#include "util/u_memory.h"
#include "util/u_inlines.h"
#include "util/u_format.h"
-#include "tgsi/tgsi_dump.h"
-#include "tgsi/tgsi_parse.h"
-#include "freedreno_lowering.h"
#include "freedreno_program.h"
#include "fd3_program.h"
#include "fd3_emit.h"
#include "fd3_texture.h"
-#include "fd3_util.h"
+#include "fd3_format.h"
static void
delete_shader_stateobj(struct fd3_shader_stateobj *so)
enum shader_t type)
{
struct fd3_shader_stateobj *so = CALLOC_STRUCT(fd3_shader_stateobj);
- so->shader = ir3_shader_create(pctx, cso->tokens, type);
+ so->shader = ir3_shader_create(pctx, cso, type);
return so;
}
}
}
-static int
-find_output(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 j;
-
- /* it seems optional to have a OUT.BCOLOR[n] for each OUT.COLOR[n]
- * in the vertex shader.. but the fragment shader doesn't know this
- * so it will always have both IN.COLOR[n] and IN.BCOLOR[n]. So
- * at link time if there is no matching OUT.BCOLOR[n], we must map
- * OUT.COLOR[n] to IN.BCOLOR[n]. And visa versa if there is only
- * a OUT.BCOLOR[n] but no matching OUT.COLOR[n]
- */
- if (sem2name(semantic) == TGSI_SEMANTIC_BCOLOR) {
- unsigned idx = sem2idx(semantic);
- semantic = ir3_semantic_name(TGSI_SEMANTIC_COLOR, idx);
- } else if (sem2name(semantic) == TGSI_SEMANTIC_COLOR) {
- unsigned idx = sem2idx(semantic);
- semantic = ir3_semantic_name(TGSI_SEMANTIC_BCOLOR, idx);
- }
-
- for (j = 0; j < so->outputs_count; j++)
- if (so->outputs[j].semantic == semantic)
- return j;
-
- debug_assert(0);
-
- return 0;
-}
-
-static int
-next_varying(const struct ir3_shader_variant *so, int i)
-{
- while (++i < so->inputs_count)
- if (so->inputs[i].compmask && so->inputs[i].bary)
- break;
- return i;
-}
-
-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);
-}
-
void
-fd3_program_emit(struct fd_ringbuffer *ring,
- struct fd_program_stateobj *prog,
- struct ir3_shader_key key,
- boolean rasterflat)
+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;
+ uint32_t pos_regid, posz_regid, psize_regid, color_regid[4] = {0};
+ int constmode;
int i, j, k;
- vp = fd3_shader_variant(prog->vp, key);
+ debug_assert(nr <= ARRAY_SIZE(color_regid));
+
+ vp = fd3_emit_get_vp(emit);
- if (key.binning_pass) {
+ 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_shader_variant(prog->fp, key);
+ fp = fd3_emit_get_fp(emit);
}
vsi = &vp->info;
fsoff = 256 - fpbuffersz;
}
- pos_regid = find_output_regid(vp,
- ir3_semantic_name(TGSI_SEMANTIC_POSITION, 0));
- posz_regid = find_output_regid(fp,
- ir3_semantic_name(TGSI_SEMANTIC_POSITION, 0));
- psize_regid = find_output_regid(vp,
- ir3_semantic_name(TGSI_SEMANTIC_PSIZE, 0));
- color_regid = find_output_regid(fp,
- ir3_semantic_name(TGSI_SEMANTIC_COLOR, 0));
+ /* 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_FS_CONTROL_REG_INSTRLENGTH(fpbuffersz));
OUT_PKT0(ring, REG_A3XX_SP_SP_CTRL_REG, 1);
- OUT_RING(ring, A3XX_SP_SP_CTRL_REG_CONSTMODE(0) |
- COND(key.binning_pass, A3XX_SP_SP_CTRL_REG_BINNING) |
+ 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_OUT_REG(i), 1);
- j = next_varying(fp, j);
+ j = ir3_next_varying(fp, j);
if (j < fp->inputs_count) {
- k = find_output(vp, fp->inputs[j].semantic);
+ 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 = next_varying(fp, j);
+ j = ir3_next_varying(fp, j);
if (j < fp->inputs_count) {
- k = find_output(vp, fp->inputs[j].semantic);
+ 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_PKT0(ring, REG_A3XX_SP_VS_VPC_DST_REG(i), 1);
- j = next_varying(fp, j);
+ j = ir3_next_varying(fp, j);
if (j < fp->inputs_count)
reg |= A3XX_SP_VS_VPC_DST_REG_OUTLOC0(fp->inputs[j].inloc);
- j = next_varying(fp, j);
+ j = ir3_next_varying(fp, j);
if (j < fp->inputs_count)
reg |= A3XX_SP_VS_VPC_DST_REG_OUTLOC1(fp->inputs[j].inloc);
- j = next_varying(fp, j);
+ j = ir3_next_varying(fp, j);
if (j < fp->inputs_count)
reg |= A3XX_SP_VS_VPC_DST_REG_OUTLOC2(fp->inputs[j].inloc);
- j = next_varying(fp, j);
+ j = ir3_next_varying(fp, j);
if (j < fp->inputs_count)
reg |= A3XX_SP_VS_VPC_DST_REG_OUTLOC3(fp->inputs[j].inloc);
A3XX_SP_VS_OBJ_OFFSET_REG_SHADEROBJOFFSET(0));
OUT_RELOC(ring, vp->bo, 0, 0, 0); /* SP_VS_OBJ_START_REG */
- if (key.binning_pass) {
+ if (emit->key.binning_pass) {
OUT_PKT0(ring, REG_A3XX_SP_FS_LENGTH_REG, 1);
OUT_RING(ring, 0x00000000);
A3XX_SP_FS_CTRL_REG1_CONSTFOOTPRINT(MAX2(fp->constlen + 1, 0)) |
A3XX_SP_FS_CTRL_REG1_HALFPRECVAROFFSET(63));
- /* NOTE: I believe VS.CONSTLEN should be <= FS.CONSTOBJOFFSET*/
- debug_assert(vp->constlen <= 128);
-
OUT_PKT0(ring, REG_A3XX_SP_FS_OBJ_OFFSET_REG, 2);
- OUT_RING(ring, A3XX_SP_FS_OBJ_OFFSET_REG_CONSTOBJECTOFFSET(128) |
+ 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);
- if (fp->writes_pos) {
- OUT_RING(ring, A3XX_SP_FS_OUTPUT_REG_DEPTH_ENABLE |
- A3XX_SP_FS_OUTPUT_REG_DEPTH_REGID(posz_regid));
- } else {
- OUT_RING(ring, 0x00000000);
- }
+ 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);
- OUT_RING(ring, A3XX_SP_FS_MRT_REG_REGID(color_regid) |
- COND(fp->key.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));
+ 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 (key.binning_pass) {
+ 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] = {0}, flatshade[2] = {0};
+ 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 = next_varying(fp, j)) < (int)fp->inputs_count; ) {
- uint32_t interp = fp->inputs[j].interpolate;
- if ((interp == TGSI_INTERPOLATE_CONSTANT) ||
- ((interp == TGSI_INTERPOLATE_COLOR) && rasterflat)) {
- /* TODO might be cleaner to just +8 in SP_VS_VPC_DST_REG
- * instead.. rather than -8 everywhere else..
- */
- uint32_t loc = fp->inputs[j].inloc - 8;
+ for (j = -1; (j = ir3_next_varying(fp, j)) < (int)fp->inputs_count; ) {
- /* currently assuming varyings aligned to 4 (not
- * packed):
- */
- debug_assert((loc % 4) == 0);
+ /* 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 ((fp->inputs[j].interpolate == INTERP_QUALIFIER_FLAT) ||
+ (fp->inputs[j].rasterflat && 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);
}
}
+
+ 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) {
+ 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, vinterp[3]); /* VPC_VARYING_INTERP[3].MODE */
OUT_PKT0(ring, REG_A3XX_VPC_VARYING_PS_REPL_MODE(0), 4);
- OUT_RING(ring, fp->shader->vpsrepl[0]); /* VPC_VARYING_PS_REPL[0].MODE */
- OUT_RING(ring, fp->shader->vpsrepl[1]); /* VPC_VARYING_PS_REPL[1].MODE */
- OUT_RING(ring, fp->shader->vpsrepl[2]); /* VPC_VARYING_PS_REPL[2].MODE */
- OUT_RING(ring, fp->shader->vpsrepl[3]); /* VPC_VARYING_PS_REPL[3].MODE */
+ 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 */
}
- 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));
-
if (vpbuffer == BUFFER)
emit_shader(ring, vp);
OUT_PKT0(ring, REG_A3XX_VFD_PERFCOUNTER0_SELECT, 1);
OUT_RING(ring, 0x00000000); /* VFD_PERFCOUNTER0_SELECT */
- if (!key.binning_pass) {
+ if (!emit->key.binning_pass) {
if (fpbuffer == BUFFER)
emit_shader(ring, fp);
}
}
-/* hack.. until we figure out how to deal w/ vpsrepl properly.. */
-static void
-fix_blit_fp(struct pipe_context *pctx)
-{
- struct fd_context *ctx = fd_context(pctx);
- struct fd3_shader_stateobj *so = ctx->blit_prog.fp;
-
- so->shader->vpsrepl[0] = 0x99999999;
- so->shader->vpsrepl[1] = 0x99999999;
- so->shader->vpsrepl[2] = 0x99999999;
- so->shader->vpsrepl[3] = 0x99999999;
-}
-
void
fd3_prog_init(struct pipe_context *pctx)
{
pctx->delete_vs_state = fd3_vp_state_delete;
fd_prog_init(pctx);
-
- fix_blit_fp(pctx);
}
projects / mesa.git / blobdiff