static void si_init_shader_ctx(struct si_shader_context *ctx,
struct si_screen *sscreen,
- struct ac_llvm_compiler *compiler);
+ struct ac_llvm_compiler *compiler,
+ unsigned wave_size,
+ bool nir);
static void si_llvm_emit_barrier(const struct lp_build_tgsi_action *action,
struct lp_build_tgsi_context *bld_base,
enum si_arg_regfile regfile, LLVMTypeRef type,
LLVMValueRef *assign, unsigned idx)
{
- MAYBE_UNUSED unsigned actual = add_arg_assign(fninfo, regfile, type, assign);
+ ASSERTED unsigned actual = add_arg_assign(fninfo, regfile, type, assign);
assert(actual == idx);
}
LLVMValueRef out[4])
{
const struct tgsi_shader_info *info = &ctx->shader->selector->info;
- unsigned vs_blit_property = info->properties[TGSI_PROPERTY_VS_BLIT_SGPRS];
+ unsigned vs_blit_property = info->properties[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD];
if (vs_blit_property) {
LLVMValueRef vertex_id = ctx->abi.vertex_id;
}
+static LLVMValueRef load_tess_level_default(struct si_shader_context *ctx,
+ unsigned semantic_name)
+{
+ LLVMValueRef buf, slot, val[4];
+ int i, offset;
+
+ slot = LLVMConstInt(ctx->i32, SI_HS_CONST_DEFAULT_TESS_LEVELS, 0);
+ buf = LLVMGetParam(ctx->main_fn, ctx->param_rw_buffers);
+ buf = ac_build_load_to_sgpr(&ctx->ac, buf, slot);
+ offset = semantic_name == TGSI_SEMANTIC_TESS_DEFAULT_INNER_LEVEL ? 4 : 0;
+
+ for (i = 0; i < 4; i++)
+ val[i] = buffer_load_const(ctx, buf,
+ LLVMConstInt(ctx->i32, (offset + i) * 4, 0));
+ return ac_build_gather_values(&ctx->ac, val, 4);
+}
+
static LLVMValueRef si_load_tess_level(struct ac_shader_abi *abi,
- unsigned varying_id)
+ unsigned varying_id,
+ bool load_default_state)
{
struct si_shader_context *ctx = si_shader_context_from_abi(abi);
unsigned semantic_name;
+ if (load_default_state) {
+ switch (varying_id) {
+ case VARYING_SLOT_TESS_LEVEL_INNER:
+ semantic_name = TGSI_SEMANTIC_TESS_DEFAULT_INNER_LEVEL;
+ break;
+ case VARYING_SLOT_TESS_LEVEL_OUTER:
+ semantic_name = TGSI_SEMANTIC_TESS_DEFAULT_OUTER_LEVEL;
+ break;
+ default:
+ unreachable("unknown tess level");
+ }
+ return load_tess_level_default(ctx, semantic_name);
+ }
+
switch (varying_id) {
case VARYING_SLOT_TESS_LEVEL_INNER:
semantic_name = TGSI_SEMANTIC_TESSINNER;
value = load_tess_level(ctx, decl->Semantic.Name);
break;
- case TGSI_SEMANTIC_DEFAULT_TESSOUTER_SI:
- case TGSI_SEMANTIC_DEFAULT_TESSINNER_SI:
- {
- LLVMValueRef buf, slot, val[4];
- int i, offset;
-
- slot = LLVMConstInt(ctx->i32, SI_HS_CONST_DEFAULT_TESS_LEVELS, 0);
- buf = LLVMGetParam(ctx->main_fn, ctx->param_rw_buffers);
- buf = ac_build_load_to_sgpr(&ctx->ac, buf, slot);
- offset = decl->Semantic.Name == TGSI_SEMANTIC_DEFAULT_TESSINNER_SI ? 4 : 0;
-
- for (i = 0; i < 4; i++)
- val[i] = buffer_load_const(ctx, buf,
- LLVMConstInt(ctx->i32, (offset + i) * 4, 0));
- value = ac_build_gather_values(&ctx->ac, val, 4);
+ case TGSI_SEMANTIC_TESS_DEFAULT_OUTER_LEVEL:
+ case TGSI_SEMANTIC_TESS_DEFAULT_INNER_LEVEL:
+ value = load_tess_level_default(ctx, decl->Semantic.Name);
break;
- }
case TGSI_SEMANTIC_PRIMID:
value = si_get_primitive_id(ctx, 0);
break;
case TGSI_SEMANTIC_SUBGROUP_SIZE:
- value = LLVMConstInt(ctx->i32, 64, 0);
+ value = LLVMConstInt(ctx->i32, ctx->ac.wave_size, 0);
break;
case TGSI_SEMANTIC_SUBGROUP_INVOCATION:
break;
}
- case TGSI_SEMANTIC_CS_USER_DATA:
+ case TGSI_SEMANTIC_CS_USER_DATA_AMD:
value = LLVMGetParam(ctx->main_fn, ctx->param_cs_user_data);
break;
struct pipe_stream_output_info *so = &sel->so;
LLVMBuilderRef builder = ctx->ac.builder;
int i;
- struct lp_build_if_state if_ctx;
/* Get bits [22:16], i.e. (so_param >> 16) & 127; */
LLVMValueRef so_vtx_count =
/* Emit the streamout code conditionally. This actually avoids
* out-of-bounds buffer access. The hw tells us via the SGPR
* (so_vtx_count) which threads are allowed to emit streamout data. */
- lp_build_if(&if_ctx, &ctx->gallivm, can_emit);
+ ac_build_ifcc(&ctx->ac, can_emit, 6501);
{
/* The buffer offset is computed as follows:
* ByteOffset = streamout_offset[buffer_id]*4 +
&so->output[i], &outputs[reg]);
}
}
- lp_build_endif(&if_ctx);
+ ac_build_endif(&ctx->ac, 6501);
}
static void si_export_param(struct si_shader_context *ctx, unsigned index,
LLVMValueRef cond = LLVMGetParam(ctx->main_fn, ctx->param_vs_state_bits);
cond = LLVMBuildTrunc(ctx->ac.builder, cond, ctx->i1, "");
- struct lp_build_if_state if_ctx;
- lp_build_if(&if_ctx, &ctx->gallivm, cond);
+ ac_build_ifcc(&ctx->ac, cond, 6502);
/* Store clamped colors to alloca variables within the conditional block. */
for (unsigned i = 0; i < noutput; i++) {
addr[i][j]);
}
}
- lp_build_endif(&if_ctx);
+ ac_build_endif(&ctx->ac, 6502);
/* Load clamped colors */
for (unsigned i = 0; i < noutput; i++) {
LLVMValueRef lds_base, lds_inner, lds_outer, byteoffset, buffer;
LLVMValueRef out[6], vec0, vec1, tf_base, inner[4], outer[4];
unsigned stride, outer_comps, inner_comps, i, offset;
- struct lp_build_if_state if_ctx, inner_if_ctx;
/* Add a barrier before loading tess factors from LDS. */
if (!shader->key.part.tcs.epilog.invoc0_tess_factors_are_def)
* This can't jump, because invocation 0 executes this. It should
* at least mask out the loads and stores for other invocations.
*/
- lp_build_if(&if_ctx, &ctx->gallivm,
- LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ,
- invocation_id, ctx->i32_0, ""));
+ ac_build_ifcc(&ctx->ac,
+ LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ,
+ invocation_id, ctx->i32_0, ""), 6503);
/* Determine the layout of one tess factor element in the buffer. */
switch (shader->key.part.tcs.epilog.prim_mode) {
byteoffset = LLVMBuildMul(ctx->ac.builder, rel_patch_id,
LLVMConstInt(ctx->i32, 4 * stride, 0), "");
- lp_build_if(&inner_if_ctx, &ctx->gallivm,
- LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ,
- rel_patch_id, ctx->i32_0, ""));
+ ac_build_ifcc(&ctx->ac,
+ LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ,
+ rel_patch_id, ctx->i32_0, ""), 6504);
/* Store the dynamic HS control word. */
offset = 0;
offset += 4;
}
- lp_build_endif(&inner_if_ctx);
+ ac_build_endif(&ctx->ac, 6504);
/* Store the tessellation factors. */
ac_build_buffer_store_dword(&ctx->ac, buffer, vec0,
}
}
- lp_build_endif(&if_ctx);
+ ac_build_endif(&ctx->ac, 6503);
}
static LLVMValueRef
if (ctx->screen->info.chip_class >= GFX9) {
LLVMBasicBlockRef blocks[2] = {
LLVMGetInsertBlock(builder),
- ctx->merged_wrap_if_state.entry_block
+ ctx->merged_wrap_if_entry_block
};
LLVMValueRef values[2];
- lp_build_endif(&ctx->merged_wrap_if_state);
+ ac_build_endif(&ctx->ac, ctx->merged_wrap_if_label);
values[0] = rel_patch_id;
values[1] = LLVMGetUndef(ctx->i32);
ret = si_insert_input_ptr(ctx, ret,
ctx->param_bindless_samplers_and_images,
8 + SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES);
- if (ctx->screen->info.chip_class >= GFX10) {
+ if (ctx->screen->use_ngg) {
ret = si_insert_input_ptr(ctx, ret, ctx->param_vs_state_bits,
8 + SI_SGPR_VS_STATE_BITS);
}
LLVMValueRef wave_idx = si_unpack_param(ctx, ctx->param_merged_wave_info, 24, 4);
vertex_idx = LLVMBuildOr(ctx->ac.builder, vertex_idx,
LLVMBuildMul(ctx->ac.builder, wave_idx,
- LLVMConstInt(ctx->i32, 64, false), ""), "");
+ LLVMConstInt(ctx->i32, ctx->ac.wave_size, false), ""), "");
lds_base = LLVMBuildMul(ctx->ac.builder, vertex_idx,
LLVMConstInt(ctx->i32, itemsize_dw, 0), "");
}
si_get_gs_wave_id(ctx));
if (ctx->screen->info.chip_class >= GFX9)
- lp_build_endif(&ctx->merged_wrap_if_state);
+ ac_build_endif(&ctx->ac, ctx->merged_wrap_if_label);
}
static void si_llvm_emit_gs_epilogue(struct ac_shader_abi *abi,
}
}
- if (ctx->ac.chip_class <= GFX9 &&
+ if (!ctx->screen->use_ngg_streamout &&
ctx->shader->selector->so.num_outputs)
si_llvm_emit_streamout(ctx, outputs, i, 0);
struct tgsi_shader_info *info = &ctx->shader->selector->info;
struct si_shader *shader = ctx->shader;
- struct lp_build_if_state if_state;
LLVMValueRef soffset = LLVMGetParam(ctx->main_fn,
ctx->param_gs2vs_offset);
LLVMValueRef gs_next_vertex;
if (use_kill) {
ac_build_kill_if_false(&ctx->ac, can_emit);
} else {
- lp_build_if(&if_state, &ctx->gallivm, can_emit);
+ ac_build_ifcc(&ctx->ac, can_emit, 6505);
}
offset = 0;
}
if (!use_kill)
- lp_build_endif(&if_state);
+ ac_build_endif(&ctx->ac, 6505);
}
/* Emit one vertex from the geometry shader */
LLVMAddTargetDependentFunctionAttr(ctx->main_fn,
"no-signed-zeros-fp-math",
"true");
-
- if (ctx->screen->debug_flags & DBG(UNSAFE_MATH)) {
- /* These were copied from some LLVM test. */
- LLVMAddTargetDependentFunctionAttr(ctx->main_fn,
- "less-precise-fpmad",
- "true");
- LLVMAddTargetDependentFunctionAttr(ctx->main_fn,
- "no-infs-fp-math",
- "true");
- LLVMAddTargetDependentFunctionAttr(ctx->main_fn,
- "no-nans-fp-math",
- "true");
- LLVMAddTargetDependentFunctionAttr(ctx->main_fn,
- "unsafe-fp-math",
- "true");
- }
}
static void declare_streamout_params(struct si_shader_context *ctx,
struct pipe_stream_output_info *so,
struct si_function_info *fninfo)
{
- if (ctx->ac.chip_class >= GFX10)
+ if (ctx->screen->use_ngg_streamout)
return;
/* Streamout SGPRs. */
add_arg_assign(fninfo, ARG_VGPR, ctx->i32, &ctx->abi.instance_id);
add_arg(fninfo, ARG_VGPR, ctx->i32); /* unused */
}
- } else if (ctx->screen->info.chip_class == GFX10 &&
- !shader->is_gs_copy_shader) {
- add_arg(fninfo, ARG_VGPR, ctx->i32); /* user vgpr */
+ } else if (ctx->screen->info.chip_class >= GFX10) {
add_arg(fninfo, ARG_VGPR, ctx->i32); /* user vgpr */
+ ctx->param_vs_prim_id = add_arg(fninfo, ARG_VGPR, ctx->i32); /* user vgpr or PrimID (legacy) */
add_arg_assign(fninfo, ARG_VGPR, ctx->i32, &ctx->abi.instance_id);
} else {
add_arg_assign(fninfo, ARG_VGPR, ctx->i32, &ctx->abi.instance_id);
unsigned num_prolog_vgprs = 0;
unsigned type = ctx->type;
unsigned vs_blit_property =
- shader->selector->info.properties[TGSI_PROPERTY_VS_BLIT_SGPRS];
+ shader->selector->info.properties[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD];
si_init_function_info(&fninfo);
ctx->param_block_size = add_arg(&fninfo, ARG_SGPR, v3i32);
unsigned cs_user_data_dwords =
- shader->selector->info.properties[TGSI_PROPERTY_CS_USER_DATA_DWORDS];
+ shader->selector->info.properties[TGSI_PROPERTY_CS_USER_DATA_COMPONENTS_AMD];
if (cs_user_data_dwords) {
ctx->param_cs_user_data = add_arg(&fninfo, ARG_SGPR,
LLVMVectorType(ctx->i32, cs_user_data_dwords));
/* Limit on the stride field for <= GFX7. */
assert(stride < (1 << 14));
- num_records = 64;
+ num_records = ctx->ac.wave_size;
ring = LLVMBuildBitCast(builder, base_ring, v2i64, "");
tmp = LLVMBuildExtractElement(builder, ring, ctx->i32_0, "");
tmp = LLVMBuildAdd(builder, tmp,
LLVMConstInt(ctx->i64,
stream_offset, 0), "");
- stream_offset += stride * 64;
+ stream_offset += stride * ctx->ac.wave_size;
ring = LLVMBuildInsertElement(builder, ring, tmp, ctx->i32_0, "");
ring = LLVMBuildBitCast(builder, ring, ctx->v4i32, "");
.halt_at_entry = screen->options.halt_shaders,
},
.shader_type = tgsi_processor_to_shader_stage(sel->type),
- .wave_size = 64,
+ .wave_size = si_get_shader_wave_size(shader),
.num_parts = num_parts,
.elf_ptrs = part_elfs,
.elf_sizes = part_sizes,
static void si_shader_dump_disassembly(struct si_screen *screen,
const struct si_shader_binary *binary,
enum pipe_shader_type shader_type,
+ unsigned wave_size,
struct pipe_debug_callback *debug,
const char *name, FILE *file)
{
if (!ac_rtld_open(&rtld_binary, (struct ac_rtld_open_info){
.info = &screen->info,
.shader_type = tgsi_processor_to_shader_stage(shader_type),
- .wave_size = 64,
+ .wave_size = wave_size,
.num_parts = 1,
.elf_ptrs = &binary->elf_buffer,
.elf_sizes = &binary->elf_size }))
unsigned max_workgroup_size =
si_get_max_workgroup_size(shader);
lds_per_wave = (conf->lds_size * lds_increment) /
- DIV_ROUND_UP(max_workgroup_size, 64);
+ DIV_ROUND_UP(max_workgroup_size,
+ sscreen->compute_wave_size);
}
break;
default:;
if (screen->options.debug_disassembly)
si_shader_dump_disassembly(screen, &shader->binary,
shader->selector->type,
+ si_get_shader_wave_size(shader),
debug, "main", NULL);
pipe_debug_message(debug, SHADER_INFO,
if (!check_debug_option ||
(si_can_dump_shader(sscreen, shader_type) &&
!(sscreen->debug_flags & DBG(NO_ASM)))) {
+ unsigned wave_size = si_get_shader_wave_size(shader);
+
fprintf(file, "\n%s:\n", si_get_shader_name(shader));
if (shader->prolog)
si_shader_dump_disassembly(sscreen, &shader->prolog->binary,
- shader_type, debug, "prolog", file);
+ shader_type, wave_size, debug, "prolog", file);
if (shader->previous_stage)
si_shader_dump_disassembly(sscreen, &shader->previous_stage->binary,
- shader_type, debug, "previous stage", file);
+ shader_type, wave_size, debug, "previous stage", file);
if (shader->prolog2)
si_shader_dump_disassembly(sscreen, &shader->prolog2->binary,
- shader_type, debug, "prolog2", file);
+ shader_type, wave_size, debug, "prolog2", file);
- si_shader_dump_disassembly(sscreen, &shader->binary, shader_type, debug, "main", file);
+ si_shader_dump_disassembly(sscreen, &shader->binary, shader_type,
+ wave_size, debug, "main", file);
if (shader->epilog)
si_shader_dump_disassembly(sscreen, &shader->epilog->binary,
- shader_type, debug, "epilog", file);
+ shader_type, wave_size, debug, "epilog", file);
fprintf(file, "\n");
}
LLVMModuleRef mod,
struct pipe_debug_callback *debug,
enum pipe_shader_type shader_type,
+ unsigned wave_size,
const char *name,
bool less_optimized)
{
if (!si_replace_shader(count, binary)) {
unsigned r = si_llvm_compile(mod, binary, compiler, debug,
- less_optimized);
+ less_optimized, wave_size);
if (r)
return r;
}
if (!ac_rtld_open(&rtld, (struct ac_rtld_open_info){
.info = &sscreen->info,
.shader_type = tgsi_processor_to_shader_stage(shader_type),
- .wave_size = 64,
+ .wave_size = wave_size,
.num_parts = 1,
.elf_ptrs = &binary->elf_buffer,
.elf_sizes = &binary->elf_size }))
shader->selector = gs_selector;
shader->is_gs_copy_shader = true;
- si_init_shader_ctx(&ctx, sscreen, compiler);
+ si_init_shader_ctx(&ctx, sscreen, compiler,
+ si_get_wave_size(sscreen, PIPE_SHADER_VERTEX, false, false),
+ false);
ctx.shader = shader;
ctx.type = PIPE_SHADER_VERTEX;
/* Fetch the vertex stream ID.*/
LLVMValueRef stream_id;
- if (ctx.ac.chip_class <= GFX9 && gs_selector->so.num_outputs)
+ if (!sscreen->use_ngg_streamout && gs_selector->so.num_outputs)
stream_id = si_unpack_param(&ctx, ctx.param_streamout_config, 24, 2);
else
stream_id = ctx.i32_0;
}
/* Streamout and exports. */
- if (ctx.ac.chip_class <= GFX9 && gs_selector->so.num_outputs) {
+ if (!sscreen->use_ngg_streamout && gs_selector->so.num_outputs) {
si_llvm_emit_streamout(&ctx, outputs,
gsinfo->num_outputs,
stream);
if (si_compile_llvm(sscreen, &ctx.shader->binary,
&ctx.shader->config, ctx.compiler,
ctx.ac.module,
- debug, PIPE_SHADER_GEOMETRY,
+ debug, PIPE_SHADER_GEOMETRY, ctx.ac.wave_size,
"GS Copy Shader", false) == 0) {
if (si_can_dump_shader(sscreen, PIPE_SHADER_GEOMETRY))
fprintf(stderr, "GS Copy Shader:\n");
"part.vs.prolog", f);
fprintf(f, " as_es = %u\n", key->as_es);
fprintf(f, " as_ls = %u\n", key->as_ls);
+ fprintf(f, " as_ngg = %u\n", key->as_ngg);
fprintf(f, " mono.u.vs_export_prim_id = %u\n",
key->mono.u.vs_export_prim_id);
fprintf(f, " opt.vs_as_prim_discard_cs = %u\n",
case PIPE_SHADER_TESS_EVAL:
fprintf(f, " as_es = %u\n", key->as_es);
+ fprintf(f, " as_ngg = %u\n", key->as_ngg);
fprintf(f, " mono.u.vs_export_prim_id = %u\n",
key->mono.u.vs_export_prim_id);
break;
"part.gs.vs_prolog", f);
}
fprintf(f, " part.gs.prolog.tri_strip_adj_fix = %u\n", key->part.gs.prolog.tri_strip_adj_fix);
+ fprintf(f, " part.gs.prolog.gfx9_prev_is_vs = %u\n", key->part.gs.prolog.gfx9_prev_is_vs);
+ fprintf(f, " as_ngg = %u\n", key->as_ngg);
break;
case PIPE_SHADER_COMPUTE:
fprintf(f, " part.ps.prolog.force_linear_center_interp = %u\n", key->part.ps.prolog.force_linear_center_interp);
fprintf(f, " part.ps.prolog.bc_optimize_for_persp = %u\n", key->part.ps.prolog.bc_optimize_for_persp);
fprintf(f, " part.ps.prolog.bc_optimize_for_linear = %u\n", key->part.ps.prolog.bc_optimize_for_linear);
+ fprintf(f, " part.ps.prolog.samplemask_log_ps_iter = %u\n", key->part.ps.prolog.samplemask_log_ps_iter);
fprintf(f, " part.ps.epilog.spi_shader_col_format = 0x%x\n", key->part.ps.epilog.spi_shader_col_format);
fprintf(f, " part.ps.epilog.color_is_int8 = 0x%X\n", key->part.ps.epilog.color_is_int8);
fprintf(f, " part.ps.epilog.color_is_int10 = 0x%X\n", key->part.ps.epilog.color_is_int10);
fprintf(f, " part.ps.epilog.alpha_to_one = %u\n", key->part.ps.epilog.alpha_to_one);
fprintf(f, " part.ps.epilog.poly_line_smoothing = %u\n", key->part.ps.epilog.poly_line_smoothing);
fprintf(f, " part.ps.epilog.clamp_color = %u\n", key->part.ps.epilog.clamp_color);
+ fprintf(f, " mono.u.ps.interpolate_at_sample_force_center = %u\n", key->mono.u.ps.interpolate_at_sample_force_center);
+ fprintf(f, " mono.u.ps.fbfetch_msaa = %u\n", key->mono.u.ps.fbfetch_msaa);
+ fprintf(f, " mono.u.ps.fbfetch_is_1D = %u\n", key->mono.u.ps.fbfetch_is_1D);
+ fprintf(f, " mono.u.ps.fbfetch_layered = %u\n", key->mono.u.ps.fbfetch_layered);
break;
default:
static void si_init_shader_ctx(struct si_shader_context *ctx,
struct si_screen *sscreen,
- struct ac_llvm_compiler *compiler)
+ struct ac_llvm_compiler *compiler,
+ unsigned wave_size,
+ bool nir)
{
struct lp_build_tgsi_context *bld_base;
- si_llvm_context_init(ctx, sscreen, compiler);
+ si_llvm_context_init(ctx, sscreen, compiler, wave_size,
+ nir ? 64 : wave_size);
bld_base = &ctx->bld_base;
bld_base->emit_fetch_funcs[TGSI_FILE_CONSTANT] = fetch_constant;
case PIPE_SHADER_TESS_CTRL:
bld_base->emit_fetch_funcs[TGSI_FILE_INPUT] = fetch_input_tcs;
ctx->abi.load_tess_varyings = si_nir_load_tcs_varyings;
+ ctx->abi.load_tess_level = si_load_tess_level;
bld_base->emit_fetch_funcs[TGSI_FILE_OUTPUT] = fetch_output_tcs;
bld_base->emit_store = store_output_tcs;
ctx->abi.store_tcs_outputs = si_nir_store_output_tcs;
ctx->abi.lookup_interp_param = si_nir_lookup_interp_param;
ctx->abi.load_sample_position = load_sample_position;
ctx->abi.load_sample_mask_in = load_sample_mask_in;
+ ctx->abi.emit_fbfetch = si_nir_emit_fbfetch;
ctx->abi.emit_kill = si_llvm_emit_kill;
break;
case PIPE_SHADER_COMPUTE:
if (shader->key.as_ngg) {
for (unsigned i = 0; i < 4; ++i) {
ctx->gs_curprim_verts[i] =
- lp_build_alloca(&ctx->gallivm, ctx->ac.i32, "");
+ ac_build_alloca(&ctx->ac, ctx->ac.i32, "");
ctx->gs_generated_prims[i] =
- lp_build_alloca(&ctx->gallivm, ctx->ac.i32, "");
+ ac_build_alloca(&ctx->ac, ctx->ac.i32, "");
}
unsigned scratch_size = 8;
}
}
- if (shader->key.as_ngg && ctx->type != PIPE_SHADER_GEOMETRY) {
+ if (ctx->type != PIPE_SHADER_GEOMETRY &&
+ (shader->key.as_ngg && !shader->key.as_es)) {
/* Unconditionally declare scratch space base for streamout and
* vertex compaction. Whether space is actually allocated is
* determined during linking / PM4 creation.
ctx->param_merged_wave_info, 0);
} else if (ctx->type == PIPE_SHADER_TESS_CTRL ||
ctx->type == PIPE_SHADER_GEOMETRY ||
- shader->key.as_ngg) {
+ (shader->key.as_ngg && !shader->key.as_es)) {
LLVMValueRef num_threads;
bool nested_barrier;
if (!shader->is_monolithic ||
(ctx->type == PIPE_SHADER_TESS_EVAL &&
- shader->key.as_ngg))
+ (shader->key.as_ngg && !shader->key.as_es)))
ac_init_exec_full_mask(&ctx->ac);
if (ctx->type == PIPE_SHADER_TESS_CTRL ||
LLVMValueRef ena =
LLVMBuildICmp(ctx->ac.builder, LLVMIntULT,
ac_get_thread_id(&ctx->ac), num_threads, "");
- lp_build_if(&ctx->merged_wrap_if_state, &ctx->gallivm, ena);
+
+ ctx->merged_wrap_if_entry_block = LLVMGetInsertBlock(ctx->ac.builder);
+ ctx->merged_wrap_if_label = 11500;
+ ac_build_ifcc(&ctx->ac, ena, ctx->merged_wrap_if_label);
if (nested_barrier) {
/* Execute a barrier before the second shader in
LLVMTypeRef function_type;
unsigned num_first_params;
unsigned num_out, initial_num_out;
- MAYBE_UNUSED unsigned num_out_sgpr; /* used in debug checks */
- MAYBE_UNUSED unsigned initial_num_out_sgpr; /* used in debug checks */
+ ASSERTED unsigned num_out_sgpr; /* used in debug checks */
+ ASSERTED unsigned initial_num_out_sgpr; /* used in debug checks */
unsigned num_sgprs, num_vgprs;
unsigned gprs;
- struct lp_build_if_state if_state;
si_init_function_info(&fninfo);
initial_num_out_sgpr = num_out_sgpr;
/* Now chain the parts. */
- LLVMValueRef ret;
+ LLVMValueRef ret = NULL;
for (unsigned part = 0; part < num_parts; ++part) {
LLVMValueRef in[48];
LLVMTypeRef ret_type;
LLVMConstInt(ctx->i32, 0x7f, 0), "");
ena = LLVMBuildICmp(builder, LLVMIntULT,
ac_get_thread_id(&ctx->ac), count, "");
- lp_build_if(&if_state, &ctx->gallivm, ena);
+ ac_build_ifcc(&ctx->ac, ena, 6506);
}
/* Derive arguments for the next part from outputs of the
if (is_multi_part_shader(ctx) &&
part + 1 == next_shader_first_part) {
- lp_build_endif(&if_state);
+ ac_build_endif(&ctx->ac, 6506);
/* The second half of the merged shader should use
* the inputs from the toplevel (wrapper) function,
si_dump_streamout(&sel->so);
}
- si_init_shader_ctx(&ctx, sscreen, compiler);
- si_llvm_context_set_tgsi(&ctx, shader);
+ si_init_shader_ctx(&ctx, sscreen, compiler, si_get_shader_wave_size(shader),
+ sel->nir != NULL);
+ si_llvm_context_set_ir(&ctx, shader);
memset(shader->info.vs_output_param_offset, AC_EXP_PARAM_UNDEFINED,
sizeof(shader->info.vs_output_param_offset));
shader_ls.key.mono = shader->key.mono;
shader_ls.key.opt = shader->key.opt;
shader_ls.is_monolithic = true;
- si_llvm_context_set_tgsi(&ctx, &shader_ls);
+ si_llvm_context_set_ir(&ctx, &shader_ls);
if (!si_compile_tgsi_main(&ctx)) {
si_llvm_dispose(&ctx);
struct si_shader shader_es = {};
shader_es.selector = es;
shader_es.key.as_es = 1;
+ shader_es.key.as_ngg = shader->key.as_ngg;
shader_es.key.mono = shader->key.mono;
shader_es.key.opt = shader->key.opt;
shader_es.is_monolithic = true;
- si_llvm_context_set_tgsi(&ctx, &shader_es);
+ si_llvm_context_set_ir(&ctx, &shader_es);
if (!si_compile_tgsi_main(&ctx)) {
si_llvm_dispose(&ctx);
/* Compile to bytecode. */
r = si_compile_llvm(sscreen, &shader->binary, &shader->config, compiler,
- ctx.ac.module, debug, ctx.type,
+ ctx.ac.module, debug, ctx.type, ctx.ac.wave_size,
si_get_shader_name(shader),
si_should_optimize_less(compiler, shader->selector));
si_llvm_dispose(&ctx);
* LLVM 3.9svn has this bug.
*/
if (sel->type == PIPE_SHADER_COMPUTE) {
- unsigned wave_size = 64;
+ unsigned wave_size = sscreen->compute_wave_size;
unsigned max_vgprs = 256;
unsigned max_sgprs = sscreen->info.chip_class >= GFX8 ? 800 : 512;
unsigned max_sgprs_per_wave = 128;
}
struct si_shader_context ctx;
- si_init_shader_ctx(&ctx, sscreen, compiler);
+ si_init_shader_ctx(&ctx, sscreen, compiler,
+ si_get_wave_size(sscreen, type, shader.key.as_ngg,
+ shader.key.as_es),
+ false);
ctx.shader = &shader;
ctx.type = type;
si_llvm_optimize_module(&ctx);
if (si_compile_llvm(sscreen, &result->binary, &result->config, compiler,
- ctx.ac.module, debug, ctx.type, name, false)) {
+ ctx.ac.module, debug, ctx.type, ctx.ac.wave_size,
+ name, false)) {
FREE(result);
result = NULL;
goto out;
key->vs_prolog.num_input_sgprs + i, "");
}
- struct lp_build_if_state wrap_if_state;
LLVMValueRef original_ret = ret;
bool wrapped = false;
+ LLVMBasicBlockRef if_entry_block = NULL;
if (key->vs_prolog.is_monolithic && key->vs_prolog.as_ngg) {
LLVMValueRef num_threads;
num_threads = si_unpack_param(ctx, 3, 0, 8);
ena = LLVMBuildICmp(ctx->ac.builder, LLVMIntULT,
ac_get_thread_id(&ctx->ac), num_threads, "");
- lp_build_if(&wrap_if_state, &ctx->gallivm, ena);
+ if_entry_block = LLVMGetInsertBlock(ctx->ac.builder);
+ ac_build_ifcc(&ctx->ac, ena, 11501);
wrapped = true;
}
}
if (wrapped) {
- lp_build_endif(&wrap_if_state);
+ LLVMBasicBlockRef bbs[2] = {
+ LLVMGetInsertBlock(ctx->ac.builder),
+ if_entry_block,
+ };
+ ac_build_endif(&ctx->ac, 11501);
LLVMValueRef values[2] = {
ret,
original_ret
};
- LLVMBasicBlockRef bbs[2] = {
- wrap_if_state.true_block,
- wrap_if_state.entry_block
- };
ret = ac_build_phi(&ctx->ac, LLVMTypeOf(ret), 2, values, bbs);
}
struct pipe_debug_callback *debug)
{
if (sscreen->info.chip_class >= GFX9) {
- struct si_shader *es_main_part =
- shader->key.part.gs.es->main_shader_part_es;
+ struct si_shader *es_main_part;
+ enum pipe_shader_type es_type = shader->key.part.gs.es->type;
+
+ if (es_type == PIPE_SHADER_TESS_EVAL && shader->key.as_ngg)
+ es_main_part = shader->key.part.gs.es->main_shader_part_ngg_es;
+ else
+ es_main_part = shader->key.part.gs.es->main_shader_part_es;
- if (shader->key.part.gs.es->type == PIPE_SHADER_VERTEX &&
+ if (es_type == PIPE_SHADER_VERTEX &&
!si_get_vs_prolog(sscreen, compiler, shader, debug, es_main_part,
&shader->key.part.gs.vs_prolog))
return false;
shader->config.num_sgprs = MAX2(shader->config.num_sgprs, min_sgprs);
if (shader->selector->type == PIPE_SHADER_COMPUTE &&
- si_get_max_workgroup_size(shader) > 64) {
+ si_get_max_workgroup_size(shader) > sscreen->compute_wave_size) {
si_multiwave_lds_size_workaround(sscreen,
&shader->config.lds_size);
}
projects / mesa.git / blobdiff