#include "radv_shader.h"
#include "radv_shader_helper.h"
#include "radv_shader_args.h"
+#include "radv_debug.h"
#include "nir/nir.h"
#include "sid.h"
static void gfx10_ngg_gs_emit_vertex(struct radv_shader_context *ctx,
unsigned stream,
+ LLVMValueRef vertexidx,
LLVMValueRef *addrs);
static void
-visit_emit_vertex(struct ac_shader_abi *abi, unsigned stream, LLVMValueRef *addrs)
+visit_emit_vertex_with_counter(struct ac_shader_abi *abi, unsigned stream,
+ LLVMValueRef vertexidx, LLVMValueRef *addrs)
{
- LLVMValueRef gs_next_vertex;
- LLVMValueRef can_emit;
unsigned offset = 0;
struct radv_shader_context *ctx = radv_shader_context_from_abi(abi);
if (ctx->args->options->key.vs_common_out.as_ngg) {
- gfx10_ngg_gs_emit_vertex(ctx, stream, addrs);
+ gfx10_ngg_gs_emit_vertex(ctx, stream, vertexidx, addrs);
return;
}
- /* Write vertex attribute values to GSVS ring */
- gs_next_vertex = LLVMBuildLoad(ctx->ac.builder,
- ctx->gs_next_vertex[stream],
- "");
-
- /* If this thread has already emitted the declared maximum number of
- * vertices, don't emit any more: excessive vertex emissions are not
- * supposed to have any effect.
- */
- can_emit = LLVMBuildICmp(ctx->ac.builder, LLVMIntULT, gs_next_vertex,
- LLVMConstInt(ctx->ac.i32, ctx->shader->info.gs.vertices_out, false), "");
-
- bool use_kill = !ctx->args->shader_info->gs.writes_memory;
- if (use_kill)
- ac_build_kill_if_false(&ctx->ac, can_emit);
- else
- ac_build_ifcc(&ctx->ac, can_emit, 6505);
-
for (unsigned i = 0; i < AC_LLVM_MAX_OUTPUTS; ++i) {
unsigned output_usage_mask =
ctx->args->shader_info->gs.output_usage_mask[i];
offset++;
- voffset = LLVMBuildAdd(ctx->ac.builder, voffset, gs_next_vertex, "");
+ voffset = LLVMBuildAdd(ctx->ac.builder, voffset, vertexidx, "");
voffset = LLVMBuildMul(ctx->ac.builder, voffset, LLVMConstInt(ctx->ac.i32, 4, false), "");
out_val = ac_to_integer(&ctx->ac, out_val);
}
}
- gs_next_vertex = LLVMBuildAdd(ctx->ac.builder, gs_next_vertex,
- ctx->ac.i32_1, "");
- LLVMBuildStore(ctx->ac.builder, gs_next_vertex, ctx->gs_next_vertex[stream]);
-
ac_build_sendmsg(&ctx->ac,
AC_SENDMSG_GS_OP_EMIT | AC_SENDMSG_GS | (stream << 8),
ctx->gs_wave_id);
-
- if (!use_kill)
- ac_build_endif(&ctx->ac, 6505);
}
static void
static void gfx10_ngg_gs_emit_vertex(struct radv_shader_context *ctx,
unsigned stream,
+ LLVMValueRef vertexidx,
LLVMValueRef *addrs)
{
LLVMBuilderRef builder = ctx->ac.builder;
LLVMValueRef tmp;
- const LLVMValueRef vertexidx =
- LLVMBuildLoad(builder, ctx->gs_next_vertex[stream], "");
-
- /* If this thread has already emitted the declared maximum number of
- * vertices, skip the write: excessive vertex emissions are not
- * supposed to have any effect.
- */
- const LLVMValueRef can_emit =
- LLVMBuildICmp(builder, LLVMIntULT, vertexidx,
- LLVMConstInt(ctx->ac.i32, ctx->shader->info.gs.vertices_out, false), "");
- ac_build_ifcc(&ctx->ac, can_emit, 9001);
-
- tmp = LLVMBuildAdd(builder, vertexidx, ctx->ac.i32_1, "");
- tmp = LLVMBuildSelect(builder, can_emit, tmp, vertexidx, "");
- LLVMBuildStore(builder, tmp, ctx->gs_next_vertex[stream]);
const LLVMValueRef vertexptr =
ngg_gs_emit_vertex_ptr(ctx, get_thread_id_in_tg(ctx), vertexidx);
}
assert(out_idx * 4 <= ctx->args->shader_info->gs.gsvs_vertex_size);
+ /* Store the current number of emitted vertices to zero out remaining
+ * primitive flags in case the geometry shader doesn't emit the maximum
+ * number of vertices.
+ */
+ tmp = LLVMBuildAdd(builder, vertexidx, ctx->ac.i32_1, "");
+ LLVMBuildStore(builder, tmp, ctx->gs_next_vertex[stream]);
+
/* Determine and store whether this vertex completed a primitive. */
const LLVMValueRef curverts = LLVMBuildLoad(builder, ctx->gs_curprim_verts[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]);
-
- ac_build_endif(&ctx->ac, 9001);
}
static void
ac_llvm_context_init(&ctx.ac, ac_llvm, args->options->chip_class,
args->options->family, float_mode,
- args->shader_info->wave_size, 64);
+ args->shader_info->wave_size,
+ args->shader_info->ballot_bit_size);
ctx.context = ctx.ac.context;
ctx.max_workgroup_size = 0;
ctx.abi.inputs = &ctx.inputs[0];
ctx.abi.emit_outputs = handle_shader_outputs_post;
- ctx.abi.emit_vertex = visit_emit_vertex;
+ ctx.abi.emit_vertex_with_counter = visit_emit_vertex_with_counter;
ctx.abi.load_ubo = radv_load_ubo;
ctx.abi.load_ssbo = radv_load_ssbo;
ctx.abi.load_sampler_desc = radv_get_sampler_desc;
free(elf_buffer);
}
-void
+static void
radv_compile_nir_shader(struct ac_llvm_compiler *ac_llvm,
struct radv_shader_binary **rbinary,
const struct radv_shader_args *args,
LLVMPositionBuilderAtEnd(ctx->ac.builder, end_bb);
}
-void
+static void
radv_compile_gs_copy_shader(struct ac_llvm_compiler *ac_llvm,
struct nir_shader *geom_shader,
struct radv_shader_binary **rbinary,
(*rbinary)->is_gs_copy_shader = true;
}
+
+void
+llvm_compile_shader(struct radv_device *device,
+ unsigned shader_count,
+ struct nir_shader *const *shaders,
+ struct radv_shader_binary **binary,
+ struct radv_shader_args *args)
+{
+ enum ac_target_machine_options tm_options = 0;
+ struct ac_llvm_compiler ac_llvm;
+ bool thread_compiler;
+
+ tm_options |= AC_TM_SUPPORTS_SPILL;
+ if (args->options->check_ir)
+ tm_options |= AC_TM_CHECK_IR;
+ if (device->instance->debug_flags & RADV_DEBUG_NO_LOAD_STORE_OPT)
+ tm_options |= AC_TM_NO_LOAD_STORE_OPT;
+
+ thread_compiler = !(device->instance->debug_flags & RADV_DEBUG_NOTHREADLLVM);
+
+ radv_init_llvm_compiler(&ac_llvm, thread_compiler,
+ args->options->family, tm_options,
+ args->shader_info->wave_size);
+
+ if (args->is_gs_copy_shader) {
+ radv_compile_gs_copy_shader(&ac_llvm, *shaders, binary, args);
+ } else {
+ radv_compile_nir_shader(&ac_llvm, binary, args,
+ shaders, shader_count);
+ }
+
+ radv_destroy_llvm_compiler(&ac_llvm, thread_compiler);
+}
projects / mesa.git / blobdiff