static LLVMValueRef get_wave_id_in_tg(struct si_shader_context *ctx)
{
- return si_unpack_param(ctx, ctx->param_merged_wave_info, 24, 4);
+ return si_unpack_param(ctx, ctx->merged_wave_info, 24, 4);
}
static LLVMValueRef get_tgsize(struct si_shader_context *ctx)
{
- return si_unpack_param(ctx, ctx->param_merged_wave_info, 28, 4);
+ return si_unpack_param(ctx, ctx->merged_wave_info, 28, 4);
}
static LLVMValueRef get_thread_id_in_tg(struct si_shader_context *ctx)
LLVMBuilderRef builder = ctx->ac.builder;
LLVMValueRef tmp;
tmp = LLVMBuildMul(builder, get_wave_id_in_tg(ctx),
- LLVMConstInt(ctx->ac.i32, 64, false), "");
+ LLVMConstInt(ctx->ac.i32, ctx->ac.wave_size, false), "");
return LLVMBuildAdd(builder, tmp, ac_get_thread_id(&ctx->ac), "");
}
static LLVMValueRef ngg_get_vtx_cnt(struct si_shader_context *ctx)
{
- return ac_build_bfe(&ctx->ac, ctx->gs_tg_info,
- LLVMConstInt(ctx->ac.i32, 12, false),
- LLVMConstInt(ctx->ac.i32, 9, false),
- false);
+ return si_unpack_param(ctx, ctx->gs_tg_info, 12, 9);
}
static LLVMValueRef ngg_get_prim_cnt(struct si_shader_context *ctx)
{
- return ac_build_bfe(&ctx->ac, ctx->gs_tg_info,
- LLVMConstInt(ctx->ac.i32, 22, false),
- LLVMConstInt(ctx->ac.i32, 9, false),
- false);
+ return si_unpack_param(ctx, ctx->gs_tg_info, 22, 9);
}
static LLVMValueRef ngg_get_ordered_id(struct si_shader_context *ctx)
{
- return ac_build_bfe(&ctx->ac, ctx->gs_tg_info,
- ctx->i32_0,
- LLVMConstInt(ctx->ac.i32, 11, false),
- false);
+ return si_unpack_param(ctx, ctx->gs_tg_info, 0, 11);
}
static LLVMValueRef ngg_get_query_buf(struct si_shader_context *ctx)
{
- LLVMValueRef buf_ptr = LLVMGetParam(ctx->main_fn,
- ctx->param_rw_buffers);
+ LLVMValueRef buf_ptr = ac_get_arg(&ctx->ac, ctx->rw_buffers);
return ac_build_load_to_sgpr(&ctx->ac, buf_ptr,
LLVMConstInt(ctx->i32, GFX10_GS_QUERY_BUF, false));
struct tgsi_shader_info *info = &ctx->shader->selector->info;
struct pipe_stream_output_info *so = &ctx->shader->selector->so;
LLVMBuilderRef builder = ctx->ac.builder;
- LLVMValueRef buf_ptr = LLVMGetParam(ctx->main_fn, ctx->param_rw_buffers);
+ LLVMValueRef buf_ptr = ac_get_arg(&ctx->ac, ctx->rw_buffers);
LLVMValueRef tid = get_thread_id_in_tg(ctx);
LLVMValueRef tmp, tmp2;
LLVMValueRef i32_2 = LLVMConstInt(ctx->i32, 2, false);
}
}
+static unsigned ngg_nogs_vertex_size(struct si_shader *shader)
+{
+ unsigned lds_vertex_size = 0;
+
+ /* The edgeflag is always stored in the last element that's also
+ * used for padding to reduce LDS bank conflicts. */
+ if (shader->selector->so.num_outputs)
+ lds_vertex_size = 4 * shader->selector->info.num_outputs + 1;
+ if (shader->selector->info.writes_edgeflag)
+ lds_vertex_size = MAX2(lds_vertex_size, 1);
+
+ return lds_vertex_size;
+}
+
/**
* Returns an `[N x i32] addrspace(LDS)*` pointing at contiguous LDS storage
* for the vertex outputs.
LLVMValueRef vtxid)
{
/* The extra dword is used to avoid LDS bank conflicts. */
- unsigned vertex_size = 4 * ctx->shader->selector->info.num_outputs + 1;
+ unsigned vertex_size = ngg_nogs_vertex_size(ctx->shader);
LLVMTypeRef ai32 = LLVMArrayType(ctx->i32, vertex_size);
LLVMTypeRef pai32 = LLVMPointerType(ai32, AC_ADDR_SPACE_LDS);
LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, ctx->esgs_ring, pai32, "");
struct si_shader_context *ctx = si_shader_context_from_abi(abi);
struct si_shader_selector *sel = ctx->shader->selector;
struct tgsi_shader_info *info = &sel->info;
- struct si_shader_output_values *outputs = NULL;
+ struct si_shader_output_values outputs[PIPE_MAX_SHADER_OUTPUTS];
LLVMBuilderRef builder = ctx->ac.builder;
- struct lp_build_if_state if_state;
LLVMValueRef tmp, tmp2;
assert(!ctx->shader->is_gs_copy_shader);
assert(info->num_outputs <= max_outputs);
- outputs = MALLOC((info->num_outputs + 1) * sizeof(outputs[0]));
-
LLVMValueRef vertex_ptr = NULL;
- if (sel->so.num_outputs)
+ if (sel->so.num_outputs || sel->info.writes_edgeflag)
vertex_ptr = ngg_nogs_vertex_ptr(ctx, get_thread_id_in_tg(ctx));
for (unsigned i = 0; i < info->num_outputs; i++) {
outputs[i].semantic_name = info->output_semantic_name[i];
outputs[i].semantic_index = info->output_semantic_index[i];
- /* This is used only by streamout. */
for (unsigned j = 0; j < 4; j++) {
- outputs[i].values[j] =
- LLVMBuildLoad(builder,
- addrs[4 * i + j],
- "");
outputs[i].vertex_stream[j] =
(info->output_streams[i] >> (2 * j)) & 3;
- if (vertex_ptr) {
+ /* TODO: we may store more outputs than streamout needs,
+ * but streamout performance isn't that important.
+ */
+ if (sel->so.num_outputs) {
tmp = ac_build_gep0(&ctx->ac, vertex_ptr,
LLVMConstInt(ctx->i32, 4 * i + j, false));
- tmp2 = ac_to_integer(&ctx->ac, outputs[i].values[j]);
+ tmp2 = LLVMBuildLoad(builder, addrs[4 * i + j], "");
+ tmp2 = ac_to_integer(&ctx->ac, tmp2);
LLVMBuildStore(builder, tmp2, tmp);
}
}
+
+ /* Store the edgeflag at the end (if streamout is enabled) */
+ if (info->output_semantic_name[i] == TGSI_SEMANTIC_EDGEFLAG &&
+ sel->info.writes_edgeflag) {
+ LLVMValueRef edgeflag = LLVMBuildLoad(builder, addrs[4 * i], "");
+ /* The output is a float, but the hw expects a 1-bit integer. */
+ edgeflag = LLVMBuildFPToUI(ctx->ac.builder, edgeflag, ctx->i32, "");
+ edgeflag = ac_build_umin(&ctx->ac, edgeflag, ctx->i32_1);
+
+ tmp = LLVMConstInt(ctx->i32, ngg_nogs_vertex_size(ctx->shader) - 1, 0);
+ tmp = ac_build_gep0(&ctx->ac, vertex_ptr, tmp);
+ LLVMBuildStore(builder, edgeflag, tmp);
+ }
}
- lp_build_endif(&ctx->merged_wrap_if_state);
+ ac_build_endif(&ctx->ac, ctx->merged_wrap_if_label);
- LLVMValueRef prims_in_wave = si_unpack_param(ctx, ctx->param_merged_wave_info, 8, 8);
- LLVMValueRef vtx_in_wave = si_unpack_param(ctx, ctx->param_merged_wave_info, 0, 8);
- LLVMValueRef is_gs_thread = LLVMBuildICmp(builder, LLVMIntULT,
- ac_get_thread_id(&ctx->ac), prims_in_wave, "");
- LLVMValueRef is_es_thread = LLVMBuildICmp(builder, LLVMIntULT,
- ac_get_thread_id(&ctx->ac), vtx_in_wave, "");
+ LLVMValueRef is_gs_thread = si_is_gs_thread(ctx);
+ LLVMValueRef is_es_thread = si_is_es_thread(ctx);
LLVMValueRef vtxindex[] = {
- si_unpack_param(ctx, ctx->param_gs_vtx01_offset, 0, 16),
- si_unpack_param(ctx, ctx->param_gs_vtx01_offset, 16, 16),
- si_unpack_param(ctx, ctx->param_gs_vtx23_offset, 0, 16),
+ si_unpack_param(ctx, ctx->gs_vtx01_offset, 0, 16),
+ si_unpack_param(ctx, ctx->gs_vtx01_offset, 16, 16),
+ si_unpack_param(ctx, ctx->gs_vtx23_offset, 0, 16),
};
/* Determine the number of vertices per primitive. */
LLVMValueRef num_vertices_val;
if (ctx->type == PIPE_SHADER_VERTEX) {
- if (info->properties[TGSI_PROPERTY_VS_BLIT_SGPRS]) {
+ if (info->properties[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD]) {
/* Blits always use axis-aligned rectangles with 3 vertices. */
num_vertices = 3;
num_vertices_val = LLVMConstInt(ctx->i32, 3, 0);
} else {
/* Extract OUTPRIM field. */
- tmp = si_unpack_param(ctx, ctx->param_vs_state_bits, 2, 2);
+ tmp = si_unpack_param(ctx, ctx->vs_state_bits, 2, 2);
num_vertices_val = LLVMBuildAdd(builder, tmp, ctx->i32_1, "");
num_vertices = 3; /* TODO: optimize for points & lines */
}
emitted_prims = nggso.emit[0];
}
+ LLVMValueRef user_edgeflags[3] = {};
+
+ if (sel->info.writes_edgeflag) {
+ /* Streamout already inserted the barrier, so don't insert it again. */
+ if (!sel->so.num_outputs)
+ ac_build_s_barrier(&ctx->ac);
+
+ ac_build_ifcc(&ctx->ac, is_gs_thread, 5400);
+ /* Load edge flags from ES threads and store them into VGPRs in GS threads. */
+ for (unsigned i = 0; i < num_vertices; i++) {
+ tmp = ngg_nogs_vertex_ptr(ctx, vtxindex[i]);
+ tmp2 = LLVMConstInt(ctx->i32, ngg_nogs_vertex_size(ctx->shader) - 1, 0);
+ tmp = ac_build_gep0(&ctx->ac, tmp, tmp2);
+ tmp = LLVMBuildLoad(builder, tmp, "");
+ tmp = LLVMBuildTrunc(builder, tmp, ctx->i1, "");
+
+ user_edgeflags[i] = ac_build_alloca_undef(&ctx->ac, ctx->i1, "");
+ LLVMBuildStore(builder, tmp, user_edgeflags[i]);
+ }
+ ac_build_endif(&ctx->ac, 5400);
+ }
+
/* Copy Primitive IDs from GS threads to the LDS address corresponding
* to the ES thread of the provoking vertex.
*/
if (ctx->type == PIPE_SHADER_VERTEX &&
ctx->shader->key.mono.u.vs_export_prim_id) {
- /* Streamout uses LDS. We need to wait for it before we can reuse it. */
- if (sel->so.num_outputs)
+ /* Streamout and edge flags use LDS. Make it idle, so that we can reuse it. */
+ if (sel->so.num_outputs || sel->info.writes_edgeflag)
ac_build_s_barrier(&ctx->ac);
ac_build_ifcc(&ctx->ac, is_gs_thread, 5400);
/* Extract the PROVOKING_VTX_INDEX field. */
LLVMValueRef provoking_vtx_in_prim =
- si_unpack_param(ctx, ctx->param_vs_state_bits, 4, 2);
+ si_unpack_param(ctx, ctx->vs_state_bits, 4, 2);
/* provoking_vtx_index = vtxindex[provoking_vtx_in_prim]; */
LLVMValueRef indices = ac_build_gather_values(&ctx->ac, vtxindex, 3);
LLVMValueRef provoking_vtx_index =
LLVMBuildExtractElement(builder, indices, provoking_vtx_in_prim, "");
- LLVMBuildStore(builder, ctx->abi.gs_prim_id,
+ LLVMBuildStore(builder, ac_get_arg(&ctx->ac, ctx->args.gs_prim_id),
ac_build_gep0(&ctx->ac, ctx->esgs_ring, provoking_vtx_index));
ac_build_endif(&ctx->ac, 5400);
}
- /* TODO: primitive culling */
-
build_sendmsg_gs_alloc_req(ctx, ngg_get_vtx_cnt(ctx), ngg_get_prim_cnt(ctx));
/* Update query buffer */
/* TODO: this won't catch 96-bit clear_buffer via transform feedback. */
- if (!info->properties[TGSI_PROPERTY_VS_BLIT_SGPRS]) {
- tmp = si_unpack_param(ctx, ctx->param_vs_state_bits, 6, 1);
+ if (!info->properties[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD]) {
+ tmp = si_unpack_param(ctx, ctx->vs_state_bits, 6, 1);
tmp = LLVMBuildTrunc(builder, tmp, ctx->i1, "");
ac_build_ifcc(&ctx->ac, tmp, 5029); /* if (STREAMOUT_QUERY_ENABLED) */
tmp = LLVMBuildICmp(builder, LLVMIntEQ, get_wave_id_in_tg(ctx), ctx->ac.i32_0, "");
* TODO: culling depends on the primitive type, so can have some
* interaction here.
*/
- lp_build_if(&if_state, &ctx->gallivm, is_gs_thread);
+ ac_build_ifcc(&ctx->ac, is_gs_thread, 6001);
{
struct ngg_prim prim = {};
memcpy(prim.index, vtxindex, sizeof(vtxindex[0]) * 3);
for (unsigned i = 0; i < num_vertices; ++i) {
- tmp = LLVMBuildLShr(builder, ctx->abi.gs_invocation_id,
+ if (ctx->type != PIPE_SHADER_VERTEX) {
+ prim.edgeflag[i] = ctx->i1false;
+ continue;
+ }
+
+ tmp = LLVMBuildLShr(builder,
+ ac_get_arg(&ctx->ac, ctx->args.gs_invocation_id),
LLVMConstInt(ctx->ac.i32, 8 + i, false), "");
prim.edgeflag[i] = LLVMBuildTrunc(builder, tmp, ctx->ac.i1, "");
+
+ if (sel->info.writes_edgeflag) {
+ tmp2 = LLVMBuildLoad(builder, user_edgeflags[i], "");
+ prim.edgeflag[i] = LLVMBuildAnd(builder, prim.edgeflag[i],
+ tmp2, "");
+ }
}
build_export_prim(ctx, &prim);
}
- lp_build_endif(&if_state);
+ ac_build_endif(&ctx->ac, 6001);
/* Export per-vertex data (positions and parameters). */
- lp_build_if(&if_state, &ctx->gallivm, is_es_thread);
+ ac_build_ifcc(&ctx->ac, is_es_thread, 6002);
{
unsigned i;
si_llvm_export_vs(ctx, outputs, i);
}
- lp_build_endif(&if_state);
-
- FREE(outputs);
+ ac_build_endif(&ctx->ac, 6002);
}
static LLVMValueRef
return ngg_gs_vertex_ptr(ctx, vertexidx);
}
+static LLVMValueRef
+ngg_gs_get_emit_output_ptr(struct si_shader_context *ctx, LLVMValueRef vertexptr,
+ unsigned out_idx)
+{
+ LLVMValueRef gep_idx[3] = {
+ ctx->ac.i32_0, /* implied C-style array */
+ ctx->ac.i32_0, /* first struct entry */
+ LLVMConstInt(ctx->ac.i32, out_idx, false),
+ };
+ return LLVMBuildGEP(ctx->ac.builder, vertexptr, gep_idx, 3, "");
+}
+
+static LLVMValueRef
+ngg_gs_get_emit_primflag_ptr(struct si_shader_context *ctx, LLVMValueRef vertexptr,
+ unsigned stream)
+{
+ LLVMValueRef gep_idx[3] = {
+ ctx->ac.i32_0, /* implied C-style array */
+ ctx->ac.i32_1, /* second struct entry */
+ LLVMConstInt(ctx->ac.i32, stream, false),
+ };
+ return LLVMBuildGEP(ctx->ac.builder, vertexptr, gep_idx, 3, "");
+}
+
void gfx10_ngg_gs_emit_vertex(struct si_shader_context *ctx,
unsigned stream,
LLVMValueRef *addrs)
const struct si_shader_selector *sel = ctx->shader->selector;
const struct tgsi_shader_info *info = &sel->info;
LLVMBuilderRef builder = ctx->ac.builder;
- struct lp_build_if_state if_state;
LLVMValueRef tmp;
const LLVMValueRef vertexidx =
LLVMBuildLoad(builder, ctx->gs_next_vertex[stream], "");
tmp = LLVMBuildSelect(builder, can_emit, tmp, vertexidx, "");
LLVMBuildStore(builder, tmp, ctx->gs_next_vertex[stream]);
- lp_build_if(&if_state, &ctx->gallivm, can_emit);
+ ac_build_ifcc(&ctx->ac, can_emit, 9001);
const LLVMValueRef vertexptr =
ngg_gs_emit_vertex_ptr(ctx, get_thread_id_in_tg(ctx), vertexidx);
continue;
LLVMValueRef out_val = LLVMBuildLoad(builder, addrs[4 * i + chan], "");
- LLVMValueRef gep_idx[3] = {
- ctx->ac.i32_0, /* implied C-style array */
- ctx->ac.i32_0, /* first entry of struct */
- LLVMConstInt(ctx->ac.i32, out_idx, false),
- };
- LLVMValueRef ptr = LLVMBuildGEP(builder, vertexptr, gep_idx, 3, "");
-
out_val = ac_to_integer(&ctx->ac, out_val);
- LLVMBuildStore(builder, out_val, ptr);
+ LLVMBuildStore(builder, out_val,
+ ngg_gs_get_emit_output_ptr(ctx, vertexptr, out_idx));
}
}
assert(out_idx * 4 == sel->gsvs_vertex_size);
const LLVMValueRef iscompleteprim =
LLVMBuildICmp(builder, LLVMIntUGE, curverts, tmp, "");
+ /* Since the geometry shader emits triangle strips, we need to
+ * track which primitive is odd and swap vertex indices to get
+ * the correct vertex order.
+ */
+ LLVMValueRef is_odd = ctx->i1false;
+ if (stream == 0 && u_vertices_per_prim(sel->gs_output_prim) == 3) {
+ tmp = LLVMBuildAnd(builder, curverts, ctx->i32_1, "");
+ is_odd = LLVMBuildICmp(builder, LLVMIntEQ, tmp, ctx->i32_1, "");
+ }
+
tmp = LLVMBuildAdd(builder, curverts, ctx->ac.i32_1, "");
LLVMBuildStore(builder, tmp, ctx->gs_curprim_verts[stream]);
- LLVMValueRef gep_idx[3] = {
- ctx->ac.i32_0, /* implied C-style array */
- ctx->ac.i32_1, /* second struct entry */
- LLVMConstInt(ctx->ac.i32, stream, false),
- };
- const LLVMValueRef primflagptr =
- LLVMBuildGEP(builder, vertexptr, gep_idx, 3, "");
-
+ /* The per-vertex primitive flag encoding:
+ * bit 0: whether this vertex finishes a primitive
+ * bit 1: whether the primitive is odd (if we are emitting triangle strips)
+ */
tmp = LLVMBuildZExt(builder, iscompleteprim, ctx->ac.i8, "");
- LLVMBuildStore(builder, tmp, primflagptr);
+ tmp = LLVMBuildOr(builder, tmp,
+ LLVMBuildShl(builder,
+ LLVMBuildZExt(builder, is_odd, ctx->ac.i8, ""),
+ ctx->ac.i8_1, ""), "");
+ LLVMBuildStore(builder, tmp, ngg_gs_get_emit_primflag_ptr(ctx, vertexptr, stream));
tmp = LLVMBuildLoad(builder, ctx->gs_generated_prims[stream], "");
tmp = LLVMBuildAdd(builder, tmp, LLVMBuildZExt(builder, iscompleteprim, ctx->ac.i32, ""), "");
LLVMBuildStore(builder, tmp, ctx->gs_generated_prims[stream]);
- lp_build_endif(&if_state);
+ ac_build_endif(&ctx->ac, 9001);
}
void gfx10_ngg_gs_emit_prologue(struct si_shader_context *ctx)
LLVMBuildStore(builder, tmp, ctx->gs_next_vertex[stream]);
tmp = ngg_gs_emit_vertex_ptr(ctx, gsthread, vertexidx);
- LLVMValueRef gep_idx[3] = {
- ctx->ac.i32_0, /* implied C-style array */
- ctx->ac.i32_1, /* second entry of struct */
- LLVMConstInt(ctx->ac.i32, stream, false),
- };
- tmp = LLVMBuildGEP(builder, tmp, gep_idx, 3, "");
- LLVMBuildStore(builder, i8_0, tmp);
+ LLVMBuildStore(builder, i8_0, ngg_gs_get_emit_primflag_ptr(ctx, tmp, stream));
ac_build_endloop(&ctx->ac, 5100);
}
LLVMValueRef numprims =
LLVMBuildLoad(builder, ctx->gs_generated_prims[stream], "");
- numprims = ac_build_reduce(&ctx->ac, numprims, nir_op_iadd, 64);
+ numprims = ac_build_reduce(&ctx->ac, numprims, nir_op_iadd, ctx->ac.wave_size);
tmp = LLVMBuildICmp(builder, LLVMIntEQ, ac_get_thread_id(&ctx->ac), ctx->i32_0, "");
ac_build_ifcc(&ctx->ac, tmp, 5105);
ac_build_endif(&ctx->ac, 5105);
}
- lp_build_endif(&ctx->merged_wrap_if_state);
+ ac_build_endif(&ctx->ac, ctx->merged_wrap_if_label);
ac_build_s_barrier(&ctx->ac);
if (!info->num_stream_output_components[stream])
continue;
- LLVMValueRef gep_idx[3] = {
- ctx->i32_0, /* implicit C-style array */
- ctx->i32_1, /* second value of struct */
- LLVMConstInt(ctx->i32, stream, false),
- };
- tmp = LLVMBuildGEP(builder, vertexptr, gep_idx, 3, "");
- tmp = LLVMBuildLoad(builder, tmp, "");
+ tmp = LLVMBuildLoad(builder, ngg_gs_get_emit_primflag_ptr(ctx, vertexptr, stream), "");
tmp = LLVMBuildTrunc(builder, tmp, ctx->i1, "");
tmp2 = LLVMBuildICmp(builder, LLVMIntULT, tid, num_emit_threads, "");
nggso.prim_enable[stream] = LLVMBuildAnd(builder, tmp, tmp2, "");
}
/* Write shader query data. */
- tmp = si_unpack_param(ctx, ctx->param_vs_state_bits, 6, 1);
+ tmp = si_unpack_param(ctx, ctx->vs_state_bits, 6, 1);
tmp = LLVMBuildTrunc(builder, tmp, ctx->i1, "");
ac_build_ifcc(&ctx->ac, tmp, 5109); /* if (STREAMOUT_QUERY_ENABLED) */
unsigned num_query_comps = sel->so.num_outputs ? 8 : 4;
/* TODO: culling */
/* Determine vertex liveness. */
- LLVMValueRef vertliveptr = lp_build_alloca(&ctx->gallivm, ctx->ac.i1, "vertexlive");
+ LLVMValueRef vertliveptr = ac_build_alloca(&ctx->ac, ctx->ac.i1, "vertexlive");
tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, num_emit_threads, "");
ac_build_ifcc(&ctx->ac, tmp, 5120);
/* Load primitive liveness */
tmp = ngg_gs_vertex_ptr(ctx, primidx);
- LLVMValueRef gep_idx[3] = {
- ctx->ac.i32_0, /* implicit C-style array */
- ctx->ac.i32_1, /* second value of struct */
- ctx->ac.i32_0, /* stream 0 */
- };
- tmp = LLVMBuildGEP(builder, tmp, gep_idx, 3, "");
- tmp = LLVMBuildLoad(builder, tmp, "");
+ tmp = LLVMBuildLoad(builder, ngg_gs_get_emit_primflag_ptr(ctx, tmp, 0), "");
const LLVMValueRef primlive =
LLVMBuildTrunc(builder, tmp, ctx->ac.i1, "");
ac_build_ifcc(&ctx->ac, vertlive, 5130);
{
tmp = ngg_gs_vertex_ptr(ctx, vertlive_scan.result_exclusive);
- LLVMValueRef gep_idx[3] = {
- ctx->ac.i32_0, /* implicit C-style array */
- ctx->ac.i32_1, /* second value of struct */
- ctx->ac.i32_1, /* stream 1 */
- };
- tmp = LLVMBuildGEP(builder, tmp, gep_idx, 3, "");
tmp2 = LLVMBuildTrunc(builder, tid, ctx->ac.i8, "");
- LLVMBuildStore(builder, tmp2, tmp);
+ LLVMBuildStore(builder, tmp2, ngg_gs_get_emit_primflag_ptr(ctx, tmp, 1));
}
ac_build_endif(&ctx->ac, 5130);
tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, num_emit_threads, "");
ac_build_ifcc(&ctx->ac, tmp, 5140);
{
+ LLVMValueRef flags;
struct ngg_prim prim = {};
prim.num_vertices = verts_per_prim;
tmp = ngg_gs_vertex_ptr(ctx, tid);
- LLVMValueRef gep_idx[3] = {
- ctx->ac.i32_0, /* implicit C-style array */
- ctx->ac.i32_1, /* second value of struct */
- ctx->ac.i32_0, /* primflag */
- };
- tmp = LLVMBuildGEP(builder, tmp, gep_idx, 3, "");
- tmp = LLVMBuildLoad(builder, tmp, "");
- prim.isnull = LLVMBuildICmp(builder, LLVMIntEQ, tmp,
- LLVMConstInt(ctx->ac.i8, 0, false), "");
+ flags = LLVMBuildLoad(builder, ngg_gs_get_emit_primflag_ptr(ctx, tmp, 0), "");
+ prim.isnull = LLVMBuildNot(builder, LLVMBuildTrunc(builder, flags, ctx->i1, ""), "");
for (unsigned i = 0; i < verts_per_prim; ++i) {
prim.index[i] = LLVMBuildSub(builder, vertlive_scan.result_exclusive,
prim.edgeflag[i] = ctx->ac.i1false;
}
+ /* Geometry shaders output triangle strips, but NGG expects triangles.
+ * We need to change the vertex order for odd triangles to get correct
+ * front/back facing by swapping 2 vertex indices, but we also have to
+ * keep the provoking vertex in the same place.
+ *
+ * If the first vertex is provoking, swap index 1 and 2.
+ * If the last vertex is provoking, swap index 0 and 1.
+ */
+ if (verts_per_prim == 3) {
+ LLVMValueRef is_odd = LLVMBuildLShr(builder, flags, ctx->ac.i8_1, "");
+ is_odd = LLVMBuildTrunc(builder, is_odd, ctx->i1, "");
+ LLVMValueRef flatshade_first =
+ LLVMBuildICmp(builder, LLVMIntEQ,
+ si_unpack_param(ctx, ctx->vs_state_bits, 4, 2),
+ ctx->i32_0, "");
+
+ struct ngg_prim in = prim;
+ prim.index[0] = LLVMBuildSelect(builder, flatshade_first,
+ in.index[0],
+ LLVMBuildSelect(builder, is_odd,
+ in.index[1], in.index[0], ""), "");
+ prim.index[1] = LLVMBuildSelect(builder, flatshade_first,
+ LLVMBuildSelect(builder, is_odd,
+ in.index[2], in.index[1], ""),
+ LLVMBuildSelect(builder, is_odd,
+ in.index[0], in.index[1], ""), "");
+ prim.index[2] = LLVMBuildSelect(builder, flatshade_first,
+ LLVMBuildSelect(builder, is_odd,
+ in.index[1], in.index[2], ""),
+ in.index[2], "");
+ }
+
build_export_prim(ctx, &prim);
}
ac_build_endif(&ctx->ac, 5140);
tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, vertlive_scan.result_reduce, "");
ac_build_ifcc(&ctx->ac, tmp, 5145);
{
- struct si_shader_output_values *outputs = NULL;
- outputs = MALLOC(info->num_outputs * sizeof(outputs[0]));
+ struct si_shader_output_values outputs[PIPE_MAX_SHADER_OUTPUTS];
tmp = ngg_gs_vertex_ptr(ctx, tid);
- LLVMValueRef gep_idx[3] = {
- ctx->ac.i32_0, /* implicit C-style array */
- ctx->ac.i32_1, /* second value of struct */
- ctx->ac.i32_1, /* stream 1: source data index */
- };
- tmp = LLVMBuildGEP(builder, tmp, gep_idx, 3, "");
- tmp = LLVMBuildLoad(builder, tmp, "");
+ tmp = LLVMBuildLoad(builder, ngg_gs_get_emit_primflag_ptr(ctx, tmp, 1), "");
tmp = LLVMBuildZExt(builder, tmp, ctx->ac.i32, "");
const LLVMValueRef vertexptr = ngg_gs_vertex_ptr(ctx, tmp);
unsigned out_idx = 0;
- gep_idx[1] = ctx->ac.i32_0;
for (unsigned i = 0; i < info->num_outputs; i++) {
outputs[i].semantic_name = info->output_semantic_name[i];
outputs[i].semantic_index = info->output_semantic_index[i];
for (unsigned j = 0; j < 4; j++, out_idx++) {
- gep_idx[2] = LLVMConstInt(ctx->ac.i32, out_idx, false);
- tmp = LLVMBuildGEP(builder, vertexptr, gep_idx, 3, "");
+ tmp = ngg_gs_get_emit_output_ptr(ctx, vertexptr, out_idx);
tmp = LLVMBuildLoad(builder, tmp, "");
outputs[i].values[j] = ac_to_float(&ctx->ac, tmp);
outputs[i].vertex_stream[j] =
}
si_llvm_export_vs(ctx, outputs, info->num_outputs);
-
- FREE(outputs);
}
ac_build_endif(&ctx->ac, 5145);
}
shader->previous_stage_sel ? shader->previous_stage_sel : gs_sel;
const enum pipe_shader_type gs_type = gs_sel->type;
const unsigned gs_num_invocations = MAX2(gs_sel->gs_num_invocations, 1);
- /* TODO: Specialize for known primitive type without GS. */
- const unsigned input_prim = gs_type == PIPE_SHADER_GEOMETRY ?
- gs_sel->info.properties[TGSI_PROPERTY_GS_INPUT_PRIM] :
- PIPE_PRIM_TRIANGLES;
+ const unsigned input_prim = si_get_input_prim(gs_sel);
const bool use_adjacency = input_prim >= PIPE_PRIM_LINES_ADJACENCY &&
input_prim <= PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY;
const unsigned max_verts_per_prim = u_vertices_per_prim(input_prim);
/* We can't allow using the whole LDS, because GS waves compete with
* other shader stages for LDS space.
*
- * Streamout can increase the ESGS buffer size later on, so be more
- * conservative with streamout and use 4K dwords. This may be suboptimal.
- *
- * Otherwise, use the limit of 7K dwords. The reason is that we need
- * to leave some headroom for the max_esverts increase at the end.
- *
* TODO: We should really take the shader's internal LDS use into
* account. The linker will fail if the size is greater than
* 8K dwords.
*/
- const unsigned max_lds_size = (gs_sel->so.num_outputs ? 4 : 7) * 1024 - 128;
+ const unsigned max_lds_size = 8 * 1024 - 768;
const unsigned target_lds_size = max_lds_size;
unsigned esvert_lds_size = 0;
unsigned gsprim_lds_size = 0;
/* All these are per subgroup: */
bool max_vert_out_per_gs_instance = false;
- unsigned max_esverts_base = 256;
+ unsigned max_esverts_base = 128;
unsigned max_gsprims_base = 128; /* default prim group size clamp */
/* Hardware has the following non-natural restrictions on the value
esvert_lds_size = es_sel->esgs_itemsize / 4;
gsprim_lds_size = (gs_sel->gsvs_vertex_size / 4 + 1) * max_out_verts_per_gsprim;
} else {
- /* TODO: This needs to be adjusted once LDS use for compaction
- * after culling is implemented. */
- if (es_sel->so.num_outputs)
- esvert_lds_size = 4 * es_sel->info.num_outputs + 1;
+ /* VS and TES. */
+ /* LDS size for passing data from ES to GS. */
+ esvert_lds_size = ngg_nogs_vertex_size(shader);
+
+ /* LDS size for passing data from GS to ES.
+ * GS stores Primitive IDs into LDS at the address corresponding
+ * to the ES thread of the provoking vertex. All ES threads
+ * load and export PrimitiveID for their thread.
+ */
+ if (gs_sel->type == PIPE_SHADER_VERTEX &&
+ shader->key.mono.u.vs_export_prim_id)
+ esvert_lds_size = MAX2(esvert_lds_size, 1);
}
unsigned max_gsprims = max_gsprims_base;
/* Round up towards full wave sizes for better ALU utilization. */
if (!max_vert_out_per_gs_instance) {
- const unsigned wavesize = 64;
+ const unsigned wavesize = gs_sel->screen->ge_wave_size;
unsigned orig_max_esverts;
unsigned orig_max_gsprims;
do {