#include "util/u_atomic.h"
#include "util/u_string.h"
#include "util/u_memory.h"
-#include "util/u_format.h"
+#include "util/format/u_format.h"
#include "drm/freedreno_drmif.h"
{
if (v->ir)
ir3_destroy(v->ir);
- if (v->bo)
- fd_bo_del(v->bo);
+ assert(!v->bo);
+ if (v->binning)
+ delete_variant(v->binning);
+ free(v->bin);
free(v);
}
}
for (i = 0; i < v->outputs_count; i++) {
+ /* for ex, VS shaders with tess don't have normal varying outs: */
+ if (!VALIDREG(v->outputs[i].regid))
+ continue;
int32_t regid = v->outputs[i].regid + 3;
if (v->outputs[i].half) {
if (gpu_id < 500) {
assemble_variant(struct ir3_shader_variant *v)
{
struct ir3_compiler *compiler = v->shader->compiler;
- struct shader_info *info = &v->shader->nir->info;
uint32_t gpu_id = compiler->gpu_id;
- uint32_t sz, *bin;
-
- bin = ir3_shader_assemble(v, gpu_id);
- sz = v->info.sizedwords * 4;
-
- v->bo = fd_bo_new(compiler->dev, sz,
- DRM_FREEDRENO_GEM_CACHE_WCOMBINE |
- DRM_FREEDRENO_GEM_TYPE_KMEM,
- "%s:%s", ir3_shader_stage(v->shader), info->name);
- memcpy(fd_bo_map(v->bo), bin, sz);
-
- if (ir3_shader_debug & IR3_DBG_DISASM) {
- struct ir3_shader_key key = v->key;
- printf("disassemble: type=%d, k={bp=%u,cts=%u,hp=%u}\n", v->type,
- v->binning_pass, key.color_two_side, key.half_precision);
- ir3_shader_disasm(v, bin, stdout);
- }
+ v->bin = ir3_shader_assemble(v, gpu_id);
if (shader_debug_enabled(v->shader->type)) {
- fprintf(stderr, "Native code for unnamed %s shader %s:\n",
- _mesa_shader_stage_to_string(v->shader->type),
- v->shader->nir->info.name);
+ fprintf(stdout, "Native code for unnamed %s shader %s:\n",
+ ir3_shader_stage(v), v->shader->nir->info.name);
if (v->shader->type == MESA_SHADER_FRAGMENT)
- fprintf(stderr, "SIMD0\n");
- ir3_shader_disasm(v, bin, stderr);
+ fprintf(stdout, "SIMD0\n");
+ ir3_shader_disasm(v, v->bin, stdout);
}
- free(bin);
-
/* no need to keep the ir around beyond this point: */
ir3_destroy(v->ir);
v->ir = NULL;
}
assemble_variant(v);
- if (!v->bo) {
+ if (!v->bin) {
debug_error("assemble failed!");
goto fail;
}
free(shader);
}
+/**
+ * Creates a bitmask of the used bits of the shader key by this particular
+ * shader. Used by the gallium driver to skip state-dependent recompiles when
+ * possible.
+ */
+static void
+ir3_setup_used_key(struct ir3_shader *shader)
+{
+ nir_shader *nir = shader->nir;
+ struct shader_info *info = &nir->info;
+ struct ir3_shader_key *key = &shader->key_mask;
+
+ /* This key flag is just used to make for a cheaper ir3_shader_key_equal
+ * check in the common case.
+ */
+ key->has_per_samp = true;
+
+ if (info->stage == MESA_SHADER_FRAGMENT) {
+ key->fsaturate_s = ~0;
+ key->fsaturate_t = ~0;
+ key->fsaturate_r = ~0;
+ key->fastc_srgb = ~0;
+ key->fsamples = ~0;
+
+ if (info->inputs_read & VARYING_BITS_COLOR) {
+ key->rasterflat = true;
+ key->color_two_side = true;
+ }
+
+ if ((info->outputs_written & ~(FRAG_RESULT_DEPTH |
+ FRAG_RESULT_STENCIL |
+ FRAG_RESULT_SAMPLE_MASK)) != 0) {
+ key->fclamp_color = true;
+ }
+
+ /* Only used for deciding on behavior of
+ * nir_intrinsic_load_barycentric_sample
+ */
+ key->msaa = info->fs.uses_sample_qualifier;
+ } else {
+ key->tessellation = ~0;
+ key->has_gs = true;
+
+ if (info->outputs_written & VARYING_BITS_COLOR)
+ key->vclamp_color = true;
+
+ if (info->stage == MESA_SHADER_VERTEX) {
+ key->vsaturate_s = ~0;
+ key->vsaturate_t = ~0;
+ key->vsaturate_r = ~0;
+ key->vastc_srgb = ~0;
+ key->vsamples = ~0;
+ }
+ }
+}
+
struct ir3_shader *
-ir3_shader_from_nir(struct ir3_compiler *compiler, nir_shader *nir)
+ir3_shader_from_nir(struct ir3_compiler *compiler, nir_shader *nir,
+ struct ir3_stream_output_info *stream_output)
{
struct ir3_shader *shader = CALLOC_STRUCT(ir3_shader);
shader->compiler = compiler;
shader->id = p_atomic_inc_return(&shader->compiler->shader_count);
shader->type = nir->info.stage;
+ if (stream_output)
+ memcpy(&shader->stream_output, stream_output, sizeof(shader->stream_output));
+
+ if (nir->info.stage == MESA_SHADER_GEOMETRY)
+ NIR_PASS_V(nir, ir3_nir_lower_gs);
NIR_PASS_V(nir, nir_lower_io, nir_var_all, ir3_glsl_type_size,
(nir_lower_io_options)0);
+ if (compiler->gpu_id >= 600 &&
+ nir->info.stage == MESA_SHADER_FRAGMENT &&
+ !(ir3_shader_debug & IR3_DBG_NOFP16))
+ NIR_PASS_V(nir, nir_lower_mediump_outputs);
+
if (nir->info.stage == MESA_SHADER_FRAGMENT) {
/* NOTE: lower load_barycentric_at_sample first, since it
* produces load_barycentric_at_offset:
nir_print_shader(shader->nir, stdout);
}
+ ir3_setup_used_key(shader);
+
return shader;
}
return "GS_HEADER";
case VARYING_SLOT_GS_VERTEX_FLAGS_IR3:
return "GS_VERTEX_FLAGS";
+ case VARYING_SLOT_TCS_HEADER_IR3:
+ return "TCS_HEADER";
default:
return gl_varying_slot_name(so->outputs[i].slot);
}
{
struct ir3 *ir = so->ir;
struct ir3_register *reg;
- const char *type = ir3_shader_stage(so->shader);
+ const char *type = ir3_shader_stage(so);
uint8_t regid;
unsigned i;
- for (i = 0; i < ir->ninputs; i++) {
- if (!ir->inputs[i]) {
- fprintf(out, "; in%d unused\n", i);
- continue;
- }
- reg = ir->inputs[i]->regs[0];
+ foreach_input_n (instr, i, ir) {
+ reg = instr->regs[0];
regid = reg->num;
- fprintf(out, "@in(%sr%d.%c)\tin%d\n",
+ fprintf(out, "@in(%sr%d.%c)\tin%d",
(reg->flags & IR3_REG_HALF) ? "h" : "",
(regid >> 2), "xyzw"[regid & 0x3], i);
+
+ if (reg->wrmask > 0x1)
+ fprintf(out, " (wrmask=0x%x)", reg->wrmask);
+ fprintf(out, "\n");
}
/* print pre-dispatch texture fetches: */
for (i = 0; i < so->num_sampler_prefetch; i++) {
const struct ir3_sampler_prefetch *fetch = &so->sampler_prefetch[i];
- fprintf(out, "@tex(%sr%d.%c)\tsrc=%u, samp=%u, tex=%u, wrmask=%x, cmd=%u\n",
+ fprintf(out, "@tex(%sr%d.%c)\tsrc=%u, samp=%u, tex=%u, wrmask=0x%x, cmd=%u\n",
fetch->half_precision ? "h" : "",
fetch->dst >> 2, "xyzw"[fetch->dst & 0x3],
fetch->src, fetch->samp_id, fetch->tex_id,
fetch->wrmask, fetch->cmd);
}
- for (i = 0; i < ir->noutputs; i++) {
- if (!ir->outputs[i]) {
- fprintf(out, "; out%d unused\n", i);
- continue;
- }
- /* kill shows up as a virtual output.. skip it! */
- if (is_kill(ir->outputs[i]))
- continue;
- reg = ir->outputs[i]->regs[0];
+ foreach_output_n (instr, i, ir) {
+ reg = instr->regs[0];
regid = reg->num;
- fprintf(out, "@out(%sr%d.%c)\tout%d\n",
+ fprintf(out, "@out(%sr%d.%c)\tout%d",
(reg->flags & IR3_REG_HALF) ? "h" : "",
(regid >> 2), "xyzw"[regid & 0x3], i);
+ if (reg->wrmask > 0x1)
+ fprintf(out, " (wrmask=0x%x)", reg->wrmask);
+ fprintf(out, "\n");
}
struct ir3_const_state *const_state = &so->shader->const_state;
fprintf(out, "; %s: outputs:", type);
for (i = 0; i < so->outputs_count; i++) {
uint8_t regid = so->outputs[i].regid;
- fprintf(out, " r%d.%c (%s)",
- (regid >> 2), "xyzw"[regid & 0x3],
+ const char *reg_type = so->outputs[i].half ? "hr" : "r";
+ fprintf(out, " %s%d.%c (%s)",
+ reg_type, (regid >> 2), "xyzw"[regid & 0x3],
output_name(so, i));
}
fprintf(out, "\n");
fprintf(out, "\n");
/* print generic shader info: */
- fprintf(out, "; %s prog %d/%d: %u instructions, %d half, %d full\n",
+ fprintf(out, "; %s prog %d/%d: %u instr, %u nops, %u non-nops, %u mov, %u cov, %u dwords\n",
type, so->shader->id, so->id,
so->info.instrs_count,
- so->info.max_half_reg + 1,
- so->info.max_reg + 1);
-
- fprintf(out, "; %u constlen\n", so->constlen);
+ so->info.nops_count,
+ so->info.instrs_count - so->info.nops_count,
+ so->info.mov_count, so->info.cov_count,
+ so->info.sizedwords);
- fprintf(out, "; %u (ss), %u (sy)\n", so->info.ss, so->info.sy);
+ fprintf(out, "; %s prog %d/%d: %u last-baryf, %d half, %d full, %u constlen\n",
+ type, so->shader->id, so->id,
+ so->info.last_baryf,
+ so->info.max_half_reg + 1,
+ so->info.max_reg + 1,
+ so->constlen);
- fprintf(out, "; max_sun=%u\n", ir->max_sun);
+ fprintf(out, "; %s prog %d/%d: %u sstall, %u (ss), %u (sy), %d max_sun, %d loops\n",
+ type, so->shader->id, so->id,
+ so->info.sstall,
+ so->info.ss,
+ so->info.sy,
+ so->max_sun,
+ so->loops);
/* print shader type specific info: */
switch (so->type) {
break;
case MESA_SHADER_FRAGMENT:
dump_reg(out, "pos (ij_pixel)",
- ir3_find_sysval_regid(so, SYSTEM_VALUE_BARYCENTRIC_PIXEL));
+ ir3_find_sysval_regid(so, SYSTEM_VALUE_BARYCENTRIC_PERSP_PIXEL));
dump_reg(out, "pos (ij_centroid)",
- ir3_find_sysval_regid(so, SYSTEM_VALUE_BARYCENTRIC_CENTROID));
+ ir3_find_sysval_regid(so, SYSTEM_VALUE_BARYCENTRIC_PERSP_CENTROID));
dump_reg(out, "pos (ij_size)",
- ir3_find_sysval_regid(so, SYSTEM_VALUE_BARYCENTRIC_SIZE));
+ ir3_find_sysval_regid(so, SYSTEM_VALUE_BARYCENTRIC_PERSP_SIZE));
dump_output(out, so, FRAG_RESULT_DEPTH, "posz");
if (so->color0_mrt) {
dump_output(out, so, FRAG_RESULT_COLOR, "color");
dump_output(out, so, FRAG_RESULT_DATA6, "data6");
dump_output(out, so, FRAG_RESULT_DATA7, "data7");
}
- /* these two are hard-coded since we don't know how to
- * program them to anything but all 0's...
- */
- if (so->frag_coord)
- fprintf(out, "; fragcoord: r0.x\n");
- if (so->frag_face)
- fprintf(out, "; fragface: hr0.x\n");
+ dump_reg(out, "fragcoord",
+ ir3_find_sysval_regid(so, SYSTEM_VALUE_FRAG_COORD));
+ dump_reg(out, "fragface",
+ ir3_find_sysval_regid(so, SYSTEM_VALUE_FRONT_FACE));
break;
default:
/* TODO */