_OUT_REG_B_COMPMASK(s[FS].v->inputs[j].compmask);
}
OUT_RING(ring, reg);
}
for (i = 0, j = -1; (i < 8) && (j < (int)s[FS].v->inputs_count); i++) {
uint32_t reg = 0;
OUT_PKT0(ring, REG_A4XX_SP_VS_VPC_DST_REG(i), 1);
j = ir3_next_varying(s[FS].v, j);
if (j < s[FS].v->inputs_count)
reg |= A4XX_SP_VS_VPC_DST_REG_OUTLOC0(s[FS].v->inputs[j].inloc);
j = ir3_next_varying(s[FS].v, j);
if (j < s[FS].v->inputs_count)
reg |= A4XX_SP_VS_VPC_DST_REG_OUTLOC1(s[FS].v->inputs[j].inloc);
j = ir3_next_varying(s[FS].v, j);
if (j < s[FS].v->inputs_count)
reg |= A4XX_SP_VS_VPC_DST_REG_OUTLOC2(s[FS].v->inputs[j].inloc);
j = ir3_next_varying(s[FS].v, j);
if (j < s[FS].v->inputs_count)
reg |= A4XX_SP_VS_VPC_DST_REG_OUTLOC3(s[FS].v->inputs[j].inloc);
OUT_RING(ring, reg);
}
OUT_PKT0(ring, REG_A4XX_SP_VS_OBJ_OFFSET_REG, 2);
OUT_RING(ring, A4XX_SP_VS_OBJ_OFFSET_REG_CONSTOBJECTOFFSET(s[VS].constoff) |
A4XX_SP_VS_OBJ_OFFSET_REG_SHADEROBJOFFSET(s[VS].instroff));
OUT_RELOC(ring, s[VS].v->bo, 0, 0, 0); /* SP_VS_OBJ_START_REG */
if (emit->key.binning_pass) {
OUT_PKT0(ring, REG_A4XX_SP_FS_LENGTH_REG, 1);
OUT_RING(ring, 0x00000000); /* SP_FS_LENGTH_REG */
OUT_PKT0(ring, REG_A4XX_SP_FS_CTRL_REG0, 2);
OUT_RING(ring, A4XX_SP_FS_CTRL_REG0_THREADMODE(MULTI) |
COND(s[FS].v->total_in > 0, A4XX_SP_FS_CTRL_REG0_VARYING) |
A4XX_SP_FS_CTRL_REG0_HALFREGFOOTPRINT(0) |
A4XX_SP_FS_CTRL_REG0_FULLREGFOOTPRINT(0) |
A4XX_SP_FS_CTRL_REG0_INOUTREGOVERLAP(1) |
A4XX_SP_FS_CTRL_REG0_THREADSIZE(fssz) |
A4XX_SP_FS_CTRL_REG0_SUPERTHREADMODE);
OUT_RING(ring, A4XX_SP_FS_CTRL_REG1_CONSTLENGTH(s[FS].constlen) |
0x80000000);
OUT_PKT0(ring, REG_A4XX_SP_FS_OBJ_OFFSET_REG, 2);
OUT_RING(ring, A4XX_SP_FS_OBJ_OFFSET_REG_CONSTOBJECTOFFSET(s[FS].constoff) |
A4XX_SP_FS_OBJ_OFFSET_REG_SHADEROBJOFFSET(s[FS].instroff));
OUT_RING(ring, 0x00000000);
} else {
OUT_PKT0(ring, REG_A4XX_SP_FS_LENGTH_REG, 1);
OUT_RING(ring, s[FS].v->instrlen); /* SP_FS_LENGTH_REG */
OUT_PKT0(ring, REG_A4XX_SP_FS_CTRL_REG0, 2);
OUT_RING(ring, A4XX_SP_FS_CTRL_REG0_THREADMODE(MULTI) |
COND(s[FS].v->total_in > 0, A4XX_SP_FS_CTRL_REG0_VARYING) |
A4XX_SP_FS_CTRL_REG0_HALFREGFOOTPRINT(s[FS].i->max_half_reg + 1) |
A4XX_SP_FS_CTRL_REG0_FULLREGFOOTPRINT(s[FS].i->max_reg + 1) |
A4XX_SP_FS_CTRL_REG0_INOUTREGOVERLAP(1) |
A4XX_SP_FS_CTRL_REG0_THREADSIZE(fssz) |
A4XX_SP_FS_CTRL_REG0_SUPERTHREADMODE |
COND(s[FS].v->has_samp, A4XX_SP_FS_CTRL_REG0_PIXLODENABLE));
OUT_RING(ring, A4XX_SP_FS_CTRL_REG1_CONSTLENGTH(s[FS].constlen) |
0x80000000 | /* XXX */
COND(s[FS].v->frag_face, A4XX_SP_FS_CTRL_REG1_FACENESS) |
COND(s[FS].v->total_in > 0, A4XX_SP_FS_CTRL_REG1_VARYING) |
COND(s[FS].v->frag_coord, A4XX_SP_FS_CTRL_REG1_FRAGCOORD));
OUT_PKT0(ring, REG_A4XX_SP_FS_OBJ_OFFSET_REG, 2);
OUT_RING(ring, A4XX_SP_FS_OBJ_OFFSET_REG_CONSTOBJECTOFFSET(s[FS].constoff) |
A4XX_SP_FS_OBJ_OFFSET_REG_SHADEROBJOFFSET(s[FS].instroff));
OUT_RELOC(ring, s[FS].v->bo, 0, 0, 0); /* SP_FS_OBJ_START_REG */
}
OUT_PKT0(ring, REG_A4XX_SP_HS_OBJ_OFFSET_REG, 1);
OUT_RING(ring, A4XX_SP_HS_OBJ_OFFSET_REG_CONSTOBJECTOFFSET(s[HS].constoff) |
A4XX_SP_HS_OBJ_OFFSET_REG_SHADEROBJOFFSET(s[HS].instroff));
OUT_PKT0(ring, REG_A4XX_SP_DS_OBJ_OFFSET_REG, 1);
OUT_RING(ring, A4XX_SP_DS_OBJ_OFFSET_REG_CONSTOBJECTOFFSET(s[DS].constoff) |
A4XX_SP_DS_OBJ_OFFSET_REG_SHADEROBJOFFSET(s[DS].instroff));
OUT_PKT0(ring, REG_A4XX_SP_GS_OBJ_OFFSET_REG, 1);
OUT_RING(ring, A4XX_SP_GS_OBJ_OFFSET_REG_CONSTOBJECTOFFSET(s[GS].constoff) |
A4XX_SP_GS_OBJ_OFFSET_REG_SHADEROBJOFFSET(s[GS].instroff));
OUT_PKT0(ring, REG_A4XX_RB_RENDER_CONTROL2, 1);
OUT_RING(ring, A4XX_RB_RENDER_CONTROL2_MSAA_SAMPLES(0) |
COND(s[FS].v->total_in > 0, A4XX_RB_RENDER_CONTROL2_VARYING) |
COND(s[FS].v->frag_face, A4XX_RB_RENDER_CONTROL2_FACENESS) |
COND(s[FS].v->frag_coord, A4XX_RB_RENDER_CONTROL2_XCOORD |
A4XX_RB_RENDER_CONTROL2_YCOORD |
A4XX_RB_RENDER_CONTROL2_ZCOORD |
A4XX_RB_RENDER_CONTROL2_WCOORD));
OUT_PKT0(ring, REG_A4XX_RB_FS_OUTPUT_REG, 1);
OUT_RING(ring, A4XX_RB_FS_OUTPUT_REG_MRT(nr) |
COND(s[FS].v->writes_pos, A4XX_RB_FS_OUTPUT_REG_FRAG_WRITES_Z));
OUT_PKT0(ring, REG_A4XX_SP_FS_OUTPUT_REG, 1);
OUT_RING(ring, A4XX_SP_FS_OUTPUT_REG_MRT(nr) |
COND(s[FS].v->writes_pos, A4XX_SP_FS_OUTPUT_REG_DEPTH_ENABLE) |
A4XX_SP_FS_OUTPUT_REG_DEPTH_REGID(posz_regid));
OUT_PKT0(ring, REG_A4XX_SP_FS_MRT_REG(0), 8);
for (i = 0; i < 8; i++) {
enum a4xx_color_fmt format = 0;
bool srgb = false;
if (i < nr) {
format = fd4_emit_format(bufs[i]);
if (bufs[i] && !emit->no_decode_srgb)
srgb = util_format_is_srgb(bufs[i]->format);
}
OUT_RING(ring, A4XX_SP_FS_MRT_REG_REGID(color_regid[i]) |
A4XX_SP_FS_MRT_REG_MRTFORMAT(format) |
COND(srgb, A4XX_SP_FS_MRT_REG_COLOR_SRGB) |
COND(emit->key.half_precision,
A4XX_SP_FS_MRT_REG_HALF_PRECISION));
}
if (emit->key.binning_pass) {
OUT_PKT0(ring, REG_A4XX_VPC_ATTR, 2);
OUT_RING(ring, A4XX_VPC_ATTR_THRDASSIGN(1) |
0x40000000 | /* XXX */
COND(s[VS].v->writes_psize, A4XX_VPC_ATTR_PSIZE));
OUT_RING(ring, 0x00000000);
} else {
uint32_t vinterp[8], vpsrepl[8];
memset(vinterp, 0, sizeof(vinterp));
memset(vpsrepl, 0, sizeof(vpsrepl));
/* looks like we need to do int varyings in the frag
* shader on a4xx (no flatshad reg? or a420.0 bug?):
*
* (sy)(ss)nop
* (sy)ldlv.u32 r0.x,l[r0.x], 1
* ldlv.u32 r0.y,l[r0.x+1], 1
* (ss)bary.f (ei)r63.x, 0, r0.x
* (ss)(rpt1)cov.s32f16 hr0.x, (r)r0.x
* (rpt5)nop
* sam (f16)(xyzw)hr0.x, hr0.x, s#0, t#0
*
* Possibly on later a4xx variants we'll be able to use
* something like the code below instead of workaround
* in the shader:
*/
/* figure out VARYING_INTERP / VARYING_PS_REPL register values: */
for (j = -1; (j = ir3_next_varying(s[FS].v, j)) < (int)s[FS].v->inputs_count; ) {
/* NOTE: varyings are packed, so if compmask is 0xb
* then first, third, and fourth component occupy
* three consecutive varying slots:
*/
unsigned compmask = s[FS].v->inputs[j].compmask;
/* TODO might be cleaner to just +8 in SP_VS_VPC_DST_REG
* instead.. rather than -8 everywhere else..
*/
uint32_t inloc = s[FS].v->inputs[j].inloc - 8;
if ((s[FS].v->inputs[j].interpolate == INTERP_MODE_FLAT) ||
(s[FS].v->inputs[j].rasterflat && emit->rasterflat)) {
uint32_t loc = inloc;
for (i = 0; i < 4; i++) {
if (compmask & (1 << i)) {
vinterp[loc / 16] |= 1 << ((loc % 16) * 2);
//flatshade[loc / 32] |= 1 << (loc % 32);
loc++;
}
}
}
gl_varying_slot slot = s[FS].v->inputs[j].slot;
/* since we don't enable PIPE_CAP_TGSI_TEXCOORD: */
if (slot >= VARYING_SLOT_VAR0) {
unsigned texmask = 1 << (slot - VARYING_SLOT_VAR0);
/* 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 & texmask) {
/* mask is two 2-bit fields, where:
* '01' -> S
* '10' -> T
* '11' -> 1 - T (flip mode)
*/
unsigned mask = emit->sprite_coord_mode ? 0b1101 : 0b1001;
uint32_t loc = inloc;
if (compmask & 0x1) {
vpsrepl[loc / 16] |= ((mask >> 0) & 0x3) << ((loc % 16) * 2);
loc++;
}
if (compmask & 0x2) {
vpsrepl[loc / 16] |= ((mask >> 2) & 0x3) << ((loc % 16) * 2);
loc++;
}
if (compmask & 0x4) {
/* .z <- 0.0f */
vinterp[loc / 16] |= 0b10 << ((loc % 16) * 2);
loc++;
}
if (compmask & 0x8) {
/* .w <- 1.0f */
vinterp[loc / 16] |= 0b11 << ((loc % 16) * 2);
loc++;
}
}
