#include "brw_nir.h"
#include "brw_shader.h"
-#include "glsl/glsl_parser_extras.h"
-#include "glsl/nir/glsl_to_nir.h"
+#include "compiler/nir/glsl_to_nir.h"
+#include "compiler/nir/nir_builder.h"
#include "program/prog_to_nir.h"
+static bool
+is_input(nir_intrinsic_instr *intrin)
+{
+ return intrin->intrinsic == nir_intrinsic_load_input ||
+ intrin->intrinsic == nir_intrinsic_load_per_vertex_input;
+}
+
+static bool
+is_output(nir_intrinsic_instr *intrin)
+{
+ return intrin->intrinsic == nir_intrinsic_load_output ||
+ intrin->intrinsic == nir_intrinsic_load_per_vertex_output ||
+ intrin->intrinsic == nir_intrinsic_store_output ||
+ intrin->intrinsic == nir_intrinsic_store_per_vertex_output;
+}
+
+/**
+ * In many cases, we just add the base and offset together, so there's no
+ * reason to keep them separate. Sometimes, combining them is essential:
+ * if a shader only accesses part of a compound variable (such as a matrix
+ * or array), the variable's base may not actually exist in the VUE map.
+ *
+ * This pass adds constant offsets to instr->const_index[0], and resets
+ * the offset source to 0. Non-constant offsets remain unchanged - since
+ * we don't know what part of a compound variable is accessed, we allocate
+ * storage for the entire thing.
+ */
+struct add_const_offset_to_base_params {
+ nir_builder b;
+ nir_variable_mode mode;
+};
+
+static bool
+add_const_offset_to_base_block(nir_block *block, void *closure)
+{
+ struct add_const_offset_to_base_params *params = closure;
+ nir_builder *b = ¶ms->b;
+
+ nir_foreach_instr_safe(block, instr) {
+ if (instr->type != nir_instr_type_intrinsic)
+ continue;
+
+ nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
+
+ if ((params->mode == nir_var_shader_in && is_input(intrin)) ||
+ (params->mode == nir_var_shader_out && is_output(intrin))) {
+ nir_src *offset = nir_get_io_offset_src(intrin);
+ nir_const_value *const_offset = nir_src_as_const_value(*offset);
+
+ if (const_offset) {
+ intrin->const_index[0] += const_offset->u32[0];
+ b->cursor = nir_before_instr(&intrin->instr);
+ nir_instr_rewrite_src(&intrin->instr, offset,
+ nir_src_for_ssa(nir_imm_int(b, 0)));
+ }
+ }
+ }
+ return true;
+}
+
+static void
+add_const_offset_to_base(nir_shader *nir, nir_variable_mode mode)
+{
+ struct add_const_offset_to_base_params params = { .mode = mode };
+
+ nir_foreach_function(nir, f) {
+ if (f->impl) {
+ nir_builder_init(¶ms.b, f->impl);
+ nir_foreach_block(f->impl, add_const_offset_to_base_block, ¶ms);
+ }
+ }
+}
+
static bool
remap_vs_attrs(nir_block *block, void *closure)
{
nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
- /* We set EmitNoIndirect for VS inputs, so there are no indirects. */
- assert(intrin->intrinsic != nir_intrinsic_load_input_indirect);
-
if (intrin->intrinsic == nir_intrinsic_load_input) {
/* Attributes come in a contiguous block, ordered by their
* gl_vert_attrib value. That means we can compute the slot
*/
int attr = intrin->const_index[0];
int slot = _mesa_bitcount_64(inputs_read & BITFIELD64_MASK(attr));
+
intrin->const_index[0] = 4 * slot;
}
}
return true;
}
-static void
-brw_nir_lower_inputs(nir_shader *nir,
- const struct brw_device_info *devinfo,
- bool is_scalar)
+static bool
+remap_inputs_with_vue_map(nir_block *block, void *closure)
{
- switch (nir->stage) {
- case MESA_SHADER_VERTEX:
- /* For now, leave the vec4 backend doing the old method. */
- if (!is_scalar) {
- nir_assign_var_locations(&nir->inputs, &nir->num_inputs,
- type_size_vec4);
- break;
+ const struct brw_vue_map *vue_map = closure;
+
+ nir_foreach_instr(block, instr) {
+ if (instr->type != nir_instr_type_intrinsic)
+ continue;
+
+ nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
+
+ if (intrin->intrinsic == nir_intrinsic_load_input ||
+ intrin->intrinsic == nir_intrinsic_load_per_vertex_input) {
+ int vue_slot = vue_map->varying_to_slot[intrin->const_index[0]];
+ assert(vue_slot != -1);
+ intrin->const_index[0] = vue_slot;
}
+ }
+ return true;
+}
- /* Start with the location of the variable's base. */
- foreach_list_typed(nir_variable, var, node, &nir->inputs) {
- var->data.driver_location = var->data.location;
+struct remap_patch_urb_offsets_state {
+ nir_builder b;
+ const struct brw_vue_map *vue_map;
+};
+
+static bool
+remap_patch_urb_offsets(nir_block *block, void *closure)
+{
+ struct remap_patch_urb_offsets_state *state = closure;
+
+ nir_foreach_instr_safe(block, instr) {
+ if (instr->type != nir_instr_type_intrinsic)
+ continue;
+
+ nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
+
+ gl_shader_stage stage = state->b.shader->stage;
+
+ if ((stage == MESA_SHADER_TESS_CTRL && is_output(intrin)) ||
+ (stage == MESA_SHADER_TESS_EVAL && is_input(intrin))) {
+ int vue_slot = state->vue_map->varying_to_slot[intrin->const_index[0]];
+ assert(vue_slot != -1);
+ intrin->const_index[0] = vue_slot;
+
+ nir_src *vertex = nir_get_io_vertex_index_src(intrin);
+ if (vertex) {
+ nir_const_value *const_vertex = nir_src_as_const_value(*vertex);
+ if (const_vertex) {
+ intrin->const_index[0] += const_vertex->u32[0] *
+ state->vue_map->num_per_vertex_slots;
+ } else {
+ state->b.cursor = nir_before_instr(&intrin->instr);
+
+ /* Multiply by the number of per-vertex slots. */
+ nir_ssa_def *vertex_offset =
+ nir_imul(&state->b,
+ nir_ssa_for_src(&state->b, *vertex, 1),
+ nir_imm_int(&state->b,
+ state->vue_map->num_per_vertex_slots));
+
+ /* Add it to the existing offset */
+ nir_src *offset = nir_get_io_offset_src(intrin);
+ nir_ssa_def *total_offset =
+ nir_iadd(&state->b, vertex_offset,
+ nir_ssa_for_src(&state->b, *offset, 1));
+
+ nir_instr_rewrite_src(&intrin->instr, offset,
+ nir_src_for_ssa(total_offset));
+ }
+ }
}
+ }
+ return true;
+}
- /* Now use nir_lower_io to walk dereference chains. Attribute arrays
- * are loaded as one vec4 per element (or matrix column), so we use
- * type_size_vec4 here.
- */
- nir_lower_io(nir, nir_var_shader_in, type_size_vec4);
+void
+brw_nir_lower_vs_inputs(nir_shader *nir,
+ const struct brw_device_info *devinfo,
+ bool is_scalar,
+ bool use_legacy_snorm_formula,
+ const uint8_t *vs_attrib_wa_flags)
+{
+ /* Start with the location of the variable's base. */
+ foreach_list_typed(nir_variable, var, node, &nir->inputs) {
+ var->data.driver_location = var->data.location;
+ }
+
+ /* Now use nir_lower_io to walk dereference chains. Attribute arrays
+ * are loaded as one vec4 per element (or matrix column), so we use
+ * type_size_vec4 here.
+ */
+ nir_lower_io(nir, nir_var_shader_in, type_size_vec4);
+
+ /* This pass needs actual constants */
+ nir_opt_constant_folding(nir);
+
+ add_const_offset_to_base(nir, nir_var_shader_in);
+
+ brw_nir_apply_attribute_workarounds(nir, use_legacy_snorm_formula,
+ vs_attrib_wa_flags);
+ if (is_scalar) {
/* Finally, translate VERT_ATTRIB_* values into the actual registers.
*
- * Note that we can use nir->info.inputs_read instead of key->inputs_read
- * since the two are identical aside from Gen4-5 edge flag differences.
+ * Note that we can use nir->info.inputs_read instead of
+ * key->inputs_read since the two are identical aside from Gen4-5
+ * edge flag differences.
*/
GLbitfield64 inputs_read = nir->info.inputs_read;
- nir_foreach_overload(nir, overload) {
- if (overload->impl) {
- nir_foreach_block(overload->impl, remap_vs_attrs, &inputs_read);
+
+ nir_foreach_function(nir, function) {
+ if (function->impl) {
+ nir_foreach_block(function->impl, remap_vs_attrs, &inputs_read);
}
}
- break;
- case MESA_SHADER_GEOMETRY: {
- if (!is_scalar) {
- foreach_list_typed(nir_variable, var, node, &nir->inputs) {
- var->data.driver_location = var->data.location;
- }
- } else {
- /* The GLSL linker will have already matched up GS inputs and
- * the outputs of prior stages. The driver does extend VS outputs
- * in some cases, but only for legacy OpenGL or Gen4-5 hardware,
- * neither of which offer geometry shader support. So we can
- * safely ignore that.
- *
- * For SSO pipelines, we use a fixed VUE map layout based on variable
- * locations, so we can rely on rendezvous-by-location to make this
- * work.
- *
- * However, we need to ignore VARYING_SLOT_PRIMITIVE_ID, as it's not
- * written by previous stages and shows up via payload magic.
- */
- struct brw_vue_map input_vue_map;
- GLbitfield64 inputs_read =
- nir->info.inputs_read & ~VARYING_BIT_PRIMITIVE_ID;
- brw_compute_vue_map(devinfo, &input_vue_map, inputs_read,
- nir->info.separate_shader);
-
- /* Start with the slot for the variable's base. */
- foreach_list_typed(nir_variable, var, node, &nir->inputs) {
- assert(input_vue_map.varying_to_slot[var->data.location] != -1);
- var->data.driver_location =
- input_vue_map.varying_to_slot[var->data.location];
+ }
+}
+
+void
+brw_nir_lower_vue_inputs(nir_shader *nir, bool is_scalar,
+ const struct brw_vue_map *vue_map)
+{
+ foreach_list_typed(nir_variable, var, node, &nir->inputs) {
+ var->data.driver_location = var->data.location;
+ }
+
+ /* Inputs are stored in vec4 slots, so use type_size_vec4(). */
+ nir_lower_io(nir, nir_var_shader_in, type_size_vec4);
+
+ if (is_scalar || nir->stage != MESA_SHADER_GEOMETRY) {
+ /* This pass needs actual constants */
+ nir_opt_constant_folding(nir);
+
+ add_const_offset_to_base(nir, nir_var_shader_in);
+
+ nir_foreach_function(nir, function) {
+ if (function->impl) {
+ nir_foreach_block(function->impl, remap_inputs_with_vue_map,
+ (void *) vue_map);
}
+ }
+ }
+}
- /* Inputs are stored in vec4 slots, so use type_size_vec4(). */
- nir_lower_io(nir, nir_var_shader_in, type_size_vec4);
+void
+brw_nir_lower_tes_inputs(nir_shader *nir, const struct brw_vue_map *vue_map)
+{
+ struct remap_patch_urb_offsets_state state;
+ state.vue_map = vue_map;
+
+ foreach_list_typed(nir_variable, var, node, &nir->inputs) {
+ var->data.driver_location = var->data.location;
+ }
+
+ nir_lower_io(nir, nir_var_shader_in, type_size_vec4);
+
+ /* This pass needs actual constants */
+ nir_opt_constant_folding(nir);
+
+ add_const_offset_to_base(nir, nir_var_shader_in);
+
+ nir_foreach_function(nir, function) {
+ if (function->impl) {
+ nir_builder_init(&state.b, function->impl);
+ nir_foreach_block(function->impl, remap_patch_urb_offsets, &state);
}
- break;
}
- case MESA_SHADER_FRAGMENT:
- assert(is_scalar);
- nir_assign_var_locations(&nir->inputs, &nir->num_inputs,
- type_size_scalar);
- break;
- case MESA_SHADER_COMPUTE:
- /* Compute shaders have no inputs. */
- assert(exec_list_is_empty(&nir->inputs));
- break;
- default:
- unreachable("unsupported shader stage");
+}
+
+void
+brw_nir_lower_fs_inputs(nir_shader *nir)
+{
+ nir_assign_var_locations(&nir->inputs, &nir->num_inputs, type_size_scalar);
+ nir_lower_io(nir, nir_var_shader_in, type_size_scalar);
+}
+
+void
+brw_nir_lower_vue_outputs(nir_shader *nir,
+ bool is_scalar)
+{
+ if (is_scalar) {
+ nir_assign_var_locations(&nir->outputs, &nir->num_outputs,
+ type_size_vec4_times_4);
+ nir_lower_io(nir, nir_var_shader_out, type_size_vec4_times_4);
+ } else {
+ nir_foreach_variable(var, &nir->outputs)
+ var->data.driver_location = var->data.location;
+ nir_lower_io(nir, nir_var_shader_out, type_size_vec4);
}
}
-static void
-brw_nir_lower_outputs(nir_shader *nir, bool is_scalar)
+void
+brw_nir_lower_tcs_outputs(nir_shader *nir, const struct brw_vue_map *vue_map)
{
- switch (nir->stage) {
- case MESA_SHADER_VERTEX:
- case MESA_SHADER_GEOMETRY:
- if (is_scalar) {
- nir_assign_var_locations(&nir->outputs, &nir->num_outputs,
- type_size_vec4_times_4);
- nir_lower_io(nir, nir_var_shader_out, type_size_vec4_times_4);
- } else {
- nir_foreach_variable(var, &nir->outputs)
- var->data.driver_location = var->data.location;
+ struct remap_patch_urb_offsets_state state;
+ state.vue_map = vue_map;
+
+ nir_foreach_variable(var, &nir->outputs) {
+ var->data.driver_location = var->data.location;
+ }
+
+ nir_lower_io(nir, nir_var_shader_out, type_size_vec4);
+
+ /* This pass needs actual constants */
+ nir_opt_constant_folding(nir);
+
+ add_const_offset_to_base(nir, nir_var_shader_out);
+
+ nir_foreach_function(nir, function) {
+ if (function->impl) {
+ nir_builder_init(&state.b, function->impl);
+ nir_foreach_block(function->impl, remap_patch_urb_offsets, &state);
}
- break;
- case MESA_SHADER_FRAGMENT:
- nir_assign_var_locations(&nir->outputs, &nir->num_outputs,
- type_size_scalar);
- break;
- case MESA_SHADER_COMPUTE:
- /* Compute shaders have no outputs. */
- assert(exec_list_is_empty(&nir->outputs));
- break;
- default:
- unreachable("unsupported shader stage");
}
}
-static bool
-should_clone_nir()
+void
+brw_nir_lower_fs_outputs(nir_shader *nir)
{
- static int should_clone = -1;
- if (should_clone < 1)
- should_clone = brw_env_var_as_boolean("NIR_TEST_CLONE", false);
+ nir_assign_var_locations(&nir->outputs, &nir->num_outputs,
+ type_size_scalar);
+ nir_lower_io(nir, nir_var_shader_out, type_size_scalar);
+}
- return should_clone;
+static int
+type_size_scalar_bytes(const struct glsl_type *type)
+{
+ return type_size_scalar(type) * 4;
}
-#define _OPT(do_pass) (({ \
- bool this_progress = true; \
- do_pass \
- nir_validate_shader(nir); \
- if (should_clone_nir()) { \
- nir_shader *clone = nir_shader_clone(ralloc_parent(nir), nir); \
- ralloc_free(nir); \
- nir = clone; \
- } \
- this_progress; \
-}))
-
-#define OPT(pass, ...) _OPT( \
- nir_metadata_set_validation_flag(nir); \
- this_progress = pass(nir ,##__VA_ARGS__); \
- if (this_progress) { \
- progress = true; \
- nir_metadata_check_validation_flag(nir); \
- } \
-)
-
-#define OPT_V(pass, ...) _OPT( \
- pass(nir, ##__VA_ARGS__); \
-)
+static int
+type_size_vec4_bytes(const struct glsl_type *type)
+{
+ return type_size_vec4(type) * 16;
+}
+
+static void
+brw_nir_lower_uniforms(nir_shader *nir, bool is_scalar)
+{
+ if (is_scalar) {
+ nir_assign_var_locations(&nir->uniforms, &nir->num_uniforms,
+ type_size_scalar_bytes);
+ nir_lower_io(nir, nir_var_uniform, type_size_scalar_bytes);
+ } else {
+ nir_assign_var_locations(&nir->uniforms, &nir->num_uniforms,
+ type_size_vec4_bytes);
+ nir_lower_io(nir, nir_var_uniform, type_size_vec4_bytes);
+ }
+}
+
+void
+brw_nir_lower_cs_shared(nir_shader *nir)
+{
+ nir_assign_var_locations(&nir->shared, &nir->num_shared,
+ type_size_scalar_bytes);
+ nir_lower_io(nir, nir_var_shared, type_size_scalar_bytes);
+}
+
+#define OPT(pass, ...) ({ \
+ bool this_progress = false; \
+ NIR_PASS(this_progress, nir, pass, ##__VA_ARGS__); \
+ if (this_progress) \
+ progress = true; \
+ this_progress; \
+})
+
+#define OPT_V(pass, ...) NIR_PASS_V(nir, pass, ##__VA_ARGS__)
static nir_shader *
nir_optimize(nir_shader *nir, bool is_scalar)
* is_scalar = true to scalarize everything prior to code gen.
*/
nir_shader *
-brw_preprocess_nir(nir_shader *nir, bool is_scalar)
+brw_preprocess_nir(const struct brw_compiler *compiler, nir_shader *nir)
{
bool progress; /* Written by OPT and OPT_V */
(void)progress;
+ const bool is_scalar = compiler->scalar_stage[nir->stage];
+
if (nir->stage == MESA_SHADER_GEOMETRY)
OPT(nir_lower_gs_intrinsics);
+ if (compiler->precise_trig)
+ OPT(brw_nir_apply_trig_workarounds);
+
static const nir_lower_tex_options tex_options = {
.lower_txp = ~0,
};
nir = nir_optimize(nir, is_scalar);
+ if (is_scalar) {
+ OPT_V(nir_lower_load_const_to_scalar);
+ }
+
/* Lower a bunch of stuff */
OPT_V(nir_lower_var_copies);
/* Get rid of split copies */
nir = nir_optimize(nir, is_scalar);
- OPT(nir_remove_dead_variables);
+ OPT(nir_remove_dead_variables, nir_var_local);
return nir;
}
-/* Lowers inputs, outputs, uniforms, and samplers for i965
- *
- * This function does all of the standard lowering prior to post-processing.
- * The lowering done is highly gen, stage, and backend-specific. The
- * shader_prog parameter is optional and is used only for lowering sampler
- * derefs and atomics for GLSL shaders.
- */
-nir_shader *
-brw_lower_nir(nir_shader *nir,
- const struct brw_device_info *devinfo,
- const struct gl_shader_program *shader_prog,
- bool is_scalar)
-{
- bool progress; /* Written by OPT and OPT_V */
- (void)progress;
-
- OPT_V(brw_nir_lower_inputs, devinfo, is_scalar);
- OPT_V(brw_nir_lower_outputs, is_scalar);
- nir_assign_var_locations(&nir->uniforms,
- &nir->num_uniforms,
- is_scalar ? type_size_scalar : type_size_vec4);
- OPT_V(nir_lower_io, nir_var_all, is_scalar ? type_size_scalar : type_size_vec4);
-
- if (shader_prog) {
- OPT_V(nir_lower_samplers, shader_prog);
- }
-
- OPT(nir_lower_system_values);
-
- if (shader_prog) {
- OPT_V(nir_lower_atomics, shader_prog);
- }
-
- return nir_optimize(nir, is_scalar);
-}
-
/* Prepare the given shader for codegen
*
* This function is intended to be called right before going into the actual
bool progress; /* Written by OPT and OPT_V */
(void)progress;
+ nir = nir_optimize(nir, is_scalar);
+
if (devinfo->gen >= 6) {
/* Try and fuse multiply-adds */
OPT(brw_nir_opt_peephole_ffma);
if (unlikely(debug_enabled)) {
/* Re-index SSA defs so we print more sensible numbers. */
- nir_foreach_overload(nir, overload) {
- if (overload->impl)
- nir_index_ssa_defs(overload->impl);
+ nir_foreach_function(nir, function) {
+ if (function->impl)
+ nir_index_ssa_defs(function->impl);
}
fprintf(stderr, "NIR (SSA form) for %s shader:\n",
bool is_scalar)
{
struct gl_context *ctx = &brw->ctx;
- const struct brw_device_info *devinfo = brw->intelScreen->devinfo;
const nir_shader_compiler_options *options =
ctx->Const.ShaderCompilerOptions[stage].NirOptions;
bool progress;
(void)progress;
- nir = brw_preprocess_nir(nir, is_scalar);
- nir = brw_lower_nir(nir, devinfo, shader_prog, is_scalar);
+ nir = brw_preprocess_nir(brw->intelScreen->compiler, nir);
+
+ OPT(nir_lower_system_values);
+ OPT_V(brw_nir_lower_uniforms, is_scalar);
+
+ if (shader_prog) {
+ OPT_V(nir_lower_samplers, shader_prog);
+ OPT_V(nir_lower_atomics, shader_prog);
+ }
return nir;
}
tex_options.saturate_r = key_tex->gl_clamp_mask[2];
}
+ /* Prior to Haswell, we have to fake texture swizzle */
+ for (unsigned s = 0; s < MAX_SAMPLERS; s++) {
+ if (key_tex->swizzles[s] == SWIZZLE_NOOP)
+ continue;
+
+ tex_options.swizzle_result |= (1 << s);
+ for (unsigned c = 0; c < 4; c++)
+ tex_options.swizzles[s][c] = GET_SWZ(key_tex->swizzles[s], c);
+ }
+
if (nir_lower_tex(nir, &tex_options)) {
nir_validate_shader(nir);
nir = nir_optimize(nir, is_scalar);
{
switch (type) {
case nir_type_uint:
+ case nir_type_uint32:
return BRW_REGISTER_TYPE_UD;
case nir_type_bool:
case nir_type_int:
+ case nir_type_bool32:
+ case nir_type_int32:
return BRW_REGISTER_TYPE_D;
case nir_type_float:
+ case nir_type_float32:
return BRW_REGISTER_TYPE_F;
+ case nir_type_float64:
+ return BRW_REGISTER_TYPE_DF;
+ case nir_type_int64:
+ case nir_type_uint64:
+ /* TODO we should only see these in moves, so for now it's ok, but when
+ * we add actual 64-bit integer support we should fix this.
+ */
+ return BRW_REGISTER_TYPE_DF;
default:
unreachable("unknown type");
}
{
switch (type) {
case nir_type_float:
+ case nir_type_float32:
return GLSL_TYPE_FLOAT;
+ case nir_type_float64:
+ return GLSL_TYPE_DOUBLE;
+
case nir_type_int:
+ case nir_type_int32:
return GLSL_TYPE_INT;
case nir_type_uint:
+ case nir_type_uint32:
return GLSL_TYPE_UINT;
default: