return nir_deref_instr_get_variable(nir_src_as_deref(instr->src[0]));
}
-static void scan_instruction(struct tgsi_shader_info *info,
+static void gather_intrinsic_load_deref_info(const nir_shader *nir,
+ const nir_intrinsic_instr *instr,
+ nir_variable *var,
+ struct tgsi_shader_info *info)
+{
+ assert(var && var->data.mode == nir_var_shader_in);
+
+ switch (nir->info.stage) {
+ case MESA_SHADER_VERTEX: {
+ unsigned i = var->data.driver_location;
+ unsigned attrib_count = glsl_count_attribute_slots(var->type, false);
+
+ for (unsigned j = 0; j < attrib_count; j++, i++) {
+ if (glsl_type_is_64bit(glsl_without_array(var->type))) {
+ /* TODO: set usage mask more accurately for doubles */
+ info->input_usage_mask[i] = TGSI_WRITEMASK_XYZW;
+ } else {
+ uint8_t mask = nir_ssa_def_components_read(&instr->dest.ssa);
+ info->input_usage_mask[i] |= mask << var->data.location_frac;
+ }
+ }
+ break;
+ }
+ default: {
+ unsigned semantic_name, semantic_index;
+ tgsi_get_gl_varying_semantic(var->data.location, true,
+ &semantic_name, &semantic_index);
+
+ if (semantic_name == TGSI_SEMANTIC_COLOR) {
+ uint8_t mask = nir_ssa_def_components_read(&instr->dest.ssa);
+ info->colors_read |= mask << (semantic_index * 4);
+ }
+ break;
+ }
+ }
+}
+
+static void scan_instruction(const struct nir_shader *nir,
+ struct tgsi_shader_info *info,
nir_instr *instr)
{
if (instr->type == nir_instr_type_alu) {
case nir_intrinsic_load_base_vertex:
info->uses_basevertex = 1;
break;
+ case nir_intrinsic_load_draw_id:
+ info->uses_drawid = 1;
+ break;
case nir_intrinsic_load_primitive_id:
info->uses_primid = 1;
break;
case nir_intrinsic_load_tess_level_outer:
info->reads_tess_factors = true;
break;
- case nir_intrinsic_image_deref_load:
- case nir_intrinsic_image_deref_size:
- case nir_intrinsic_image_deref_samples: {
- nir_variable *var = intrinsic_get_var(intr);
- if (var->data.bindless)
- info->uses_bindless_images = true;
+ case nir_intrinsic_bindless_image_load:
+ info->uses_bindless_images = true;
+ if (nir_intrinsic_image_dim(intr) == GLSL_SAMPLER_DIM_BUF)
+ info->uses_bindless_buffer_load = true;
+ else
+ info->uses_bindless_image_load = true;
+ break;
+ case nir_intrinsic_bindless_image_size:
+ case nir_intrinsic_bindless_image_samples:
+ info->uses_bindless_images = true;
+ break;
+ case nir_intrinsic_bindless_image_store:
+ info->uses_bindless_images = true;
+
+ if (nir_intrinsic_image_dim(intr) == GLSL_SAMPLER_DIM_BUF)
+ info->uses_bindless_buffer_store = true;
+ else
+ info->uses_bindless_image_store = true;
+
+ info->writes_memory = true;
break;
- }
case nir_intrinsic_image_deref_store:
+ info->writes_memory = true;
+ break;
+ case nir_intrinsic_bindless_image_atomic_add:
+ case nir_intrinsic_bindless_image_atomic_min:
+ case nir_intrinsic_bindless_image_atomic_max:
+ case nir_intrinsic_bindless_image_atomic_and:
+ case nir_intrinsic_bindless_image_atomic_or:
+ case nir_intrinsic_bindless_image_atomic_xor:
+ case nir_intrinsic_bindless_image_atomic_exchange:
+ case nir_intrinsic_bindless_image_atomic_comp_swap:
+ info->uses_bindless_images = true;
+
+ if (nir_intrinsic_image_dim(intr) == GLSL_SAMPLER_DIM_BUF)
+ info->uses_bindless_buffer_atomic = true;
+ else
+ info->uses_bindless_image_atomic = true;
+
+ info->writes_memory = true;
+ break;
case nir_intrinsic_image_deref_atomic_add:
case nir_intrinsic_image_deref_atomic_min:
case nir_intrinsic_image_deref_atomic_max:
case nir_intrinsic_image_deref_atomic_or:
case nir_intrinsic_image_deref_atomic_xor:
case nir_intrinsic_image_deref_atomic_exchange:
- case nir_intrinsic_image_deref_atomic_comp_swap: {
- nir_variable *var = intrinsic_get_var(intr);
- if (var->data.bindless)
- info->uses_bindless_images = true;
-
- /* fall-through */
- }
+ case nir_intrinsic_image_deref_atomic_comp_swap:
+ info->writes_memory = true;
+ break;
case nir_intrinsic_store_ssbo:
case nir_intrinsic_ssbo_atomic_add:
case nir_intrinsic_ssbo_atomic_imin:
glsl_get_base_type(glsl_without_array(var->type));
if (mode == nir_var_shader_in) {
+ gather_intrinsic_load_deref_info(nir, intr, var, info);
+
switch (var->data.interpolation) {
case INTERP_MODE_NONE:
if (glsl_base_type_is_integer(base_type))
}
void si_nir_scan_tess_ctrl(const struct nir_shader *nir,
- const struct tgsi_shader_info *info,
struct tgsi_tessctrl_info *out)
{
memset(out, 0, sizeof(*out));
if (nir->info.stage != MESA_SHADER_TESS_CTRL)
return;
- /* Initial value = true. Here the pass will accumulate results from
- * multiple segments surrounded by barriers. If tess factors aren't
- * written at all, it's a shader bug and we don't care if this will be
- * true.
- */
- out->tessfactors_are_def_in_all_invocs = true;
-
- /* TODO: Implement scanning of tess factors, see tgsi backend. */
+ out->tessfactors_are_def_in_all_invocs =
+ ac_are_tessfactors_def_in_all_invocs(nir);
}
void si_nir_scan_shader(const struct nir_shader *nir,
info->properties[TGSI_PROPERTY_NEXT_SHADER] =
pipe_shader_type_from_mesa(nir->info.next_stage);
+ if (nir->info.stage == MESA_SHADER_VERTEX) {
+ info->properties[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION] =
+ nir->info.vs.window_space_position;
+ }
+
if (nir->info.stage == MESA_SHADER_TESS_CTRL) {
info->properties[TGSI_PROPERTY_TCS_VERTICES_OUT] =
nir->info.tess.tcs_vertices_out;
}
}
- if (nir->info.stage == MESA_SHADER_COMPUTE) {
+ if (gl_shader_stage_is_compute(nir->info.stage)) {
info->properties[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH] = nir->info.cs.local_size[0];
info->properties[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT] = nir->info.cs.local_size[1];
info->properties[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH] = nir->info.cs.local_size[2];
* variable->data.driver_location.
*/
if (nir->info.stage == MESA_SHADER_VERTEX) {
- /* TODO: gather the actual input useage and remove this. */
- info->input_usage_mask[i] = TGSI_WRITEMASK_XYZW;
-
- if (glsl_type_is_dual_slot(variable->type)) {
- num_inputs += 2;
-
- /* TODO: gather the actual input useage and remove this. */
- info->input_usage_mask[i+1] = TGSI_WRITEMASK_XYZW;
- } else
+ if (glsl_type_is_dual_slot(glsl_without_array(variable->type)))
num_inputs++;
+
+ num_inputs++;
continue;
}
/* Fragment shader position is a system value. */
if (nir->info.stage == MESA_SHADER_FRAGMENT &&
variable->data.location == VARYING_SLOT_POS) {
- if (variable->data.pixel_center_integer)
+ if (nir->info.fs.pixel_center_integer)
info->properties[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER] =
TGSI_FS_COORD_PIXEL_CENTER_INTEGER;
info->input_interpolate[i] = TGSI_INTERPOLATE_CONSTANT;
break;
}
-
- /* TODO make this more precise */
- if (variable->data.location == VARYING_SLOT_COL0)
- info->colors_read |= 0x0f;
- else if (variable->data.location == VARYING_SLOT_COL1)
- info->colors_read |= 0xf0;
}
}
struct set *ubo_set = _mesa_set_create(NULL, _mesa_hash_pointer,
_mesa_key_pointer_equal);
+ struct set *ssbo_set = _mesa_set_create(NULL, _mesa_hash_pointer,
+ _mesa_key_pointer_equal);
/* Intialise const_file_max[0] */
info->const_file_max[0] = -1;
+ /* The first 8 are reserved for atomic counters using ssbo */
+ unsigned ssbo_idx = 8;
+
unsigned ubo_idx = 1;
nir_foreach_variable(variable, &nir->uniforms) {
const struct glsl_type *type = variable->type;
enum glsl_base_type base_type =
glsl_get_base_type(glsl_without_array(type));
unsigned aoa_size = MAX2(1, glsl_get_aoa_size(type));
+ unsigned loc = variable->data.location;
+ int slot_count = glsl_count_attribute_slots(type, false);
+ int max_slot = MAX2(info->const_file_max[0], (int) loc) + slot_count;
/* Gather buffers declared bitmasks. Note: radeonsi doesn't
* really use the mask (other than ubo_idx == 1 for regular
* so we don't need to worry about the ordering.
*/
if (variable->interface_type != NULL) {
- if (variable->data.mode == nir_var_uniform) {
+ if (variable->data.mode == nir_var_uniform ||
+ variable->data.mode == nir_var_mem_ubo ||
+ variable->data.mode == nir_var_mem_ssbo) {
+
+ struct set *buf_set = variable->data.mode == nir_var_mem_ssbo ?
+ ssbo_set : ubo_set;
unsigned block_count;
if (base_type != GLSL_TYPE_INTERFACE) {
struct set_entry *entry =
- _mesa_set_search(ubo_set, variable->interface_type);
+ _mesa_set_search(buf_set, variable->interface_type);
/* Check if we have already processed
* a member from this ubo.
block_count = aoa_size;
}
- info->const_buffers_declared |= u_bit_consecutive(ubo_idx, block_count);
- ubo_idx += block_count;
+ if (variable->data.mode == nir_var_uniform ||
+ variable->data.mode == nir_var_mem_ubo) {
+ info->const_buffers_declared |= u_bit_consecutive(ubo_idx, block_count);
+ ubo_idx += block_count;
+ } else {
+ assert(variable->data.mode == nir_var_mem_ssbo);
- _mesa_set_add(ubo_set, variable->interface_type);
- }
+ info->shader_buffers_declared |= u_bit_consecutive(ssbo_idx, block_count);
+ ssbo_idx += block_count;
+ }
- if (variable->data.mode == nir_var_shader_storage) {
- /* TODO: make this more accurate */
- info->shader_buffers_declared =
- u_bit_consecutive(0, SI_NUM_SHADER_BUFFERS);
+ _mesa_set_add(buf_set, variable->interface_type);
}
continue;
if (base_type == GLSL_TYPE_SAMPLER) {
if (variable->data.bindless) {
info->const_buffers_declared |= 1;
- info->const_file_max[0] +=
- glsl_count_attribute_slots(type, false);
+ info->const_file_max[0] = max_slot;
} else {
info->samplers_declared |=
u_bit_consecutive(variable->data.binding, aoa_size);
} else if (base_type == GLSL_TYPE_IMAGE) {
if (variable->data.bindless) {
info->const_buffers_declared |= 1;
- info->const_file_max[0] +=
- glsl_count_attribute_slots(type, false);
+ info->const_file_max[0] = max_slot;
} else {
info->images_declared |=
u_bit_consecutive(variable->data.binding, aoa_size);
u_bit_consecutive(0, SI_NUM_CONST_BUFFERS);
} else {
info->const_buffers_declared |= 1;
- info->const_file_max[0] +=
- glsl_count_attribute_slots(type, false);
+ info->const_file_max[0] = max_slot;
}
}
}
_mesa_set_destroy(ubo_set, NULL);
+ _mesa_set_destroy(ssbo_set, NULL);
info->num_written_clipdistance = nir->info.clip_distance_array_size;
info->num_written_culldistance = nir->info.cull_distance_array_size;
func = (struct nir_function *)exec_list_get_head_const(&nir->functions);
nir_foreach_block(block, func->impl) {
nir_foreach_instr(instr, block)
- scan_instruction(info, instr);
+ scan_instruction(nir, info, instr);
}
}
+void
+si_nir_opts(struct nir_shader *nir)
+{
+ bool progress;
+ unsigned lower_flrp =
+ (nir->options->lower_flrp16 ? 16 : 0) |
+ (nir->options->lower_flrp32 ? 32 : 0) |
+ (nir->options->lower_flrp64 ? 64 : 0);
+
+ do {
+ progress = false;
+
+ NIR_PASS_V(nir, nir_lower_vars_to_ssa);
+
+ NIR_PASS(progress, nir, nir_opt_copy_prop_vars);
+ NIR_PASS(progress, nir, nir_opt_dead_write_vars);
+
+ NIR_PASS_V(nir, nir_lower_alu_to_scalar, NULL);
+ NIR_PASS_V(nir, nir_lower_phis_to_scalar);
+
+ /* (Constant) copy propagation is needed for txf with offsets. */
+ NIR_PASS(progress, nir, nir_copy_prop);
+ NIR_PASS(progress, nir, nir_opt_remove_phis);
+ NIR_PASS(progress, nir, nir_opt_dce);
+ if (nir_opt_trivial_continues(nir)) {
+ progress = true;
+ NIR_PASS(progress, nir, nir_copy_prop);
+ NIR_PASS(progress, nir, nir_opt_dce);
+ }
+ NIR_PASS(progress, nir, nir_opt_if, true);
+ NIR_PASS(progress, nir, nir_opt_dead_cf);
+ NIR_PASS(progress, nir, nir_opt_cse);
+ NIR_PASS(progress, nir, nir_opt_peephole_select, 8, true, true);
+
+ /* Needed for algebraic lowering */
+ NIR_PASS(progress, nir, nir_opt_algebraic);
+ NIR_PASS(progress, nir, nir_opt_constant_folding);
+
+ if (lower_flrp != 0) {
+ bool lower_flrp_progress = false;
+
+ NIR_PASS(lower_flrp_progress, nir, nir_lower_flrp,
+ lower_flrp,
+ false /* always_precise */,
+ nir->options->lower_ffma);
+ if (lower_flrp_progress) {
+ NIR_PASS(progress, nir,
+ nir_opt_constant_folding);
+ progress = true;
+ }
+
+ /* Nothing should rematerialize any flrps, so we only
+ * need to do this lowering once.
+ */
+ lower_flrp = 0;
+ }
+
+ NIR_PASS(progress, nir, nir_opt_undef);
+ NIR_PASS(progress, nir, nir_opt_conditional_discard);
+ if (nir->options->max_unroll_iterations) {
+ NIR_PASS(progress, nir, nir_opt_loop_unroll, 0);
+ }
+ } while (progress);
+}
+
/**
* Perform "lowering" operations on the NIR that are run once when the shader
* selector is created.
* - ensure constant offsets for texture instructions are folded
* and copy-propagated
*/
- NIR_PASS_V(sel->nir, nir_lower_returns);
- NIR_PASS_V(sel->nir, nir_lower_vars_to_ssa);
- NIR_PASS_V(sel->nir, nir_lower_alu_to_scalar);
- NIR_PASS_V(sel->nir, nir_lower_phis_to_scalar);
static const struct nir_lower_tex_options lower_tex_options = {
.lower_txp = ~0u,
ac_lower_indirect_derefs(sel->nir, sel->screen->info.chip_class);
- NIR_PASS_V(sel->nir, nir_lower_load_const_to_scalar);
+ si_nir_opts(sel->nir);
- bool progress;
- do {
- progress = false;
+ NIR_PASS_V(sel->nir, nir_lower_bool_to_int32);
- /* (Constant) copy propagation is needed for txf with offsets. */
- NIR_PASS(progress, sel->nir, nir_copy_prop);
- NIR_PASS(progress, sel->nir, nir_opt_remove_phis);
- NIR_PASS(progress, sel->nir, nir_opt_dce);
- if (nir_opt_trivial_continues(sel->nir)) {
- progress = true;
- NIR_PASS(progress, sel->nir, nir_copy_prop);
- NIR_PASS(progress, sel->nir, nir_opt_dce);
- }
- NIR_PASS(progress, sel->nir, nir_opt_if);
- NIR_PASS(progress, sel->nir, nir_opt_dead_cf);
- NIR_PASS(progress, sel->nir, nir_opt_cse);
- NIR_PASS(progress, sel->nir, nir_opt_peephole_select, 8);
-
- /* Needed for algebraic lowering */
- NIR_PASS(progress, sel->nir, nir_opt_algebraic);
- NIR_PASS(progress, sel->nir, nir_opt_constant_folding);
-
- NIR_PASS(progress, sel->nir, nir_opt_undef);
- NIR_PASS(progress, sel->nir, nir_opt_conditional_discard);
- if (sel->nir->options->max_unroll_iterations) {
- NIR_PASS(progress, sel->nir, nir_opt_loop_unroll, 0);
- }
- } while (progress);
+ /* Strip the resulting shader so that the shader cache is more likely
+ * to hit from other similar shaders.
+ */
+ nir_strip(sel->nir);
}
static void declare_nir_input_vs(struct si_shader_context *ctx,
bool write, bool bindless)
{
struct si_shader_context *ctx = si_shader_context_from_abi(abi);
+ const struct tgsi_shader_info *info = &ctx->shader->selector->info;
LLVMBuilderRef builder = ctx->ac.builder;
- LLVMValueRef list = LLVMGetParam(ctx->main_fn, ctx->param_samplers_and_images);
- LLVMValueRef index;
+ unsigned const_index = base_index + constant_index;
+ bool dcc_off = write;
+
+ /* TODO: images_store and images_atomic are not set */
+ if (!dynamic_index && image &&
+ (info->images_store | info->images_atomic) & (1 << const_index))
+ dcc_off = true;
assert(!descriptor_set);
+ assert(!image || desc_type == AC_DESC_IMAGE || desc_type == AC_DESC_BUFFER);
- dynamic_index = dynamic_index ? dynamic_index : ctx->ac.i32_0;
- index = LLVMBuildAdd(builder, dynamic_index,
- LLVMConstInt(ctx->ac.i32, base_index + constant_index, false),
- "");
+ if (bindless) {
+ LLVMValueRef list =
+ LLVMGetParam(ctx->main_fn, ctx->param_bindless_samplers_and_images);
- if (image) {
- assert(desc_type == AC_DESC_IMAGE || desc_type == AC_DESC_BUFFER);
- assert(base_index + constant_index < ctx->num_images);
+ /* dynamic_index is the bindless handle */
+ if (image) {
+ /* For simplicity, bindless image descriptors use fixed
+ * 16-dword slots for now.
+ */
+ dynamic_index = LLVMBuildMul(ctx->ac.builder, dynamic_index,
+ LLVMConstInt(ctx->i32, 2, 0), "");
- if (dynamic_index)
- index = si_llvm_bound_index(ctx, index, ctx->num_images);
+ return si_load_image_desc(ctx, list, dynamic_index, desc_type,
+ dcc_off, true);
+ }
+
+ /* Since bindless handle arithmetic can contain an unsigned integer
+ * wraparound and si_load_sampler_desc assumes there isn't any,
+ * use GEP without "inbounds" (inside ac_build_pointer_add)
+ * to prevent incorrect code generation and hangs.
+ */
+ dynamic_index = LLVMBuildMul(ctx->ac.builder, dynamic_index,
+ LLVMConstInt(ctx->i32, 2, 0), "");
+ list = ac_build_pointer_add(&ctx->ac, list, dynamic_index);
+ return si_load_sampler_desc(ctx, list, ctx->i32_0, desc_type);
+ }
+
+ unsigned num_slots = image ? ctx->num_images : ctx->num_samplers;
+ assert(const_index < num_slots);
+
+ LLVMValueRef list = LLVMGetParam(ctx->main_fn, ctx->param_samplers_and_images);
+ LLVMValueRef index = LLVMConstInt(ctx->ac.i32, const_index, false);
+
+ if (dynamic_index) {
+ index = LLVMBuildAdd(builder, index, dynamic_index, "");
+
+ /* From the GL_ARB_shader_image_load_store extension spec:
+ *
+ * If a shader performs an image load, store, or atomic
+ * operation using an image variable declared as an array,
+ * and if the index used to select an individual element is
+ * negative or greater than or equal to the size of the
+ * array, the results of the operation are undefined but may
+ * not lead to termination.
+ */
+ index = si_llvm_bound_index(ctx, index, num_slots);
+ }
+ if (image) {
index = LLVMBuildSub(ctx->ac.builder,
LLVMConstInt(ctx->i32, SI_NUM_IMAGES - 1, 0),
index, "");
-
- /* TODO: be smarter about when we use dcc_off */
- return si_load_image_desc(ctx, list, index, desc_type, write, bindless);
+ return si_load_image_desc(ctx, list, index, desc_type, dcc_off, false);
}
- assert(base_index + constant_index < ctx->num_samplers);
-
- if (dynamic_index)
- index = si_llvm_bound_index(ctx, index, ctx->num_samplers);
-
index = LLVMBuildAdd(ctx->ac.builder, index,
LLVMConstInt(ctx->i32, SI_NUM_IMAGES / 2, 0), "");
-
- if (bindless) {
- /* Since bindless handle arithmetic can contain an unsigned integer
- * wraparound and si_load_sampler_desc assumes there isn't any,
- * use GEP without "inbounds" (inside ac_build_pointer_add)
- * to prevent incorrect code generation and hangs.
- */
- index = LLVMBuildMul(ctx->ac.builder, index, LLVMConstInt(ctx->i32, 2, 0), "");
- list = ac_build_pointer_add(&ctx->ac, list, index);
- index = ctx->i32_0;
- }
return si_load_sampler_desc(ctx, list, index, desc_type);
}
LLVMValueRef data[4];
unsigned loc = variable->data.location;
+ if (loc >= VARYING_SLOT_VAR0 && nir->info.stage == MESA_SHADER_FRAGMENT)
+ ctx->abi.fs_input_attr_indices[loc - VARYING_SLOT_VAR0] = input_idx / 4;
+
for (unsigned i = 0; i < attrib_count; i++) {
/* Packed components share the same location so skip
* them if we have already processed the location.
ctx->num_images = util_last_bit(info->images_declared);
if (ctx->shader->selector->info.properties[TGSI_PROPERTY_CS_LOCAL_SIZE]) {
- assert(nir->info.stage == MESA_SHADER_COMPUTE);
+ assert(gl_shader_stage_is_compute(nir->info.stage));
si_declare_compute_memory(ctx);
}
ac_nir_translate(&ctx->ac, &ctx->abi, nir);