#include "intel/compiler/brw_compiler.h"
#include "intel/compiler/brw_nir.h"
#include "iris_context.h"
+#include "nir/tgsi_to_nir.h"
-#define ALL_SAMPLERS_XYZW .tex.swizzles[0 ... MAX_SAMPLERS - 1] = 0x688
-#define KEY_INIT .program_string_id = ish->program_id, ALL_SAMPLERS_XYZW
-
-static struct iris_compiled_shader *
-iris_compile_vs(struct iris_context *, struct iris_uncompiled_shader *,
- const struct brw_vs_prog_key *);
-static struct iris_compiled_shader *
-iris_compile_tcs(struct iris_context *, struct iris_uncompiled_shader *,
- const struct brw_tcs_prog_key *);
-static struct iris_compiled_shader *
-iris_compile_tes(struct iris_context *, struct iris_uncompiled_shader *,
- const struct brw_tes_prog_key *);
-static struct iris_compiled_shader *
-iris_compile_gs(struct iris_context *, struct iris_uncompiled_shader *,
- const struct brw_gs_prog_key *);
-static struct iris_compiled_shader *
-iris_compile_fs(struct iris_context *, struct iris_uncompiled_shader *,
- const struct brw_wm_prog_key *, struct brw_vue_map *);
-static struct iris_compiled_shader *
-iris_compile_cs(struct iris_context *, struct iris_uncompiled_shader *,
- const struct brw_cs_prog_key *);
-
+#define KEY_INIT_NO_ID(gen) \
+ .tex.swizzles[0 ... MAX_SAMPLERS - 1] = 0x688, \
+ .tex.compressed_multisample_layout_mask = ~0, \
+ .tex.msaa_16 = (gen >= 9 ? ~0 : 0)
+#define KEY_INIT(gen) .program_string_id = ish->program_id, KEY_INIT_NO_ID(gen)
static unsigned
get_new_program_id(struct iris_screen *screen)
// XXX: need unify_interfaces() at link time...
+/**
+ * Fix an uncompiled shader's stream output info.
+ *
+ * Core Gallium stores output->register_index as a "slot" number, where
+ * slots are assigned consecutively to all outputs in info->outputs_written.
+ * This naive packing of outputs doesn't work for us - we too have slots,
+ * but the layout is defined by the VUE map, which we won't have until we
+ * compile a specific shader variant. So, we remap these and simply store
+ * VARYING_SLOT_* in our copy's output->register_index fields.
+ *
+ * We also fix up VARYING_SLOT_{LAYER,VIEWPORT,PSIZ} to select the Y/Z/W
+ * components of our VUE header. See brw_vue_map.c for the layout.
+ */
static void
-update_so_info(struct pipe_stream_output_info *so_info)
+update_so_info(struct pipe_stream_output_info *so_info,
+ uint64_t outputs_written)
{
+ uint8_t reverse_map[64] = {};
+ unsigned slot = 0;
+ while (outputs_written) {
+ reverse_map[slot++] = u_bit_scan64(&outputs_written);
+ }
+
for (unsigned i = 0; i < so_info->num_outputs; i++) {
struct pipe_stream_output *output = &so_info->output[i];
+ /* Map Gallium's condensed "slots" back to real VARYING_SLOT_* enums */
+ output->register_index = reverse_map[output->register_index];
+
/* The VUE header contains three scalar fields packed together:
* - gl_PointSize is stored in VARYING_SLOT_PSIZ.w
* - gl_Layer is stored in VARYING_SLOT_PSIZ.y
}
/**
- * The pipe->create_[stage]_state() driver hooks.
- *
- * Performs basic NIR preprocessing, records any state dependencies, and
- * returns an iris_uncompiled_shader as the Gallium CSO.
+ * Sets up the starting offsets for the groups of binding table entries
+ * common to all pipeline stages.
*
- * Actual shader compilation to assembly happens later, at first use.
+ * Unused groups are initialized to 0xd0d0d0d0 to make it obvious that they're
+ * unused but also make sure that addition of small offsets to them will
+ * trigger some of our asserts that surface indices are < BRW_MAX_SURFACES.
*/
-static void *
-iris_create_uncompiled_shader(struct pipe_context *ctx,
- nir_shader *nir,
- const struct pipe_stream_output_info *so_info)
+static uint32_t
+assign_common_binding_table_offsets(const struct gen_device_info *devinfo,
+ const struct nir_shader *nir,
+ struct brw_stage_prog_data *prog_data,
+ uint32_t next_binding_table_offset,
+ unsigned num_system_values,
+ unsigned num_cbufs)
{
- struct iris_screen *screen = (struct iris_screen *)ctx->screen;
- const struct gen_device_info *devinfo = &screen->devinfo;
-
- struct iris_uncompiled_shader *ish =
- calloc(1, sizeof(struct iris_uncompiled_shader));
- if (!ish)
- return NULL;
+ const struct shader_info *info = &nir->info;
- nir = brw_preprocess_nir(screen->compiler, nir);
+ unsigned num_textures = util_last_bit(info->textures_used);
- NIR_PASS_V(nir, brw_nir_lower_image_load_store, devinfo);
- NIR_PASS_V(nir, iris_lower_storage_image_derefs);
+ if (num_textures) {
+ prog_data->binding_table.texture_start = next_binding_table_offset;
+ prog_data->binding_table.gather_texture_start = next_binding_table_offset;
+ next_binding_table_offset += num_textures;
+ } else {
+ prog_data->binding_table.texture_start = 0xd0d0d0d0;
+ prog_data->binding_table.gather_texture_start = 0xd0d0d0d0;
+ }
- ish->program_id = get_new_program_id(screen);
- ish->nir = nir;
- if (so_info) {
- memcpy(&ish->stream_output, so_info, sizeof(*so_info));
- update_so_info(&ish->stream_output);
+ if (info->num_images) {
+ prog_data->binding_table.image_start = next_binding_table_offset;
+ next_binding_table_offset += info->num_images;
+ } else {
+ prog_data->binding_table.image_start = 0xd0d0d0d0;
}
- return ish;
-}
+ if (num_cbufs) {
+ //assert(info->num_ubos <= BRW_MAX_UBO);
+ prog_data->binding_table.ubo_start = next_binding_table_offset;
+ next_binding_table_offset += num_cbufs;
+ } else {
+ prog_data->binding_table.ubo_start = 0xd0d0d0d0;
+ }
-static struct iris_uncompiled_shader *
-iris_create_shader_state(struct pipe_context *ctx,
- const struct pipe_shader_state *state)
-{
- assert(state->type == PIPE_SHADER_IR_NIR);
+ if (info->num_ssbos || info->num_abos) {
+ prog_data->binding_table.ssbo_start = next_binding_table_offset;
+ // XXX: see iris_state "wasting 16 binding table slots for ABOs" comment
+ next_binding_table_offset += IRIS_MAX_ABOS + info->num_ssbos;
+ } else {
+ prog_data->binding_table.ssbo_start = 0xd0d0d0d0;
+ }
- return iris_create_uncompiled_shader(ctx, state->ir.nir,
- &state->stream_output);
-}
+ prog_data->binding_table.shader_time_start = 0xd0d0d0d0;
-static void *
-iris_create_vs_state(struct pipe_context *ctx,
- const struct pipe_shader_state *state)
-{
- struct iris_context *ice = (void *) ctx;
- struct iris_screen *screen = (void *) ctx->screen;
- struct iris_uncompiled_shader *ish = iris_create_shader_state(ctx, state);
+ /* Plane 0 is just the regular texture section */
+ prog_data->binding_table.plane_start[0] = prog_data->binding_table.texture_start;
- /* User clip planes */
- if (ish->nir->info.clip_distance_array_size == 0)
- ish->nos |= (1ull << IRIS_NOS_RASTERIZER);
+ prog_data->binding_table.plane_start[1] = next_binding_table_offset;
+ next_binding_table_offset += num_textures;
- if (screen->precompile) {
- struct brw_vs_prog_key key = { KEY_INIT };
+ prog_data->binding_table.plane_start[2] = next_binding_table_offset;
+ next_binding_table_offset += num_textures;
- iris_compile_vs(ice, ish, &key);
- }
+ /* Set the binding table size */
+ prog_data->binding_table.size_bytes = next_binding_table_offset * 4;
- return ish;
+ return next_binding_table_offset;
}
-static void *
-iris_create_tcs_state(struct pipe_context *ctx,
- const struct pipe_shader_state *state)
+static void
+setup_vec4_image_sysval(uint32_t *sysvals, uint32_t idx,
+ unsigned offset, unsigned n)
{
- struct iris_context *ice = (void *) ctx;
- struct iris_screen *screen = (void *) ctx->screen;
- struct iris_uncompiled_shader *ish = iris_create_shader_state(ctx, state);
- struct shader_info *info = &ish->nir->info;
-
- // XXX: NOS?
-
- if (screen->precompile) {
- const unsigned _GL_TRIANGLES = 0x0004;
- struct brw_tcs_prog_key key = {
- KEY_INIT,
- // XXX: make sure the linker fills this out from the TES...
- .tes_primitive_mode =
- info->tess.primitive_mode ? info->tess.primitive_mode
- : _GL_TRIANGLES,
- .outputs_written = info->outputs_written,
- .patch_outputs_written = info->patch_outputs_written,
- };
+ assert(offset % sizeof(uint32_t) == 0);
- iris_compile_tcs(ice, ish, &key);
- }
+ for (unsigned i = 0; i < n; ++i)
+ sysvals[i] = BRW_PARAM_IMAGE(idx, offset / sizeof(uint32_t) + i);
- return ish;
+ for (unsigned i = n; i < 4; ++i)
+ sysvals[i] = BRW_PARAM_BUILTIN_ZERO;
}
-static void *
-iris_create_tes_state(struct pipe_context *ctx,
- const struct pipe_shader_state *state)
+/**
+ * Associate NIR uniform variables with the prog_data->param[] mechanism
+ * used by the backend. Also, decide which UBOs we'd like to push in an
+ * ideal situation (though the backend can reduce this).
+ */
+static void
+iris_setup_uniforms(const struct brw_compiler *compiler,
+ void *mem_ctx,
+ nir_shader *nir,
+ struct brw_stage_prog_data *prog_data,
+ enum brw_param_builtin **out_system_values,
+ unsigned *out_num_system_values,
+ unsigned *out_num_cbufs)
{
- struct iris_context *ice = (void *) ctx;
- struct iris_screen *screen = (void *) ctx->screen;
- struct iris_uncompiled_shader *ish = iris_create_shader_state(ctx, state);
- struct shader_info *info = &ish->nir->info;
+ const struct gen_device_info *devinfo = compiler->devinfo;
- // XXX: NOS?
+ /* The intel compiler assumes that num_uniforms is in bytes. For
+ * scalar that means 4 bytes per uniform slot.
+ *
+ * Ref: brw_nir_lower_uniforms, type_size_scalar_bytes.
+ */
+ nir->num_uniforms *= 4;
- if (screen->precompile) {
- struct brw_tes_prog_key key = {
- KEY_INIT,
- // XXX: not ideal, need TCS output/TES input unification
- .inputs_read = info->inputs_read,
- .patch_inputs_read = info->patch_inputs_read,
- };
+ const unsigned IRIS_MAX_SYSTEM_VALUES =
+ PIPE_MAX_SHADER_IMAGES * BRW_IMAGE_PARAM_SIZE;
+ enum brw_param_builtin *system_values =
+ rzalloc_array(mem_ctx, enum brw_param_builtin, IRIS_MAX_SYSTEM_VALUES);
+ unsigned num_system_values = 0;
- iris_compile_tes(ice, ish, &key);
- }
+ unsigned patch_vert_idx = -1;
+ unsigned ucp_idx[IRIS_MAX_CLIP_PLANES];
+ unsigned img_idx[PIPE_MAX_SHADER_IMAGES];
+ memset(ucp_idx, -1, sizeof(ucp_idx));
+ memset(img_idx, -1, sizeof(img_idx));
- return ish;
-}
+ nir_function_impl *impl = nir_shader_get_entrypoint(nir);
-static void *
-iris_create_gs_state(struct pipe_context *ctx,
- const struct pipe_shader_state *state)
-{
- struct iris_context *ice = (void *) ctx;
- struct iris_screen *screen = (void *) ctx->screen;
- struct iris_uncompiled_shader *ish = iris_create_shader_state(ctx, state);
+ nir_builder b;
+ nir_builder_init(&b, impl);
- // XXX: NOS?
+ b.cursor = nir_before_block(nir_start_block(impl));
+ nir_ssa_def *temp_ubo_name = nir_ssa_undef(&b, 1, 32);
- if (screen->precompile) {
- struct brw_gs_prog_key key = { KEY_INIT };
+ /* Turn system value intrinsics into uniforms */
+ nir_foreach_block(block, impl) {
+ nir_foreach_instr_safe(instr, block) {
+ if (instr->type != nir_instr_type_intrinsic)
+ continue;
- iris_compile_gs(ice, ish, &key);
- }
+ nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
+ nir_ssa_def *offset;
- return ish;
-}
+ switch (intrin->intrinsic) {
+ case nir_intrinsic_load_user_clip_plane: {
+ unsigned ucp = nir_intrinsic_ucp_id(intrin);
-static void *
-iris_create_fs_state(struct pipe_context *ctx,
- const struct pipe_shader_state *state)
-{
- struct iris_context *ice = (void *) ctx;
- struct iris_screen *screen = (void *) ctx->screen;
- struct iris_uncompiled_shader *ish = iris_create_shader_state(ctx, state);
- struct shader_info *info = &ish->nir->info;
+ if (ucp_idx[ucp] == -1) {
+ ucp_idx[ucp] = num_system_values;
+ num_system_values += 4;
+ }
- ish->nos |= (1ull << IRIS_NOS_FRAMEBUFFER) |
- (1ull << IRIS_NOS_DEPTH_STENCIL_ALPHA) |
- (1ull << IRIS_NOS_RASTERIZER) |
- (1ull << IRIS_NOS_BLEND);
+ for (int i = 0; i < 4; i++) {
+ system_values[ucp_idx[ucp] + i] =
+ BRW_PARAM_BUILTIN_CLIP_PLANE(ucp, i);
+ }
- /* The program key needs the VUE map if there are > 16 inputs */
- if (util_bitcount64(ish->nir->info.inputs_read &
- BRW_FS_VARYING_INPUT_MASK) > 16) {
- ish->nos |= (1ull << IRIS_NOS_LAST_VUE_MAP);
- }
+ b.cursor = nir_before_instr(instr);
+ offset = nir_imm_int(&b, ucp_idx[ucp] * sizeof(uint32_t));
+ break;
+ }
+ case nir_intrinsic_load_patch_vertices_in:
+ if (patch_vert_idx == -1)
+ patch_vert_idx = num_system_values++;
- if (screen->precompile) {
- const uint64_t color_outputs = info->outputs_written &
- ~(BITFIELD64_BIT(FRAG_RESULT_DEPTH) |
- BITFIELD64_BIT(FRAG_RESULT_STENCIL) |
- BITFIELD64_BIT(FRAG_RESULT_SAMPLE_MASK));
+ system_values[patch_vert_idx] =
+ BRW_PARAM_BUILTIN_PATCH_VERTICES_IN;
- bool can_rearrange_varyings =
- util_bitcount64(info->inputs_read & BRW_FS_VARYING_INPUT_MASK) <= 16;
+ b.cursor = nir_before_instr(instr);
+ offset = nir_imm_int(&b, patch_vert_idx * sizeof(uint32_t));
+ break;
+ case nir_intrinsic_image_deref_load_param_intel: {
+ assert(devinfo->gen < 9);
+ nir_deref_instr *deref = nir_src_as_deref(intrin->src[0]);
+ nir_variable *var = nir_deref_instr_get_variable(deref);
- struct brw_wm_prog_key key = {
- KEY_INIT,
- .nr_color_regions = util_bitcount(color_outputs),
- .coherent_fb_fetch = true,
- .input_slots_valid =
- can_rearrange_varyings ? 0 : info->inputs_read | VARYING_BIT_POS,
- };
+ /* XXX: var->data.binding is not set properly. We need to run
+ * some form of gl_nir_lower_samplers_as_deref() to get it.
+ * This breaks tests which use more than one image.
+ */
+ if (img_idx[var->data.binding] == -1) {
+ /* GL only allows arrays of arrays of images. */
+ assert(glsl_type_is_image(glsl_without_array(var->type)));
+ unsigned num_images = MAX2(1, glsl_get_aoa_size(var->type));
- iris_compile_fs(ice, ish, &key, NULL);
- }
+ for (int i = 0; i < num_images; i++) {
+ const unsigned img = var->data.binding + i;
- return ish;
-}
+ img_idx[img] = num_system_values;
+ num_system_values += BRW_IMAGE_PARAM_SIZE;
-static void *
-iris_create_compute_state(struct pipe_context *ctx,
- const struct pipe_compute_state *state)
-{
- assert(state->ir_type == PIPE_SHADER_IR_NIR);
+ uint32_t *img_sv = &system_values[img_idx[img]];
- struct iris_context *ice = (void *) ctx;
- struct iris_screen *screen = (void *) ctx->screen;
- struct iris_uncompiled_shader *ish =
- iris_create_uncompiled_shader(ctx, (void *) state->prog, NULL);
+ setup_vec4_image_sysval(
+ img_sv + BRW_IMAGE_PARAM_OFFSET_OFFSET, img,
+ offsetof(struct brw_image_param, offset), 2);
+ setup_vec4_image_sysval(
+ img_sv + BRW_IMAGE_PARAM_SIZE_OFFSET, img,
+ offsetof(struct brw_image_param, size), 3);
+ setup_vec4_image_sysval(
+ img_sv + BRW_IMAGE_PARAM_STRIDE_OFFSET, img,
+ offsetof(struct brw_image_param, stride), 4);
+ setup_vec4_image_sysval(
+ img_sv + BRW_IMAGE_PARAM_TILING_OFFSET, img,
+ offsetof(struct brw_image_param, tiling), 3);
+ setup_vec4_image_sysval(
+ img_sv + BRW_IMAGE_PARAM_SWIZZLING_OFFSET, img,
+ offsetof(struct brw_image_param, swizzling), 2);
+ }
+ }
- // XXX: disallow more than 64KB of shared variables
+ b.cursor = nir_before_instr(instr);
+ offset = nir_iadd(&b,
+ get_aoa_deref_offset(&b, deref, BRW_IMAGE_PARAM_SIZE * 4),
+ nir_imm_int(&b, img_idx[var->data.binding] * 4 +
+ nir_intrinsic_base(intrin) * 16));
+ break;
+ }
+ default:
+ continue;
+ }
- if (screen->precompile) {
- struct brw_cs_prog_key key = { KEY_INIT };
+ unsigned comps = nir_intrinsic_dest_components(intrin);
- iris_compile_cs(ice, ish, &key);
+ nir_intrinsic_instr *load =
+ nir_intrinsic_instr_create(nir, nir_intrinsic_load_ubo);
+ load->num_components = comps;
+ load->src[0] = nir_src_for_ssa(temp_ubo_name);
+ load->src[1] = nir_src_for_ssa(offset);
+ nir_ssa_dest_init(&load->instr, &load->dest, comps, 32, NULL);
+ nir_builder_instr_insert(&b, &load->instr);
+ nir_ssa_def_rewrite_uses(&intrin->dest.ssa,
+ nir_src_for_ssa(&load->dest.ssa));
+ nir_instr_remove(instr);
+ }
}
- return ish;
-}
-
-/**
- * The pipe->delete_[stage]_state() driver hooks.
- *
- * Frees the iris_uncompiled_shader.
- */
-static void
-iris_delete_shader_state(struct pipe_context *ctx, void *state)
-{
- struct iris_uncompiled_shader *ish = state;
+ nir_validate_shader(nir, "before remapping");
- ralloc_free(ish->nir);
- free(ish);
-}
+ /* Place the new params at the front of constant buffer 0. */
+ if (num_system_values > 0) {
+ nir->num_uniforms += num_system_values * sizeof(uint32_t);
-/**
- * The pipe->bind_[stage]_state() driver hook.
- *
- * Binds an uncompiled shader as the current one for a particular stage.
- * Updates dirty tracking to account for the shader's NOS.
- */
-static void
-bind_state(struct iris_context *ice,
- struct iris_uncompiled_shader *ish,
- gl_shader_stage stage)
-{
- uint64_t dirty_bit = IRIS_DIRTY_UNCOMPILED_VS << stage;
- const uint64_t nos = ish ? ish->nos : 0;
+ system_values = reralloc(mem_ctx, system_values, enum brw_param_builtin,
+ num_system_values);
- ice->shaders.uncompiled[stage] = ish;
- ice->state.dirty |= dirty_bit;
+ nir_foreach_block(block, impl) {
+ nir_foreach_instr_safe(instr, block) {
+ if (instr->type != nir_instr_type_intrinsic)
+ continue;
- /* Record that CSOs need to mark IRIS_DIRTY_UNCOMPILED_XS when they change
- * (or that they no longer need to do so).
- */
- for (int i = 0; i < IRIS_NOS_COUNT; i++) {
- if (nos & (1 << i))
- ice->state.dirty_for_nos[i] |= dirty_bit;
- else
- ice->state.dirty_for_nos[i] &= ~dirty_bit;
- }
-}
+ nir_intrinsic_instr *load = nir_instr_as_intrinsic(instr);
-static void
-iris_bind_vs_state(struct pipe_context *ctx, void *state)
-{
- bind_state((void *) ctx, state, MESA_SHADER_VERTEX);
-}
+ if (load->intrinsic != nir_intrinsic_load_ubo)
+ continue;
-static void
-iris_bind_tcs_state(struct pipe_context *ctx, void *state)
-{
- bind_state((void *) ctx, state, MESA_SHADER_TESS_CTRL);
-}
+ b.cursor = nir_before_instr(instr);
-static void
-iris_bind_tes_state(struct pipe_context *ctx, void *state)
-{
- struct iris_context *ice = (struct iris_context *)ctx;
+ assert(load->src[0].is_ssa);
- /* Enabling/disabling optional stages requires a URB reconfiguration. */
- if (!!state != !!ice->shaders.uncompiled[MESA_SHADER_TESS_EVAL])
- ice->state.dirty |= IRIS_DIRTY_URB;
+ if (load->src[0].ssa == temp_ubo_name) {
+ nir_instr_rewrite_src(instr, &load->src[0],
+ nir_src_for_ssa(nir_imm_int(&b, 0)));
+ } else if (nir_src_as_uint(load->src[0]) == 0) {
+ nir_ssa_def *offset =
+ nir_iadd(&b, load->src[1].ssa,
+ nir_imm_int(&b, 4 * num_system_values));
+ nir_instr_rewrite_src(instr, &load->src[1],
+ nir_src_for_ssa(offset));
+ }
+ }
+ }
- bind_state((void *) ctx, state, MESA_SHADER_TESS_EVAL);
-}
+ /* We need to fold the new iadds for brw_nir_analyze_ubo_ranges */
+ nir_opt_constant_folding(nir);
+ } else {
+ ralloc_free(system_values);
+ system_values = NULL;
+ }
-static void
-iris_bind_gs_state(struct pipe_context *ctx, void *state)
-{
- struct iris_context *ice = (struct iris_context *)ctx;
+ nir_validate_shader(nir, "after remap");
- /* Enabling/disabling optional stages requires a URB reconfiguration. */
- if (!!state != !!ice->shaders.uncompiled[MESA_SHADER_GEOMETRY])
- ice->state.dirty |= IRIS_DIRTY_URB;
+ if (nir->info.stage != MESA_SHADER_COMPUTE)
+ brw_nir_analyze_ubo_ranges(compiler, nir, NULL, prog_data->ubo_ranges);
- bind_state((void *) ctx, state, MESA_SHADER_GEOMETRY);
-}
+ /* We don't use params[], but fs_visitor::nir_setup_uniforms() asserts
+ * about it for compute shaders, so go ahead and make some fake ones
+ * which the backend will dead code eliminate.
+ */
+ prog_data->nr_params = nir->num_uniforms / 4;
+ prog_data->param = rzalloc_array(mem_ctx, uint32_t, prog_data->nr_params);
-static void
-iris_bind_fs_state(struct pipe_context *ctx, void *state)
-{
- bind_state((void *) ctx, state, MESA_SHADER_FRAGMENT);
-}
+ /* System values and uniforms are stored in constant buffer 0, the
+ * user-facing UBOs are indexed by one. So if any constant buffer is
+ * needed, the constant buffer 0 will be needed, so account for it.
+ */
+ unsigned num_cbufs = nir->info.num_ubos;
+ if (num_cbufs || num_system_values || nir->num_uniforms)
+ num_cbufs++;
-static void
-iris_bind_cs_state(struct pipe_context *ctx, void *state)
-{
- bind_state((void *) ctx, state, MESA_SHADER_COMPUTE);
+ *out_system_values = system_values;
+ *out_num_system_values = num_system_values;
+ *out_num_cbufs = num_cbufs;
}
/**
- * Sets up the starting offsets for the groups of binding table entries
- * common to all pipeline stages.
- *
- * Unused groups are initialized to 0xd0d0d0d0 to make it obvious that they're
- * unused but also make sure that addition of small offsets to them will
- * trigger some of our asserts that surface indices are < BRW_MAX_SURFACES.
+ * Compile a vertex shader, and upload the assembly.
*/
-static uint32_t
-assign_common_binding_table_offsets(const struct gen_device_info *devinfo,
- const struct nir_shader *nir,
- struct brw_stage_prog_data *prog_data,
- uint32_t next_binding_table_offset,
- unsigned num_system_values)
+static struct iris_compiled_shader *
+iris_compile_vs(struct iris_context *ice,
+ struct iris_uncompiled_shader *ish,
+ const struct brw_vs_prog_key *key)
{
- const struct shader_info *info = &nir->info;
+ struct iris_screen *screen = (struct iris_screen *)ice->ctx.screen;
+ const struct brw_compiler *compiler = screen->compiler;
+ const struct gen_device_info *devinfo = &screen->devinfo;
+ void *mem_ctx = ralloc_context(NULL);
+ struct brw_vs_prog_data *vs_prog_data =
+ rzalloc(mem_ctx, struct brw_vs_prog_data);
+ struct brw_vue_prog_data *vue_prog_data = &vs_prog_data->base;
+ struct brw_stage_prog_data *prog_data = &vue_prog_data->base;
+ enum brw_param_builtin *system_values;
+ unsigned num_system_values;
+ unsigned num_cbufs;
- if (info->num_textures) {
- prog_data->binding_table.texture_start = next_binding_table_offset;
- prog_data->binding_table.gather_texture_start = next_binding_table_offset;
- next_binding_table_offset += info->num_textures;
- } else {
- prog_data->binding_table.texture_start = 0xd0d0d0d0;
- prog_data->binding_table.gather_texture_start = 0xd0d0d0d0;
- }
+ nir_shader *nir = nir_shader_clone(mem_ctx, ish->nir);
- if (info->num_images) {
- prog_data->binding_table.image_start = next_binding_table_offset;
- next_binding_table_offset += info->num_images;
- } else {
- prog_data->binding_table.image_start = 0xd0d0d0d0;
+ if (key->nr_userclip_plane_consts) {
+ nir_function_impl *impl = nir_shader_get_entrypoint(nir);
+ nir_lower_clip_vs(nir, (1 << key->nr_userclip_plane_consts) - 1, true);
+ nir_lower_io_to_temporaries(nir, impl, true, false);
+ nir_lower_global_vars_to_local(nir);
+ nir_lower_vars_to_ssa(nir);
+ nir_shader_gather_info(nir, impl);
}
- int num_ubos = info->num_ubos +
- ((nir->num_uniforms || num_system_values) ? 1 : 0);
-
- if (num_ubos) {
- //assert(info->num_ubos <= BRW_MAX_UBO);
- prog_data->binding_table.ubo_start = next_binding_table_offset;
- next_binding_table_offset += num_ubos;
- } else {
- prog_data->binding_table.ubo_start = 0xd0d0d0d0;
- }
+ if (nir->info.name && strncmp(nir->info.name, "ARB", 3) == 0)
+ prog_data->use_alt_mode = true;
- if (info->num_ssbos || info->num_abos) {
- prog_data->binding_table.ssbo_start = next_binding_table_offset;
- // XXX: see iris_state "wasting 16 binding table slots for ABOs" comment
- next_binding_table_offset += IRIS_MAX_ABOS + info->num_ssbos;
- } else {
- prog_data->binding_table.ssbo_start = 0xd0d0d0d0;
- }
+ iris_setup_uniforms(compiler, mem_ctx, nir, prog_data, &system_values,
+ &num_system_values, &num_cbufs);
- prog_data->binding_table.shader_time_start = 0xd0d0d0d0;
+ assign_common_binding_table_offsets(devinfo, nir, prog_data, 0,
+ num_system_values, num_cbufs);
- /* Plane 0 is just the regular texture section */
- prog_data->binding_table.plane_start[0] = prog_data->binding_table.texture_start;
+ brw_compute_vue_map(devinfo,
+ &vue_prog_data->vue_map, nir->info.outputs_written,
+ nir->info.separate_shader);
- prog_data->binding_table.plane_start[1] = next_binding_table_offset;
- next_binding_table_offset += info->num_textures;
+ /* Don't tell the backend about our clip plane constants, we've already
+ * lowered them in NIR and we don't want it doing it again.
+ */
+ struct brw_vs_prog_key key_no_ucp = *key;
+ key_no_ucp.nr_userclip_plane_consts = 0;
- prog_data->binding_table.plane_start[2] = next_binding_table_offset;
- next_binding_table_offset += info->num_textures;
+ char *error_str = NULL;
+ const unsigned *program =
+ brw_compile_vs(compiler, &ice->dbg, mem_ctx, &key_no_ucp, vs_prog_data,
+ nir, -1, &error_str);
+ if (program == NULL) {
+ dbg_printf("Failed to compile vertex shader: %s\n", error_str);
+ ralloc_free(mem_ctx);
+ return false;
+ }
- /* Set the binding table size */
- prog_data->binding_table.size_bytes = next_binding_table_offset * 4;
+ uint32_t *so_decls =
+ ice->vtbl.create_so_decl_list(&ish->stream_output,
+ &vue_prog_data->vue_map);
- return next_binding_table_offset;
+ struct iris_compiled_shader *shader =
+ iris_upload_shader(ice, IRIS_CACHE_VS, sizeof(*key), key, program,
+ prog_data, so_decls, system_values, num_system_values,
+ num_cbufs);
+
+ if (ish->compiled_once) {
+ perf_debug(&ice->dbg, "Recompiling vertex shader\n");
+ } else {
+ ish->compiled_once = true;
+ }
+
+ ralloc_free(mem_ctx);
+ return shader;
}
+/**
+ * Update the current vertex shader variant.
+ *
+ * Fill out the key, look in the cache, compile and bind if needed.
+ */
static void
-setup_vec4_image_sysval(uint32_t *sysvals, uint32_t idx,
- unsigned offset, unsigned n)
+iris_update_compiled_vs(struct iris_context *ice)
{
- assert(offset % sizeof(uint32_t) == 0);
+ struct iris_uncompiled_shader *ish =
+ ice->shaders.uncompiled[MESA_SHADER_VERTEX];
+ struct iris_screen *screen = (struct iris_screen *)ice->ctx.screen;
+ const struct gen_device_info *devinfo = &screen->devinfo;
- for (unsigned i = 0; i < n; ++i)
- sysvals[i] = BRW_PARAM_IMAGE(idx, offset / sizeof(uint32_t) + i);
+ struct brw_vs_prog_key key = { KEY_INIT(devinfo->gen) };
+ ice->vtbl.populate_vs_key(ice, &ish->nir->info, &key);
- for (unsigned i = n; i < 4; ++i)
- sysvals[i] = BRW_PARAM_BUILTIN_ZERO;
+ struct iris_compiled_shader *old = ice->shaders.prog[IRIS_CACHE_VS];
+ struct iris_compiled_shader *shader =
+ iris_find_cached_shader(ice, IRIS_CACHE_VS, sizeof(key), &key);
+
+ if (!shader)
+ shader = iris_compile_vs(ice, ish, &key);
+
+ if (old != shader) {
+ ice->shaders.prog[IRIS_CACHE_VS] = shader;
+ ice->state.dirty |= IRIS_DIRTY_VS |
+ IRIS_DIRTY_BINDINGS_VS |
+ IRIS_DIRTY_CONSTANTS_VS |
+ IRIS_DIRTY_VF_SGVS;
+ const struct brw_vs_prog_data *vs_prog_data =
+ (void *) shader->prog_data;
+ const bool uses_draw_params = vs_prog_data->uses_firstvertex ||
+ vs_prog_data->uses_baseinstance;
+ const bool uses_derived_draw_params = vs_prog_data->uses_drawid ||
+ vs_prog_data->uses_is_indexed_draw;
+ const bool needs_sgvs_element = uses_draw_params ||
+ vs_prog_data->uses_instanceid ||
+ vs_prog_data->uses_vertexid;
+ bool needs_edge_flag = false;
+ nir_foreach_variable(var, &ish->nir->inputs) {
+ if (var->data.location == VERT_ATTRIB_EDGEFLAG)
+ needs_edge_flag = true;
+ }
+
+ if (ice->state.vs_uses_draw_params != uses_draw_params ||
+ ice->state.vs_uses_derived_draw_params != uses_derived_draw_params ||
+ ice->state.vs_needs_edge_flag != needs_edge_flag) {
+ ice->state.dirty |= IRIS_DIRTY_VERTEX_BUFFERS |
+ IRIS_DIRTY_VERTEX_ELEMENTS;
+ }
+ ice->state.vs_uses_draw_params = uses_draw_params;
+ ice->state.vs_uses_derived_draw_params = uses_derived_draw_params;
+ ice->state.vs_needs_sgvs_element = needs_sgvs_element;
+ ice->state.vs_needs_edge_flag = needs_edge_flag;
+ }
}
/**
- * Associate NIR uniform variables with the prog_data->param[] mechanism
- * used by the backend. Also, decide which UBOs we'd like to push in an
- * ideal situation (though the backend can reduce this).
+ * Get the shader_info for a given stage, or NULL if the stage is disabled.
*/
-static void
-iris_setup_uniforms(const struct brw_compiler *compiler,
- void *mem_ctx,
- nir_shader *nir,
- struct brw_stage_prog_data *prog_data,
- enum brw_param_builtin **out_system_values,
- unsigned *out_num_system_values)
+const struct shader_info *
+iris_get_shader_info(const struct iris_context *ice, gl_shader_stage stage)
{
- const struct gen_device_info *devinfo = compiler->devinfo;
-
- /* The intel compiler assumes that num_uniforms is in bytes. For
- * scalar that means 4 bytes per uniform slot.
- *
- * Ref: brw_nir_lower_uniforms, type_size_scalar_bytes.
- */
- nir->num_uniforms *= 4;
-
- const unsigned IRIS_MAX_SYSTEM_VALUES =
- PIPE_MAX_SHADER_IMAGES * BRW_IMAGE_PARAM_SIZE;
- enum brw_param_builtin *system_values =
- rzalloc_array(mem_ctx, enum brw_param_builtin, IRIS_MAX_SYSTEM_VALUES);
- unsigned num_system_values = 0;
+ const struct iris_uncompiled_shader *ish = ice->shaders.uncompiled[stage];
- unsigned patch_vert_idx = -1;
- unsigned ucp_idx[IRIS_MAX_CLIP_PLANES];
- unsigned img_idx[PIPE_MAX_SHADER_IMAGES];
- memset(ucp_idx, -1, sizeof(ucp_idx));
- memset(img_idx, -1, sizeof(img_idx));
+ if (!ish)
+ return NULL;
- nir_function_impl *impl = nir_shader_get_entrypoint(nir);
+ const nir_shader *nir = ish->nir;
+ return &nir->info;
+}
- nir_builder b;
- nir_builder_init(&b, impl);
+/**
+ * Get the union of TCS output and TES input slots.
+ *
+ * TCS and TES need to agree on a common URB entry layout. In particular,
+ * the data for all patch vertices is stored in a single URB entry (unlike
+ * GS which has one entry per input vertex). This means that per-vertex
+ * array indexing needs a stride.
+ *
+ * SSO requires locations to match, but doesn't require the number of
+ * outputs/inputs to match (in fact, the TCS often has extra outputs).
+ * So, we need to take the extra step of unifying these on the fly.
+ */
+static void
+get_unified_tess_slots(const struct iris_context *ice,
+ uint64_t *per_vertex_slots,
+ uint32_t *per_patch_slots)
+{
+ const struct shader_info *tcs =
+ iris_get_shader_info(ice, MESA_SHADER_TESS_CTRL);
+ const struct shader_info *tes =
+ iris_get_shader_info(ice, MESA_SHADER_TESS_EVAL);
- b.cursor = nir_before_block(nir_start_block(impl));
- nir_ssa_def *temp_ubo_name = nir_ssa_undef(&b, 1, 32);
+ *per_vertex_slots = tes->inputs_read;
+ *per_patch_slots = tes->patch_inputs_read;
- /* Turn system value intrinsics into uniforms */
- nir_foreach_block(block, impl) {
- nir_foreach_instr_safe(instr, block) {
- if (instr->type != nir_instr_type_intrinsic)
- continue;
+ if (tcs) {
+ *per_vertex_slots |= tcs->outputs_written;
+ *per_patch_slots |= tcs->patch_outputs_written;
+ }
+}
- nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
- nir_ssa_def *offset;
+/**
+ * Compile a tessellation control shader, and upload the assembly.
+ */
+static struct iris_compiled_shader *
+iris_compile_tcs(struct iris_context *ice,
+ struct iris_uncompiled_shader *ish,
+ const struct brw_tcs_prog_key *key)
+{
+ struct iris_screen *screen = (struct iris_screen *)ice->ctx.screen;
+ const struct brw_compiler *compiler = screen->compiler;
+ const struct nir_shader_compiler_options *options =
+ compiler->glsl_compiler_options[MESA_SHADER_TESS_CTRL].NirOptions;
+ const struct gen_device_info *devinfo = &screen->devinfo;
+ void *mem_ctx = ralloc_context(NULL);
+ struct brw_tcs_prog_data *tcs_prog_data =
+ rzalloc(mem_ctx, struct brw_tcs_prog_data);
+ struct brw_vue_prog_data *vue_prog_data = &tcs_prog_data->base;
+ struct brw_stage_prog_data *prog_data = &vue_prog_data->base;
+ enum brw_param_builtin *system_values = NULL;
+ unsigned num_system_values = 0;
+ unsigned num_cbufs;
- switch (intrin->intrinsic) {
- case nir_intrinsic_load_user_clip_plane: {
- unsigned ucp = nir_intrinsic_ucp_id(intrin);
+ nir_shader *nir;
- if (ucp_idx[ucp] == -1) {
- ucp_idx[ucp] = num_system_values;
- num_system_values += 4;
- }
+ if (ish) {
+ nir = nir_shader_clone(mem_ctx, ish->nir);
- for (int i = 0; i < 4; i++) {
- system_values[ucp_idx[ucp] + i] =
- BRW_PARAM_BUILTIN_CLIP_PLANE(ucp, i);
- }
+ iris_setup_uniforms(compiler, mem_ctx, nir, prog_data, &system_values,
+ &num_system_values, &num_cbufs);
+ assign_common_binding_table_offsets(devinfo, nir, prog_data, 0,
+ num_system_values, num_cbufs);
+ } else {
+ nir = brw_nir_create_passthrough_tcs(mem_ctx, compiler, options, key);
- b.cursor = nir_before_instr(instr);
- offset = nir_imm_int(&b, ucp_idx[ucp] * sizeof(uint32_t));
- break;
- }
- case nir_intrinsic_load_patch_vertices_in:
- if (patch_vert_idx == -1)
- patch_vert_idx = num_system_values++;
+ /* Reserve space for passing the default tess levels as constants. */
+ num_system_values = 8;
+ system_values =
+ rzalloc_array(mem_ctx, enum brw_param_builtin, num_system_values);
+ prog_data->param = rzalloc_array(mem_ctx, uint32_t, num_system_values);
+ prog_data->nr_params = num_system_values;
+
+ if (key->tes_primitive_mode == GL_QUADS) {
+ for (int i = 0; i < 4; i++)
+ system_values[7 - i] = BRW_PARAM_BUILTIN_TESS_LEVEL_OUTER_X + i;
+
+ system_values[3] = BRW_PARAM_BUILTIN_TESS_LEVEL_INNER_X;
+ system_values[2] = BRW_PARAM_BUILTIN_TESS_LEVEL_INNER_Y;
+ } else if (key->tes_primitive_mode == GL_TRIANGLES) {
+ for (int i = 0; i < 3; i++)
+ system_values[7 - i] = BRW_PARAM_BUILTIN_TESS_LEVEL_OUTER_X + i;
+
+ system_values[4] = BRW_PARAM_BUILTIN_TESS_LEVEL_INNER_X;
+ } else {
+ assert(key->tes_primitive_mode == GL_ISOLINES);
+ system_values[7] = BRW_PARAM_BUILTIN_TESS_LEVEL_OUTER_Y;
+ system_values[6] = BRW_PARAM_BUILTIN_TESS_LEVEL_OUTER_X;
+ }
- system_values[patch_vert_idx] =
- BRW_PARAM_BUILTIN_PATCH_VERTICES_IN;
+ prog_data->ubo_ranges[0].length = 1;
+ }
- b.cursor = nir_before_instr(instr);
- offset = nir_imm_int(&b, patch_vert_idx * sizeof(uint32_t));
- break;
- case nir_intrinsic_image_deref_load_param_intel: {
- assert(devinfo->gen < 9);
- nir_deref_instr *deref = nir_src_as_deref(intrin->src[0]);
- nir_variable *var = nir_deref_instr_get_variable(deref);
+ char *error_str = NULL;
+ const unsigned *program =
+ brw_compile_tcs(compiler, &ice->dbg, mem_ctx, key, tcs_prog_data, nir,
+ -1, &error_str);
+ if (program == NULL) {
+ dbg_printf("Failed to compile control shader: %s\n", error_str);
+ ralloc_free(mem_ctx);
+ return false;
+ }
- if (img_idx[var->data.binding] == -1) {
- /* GL only allows arrays of arrays of images. */
- assert(glsl_type_is_image(glsl_without_array(var->type)));
- unsigned num_images = MAX2(1, glsl_get_aoa_size(var->type));
+ struct iris_compiled_shader *shader =
+ iris_upload_shader(ice, IRIS_CACHE_TCS, sizeof(*key), key, program,
+ prog_data, NULL, system_values, num_system_values,
+ num_cbufs);
- for (int i = 0; i < num_images; i++) {
- const unsigned img = var->data.binding + i;
+ if (ish) {
+ if (ish->compiled_once) {
+ perf_debug(&ice->dbg, "Recompiling tessellation control shader\n");
+ } else {
+ ish->compiled_once = true;
+ }
+ }
- img_idx[img] = num_system_values;
- num_system_values += BRW_IMAGE_PARAM_SIZE;
+ ralloc_free(mem_ctx);
+ return shader;
+}
- uint32_t *img_sv = &system_values[img_idx[img]];
+/**
+ * Update the current tessellation control shader variant.
+ *
+ * Fill out the key, look in the cache, compile and bind if needed.
+ */
+static void
+iris_update_compiled_tcs(struct iris_context *ice)
+{
+ struct iris_uncompiled_shader *tcs =
+ ice->shaders.uncompiled[MESA_SHADER_TESS_CTRL];
+ struct iris_screen *screen = (struct iris_screen *)ice->ctx.screen;
+ const struct gen_device_info *devinfo = &screen->devinfo;
- setup_vec4_image_sysval(
- img_sv + BRW_IMAGE_PARAM_OFFSET_OFFSET, img,
- offsetof(struct brw_image_param, offset), 2);
- setup_vec4_image_sysval(
- img_sv + BRW_IMAGE_PARAM_SIZE_OFFSET, img,
- offsetof(struct brw_image_param, size), 3);
- setup_vec4_image_sysval(
- img_sv + BRW_IMAGE_PARAM_STRIDE_OFFSET, img,
- offsetof(struct brw_image_param, stride), 4);
- setup_vec4_image_sysval(
- img_sv + BRW_IMAGE_PARAM_TILING_OFFSET, img,
- offsetof(struct brw_image_param, tiling), 3);
- setup_vec4_image_sysval(
- img_sv + BRW_IMAGE_PARAM_SWIZZLING_OFFSET, img,
- offsetof(struct brw_image_param, swizzling), 2);
- }
- }
+ const struct shader_info *tes_info =
+ iris_get_shader_info(ice, MESA_SHADER_TESS_EVAL);
+ struct brw_tcs_prog_key key = {
+ KEY_INIT_NO_ID(devinfo->gen),
+ .program_string_id = tcs ? tcs->program_id : 0,
+ .tes_primitive_mode = tes_info->tess.primitive_mode,
+ .input_vertices = ice->state.vertices_per_patch,
+ };
+ get_unified_tess_slots(ice, &key.outputs_written,
+ &key.patch_outputs_written);
+ ice->vtbl.populate_tcs_key(ice, &key);
- b.cursor = nir_before_instr(instr);
- offset = nir_iadd(&b,
- get_aoa_deref_offset(&b, deref, BRW_IMAGE_PARAM_SIZE * 4),
- nir_imm_int(&b, img_idx[var->data.binding] * 4 +
- nir_intrinsic_base(intrin) * 16));
- break;
- }
- default:
- continue;
- }
+ struct iris_compiled_shader *old = ice->shaders.prog[IRIS_CACHE_TCS];
+ struct iris_compiled_shader *shader =
+ iris_find_cached_shader(ice, IRIS_CACHE_TCS, sizeof(key), &key);
- unsigned comps = nir_intrinsic_dest_components(intrin);
+ if (!shader)
+ shader = iris_compile_tcs(ice, tcs, &key);
- nir_intrinsic_instr *load =
- nir_intrinsic_instr_create(nir, nir_intrinsic_load_ubo);
- load->num_components = comps;
- load->src[0] = nir_src_for_ssa(temp_ubo_name);
- load->src[1] = nir_src_for_ssa(offset);
- nir_ssa_dest_init(&load->instr, &load->dest, comps, 32, NULL);
- nir_builder_instr_insert(&b, &load->instr);
- nir_ssa_def_rewrite_uses(&intrin->dest.ssa,
- nir_src_for_ssa(&load->dest.ssa));
- nir_instr_remove(instr);
- }
+ if (old != shader) {
+ ice->shaders.prog[IRIS_CACHE_TCS] = shader;
+ ice->state.dirty |= IRIS_DIRTY_TCS |
+ IRIS_DIRTY_BINDINGS_TCS |
+ IRIS_DIRTY_CONSTANTS_TCS;
}
+}
- nir_validate_shader(nir, "before remapping");
+/**
+ * Compile a tessellation evaluation shader, and upload the assembly.
+ */
+static struct iris_compiled_shader *
+iris_compile_tes(struct iris_context *ice,
+ struct iris_uncompiled_shader *ish,
+ const struct brw_tes_prog_key *key)
+{
+ struct iris_screen *screen = (struct iris_screen *)ice->ctx.screen;
+ const struct brw_compiler *compiler = screen->compiler;
+ const struct gen_device_info *devinfo = &screen->devinfo;
+ void *mem_ctx = ralloc_context(NULL);
+ struct brw_tes_prog_data *tes_prog_data =
+ rzalloc(mem_ctx, struct brw_tes_prog_data);
+ struct brw_vue_prog_data *vue_prog_data = &tes_prog_data->base;
+ struct brw_stage_prog_data *prog_data = &vue_prog_data->base;
+ enum brw_param_builtin *system_values;
+ unsigned num_system_values;
+ unsigned num_cbufs;
- /* Place the new params at the front of constant buffer 0. */
- if (num_system_values > 0) {
- nir->num_uniforms += num_system_values * sizeof(uint32_t);
+ nir_shader *nir = nir_shader_clone(mem_ctx, ish->nir);
- system_values = reralloc(mem_ctx, system_values, enum brw_param_builtin,
- num_system_values);
+ iris_setup_uniforms(compiler, mem_ctx, nir, prog_data, &system_values,
+ &num_system_values, &num_cbufs);
- nir_foreach_block(block, impl) {
- nir_foreach_instr_safe(instr, block) {
- if (instr->type != nir_instr_type_intrinsic)
- continue;
+ assign_common_binding_table_offsets(devinfo, nir, prog_data, 0,
+ num_system_values, num_cbufs);
- nir_intrinsic_instr *load = nir_instr_as_intrinsic(instr);
+ struct brw_vue_map input_vue_map;
+ brw_compute_tess_vue_map(&input_vue_map, key->inputs_read,
+ key->patch_inputs_read);
- if (load->intrinsic != nir_intrinsic_load_ubo)
- continue;
+ char *error_str = NULL;
+ const unsigned *program =
+ brw_compile_tes(compiler, &ice->dbg, mem_ctx, key, &input_vue_map,
+ tes_prog_data, nir, NULL, -1, &error_str);
+ if (program == NULL) {
+ dbg_printf("Failed to compile evaluation shader: %s\n", error_str);
+ ralloc_free(mem_ctx);
+ return false;
+ }
- b.cursor = nir_before_instr(instr);
+ uint32_t *so_decls =
+ ice->vtbl.create_so_decl_list(&ish->stream_output,
+ &vue_prog_data->vue_map);
- assert(load->src[0].is_ssa);
- if (load->src[0].ssa == temp_ubo_name) {
- nir_instr_rewrite_src(instr, &load->src[0],
- nir_src_for_ssa(nir_imm_int(&b, 0)));
- } else if (nir_src_as_uint(load->src[0]) == 0) {
- nir_ssa_def *offset =
- nir_iadd(&b, load->src[1].ssa,
- nir_imm_int(&b, 4 * num_system_values));
- nir_instr_rewrite_src(instr, &load->src[1],
- nir_src_for_ssa(offset));
- }
- }
- }
+ struct iris_compiled_shader *shader =
+ iris_upload_shader(ice, IRIS_CACHE_TES, sizeof(*key), key, program,
+ prog_data, so_decls, system_values, num_system_values,
+ num_cbufs);
- /* We need to fold the new iadds for brw_nir_analyze_ubo_ranges */
- nir_opt_constant_folding(nir);
+ if (ish->compiled_once) {
+ perf_debug(&ice->dbg, "Recompiling tessellation evaluation shader\n");
} else {
- ralloc_free(system_values);
- system_values = NULL;
+ ish->compiled_once = true;
}
- nir_validate_shader(nir, "after remap");
+ ralloc_free(mem_ctx);
+ return shader;
+}
- if (nir->info.stage != MESA_SHADER_COMPUTE)
- brw_nir_analyze_ubo_ranges(compiler, nir, NULL, prog_data->ubo_ranges);
+/**
+ * Update the current tessellation evaluation shader variant.
+ *
+ * Fill out the key, look in the cache, compile and bind if needed.
+ */
+static void
+iris_update_compiled_tes(struct iris_context *ice)
+{
+ struct iris_uncompiled_shader *ish =
+ ice->shaders.uncompiled[MESA_SHADER_TESS_EVAL];
+ struct iris_screen *screen = (struct iris_screen *)ice->ctx.screen;
+ const struct gen_device_info *devinfo = &screen->devinfo;
- /* We don't use params[], but fs_visitor::nir_setup_uniforms() asserts
- * about it for compute shaders, so go ahead and make some fake ones
- * which the backend will dead code eliminate.
- */
- prog_data->nr_params = nir->num_uniforms / 4;
- prog_data->param = rzalloc_array(mem_ctx, uint32_t, prog_data->nr_params);
+ struct brw_tes_prog_key key = { KEY_INIT(devinfo->gen) };
+ get_unified_tess_slots(ice, &key.inputs_read, &key.patch_inputs_read);
+ ice->vtbl.populate_tes_key(ice, &key);
- *out_system_values = system_values;
- *out_num_system_values = num_system_values;
+ struct iris_compiled_shader *old = ice->shaders.prog[IRIS_CACHE_TES];
+ struct iris_compiled_shader *shader =
+ iris_find_cached_shader(ice, IRIS_CACHE_TES, sizeof(key), &key);
+
+ if (!shader)
+ shader = iris_compile_tes(ice, ish, &key);
+
+ if (old != shader) {
+ ice->shaders.prog[IRIS_CACHE_TES] = shader;
+ ice->state.dirty |= IRIS_DIRTY_TES |
+ IRIS_DIRTY_BINDINGS_TES |
+ IRIS_DIRTY_CONSTANTS_TES;
+ }
+
+ /* TODO: Could compare and avoid flagging this. */
+ const struct shader_info *tes_info = &ish->nir->info;
+ if (tes_info->system_values_read & (1ull << SYSTEM_VALUE_VERTICES_IN)) {
+ ice->state.dirty |= IRIS_DIRTY_CONSTANTS_TES;
+ ice->state.shaders[MESA_SHADER_TESS_EVAL].cbuf0_needs_upload = true;
+ }
}
/**
- * Compile a vertex shader, and upload the assembly.
+ * Compile a geometry shader, and upload the assembly.
*/
static struct iris_compiled_shader *
-iris_compile_vs(struct iris_context *ice,
+iris_compile_gs(struct iris_context *ice,
struct iris_uncompiled_shader *ish,
- const struct brw_vs_prog_key *key)
+ const struct brw_gs_prog_key *key)
{
struct iris_screen *screen = (struct iris_screen *)ice->ctx.screen;
const struct brw_compiler *compiler = screen->compiler;
const struct gen_device_info *devinfo = &screen->devinfo;
void *mem_ctx = ralloc_context(NULL);
- struct brw_vs_prog_data *vs_prog_data =
- rzalloc(mem_ctx, struct brw_vs_prog_data);
- struct brw_vue_prog_data *vue_prog_data = &vs_prog_data->base;
+ struct brw_gs_prog_data *gs_prog_data =
+ rzalloc(mem_ctx, struct brw_gs_prog_data);
+ struct brw_vue_prog_data *vue_prog_data = &gs_prog_data->base;
struct brw_stage_prog_data *prog_data = &vue_prog_data->base;
enum brw_param_builtin *system_values;
unsigned num_system_values;
+ unsigned num_cbufs;
nir_shader *nir = nir_shader_clone(mem_ctx, ish->nir);
- if (key->nr_userclip_plane_consts) {
- nir_function_impl *impl = nir_shader_get_entrypoint(nir);
- nir_lower_clip_vs(nir, (1 << key->nr_userclip_plane_consts) - 1, true);
- nir_lower_io_to_temporaries(nir, impl, true, false);
- nir_lower_global_vars_to_local(nir);
- nir_lower_vars_to_ssa(nir);
- nir_shader_gather_info(nir, impl);
- }
-
- // XXX: alt mode
-
iris_setup_uniforms(compiler, mem_ctx, nir, prog_data, &system_values,
- &num_system_values);
+ &num_system_values, &num_cbufs);
assign_common_binding_table_offsets(devinfo, nir, prog_data, 0,
- num_system_values);
+ num_system_values, num_cbufs);
brw_compute_vue_map(devinfo,
&vue_prog_data->vue_map, nir->info.outputs_written,
nir->info.separate_shader);
- /* Don't tell the backend about our clip plane constants, we've already
- * lowered them in NIR and we don't want it doing it again.
- */
- struct brw_vs_prog_key key_no_ucp = *key;
- key_no_ucp.nr_userclip_plane_consts = 0;
-
char *error_str = NULL;
const unsigned *program =
- brw_compile_vs(compiler, &ice->dbg, mem_ctx, &key_no_ucp, vs_prog_data,
- nir, -1, &error_str);
+ brw_compile_gs(compiler, &ice->dbg, mem_ctx, key, gs_prog_data, nir,
+ NULL, -1, &error_str);
if (program == NULL) {
- dbg_printf("Failed to compile vertex shader: %s\n", error_str);
+ dbg_printf("Failed to compile geometry shader: %s\n", error_str);
ralloc_free(mem_ctx);
return false;
}
&vue_prog_data->vue_map);
struct iris_compiled_shader *shader =
- iris_upload_shader(ice, IRIS_CACHE_VS, sizeof(*key), key, program,
- prog_data, so_decls, system_values, num_system_values);
+ iris_upload_shader(ice, IRIS_CACHE_GS, sizeof(*key), key, program,
+ prog_data, so_decls, system_values, num_system_values,
+ num_cbufs);
if (ish->compiled_once) {
- perf_debug(&ice->dbg, "Recompiling vertex shader\n");
+ perf_debug(&ice->dbg, "Recompiling geometry shader\n");
} else {
ish->compiled_once = true;
}
}
/**
- * Update the current vertex shader variant.
+ * Update the current geometry shader variant.
*
* Fill out the key, look in the cache, compile and bind if needed.
*/
static void
-iris_update_compiled_vs(struct iris_context *ice)
+iris_update_compiled_gs(struct iris_context *ice)
{
struct iris_uncompiled_shader *ish =
- ice->shaders.uncompiled[MESA_SHADER_VERTEX];
-
- struct brw_vs_prog_key key = { KEY_INIT };
- ice->vtbl.populate_vs_key(ice, &ish->nir->info, &key);
-
- struct iris_compiled_shader *old = ice->shaders.prog[IRIS_CACHE_VS];
- struct iris_compiled_shader *shader =
- iris_find_cached_shader(ice, IRIS_CACHE_VS, sizeof(key), &key);
-
- if (!shader)
- shader = iris_compile_vs(ice, ish, &key);
-
- if (old != shader) {
- ice->shaders.prog[IRIS_CACHE_VS] = shader;
- ice->state.dirty |= IRIS_DIRTY_VS |
- IRIS_DIRTY_BINDINGS_VS |
- IRIS_DIRTY_CONSTANTS_VS |
- IRIS_DIRTY_VF_SGVS;
- }
-}
-
-/**
- * Get the shader_info for a given stage, or NULL if the stage is disabled.
- */
-const struct shader_info *
-iris_get_shader_info(const struct iris_context *ice, gl_shader_stage stage)
-{
- const struct iris_uncompiled_shader *ish = ice->shaders.uncompiled[stage];
-
- if (!ish)
- return NULL;
-
- const nir_shader *nir = ish->nir;
- return &nir->info;
-}
-
-// XXX: this function is gross
-unsigned
-iris_get_shader_num_ubos(const struct iris_context *ice, gl_shader_stage stage)
-{
- const struct iris_uncompiled_shader *ish = ice->shaders.uncompiled[stage];
- const struct iris_compiled_shader *shader = ice->shaders.prog[stage];
+ ice->shaders.uncompiled[MESA_SHADER_GEOMETRY];
+ struct iris_compiled_shader *old = ice->shaders.prog[IRIS_CACHE_GS];
+ struct iris_compiled_shader *shader = NULL;
if (ish) {
- const nir_shader *nir = ish->nir;
- /* see assign_common_binding_table_offsets */
- return nir->info.num_ubos +
- ((nir->num_uniforms || shader->num_system_values) ? 1 : 0);
- }
- return 0;
-}
+ struct iris_screen *screen = (struct iris_screen *)ice->ctx.screen;
+ const struct gen_device_info *devinfo = &screen->devinfo;
+ struct brw_gs_prog_key key = { KEY_INIT(devinfo->gen) };
+ ice->vtbl.populate_gs_key(ice, &key);
-/**
- * Get the union of TCS output and TES input slots.
- *
- * TCS and TES need to agree on a common URB entry layout. In particular,
- * the data for all patch vertices is stored in a single URB entry (unlike
- * GS which has one entry per input vertex). This means that per-vertex
- * array indexing needs a stride.
- *
- * SSO requires locations to match, but doesn't require the number of
- * outputs/inputs to match (in fact, the TCS often has extra outputs).
- * So, we need to take the extra step of unifying these on the fly.
- */
-static void
-get_unified_tess_slots(const struct iris_context *ice,
- uint64_t *per_vertex_slots,
- uint32_t *per_patch_slots)
-{
- const struct shader_info *tcs =
- iris_get_shader_info(ice, MESA_SHADER_TESS_CTRL);
- const struct shader_info *tes =
- iris_get_shader_info(ice, MESA_SHADER_TESS_EVAL);
+ shader =
+ iris_find_cached_shader(ice, IRIS_CACHE_GS, sizeof(key), &key);
- *per_vertex_slots = tes->inputs_read;
- *per_patch_slots = tes->patch_inputs_read;
+ if (!shader)
+ shader = iris_compile_gs(ice, ish, &key);
+ }
- if (tcs) {
- *per_vertex_slots |= tcs->outputs_written;
- *per_patch_slots |= tcs->patch_outputs_written;
+ if (old != shader) {
+ ice->shaders.prog[IRIS_CACHE_GS] = shader;
+ ice->state.dirty |= IRIS_DIRTY_GS |
+ IRIS_DIRTY_BINDINGS_GS |
+ IRIS_DIRTY_CONSTANTS_GS;
}
}
/**
- * Compile a tessellation control shader, and upload the assembly.
+ * Compile a fragment (pixel) shader, and upload the assembly.
*/
static struct iris_compiled_shader *
-iris_compile_tcs(struct iris_context *ice,
- struct iris_uncompiled_shader *ish,
- const struct brw_tcs_prog_key *key)
+iris_compile_fs(struct iris_context *ice,
+ struct iris_uncompiled_shader *ish,
+ const struct brw_wm_prog_key *key,
+ struct brw_vue_map *vue_map)
{
struct iris_screen *screen = (struct iris_screen *)ice->ctx.screen;
const struct brw_compiler *compiler = screen->compiler;
- const struct nir_shader_compiler_options *options =
- compiler->glsl_compiler_options[MESA_SHADER_TESS_CTRL].NirOptions;
const struct gen_device_info *devinfo = &screen->devinfo;
void *mem_ctx = ralloc_context(NULL);
- struct brw_tcs_prog_data *tcs_prog_data =
- rzalloc(mem_ctx, struct brw_tcs_prog_data);
- struct brw_vue_prog_data *vue_prog_data = &tcs_prog_data->base;
- struct brw_stage_prog_data *prog_data = &vue_prog_data->base;
- enum brw_param_builtin *system_values = NULL;
- unsigned num_system_values = 0;
-
- nir_shader *nir;
+ struct brw_wm_prog_data *fs_prog_data =
+ rzalloc(mem_ctx, struct brw_wm_prog_data);
+ struct brw_stage_prog_data *prog_data = &fs_prog_data->base;
+ enum brw_param_builtin *system_values;
+ unsigned num_system_values;
+ unsigned num_cbufs;
- if (ish) {
- nir = nir_shader_clone(mem_ctx, ish->nir);
+ nir_shader *nir = nir_shader_clone(mem_ctx, ish->nir);
- iris_setup_uniforms(compiler, mem_ctx, nir, prog_data, &system_values,
- &num_system_values);
- assign_common_binding_table_offsets(devinfo, nir, prog_data, 0,
- num_system_values);
- } else {
- nir = brw_nir_create_passthrough_tcs(mem_ctx, compiler, options, key);
+ if (nir->info.name && strncmp(nir->info.name, "ARB", 3) == 0)
+ prog_data->use_alt_mode = true;
- /* Reserve space for passing the default tess levels as constants. */
- prog_data->param = rzalloc_array(mem_ctx, uint32_t, 8);
- prog_data->nr_params = 8;
- prog_data->ubo_ranges[0].length = 1;
- }
+ iris_setup_uniforms(compiler, mem_ctx, nir, prog_data, &system_values,
+ &num_system_values, &num_cbufs);
+ assign_common_binding_table_offsets(devinfo, nir, prog_data,
+ MAX2(key->nr_color_regions, 1),
+ num_system_values, num_cbufs);
char *error_str = NULL;
const unsigned *program =
- brw_compile_tcs(compiler, &ice->dbg, mem_ctx, key, tcs_prog_data, nir,
- -1, &error_str);
+ brw_compile_fs(compiler, &ice->dbg, mem_ctx, key, fs_prog_data,
+ nir, NULL, -1, -1, -1, true, false, vue_map, &error_str);
if (program == NULL) {
- dbg_printf("Failed to compile control shader: %s\n", error_str);
+ dbg_printf("Failed to compile fragment shader: %s\n", error_str);
ralloc_free(mem_ctx);
return false;
}
struct iris_compiled_shader *shader =
- iris_upload_shader(ice, IRIS_CACHE_TCS, sizeof(*key), key, program,
- prog_data, NULL, system_values, num_system_values);
+ iris_upload_shader(ice, IRIS_CACHE_FS, sizeof(*key), key, program,
+ prog_data, NULL, system_values, num_system_values,
+ num_cbufs);
- if (ish) {
- if (ish->compiled_once) {
- perf_debug(&ice->dbg, "Recompiling tessellation control shader\n");
- } else {
- ish->compiled_once = true;
- }
+ if (ish->compiled_once) {
+ perf_debug(&ice->dbg, "Recompiling fragment shader\n");
+ } else {
+ ish->compiled_once = true;
}
ralloc_free(mem_ctx);
}
/**
- * Update the current tessellation control shader variant.
+ * Update the current fragment shader variant.
*
* Fill out the key, look in the cache, compile and bind if needed.
*/
static void
-iris_update_compiled_tcs(struct iris_context *ice)
+iris_update_compiled_fs(struct iris_context *ice)
{
- struct iris_uncompiled_shader *tcs =
- ice->shaders.uncompiled[MESA_SHADER_TESS_CTRL];
+ struct iris_uncompiled_shader *ish =
+ ice->shaders.uncompiled[MESA_SHADER_FRAGMENT];
+ struct iris_screen *screen = (struct iris_screen *)ice->ctx.screen;
+ const struct gen_device_info *devinfo = &screen->devinfo;
+ struct brw_wm_prog_key key = { KEY_INIT(devinfo->gen) };
+ ice->vtbl.populate_fs_key(ice, &key);
- const struct shader_info *tes_info =
- iris_get_shader_info(ice, MESA_SHADER_TESS_EVAL);
- struct brw_tcs_prog_key key = {
- ALL_SAMPLERS_XYZW,
- .program_string_id = tcs ? tcs->program_id : 0,
- .tes_primitive_mode = tes_info->tess.primitive_mode,
- .input_vertices = ice->state.vertices_per_patch,
- };
- get_unified_tess_slots(ice, &key.outputs_written,
- &key.patch_outputs_written);
- ice->vtbl.populate_tcs_key(ice, &key);
+ if (ish->nos & (1ull << IRIS_NOS_LAST_VUE_MAP))
+ key.input_slots_valid = ice->shaders.last_vue_map->slots_valid;
- struct iris_compiled_shader *old = ice->shaders.prog[IRIS_CACHE_TCS];
+ struct iris_compiled_shader *old = ice->shaders.prog[IRIS_CACHE_FS];
struct iris_compiled_shader *shader =
- iris_find_cached_shader(ice, IRIS_CACHE_TCS, sizeof(key), &key);
+ iris_find_cached_shader(ice, IRIS_CACHE_FS, sizeof(key), &key);
if (!shader)
- shader = iris_compile_tcs(ice, tcs, &key);
+ shader = iris_compile_fs(ice, ish, &key, ice->shaders.last_vue_map);
+
+ if (old != shader) {
+ // XXX: only need to flag CLIP if barycentric has NONPERSPECTIVE
+ // toggles. might be able to avoid flagging SBE too.
+ ice->shaders.prog[IRIS_CACHE_FS] = shader;
+ ice->state.dirty |= IRIS_DIRTY_FS |
+ IRIS_DIRTY_BINDINGS_FS |
+ IRIS_DIRTY_CONSTANTS_FS |
+ IRIS_DIRTY_WM |
+ IRIS_DIRTY_CLIP |
+ IRIS_DIRTY_SBE;
+ }
+}
+
+/**
+ * Get the compiled shader for the last enabled geometry stage.
+ *
+ * This stage is the one which will feed stream output and the rasterizer.
+ */
+static gl_shader_stage
+last_vue_stage(struct iris_context *ice)
+{
+ if (ice->shaders.prog[MESA_SHADER_GEOMETRY])
+ return MESA_SHADER_GEOMETRY;
+
+ if (ice->shaders.prog[MESA_SHADER_TESS_EVAL])
+ return MESA_SHADER_TESS_EVAL;
+
+ return MESA_SHADER_VERTEX;
+}
+
+/**
+ * Update the last enabled stage's VUE map.
+ *
+ * When the shader feeding the rasterizer's output interface changes, we
+ * need to re-emit various packets.
+ */
+static void
+update_last_vue_map(struct iris_context *ice,
+ struct brw_stage_prog_data *prog_data)
+{
+ struct brw_vue_prog_data *vue_prog_data = (void *) prog_data;
+ struct brw_vue_map *vue_map = &vue_prog_data->vue_map;
+ struct brw_vue_map *old_map = ice->shaders.last_vue_map;
+ const uint64_t changed_slots =
+ (old_map ? old_map->slots_valid : 0ull) ^ vue_map->slots_valid;
+
+ if (changed_slots & VARYING_BIT_VIEWPORT) {
+ // XXX: could use ctx->Const.MaxViewports for old API efficiency
+ ice->state.num_viewports =
+ (vue_map->slots_valid & VARYING_BIT_VIEWPORT) ? IRIS_MAX_VIEWPORTS : 1;
+ ice->state.dirty |= IRIS_DIRTY_CLIP |
+ IRIS_DIRTY_SF_CL_VIEWPORT |
+ IRIS_DIRTY_CC_VIEWPORT |
+ IRIS_DIRTY_SCISSOR_RECT |
+ IRIS_DIRTY_UNCOMPILED_FS |
+ ice->state.dirty_for_nos[IRIS_NOS_LAST_VUE_MAP];
+ // XXX: CC_VIEWPORT?
+ }
- if (old != shader) {
- ice->shaders.prog[IRIS_CACHE_TCS] = shader;
- ice->state.dirty |= IRIS_DIRTY_TCS |
- IRIS_DIRTY_BINDINGS_TCS |
- IRIS_DIRTY_CONSTANTS_TCS;
+ if (changed_slots || (old_map && old_map->separate != vue_map->separate)) {
+ ice->state.dirty |= IRIS_DIRTY_SBE;
}
+
+ ice->shaders.last_vue_map = &vue_prog_data->vue_map;
}
/**
- * Compile a tessellation evaluation shader, and upload the assembly.
+ * Get the prog_data for a given stage, or NULL if the stage is disabled.
*/
-static struct iris_compiled_shader *
-iris_compile_tes(struct iris_context *ice,
- struct iris_uncompiled_shader *ish,
- const struct brw_tes_prog_key *key)
+static struct brw_vue_prog_data *
+get_vue_prog_data(struct iris_context *ice, gl_shader_stage stage)
{
- struct iris_screen *screen = (struct iris_screen *)ice->ctx.screen;
- const struct brw_compiler *compiler = screen->compiler;
- const struct gen_device_info *devinfo = &screen->devinfo;
- void *mem_ctx = ralloc_context(NULL);
- struct brw_tes_prog_data *tes_prog_data =
- rzalloc(mem_ctx, struct brw_tes_prog_data);
- struct brw_vue_prog_data *vue_prog_data = &tes_prog_data->base;
- struct brw_stage_prog_data *prog_data = &vue_prog_data->base;
- enum brw_param_builtin *system_values;
- unsigned num_system_values;
-
- nir_shader *nir = nir_shader_clone(mem_ctx, ish->nir);
+ if (!ice->shaders.prog[stage])
+ return NULL;
- iris_setup_uniforms(compiler, mem_ctx, nir, prog_data, &system_values,
- &num_system_values);
+ return (void *) ice->shaders.prog[stage]->prog_data;
+}
- assign_common_binding_table_offsets(devinfo, nir, prog_data, 0,
- num_system_values);
+// XXX: iris_compiled_shaders are space-leaking :(
+// XXX: do remember to unbind them if deleting them.
- struct brw_vue_map input_vue_map;
- brw_compute_tess_vue_map(&input_vue_map, key->inputs_read,
- key->patch_inputs_read);
+/**
+ * Update the current shader variants for the given state.
+ *
+ * This should be called on every draw call to ensure that the correct
+ * shaders are bound. It will also flag any dirty state triggered by
+ * swapping out those shaders.
+ */
+void
+iris_update_compiled_shaders(struct iris_context *ice)
+{
+ const uint64_t dirty = ice->state.dirty;
- char *error_str = NULL;
- const unsigned *program =
- brw_compile_tes(compiler, &ice->dbg, mem_ctx, key, &input_vue_map,
- tes_prog_data, nir, NULL, -1, &error_str);
- if (program == NULL) {
- dbg_printf("Failed to compile evaluation shader: %s\n", error_str);
- ralloc_free(mem_ctx);
- return false;
+ struct brw_vue_prog_data *old_prog_datas[4];
+ if (!(dirty & IRIS_DIRTY_URB)) {
+ for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++)
+ old_prog_datas[i] = get_vue_prog_data(ice, i);
}
- uint32_t *so_decls =
- ice->vtbl.create_so_decl_list(&ish->stream_output,
- &vue_prog_data->vue_map);
-
-
- struct iris_compiled_shader *shader =
- iris_upload_shader(ice, IRIS_CACHE_TES, sizeof(*key), key, program,
- prog_data, so_decls, system_values, num_system_values);
-
- if (ish->compiled_once) {
- perf_debug(&ice->dbg, "Recompiling tessellation evaluation shader\n");
- } else {
- ish->compiled_once = true;
+ if (dirty & (IRIS_DIRTY_UNCOMPILED_TCS | IRIS_DIRTY_UNCOMPILED_TES)) {
+ struct iris_uncompiled_shader *tes =
+ ice->shaders.uncompiled[MESA_SHADER_TESS_EVAL];
+ if (tes) {
+ iris_update_compiled_tcs(ice);
+ iris_update_compiled_tes(ice);
+ } else {
+ ice->shaders.prog[IRIS_CACHE_TCS] = NULL;
+ ice->shaders.prog[IRIS_CACHE_TES] = NULL;
+ ice->state.dirty |=
+ IRIS_DIRTY_TCS | IRIS_DIRTY_TES |
+ IRIS_DIRTY_BINDINGS_TCS | IRIS_DIRTY_BINDINGS_TES |
+ IRIS_DIRTY_CONSTANTS_TCS | IRIS_DIRTY_CONSTANTS_TES;
+ }
}
- ralloc_free(mem_ctx);
- return shader;
-}
-
-/**
- * Update the current tessellation evaluation shader variant.
- *
- * Fill out the key, look in the cache, compile and bind if needed.
- */
-static void
-iris_update_compiled_tes(struct iris_context *ice)
-{
- struct iris_uncompiled_shader *ish =
- ice->shaders.uncompiled[MESA_SHADER_TESS_EVAL];
+ if (dirty & IRIS_DIRTY_UNCOMPILED_VS)
+ iris_update_compiled_vs(ice);
+ if (dirty & IRIS_DIRTY_UNCOMPILED_GS)
+ iris_update_compiled_gs(ice);
- struct brw_tes_prog_key key = { KEY_INIT };
- get_unified_tess_slots(ice, &key.inputs_read, &key.patch_inputs_read);
- ice->vtbl.populate_tes_key(ice, &key);
+ gl_shader_stage last_stage = last_vue_stage(ice);
+ struct iris_compiled_shader *shader = ice->shaders.prog[last_stage];
+ struct iris_uncompiled_shader *ish = ice->shaders.uncompiled[last_stage];
+ update_last_vue_map(ice, shader->prog_data);
+ if (ice->state.streamout != shader->streamout) {
+ ice->state.streamout = shader->streamout;
+ ice->state.dirty |= IRIS_DIRTY_SO_DECL_LIST | IRIS_DIRTY_STREAMOUT;
+ }
- struct iris_compiled_shader *old = ice->shaders.prog[IRIS_CACHE_TES];
- struct iris_compiled_shader *shader =
- iris_find_cached_shader(ice, IRIS_CACHE_TES, sizeof(key), &key);
+ if (ice->state.streamout_active) {
+ for (int i = 0; i < PIPE_MAX_SO_BUFFERS; i++) {
+ struct iris_stream_output_target *so =
+ (void *) ice->state.so_target[i];
+ if (so)
+ so->stride = ish->stream_output.stride[i];
+ }
+ }
- if (!shader)
- shader = iris_compile_tes(ice, ish, &key);
+ if (dirty & IRIS_DIRTY_UNCOMPILED_FS)
+ iris_update_compiled_fs(ice);
+ // ...
- if (old != shader) {
- ice->shaders.prog[IRIS_CACHE_TES] = shader;
- ice->state.dirty |= IRIS_DIRTY_TES |
- IRIS_DIRTY_BINDINGS_TES |
- IRIS_DIRTY_CONSTANTS_TES;
+ /* Changing shader interfaces may require a URB configuration. */
+ if (!(dirty & IRIS_DIRTY_URB)) {
+ for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) {
+ struct brw_vue_prog_data *old = old_prog_datas[i];
+ struct brw_vue_prog_data *new = get_vue_prog_data(ice, i);
+ if (!!old != !!new ||
+ (new && new->urb_entry_size != old->urb_entry_size)) {
+ ice->state.dirty |= IRIS_DIRTY_URB;
+ break;
+ }
+ }
}
}
-/**
- * Compile a geometry shader, and upload the assembly.
- */
static struct iris_compiled_shader *
-iris_compile_gs(struct iris_context *ice,
+iris_compile_cs(struct iris_context *ice,
struct iris_uncompiled_shader *ish,
- const struct brw_gs_prog_key *key)
+ const struct brw_cs_prog_key *key)
{
struct iris_screen *screen = (struct iris_screen *)ice->ctx.screen;
const struct brw_compiler *compiler = screen->compiler;
const struct gen_device_info *devinfo = &screen->devinfo;
void *mem_ctx = ralloc_context(NULL);
- struct brw_gs_prog_data *gs_prog_data =
- rzalloc(mem_ctx, struct brw_gs_prog_data);
- struct brw_vue_prog_data *vue_prog_data = &gs_prog_data->base;
- struct brw_stage_prog_data *prog_data = &vue_prog_data->base;
+ struct brw_cs_prog_data *cs_prog_data =
+ rzalloc(mem_ctx, struct brw_cs_prog_data);
+ struct brw_stage_prog_data *prog_data = &cs_prog_data->base;
enum brw_param_builtin *system_values;
unsigned num_system_values;
+ unsigned num_cbufs;
nir_shader *nir = nir_shader_clone(mem_ctx, ish->nir);
- iris_setup_uniforms(compiler, mem_ctx, nir, prog_data, &system_values,
- &num_system_values);
+ cs_prog_data->binding_table.work_groups_start = 0;
- assign_common_binding_table_offsets(devinfo, nir, prog_data, 0,
- num_system_values);
+ prog_data->total_shared = nir->info.cs.shared_size;
- brw_compute_vue_map(devinfo,
- &vue_prog_data->vue_map, nir->info.outputs_written,
- nir->info.separate_shader);
+ iris_setup_uniforms(compiler, mem_ctx, nir, prog_data, &system_values,
+ &num_system_values, &num_cbufs);
+
+ assign_common_binding_table_offsets(devinfo, nir, prog_data, 1,
+ num_system_values, num_cbufs);
char *error_str = NULL;
const unsigned *program =
- brw_compile_gs(compiler, &ice->dbg, mem_ctx, key, gs_prog_data, nir,
- NULL, -1, &error_str);
+ brw_compile_cs(compiler, &ice->dbg, mem_ctx, key, cs_prog_data,
+ nir, -1, &error_str);
if (program == NULL) {
- dbg_printf("Failed to compile geometry shader: %s\n", error_str);
+ dbg_printf("Failed to compile compute shader: %s\n", error_str);
ralloc_free(mem_ctx);
return false;
}
- uint32_t *so_decls =
- ice->vtbl.create_so_decl_list(&ish->stream_output,
- &vue_prog_data->vue_map);
-
struct iris_compiled_shader *shader =
- iris_upload_shader(ice, IRIS_CACHE_GS, sizeof(*key), key, program,
- prog_data, so_decls, system_values, num_system_values);
+ iris_upload_shader(ice, IRIS_CACHE_CS, sizeof(*key), key, program,
+ prog_data, NULL, system_values, num_system_values,
+ num_cbufs);
if (ish->compiled_once) {
- perf_debug(&ice->dbg, "Recompiling geometry shader\n");
+ perf_debug(&ice->dbg, "Recompiling compute shader\n");
} else {
ish->compiled_once = true;
}
return shader;
}
-/**
- * Update the current geometry shader variant.
- *
- * Fill out the key, look in the cache, compile and bind if needed.
- */
-static void
-iris_update_compiled_gs(struct iris_context *ice)
+void
+iris_update_compiled_compute_shader(struct iris_context *ice)
{
struct iris_uncompiled_shader *ish =
- ice->shaders.uncompiled[MESA_SHADER_GEOMETRY];
- struct iris_compiled_shader *old = ice->shaders.prog[IRIS_CACHE_GS];
- struct iris_compiled_shader *shader = NULL;
+ ice->shaders.uncompiled[MESA_SHADER_COMPUTE];
- if (ish) {
- struct brw_gs_prog_key key = { KEY_INIT };
- ice->vtbl.populate_gs_key(ice, &key);
+ struct iris_screen *screen = (struct iris_screen *)ice->ctx.screen;
+ const struct gen_device_info *devinfo = &screen->devinfo;
+ struct brw_cs_prog_key key = { KEY_INIT(devinfo->gen) };
+ ice->vtbl.populate_cs_key(ice, &key);
- shader =
- iris_find_cached_shader(ice, IRIS_CACHE_GS, sizeof(key), &key);
+ struct iris_compiled_shader *old = ice->shaders.prog[IRIS_CACHE_CS];
+ struct iris_compiled_shader *shader =
+ iris_find_cached_shader(ice, IRIS_CACHE_CS, sizeof(key), &key);
- if (!shader)
- shader = iris_compile_gs(ice, ish, &key);
- }
+ if (!shader)
+ shader = iris_compile_cs(ice, ish, &key);
if (old != shader) {
- ice->shaders.prog[IRIS_CACHE_GS] = shader;
- ice->state.dirty |= IRIS_DIRTY_GS |
- IRIS_DIRTY_BINDINGS_GS |
- IRIS_DIRTY_CONSTANTS_GS;
+ ice->shaders.prog[IRIS_CACHE_CS] = shader;
+ ice->state.dirty |= IRIS_DIRTY_CS |
+ IRIS_DIRTY_BINDINGS_CS |
+ IRIS_DIRTY_CONSTANTS_CS;
}
}
+void
+iris_fill_cs_push_const_buffer(struct brw_cs_prog_data *cs_prog_data,
+ uint32_t *dst)
+{
+ struct brw_stage_prog_data *prog_data = &cs_prog_data->base;
+ assert(cs_prog_data->push.total.size > 0);
+ assert(cs_prog_data->push.cross_thread.size == 0);
+ assert(cs_prog_data->push.per_thread.dwords == 1);
+ assert(prog_data->param[0] == BRW_PARAM_BUILTIN_SUBGROUP_ID);
+ for (unsigned t = 0; t < cs_prog_data->threads; t++)
+ dst[8 * t] = t;
+}
+
/**
- * Compile a fragment (pixel) shader, and upload the assembly.
+ * Allocate scratch BOs as needed for the given per-thread size and stage.
*/
-static struct iris_compiled_shader *
-iris_compile_fs(struct iris_context *ice,
- struct iris_uncompiled_shader *ish,
- const struct brw_wm_prog_key *key,
- struct brw_vue_map *vue_map)
+struct iris_bo *
+iris_get_scratch_space(struct iris_context *ice,
+ unsigned per_thread_scratch,
+ gl_shader_stage stage)
{
struct iris_screen *screen = (struct iris_screen *)ice->ctx.screen;
- const struct brw_compiler *compiler = screen->compiler;
- const struct gen_device_info *devinfo = &screen->devinfo;
- void *mem_ctx = ralloc_context(NULL);
- struct brw_wm_prog_data *fs_prog_data =
- rzalloc(mem_ctx, struct brw_wm_prog_data);
- struct brw_stage_prog_data *prog_data = &fs_prog_data->base;
- enum brw_param_builtin *system_values;
- unsigned num_system_values;
+ struct iris_bufmgr *bufmgr = screen->bufmgr;
+ const struct gen_device_info *devinfo = &screen->devinfo;
- nir_shader *nir = nir_shader_clone(mem_ctx, ish->nir);
+ unsigned encoded_size = ffs(per_thread_scratch) - 11;
+ assert(encoded_size < (1 << 16));
- // XXX: alt mode
+ struct iris_bo **bop = &ice->shaders.scratch_bos[encoded_size][stage];
- iris_setup_uniforms(compiler, mem_ctx, nir, prog_data, &system_values,
- &num_system_values);
+ /* The documentation for 3DSTATE_PS "Scratch Space Base Pointer" says:
+ *
+ * "Scratch Space per slice is computed based on 4 sub-slices. SW
+ * must allocate scratch space enough so that each slice has 4
+ * slices allowed."
+ *
+ * According to the other driver team, this applies to compute shaders
+ * as well. This is not currently documented at all.
+ *
+ * This hack is no longer necessary on Gen11+.
+ */
+ unsigned subslice_total = screen->subslice_total;
+ if (devinfo->gen < 11)
+ subslice_total = 4 * devinfo->num_slices;
+ assert(subslice_total >= screen->subslice_total);
- assign_common_binding_table_offsets(devinfo, nir, prog_data,
- MAX2(key->nr_color_regions, 1),
- num_system_values);
- char *error_str = NULL;
- const unsigned *program =
- brw_compile_fs(compiler, &ice->dbg, mem_ctx, key, fs_prog_data,
- nir, NULL, -1, -1, -1, true, false, vue_map, &error_str);
- if (program == NULL) {
- dbg_printf("Failed to compile fragment shader: %s\n", error_str);
- ralloc_free(mem_ctx);
- return false;
- }
+ if (!*bop) {
+ unsigned scratch_ids_per_subslice = devinfo->max_cs_threads;
+ uint32_t max_threads[] = {
+ [MESA_SHADER_VERTEX] = devinfo->max_vs_threads,
+ [MESA_SHADER_TESS_CTRL] = devinfo->max_tcs_threads,
+ [MESA_SHADER_TESS_EVAL] = devinfo->max_tes_threads,
+ [MESA_SHADER_GEOMETRY] = devinfo->max_gs_threads,
+ [MESA_SHADER_FRAGMENT] = devinfo->max_wm_threads,
+ [MESA_SHADER_COMPUTE] = scratch_ids_per_subslice * subslice_total,
+ };
- struct iris_compiled_shader *shader =
- iris_upload_shader(ice, IRIS_CACHE_FS, sizeof(*key), key, program,
- prog_data, NULL, system_values, num_system_values);
+ uint32_t size = per_thread_scratch * max_threads[stage];
- if (ish->compiled_once) {
- perf_debug(&ice->dbg, "Recompiling fragment shader\n");
- } else {
- ish->compiled_once = true;
+ *bop = iris_bo_alloc(bufmgr, "scratch", size, IRIS_MEMZONE_SHADER);
}
- ralloc_free(mem_ctx);
- return shader;
+ return *bop;
}
+/* ------------------------------------------------------------------- */
+
/**
- * Update the current fragment shader variant.
+ * The pipe->create_[stage]_state() driver hooks.
*
- * Fill out the key, look in the cache, compile and bind if needed.
+ * Performs basic NIR preprocessing, records any state dependencies, and
+ * returns an iris_uncompiled_shader as the Gallium CSO.
+ *
+ * Actual shader compilation to assembly happens later, at first use.
*/
-static void
-iris_update_compiled_fs(struct iris_context *ice)
+static void *
+iris_create_uncompiled_shader(struct pipe_context *ctx,
+ nir_shader *nir,
+ const struct pipe_stream_output_info *so_info)
{
- struct iris_uncompiled_shader *ish =
- ice->shaders.uncompiled[MESA_SHADER_FRAGMENT];
- struct brw_wm_prog_key key = { KEY_INIT };
- ice->vtbl.populate_fs_key(ice, &key);
+ struct iris_screen *screen = (struct iris_screen *)ctx->screen;
+ const struct gen_device_info *devinfo = &screen->devinfo;
- if (ish->nos & (1ull << IRIS_NOS_LAST_VUE_MAP))
- key.input_slots_valid = ice->shaders.last_vue_map->slots_valid;
+ struct iris_uncompiled_shader *ish =
+ calloc(1, sizeof(struct iris_uncompiled_shader));
+ if (!ish)
+ return NULL;
- struct iris_compiled_shader *old = ice->shaders.prog[IRIS_CACHE_FS];
- struct iris_compiled_shader *shader =
- iris_find_cached_shader(ice, IRIS_CACHE_FS, sizeof(key), &key);
+ nir = brw_preprocess_nir(screen->compiler, nir, NULL);
- if (!shader)
- shader = iris_compile_fs(ice, ish, &key, ice->shaders.last_vue_map);
+ NIR_PASS_V(nir, brw_nir_lower_image_load_store, devinfo);
+ NIR_PASS_V(nir, iris_lower_storage_image_derefs);
- if (old != shader) {
- // XXX: only need to flag CLIP if barycentric has NONPERSPECTIVE
- // toggles. might be able to avoid flagging SBE too.
- ice->shaders.prog[IRIS_CACHE_FS] = shader;
- ice->state.dirty |= IRIS_DIRTY_FS |
- IRIS_DIRTY_BINDINGS_FS |
- IRIS_DIRTY_CONSTANTS_FS |
- IRIS_DIRTY_WM |
- IRIS_DIRTY_CLIP |
- IRIS_DIRTY_SBE;
+ ish->program_id = get_new_program_id(screen);
+ ish->nir = nir;
+ if (so_info) {
+ memcpy(&ish->stream_output, so_info, sizeof(*so_info));
+ update_so_info(&ish->stream_output, nir->info.outputs_written);
}
+
+ return ish;
}
-/**
- * Get the compiled shader for the last enabled geometry stage.
- *
- * This stage is the one which will feed stream output and the rasterizer.
- */
-static gl_shader_stage
-last_vue_stage(struct iris_context *ice)
+static struct iris_uncompiled_shader *
+iris_create_shader_state(struct pipe_context *ctx,
+ const struct pipe_shader_state *state)
{
- if (ice->shaders.prog[MESA_SHADER_GEOMETRY])
- return MESA_SHADER_GEOMETRY;
+ struct nir_shader *nir;
- if (ice->shaders.prog[MESA_SHADER_TESS_EVAL])
- return MESA_SHADER_TESS_EVAL;
+ if (state->type == PIPE_SHADER_IR_TGSI)
+ nir = tgsi_to_nir(state->tokens, ctx->screen);
+ else
+ nir = state->ir.nir;
- return MESA_SHADER_VERTEX;
+ return iris_create_uncompiled_shader(ctx, nir, &state->stream_output);
}
-/**
- * Update the last enabled stage's VUE map.
- *
- * When the shader feeding the rasterizer's output interface changes, we
- * need to re-emit various packets.
- */
-static void
-update_last_vue_map(struct iris_context *ice,
- struct brw_stage_prog_data *prog_data)
+static void *
+iris_create_vs_state(struct pipe_context *ctx,
+ const struct pipe_shader_state *state)
{
- struct brw_vue_prog_data *vue_prog_data = (void *) prog_data;
- struct brw_vue_map *vue_map = &vue_prog_data->vue_map;
- struct brw_vue_map *old_map = ice->shaders.last_vue_map;
- const uint64_t changed_slots =
- (old_map ? old_map->slots_valid : 0ull) ^ vue_map->slots_valid;
+ struct iris_context *ice = (void *) ctx;
+ struct iris_screen *screen = (void *) ctx->screen;
+ struct iris_uncompiled_shader *ish = iris_create_shader_state(ctx, state);
- if (changed_slots & VARYING_BIT_VIEWPORT) {
- // XXX: could use ctx->Const.MaxViewports for old API efficiency
- ice->state.num_viewports =
- (vue_map->slots_valid & VARYING_BIT_VIEWPORT) ? IRIS_MAX_VIEWPORTS : 1;
- ice->state.dirty |= IRIS_DIRTY_CLIP |
- IRIS_DIRTY_SF_CL_VIEWPORT |
- IRIS_DIRTY_CC_VIEWPORT |
- IRIS_DIRTY_SCISSOR_RECT |
- IRIS_DIRTY_UNCOMPILED_FS |
- ice->state.dirty_for_nos[IRIS_NOS_LAST_VUE_MAP];
- // XXX: CC_VIEWPORT?
- }
+ /* User clip planes */
+ if (ish->nir->info.clip_distance_array_size == 0)
+ ish->nos |= (1ull << IRIS_NOS_RASTERIZER);
- if (changed_slots || (old_map && old_map->separate != vue_map->separate)) {
- ice->state.dirty |= IRIS_DIRTY_SBE;
+ if (screen->precompile) {
+ const struct gen_device_info *devinfo = &screen->devinfo;
+ struct brw_vs_prog_key key = { KEY_INIT(devinfo->gen) };
+
+ iris_compile_vs(ice, ish, &key);
}
- ice->shaders.last_vue_map = &vue_prog_data->vue_map;
+ return ish;
}
-/**
- * Get the prog_data for a given stage, or NULL if the stage is disabled.
- */
-static struct brw_vue_prog_data *
-get_vue_prog_data(struct iris_context *ice, gl_shader_stage stage)
+static void *
+iris_create_tcs_state(struct pipe_context *ctx,
+ const struct pipe_shader_state *state)
{
- if (!ice->shaders.prog[stage])
- return NULL;
-
- return (void *) ice->shaders.prog[stage]->prog_data;
-}
-
-// XXX: iris_compiled_shaders are space-leaking :(
-// XXX: do remember to unbind them if deleting them.
+ struct iris_context *ice = (void *) ctx;
+ struct iris_screen *screen = (void *) ctx->screen;
+ struct iris_uncompiled_shader *ish = iris_create_shader_state(ctx, state);
+ struct shader_info *info = &ish->nir->info;
-/**
- * Update the current shader variants for the given state.
- *
- * This should be called on every draw call to ensure that the correct
- * shaders are bound. It will also flag any dirty state triggered by
- * swapping out those shaders.
- */
-void
-iris_update_compiled_shaders(struct iris_context *ice)
-{
- const uint64_t dirty = ice->state.dirty;
+ // XXX: NOS?
- struct brw_vue_prog_data *old_prog_datas[4];
- if (!(dirty & IRIS_DIRTY_URB)) {
- for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++)
- old_prog_datas[i] = get_vue_prog_data(ice, i);
- }
+ if (screen->precompile) {
+ const unsigned _GL_TRIANGLES = 0x0004;
+ const struct gen_device_info *devinfo = &screen->devinfo;
+ struct brw_tcs_prog_key key = {
+ KEY_INIT(devinfo->gen),
+ // XXX: make sure the linker fills this out from the TES...
+ .tes_primitive_mode =
+ info->tess.primitive_mode ? info->tess.primitive_mode
+ : _GL_TRIANGLES,
+ .outputs_written = info->outputs_written,
+ .patch_outputs_written = info->patch_outputs_written,
+ };
- if (dirty & (IRIS_DIRTY_UNCOMPILED_TCS | IRIS_DIRTY_UNCOMPILED_TES)) {
- struct iris_uncompiled_shader *tes =
- ice->shaders.uncompiled[MESA_SHADER_TESS_EVAL];
- if (tes) {
- iris_update_compiled_tcs(ice);
- iris_update_compiled_tes(ice);
- } else {
- ice->shaders.prog[IRIS_CACHE_TCS] = NULL;
- ice->shaders.prog[IRIS_CACHE_TES] = NULL;
- ice->state.dirty |=
- IRIS_DIRTY_TCS | IRIS_DIRTY_TES |
- IRIS_DIRTY_BINDINGS_TCS | IRIS_DIRTY_BINDINGS_TES |
- IRIS_DIRTY_CONSTANTS_TCS | IRIS_DIRTY_CONSTANTS_TES;
- }
+ iris_compile_tcs(ice, ish, &key);
}
- if (dirty & IRIS_DIRTY_UNCOMPILED_VS)
- iris_update_compiled_vs(ice);
- if (dirty & IRIS_DIRTY_UNCOMPILED_GS)
- iris_update_compiled_gs(ice);
+ return ish;
+}
- gl_shader_stage last_stage = last_vue_stage(ice);
- struct iris_compiled_shader *shader = ice->shaders.prog[last_stage];
- struct iris_uncompiled_shader *ish = ice->shaders.uncompiled[last_stage];
- update_last_vue_map(ice, shader->prog_data);
- if (ice->state.streamout != shader->streamout) {
- ice->state.streamout = shader->streamout;
- ice->state.dirty |= IRIS_DIRTY_SO_DECL_LIST | IRIS_DIRTY_STREAMOUT;
- }
+static void *
+iris_create_tes_state(struct pipe_context *ctx,
+ const struct pipe_shader_state *state)
+{
+ struct iris_context *ice = (void *) ctx;
+ struct iris_screen *screen = (void *) ctx->screen;
+ struct iris_uncompiled_shader *ish = iris_create_shader_state(ctx, state);
+ struct shader_info *info = &ish->nir->info;
- if (ice->state.streamout_active) {
- for (int i = 0; i < PIPE_MAX_SO_BUFFERS; i++) {
- struct iris_stream_output_target *so =
- (void *) ice->state.so_target[i];
- if (so)
- so->stride = ish->stream_output.stride[i];
- }
- }
+ // XXX: NOS?
- if (dirty & IRIS_DIRTY_UNCOMPILED_FS)
- iris_update_compiled_fs(ice);
- // ...
+ if (screen->precompile) {
+ const struct gen_device_info *devinfo = &screen->devinfo;
+ struct brw_tes_prog_key key = {
+ KEY_INIT(devinfo->gen),
+ // XXX: not ideal, need TCS output/TES input unification
+ .inputs_read = info->inputs_read,
+ .patch_inputs_read = info->patch_inputs_read,
+ };
- /* Changing shader interfaces may require a URB configuration. */
- if (!(dirty & IRIS_DIRTY_URB)) {
- for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) {
- struct brw_vue_prog_data *old = old_prog_datas[i];
- struct brw_vue_prog_data *new = get_vue_prog_data(ice, i);
- if (!!old != !!new ||
- (new && new->urb_entry_size != old->urb_entry_size)) {
- ice->state.dirty |= IRIS_DIRTY_URB;
- break;
- }
- }
+ iris_compile_tes(ice, ish, &key);
}
+
+ return ish;
}
-static struct iris_compiled_shader *
-iris_compile_cs(struct iris_context *ice,
- struct iris_uncompiled_shader *ish,
- const struct brw_cs_prog_key *key)
+static void *
+iris_create_gs_state(struct pipe_context *ctx,
+ const struct pipe_shader_state *state)
{
- struct iris_screen *screen = (struct iris_screen *)ice->ctx.screen;
- const struct brw_compiler *compiler = screen->compiler;
- const struct gen_device_info *devinfo = &screen->devinfo;
- void *mem_ctx = ralloc_context(NULL);
- struct brw_cs_prog_data *cs_prog_data =
- rzalloc(mem_ctx, struct brw_cs_prog_data);
- struct brw_stage_prog_data *prog_data = &cs_prog_data->base;
- enum brw_param_builtin *system_values;
- unsigned num_system_values;
+ struct iris_context *ice = (void *) ctx;
+ struct iris_screen *screen = (void *) ctx->screen;
+ struct iris_uncompiled_shader *ish = iris_create_shader_state(ctx, state);
- nir_shader *nir = nir_shader_clone(mem_ctx, ish->nir);
+ // XXX: NOS?
- cs_prog_data->binding_table.work_groups_start = 0;
+ if (screen->precompile) {
+ const struct gen_device_info *devinfo = &screen->devinfo;
+ struct brw_gs_prog_key key = { KEY_INIT(devinfo->gen) };
- prog_data->total_shared = nir->info.cs.shared_size;
+ iris_compile_gs(ice, ish, &key);
+ }
- iris_setup_uniforms(compiler, mem_ctx, nir, prog_data, &system_values,
- &num_system_values);
+ return ish;
+}
- assign_common_binding_table_offsets(devinfo, nir, prog_data, 1,
- num_system_values);
+static void *
+iris_create_fs_state(struct pipe_context *ctx,
+ const struct pipe_shader_state *state)
+{
+ struct iris_context *ice = (void *) ctx;
+ struct iris_screen *screen = (void *) ctx->screen;
+ struct iris_uncompiled_shader *ish = iris_create_shader_state(ctx, state);
+ struct shader_info *info = &ish->nir->info;
- char *error_str = NULL;
- const unsigned *program =
- brw_compile_cs(compiler, &ice->dbg, mem_ctx, key, cs_prog_data,
- nir, -1, &error_str);
- if (program == NULL) {
- dbg_printf("Failed to compile compute shader: %s\n", error_str);
- ralloc_free(mem_ctx);
- return false;
+ ish->nos |= (1ull << IRIS_NOS_FRAMEBUFFER) |
+ (1ull << IRIS_NOS_DEPTH_STENCIL_ALPHA) |
+ (1ull << IRIS_NOS_RASTERIZER) |
+ (1ull << IRIS_NOS_BLEND);
+
+ /* The program key needs the VUE map if there are > 16 inputs */
+ if (util_bitcount64(ish->nir->info.inputs_read &
+ BRW_FS_VARYING_INPUT_MASK) > 16) {
+ ish->nos |= (1ull << IRIS_NOS_LAST_VUE_MAP);
}
- struct iris_compiled_shader *shader =
- iris_upload_shader(ice, IRIS_CACHE_CS, sizeof(*key), key, program,
- prog_data, NULL, system_values, num_system_values);
+ if (screen->precompile) {
+ const uint64_t color_outputs = info->outputs_written &
+ ~(BITFIELD64_BIT(FRAG_RESULT_DEPTH) |
+ BITFIELD64_BIT(FRAG_RESULT_STENCIL) |
+ BITFIELD64_BIT(FRAG_RESULT_SAMPLE_MASK));
- if (ish->compiled_once) {
- perf_debug(&ice->dbg, "Recompiling compute shader\n");
- } else {
- ish->compiled_once = true;
+ bool can_rearrange_varyings =
+ util_bitcount64(info->inputs_read & BRW_FS_VARYING_INPUT_MASK) <= 16;
+
+ const struct gen_device_info *devinfo = &screen->devinfo;
+ struct brw_wm_prog_key key = {
+ KEY_INIT(devinfo->gen),
+ .nr_color_regions = util_bitcount(color_outputs),
+ .coherent_fb_fetch = true,
+ .input_slots_valid =
+ can_rearrange_varyings ? 0 : info->inputs_read | VARYING_BIT_POS,
+ };
+
+ iris_compile_fs(ice, ish, &key, NULL);
}
- ralloc_free(mem_ctx);
- return shader;
+ return ish;
}
-void
-iris_update_compiled_compute_shader(struct iris_context *ice)
+static void *
+iris_create_compute_state(struct pipe_context *ctx,
+ const struct pipe_compute_state *state)
{
- struct iris_uncompiled_shader *ish =
- ice->shaders.uncompiled[MESA_SHADER_COMPUTE];
+ assert(state->ir_type == PIPE_SHADER_IR_NIR);
- struct brw_cs_prog_key key = { KEY_INIT };
- ice->vtbl.populate_cs_key(ice, &key);
+ struct iris_context *ice = (void *) ctx;
+ struct iris_screen *screen = (void *) ctx->screen;
+ struct iris_uncompiled_shader *ish =
+ iris_create_uncompiled_shader(ctx, (void *) state->prog, NULL);
- struct iris_compiled_shader *old = ice->shaders.prog[IRIS_CACHE_CS];
- struct iris_compiled_shader *shader =
- iris_find_cached_shader(ice, IRIS_CACHE_CS, sizeof(key), &key);
+ // XXX: disallow more than 64KB of shared variables
- if (!shader)
- shader = iris_compile_cs(ice, ish, &key);
+ if (screen->precompile) {
+ const struct gen_device_info *devinfo = &screen->devinfo;
+ struct brw_cs_prog_key key = { KEY_INIT(devinfo->gen) };
- if (old != shader) {
- ice->shaders.prog[IRIS_CACHE_CS] = shader;
- ice->state.dirty |= IRIS_DIRTY_CS |
- IRIS_DIRTY_BINDINGS_CS |
- IRIS_DIRTY_CONSTANTS_CS;
+ iris_compile_cs(ice, ish, &key);
}
+
+ return ish;
}
-void
-iris_fill_cs_push_const_buffer(struct brw_cs_prog_data *cs_prog_data,
- uint32_t *dst)
+/**
+ * The pipe->delete_[stage]_state() driver hooks.
+ *
+ * Frees the iris_uncompiled_shader.
+ */
+static void
+iris_delete_shader_state(struct pipe_context *ctx, void *state)
{
- struct brw_stage_prog_data *prog_data = &cs_prog_data->base;
- assert(cs_prog_data->push.total.size > 0);
- assert(cs_prog_data->push.cross_thread.size == 0);
- assert(cs_prog_data->push.per_thread.dwords == 1);
- assert(prog_data->param[0] == BRW_PARAM_BUILTIN_SUBGROUP_ID);
- for (unsigned t = 0; t < cs_prog_data->threads; t++)
- dst[8 * t] = t;
+ struct iris_uncompiled_shader *ish = state;
+
+ ralloc_free(ish->nir);
+ free(ish);
}
/**
- * Allocate scratch BOs as needed for the given per-thread size and stage.
+ * The pipe->bind_[stage]_state() driver hook.
+ *
+ * Binds an uncompiled shader as the current one for a particular stage.
+ * Updates dirty tracking to account for the shader's NOS.
*/
-struct iris_bo *
-iris_get_scratch_space(struct iris_context *ice,
- unsigned per_thread_scratch,
- gl_shader_stage stage)
+static void
+bind_state(struct iris_context *ice,
+ struct iris_uncompiled_shader *ish,
+ gl_shader_stage stage)
{
- struct iris_screen *screen = (struct iris_screen *)ice->ctx.screen;
- struct iris_bufmgr *bufmgr = screen->bufmgr;
- const struct gen_device_info *devinfo = &screen->devinfo;
+ uint64_t dirty_bit = IRIS_DIRTY_UNCOMPILED_VS << stage;
+ const uint64_t nos = ish ? ish->nos : 0;
- unsigned encoded_size = ffs(per_thread_scratch) - 11;
- assert(encoded_size < (1 << 16));
+ const struct shader_info *old_info = iris_get_shader_info(ice, stage);
+ const struct shader_info *new_info = ish ? &ish->nir->info : NULL;
- struct iris_bo **bop = &ice->shaders.scratch_bos[encoded_size][stage];
+ if ((old_info ? util_last_bit(old_info->textures_used) : 0) !=
+ (new_info ? util_last_bit(new_info->textures_used) : 0)) {
+ ice->state.dirty |= IRIS_DIRTY_SAMPLER_STATES_VS << stage;
+ }
- /* The documentation for 3DSTATE_PS "Scratch Space Base Pointer" says:
- *
- * "Scratch Space per slice is computed based on 4 sub-slices. SW
- * must allocate scratch space enough so that each slice has 4
- * slices allowed."
- *
- * According to the other driver team, this applies to compute shaders
- * as well. This is not currently documented at all.
- *
- * This hack is no longer necessary on Gen11+.
+ ice->shaders.uncompiled[stage] = ish;
+ ice->state.dirty |= dirty_bit;
+
+ /* Record that CSOs need to mark IRIS_DIRTY_UNCOMPILED_XS when they change
+ * (or that they no longer need to do so).
*/
- unsigned subslice_total = screen->subslice_total;
- if (devinfo->gen < 11)
- subslice_total = 4 * devinfo->num_slices;
- assert(subslice_total >= screen->subslice_total);
+ for (int i = 0; i < IRIS_NOS_COUNT; i++) {
+ if (nos & (1 << i))
+ ice->state.dirty_for_nos[i] |= dirty_bit;
+ else
+ ice->state.dirty_for_nos[i] &= ~dirty_bit;
+ }
+}
- if (!*bop) {
- unsigned scratch_ids_per_subslice = devinfo->max_cs_threads;
- uint32_t max_threads[] = {
- [MESA_SHADER_VERTEX] = devinfo->max_vs_threads,
- [MESA_SHADER_TESS_CTRL] = devinfo->max_tcs_threads,
- [MESA_SHADER_TESS_EVAL] = devinfo->max_tes_threads,
- [MESA_SHADER_GEOMETRY] = devinfo->max_gs_threads,
- [MESA_SHADER_FRAGMENT] = devinfo->max_wm_threads,
- [MESA_SHADER_COMPUTE] = scratch_ids_per_subslice * subslice_total,
- };
+static void
+iris_bind_vs_state(struct pipe_context *ctx, void *state)
+{
+ bind_state((void *) ctx, state, MESA_SHADER_VERTEX);
+}
- uint32_t size = per_thread_scratch * max_threads[stage];
+static void
+iris_bind_tcs_state(struct pipe_context *ctx, void *state)
+{
+ bind_state((void *) ctx, state, MESA_SHADER_TESS_CTRL);
+}
- *bop = iris_bo_alloc(bufmgr, "scratch", size, IRIS_MEMZONE_SHADER);
- }
+static void
+iris_bind_tes_state(struct pipe_context *ctx, void *state)
+{
+ struct iris_context *ice = (struct iris_context *)ctx;
- return *bop;
+ /* Enabling/disabling optional stages requires a URB reconfiguration. */
+ if (!!state != !!ice->shaders.uncompiled[MESA_SHADER_TESS_EVAL])
+ ice->state.dirty |= IRIS_DIRTY_URB;
+
+ bind_state((void *) ctx, state, MESA_SHADER_TESS_EVAL);
+}
+
+static void
+iris_bind_gs_state(struct pipe_context *ctx, void *state)
+{
+ struct iris_context *ice = (struct iris_context *)ctx;
+
+ /* Enabling/disabling optional stages requires a URB reconfiguration. */
+ if (!!state != !!ice->shaders.uncompiled[MESA_SHADER_GEOMETRY])
+ ice->state.dirty |= IRIS_DIRTY_URB;
+
+ bind_state((void *) ctx, state, MESA_SHADER_GEOMETRY);
+}
+
+static void
+iris_bind_fs_state(struct pipe_context *ctx, void *state)
+{
+ struct iris_context *ice = (struct iris_context *) ctx;
+ struct iris_uncompiled_shader *old_ish =
+ ice->shaders.uncompiled[MESA_SHADER_FRAGMENT];
+ struct iris_uncompiled_shader *new_ish = state;
+
+ const unsigned color_bits =
+ BITFIELD64_BIT(FRAG_RESULT_COLOR) |
+ BITFIELD64_RANGE(FRAG_RESULT_DATA0, BRW_MAX_DRAW_BUFFERS);
+
+ /* Fragment shader outputs influence HasWriteableRT */
+ if (!old_ish || !new_ish ||
+ (old_ish->nir->info.outputs_written & color_bits) !=
+ (new_ish->nir->info.outputs_written & color_bits))
+ ice->state.dirty |= IRIS_DIRTY_PS_BLEND;
+
+ bind_state((void *) ctx, state, MESA_SHADER_FRAGMENT);
+}
+
+static void
+iris_bind_cs_state(struct pipe_context *ctx, void *state)
+{
+ bind_state((void *) ctx, state, MESA_SHADER_COMPUTE);
}
void