#include "../vulkan/radv_descriptor_set.h"
#include "util/bitscan.h"
#include <llvm-c/Transforms/Scalar.h>
+#include "ac_shader_info.h"
+#include "ac_exp_param.h"
enum radeon_llvm_calling_convention {
RADEON_LLVM_AMDGPU_VS = 87,
struct ac_llvm_context ac;
const struct ac_nir_compiler_options *options;
struct ac_shader_variant_info *shader_info;
-
+ unsigned max_workgroup_size;
LLVMContextRef context;
LLVMModuleRef module;
LLVMBuilderRef builder;
LLVMValueRef rel_auto_id;
LLVMValueRef vs_prim_id;
LLVMValueRef instance_id;
-
+ LLVMValueRef ls_out_layout;
LLVMValueRef es2gs_offset;
+ LLVMValueRef tcs_offchip_layout;
+ LLVMValueRef tcs_out_offsets;
+ LLVMValueRef tcs_out_layout;
+ LLVMValueRef tcs_in_layout;
+ LLVMValueRef oc_lds;
+ LLVMValueRef tess_factor_offset;
+ LLVMValueRef tcs_patch_id;
+ LLVMValueRef tcs_rel_ids;
+ LLVMValueRef tes_rel_patch_id;
+ LLVMValueRef tes_patch_id;
+ LLVMValueRef tes_u;
+ LLVMValueRef tes_v;
+
LLVMValueRef gsvs_ring_stride;
LLVMValueRef gsvs_num_entries;
LLVMValueRef gs2vs_offset;
LLVMValueRef esgs_ring;
LLVMValueRef gsvs_ring;
+ LLVMValueRef hs_ring_tess_offchip;
+ LLVMValueRef hs_ring_tess_factor;
LLVMValueRef prim_mask;
- LLVMValueRef sample_positions;
+ LLVMValueRef sample_pos_offset;
LLVMValueRef persp_sample, persp_center, persp_centroid;
LLVMValueRef linear_sample, linear_center, linear_centroid;
LLVMValueRef front_face;
bool is_gs_copy_shader;
LLVMValueRef gs_next_vertex;
unsigned gs_max_out_vertices;
+
+ unsigned tes_primitive_mode;
+ uint64_t tess_outputs_written;
+ uint64_t tess_patch_outputs_written;
};
static LLVMValueRef get_sampler_desc(struct nir_to_llvm_context *ctx,
static unsigned shader_io_get_unique_index(gl_varying_slot slot)
{
+ /* handle patch indices separate */
+ if (slot == VARYING_SLOT_TESS_LEVEL_OUTER)
+ return 0;
+ if (slot == VARYING_SLOT_TESS_LEVEL_INNER)
+ return 1;
+ if (slot >= VARYING_SLOT_PATCH0 && slot <= VARYING_SLOT_TESS_MAX)
+ return 2 + (slot - VARYING_SLOT_PATCH0);
+
if (slot == VARYING_SLOT_POS)
return 0;
if (slot == VARYING_SLOT_PSIZ)
LLVMBuilderRef builder, LLVMTypeRef *return_types,
unsigned num_return_elems, LLVMTypeRef *param_types,
unsigned param_count, unsigned array_params_mask,
- unsigned sgpr_params, bool unsafe_math)
+ unsigned sgpr_params, unsigned max_workgroup_size,
+ bool unsafe_math)
{
LLVMTypeRef main_function_type, ret_type;
LLVMBasicBlockRef main_function_body;
}
}
+ if (max_workgroup_size) {
+ ac_llvm_add_target_dep_function_attr(main_function,
+ "amdgpu-max-work-group-size",
+ max_workgroup_size);
+ }
if (unsafe_math) {
/* These were copied from some LLVM test. */
LLVMAddTargetDependentFunctionAttr(main_function,
return value;
}
+static LLVMValueRef get_rel_patch_id(struct nir_to_llvm_context *ctx)
+{
+ switch (ctx->stage) {
+ case MESA_SHADER_TESS_CTRL:
+ return unpack_param(ctx, ctx->tcs_rel_ids, 0, 8);
+ case MESA_SHADER_TESS_EVAL:
+ return ctx->tes_rel_patch_id;
+ break;
+ default:
+ unreachable("Illegal stage");
+ }
+}
+
+/* Tessellation shaders pass outputs to the next shader using LDS.
+ *
+ * LS outputs = TCS inputs
+ * TCS outputs = TES inputs
+ *
+ * The LDS layout is:
+ * - TCS inputs for patch 0
+ * - TCS inputs for patch 1
+ * - TCS inputs for patch 2 = get_tcs_in_current_patch_offset (if RelPatchID==2)
+ * - ...
+ * - TCS outputs for patch 0 = get_tcs_out_patch0_offset
+ * - Per-patch TCS outputs for patch 0 = get_tcs_out_patch0_patch_data_offset
+ * - TCS outputs for patch 1
+ * - Per-patch TCS outputs for patch 1
+ * - TCS outputs for patch 2 = get_tcs_out_current_patch_offset (if RelPatchID==2)
+ * - Per-patch TCS outputs for patch 2 = get_tcs_out_current_patch_data_offset (if RelPatchID==2)
+ * - ...
+ *
+ * All three shaders VS(LS), TCS, TES share the same LDS space.
+ */
+static LLVMValueRef
+get_tcs_in_patch_stride(struct nir_to_llvm_context *ctx)
+{
+ if (ctx->stage == MESA_SHADER_VERTEX)
+ return unpack_param(ctx, ctx->ls_out_layout, 0, 13);
+ else if (ctx->stage == MESA_SHADER_TESS_CTRL)
+ return unpack_param(ctx, ctx->tcs_in_layout, 0, 13);
+ else {
+ assert(0);
+ return NULL;
+ }
+}
+
+static LLVMValueRef
+get_tcs_out_patch_stride(struct nir_to_llvm_context *ctx)
+{
+ return unpack_param(ctx, ctx->tcs_out_layout, 0, 13);
+}
+
+static LLVMValueRef
+get_tcs_out_patch0_offset(struct nir_to_llvm_context *ctx)
+{
+ return LLVMBuildMul(ctx->builder,
+ unpack_param(ctx, ctx->tcs_out_offsets, 0, 16),
+ LLVMConstInt(ctx->i32, 4, false), "");
+}
+
+static LLVMValueRef
+get_tcs_out_patch0_patch_data_offset(struct nir_to_llvm_context *ctx)
+{
+ return LLVMBuildMul(ctx->builder,
+ unpack_param(ctx, ctx->tcs_out_offsets, 16, 16),
+ LLVMConstInt(ctx->i32, 4, false), "");
+}
+
+static LLVMValueRef
+get_tcs_in_current_patch_offset(struct nir_to_llvm_context *ctx)
+{
+ LLVMValueRef patch_stride = get_tcs_in_patch_stride(ctx);
+ LLVMValueRef rel_patch_id = get_rel_patch_id(ctx);
+
+ return LLVMBuildMul(ctx->builder, patch_stride, rel_patch_id, "");
+}
+
+static LLVMValueRef
+get_tcs_out_current_patch_offset(struct nir_to_llvm_context *ctx)
+{
+ LLVMValueRef patch0_offset = get_tcs_out_patch0_offset(ctx);
+ LLVMValueRef patch_stride = get_tcs_out_patch_stride(ctx);
+ LLVMValueRef rel_patch_id = get_rel_patch_id(ctx);
+
+ return LLVMBuildAdd(ctx->builder, patch0_offset,
+ LLVMBuildMul(ctx->builder, patch_stride,
+ rel_patch_id, ""),
+ "");
+}
+
+static LLVMValueRef
+get_tcs_out_current_patch_data_offset(struct nir_to_llvm_context *ctx)
+{
+ LLVMValueRef patch0_patch_data_offset =
+ get_tcs_out_patch0_patch_data_offset(ctx);
+ LLVMValueRef patch_stride = get_tcs_out_patch_stride(ctx);
+ LLVMValueRef rel_patch_id = get_rel_patch_id(ctx);
+
+ return LLVMBuildAdd(ctx->builder, patch0_patch_data_offset,
+ LLVMBuildMul(ctx->builder, patch_stride,
+ rel_patch_id, ""),
+ "");
+}
+
static void set_userdata_location(struct ac_userdata_info *ud_info, uint8_t sgpr_idx, uint8_t num_sgprs)
{
ud_info->sgpr_idx = sgpr_idx;
set_userdata_location(&ctx->shader_info->user_sgprs_locs.shader_data[idx], sgpr_idx, num_sgprs);
}
-#if 0
+
static void set_userdata_location_indirect(struct ac_userdata_info *ud_info, uint8_t sgpr_idx, uint8_t num_sgprs,
uint32_t indirect_offset)
{
ud_info->indirect = true;
ud_info->indirect_offset = indirect_offset;
}
-#endif
+
+static void declare_tess_lds(struct nir_to_llvm_context *ctx)
+{
+ unsigned lds_size = ctx->options->chip_class >= CIK ? 65536 : 32768;
+ ctx->lds = LLVMBuildIntToPtr(ctx->builder, ctx->i32zero,
+ LLVMPointerType(LLVMArrayType(ctx->i32, lds_size / 4), LOCAL_ADDR_SPACE),
+ "tess_lds");
+}
+
+struct user_sgpr_info {
+ bool need_ring_offsets;
+ uint8_t sgpr_count;
+ bool indirect_all_descriptor_sets;
+};
+
+static void allocate_user_sgprs(struct nir_to_llvm_context *ctx,
+ struct user_sgpr_info *user_sgpr_info)
+{
+ memset(user_sgpr_info, 0, sizeof(struct user_sgpr_info));
+
+ /* until we sort out scratch/global buffers always assign ring offsets for gs/vs/es */
+ if (ctx->stage == MESA_SHADER_GEOMETRY ||
+ ctx->stage == MESA_SHADER_VERTEX ||
+ ctx->stage == MESA_SHADER_TESS_CTRL ||
+ ctx->stage == MESA_SHADER_TESS_EVAL ||
+ ctx->is_gs_copy_shader)
+ user_sgpr_info->need_ring_offsets = true;
+
+ if (ctx->stage == MESA_SHADER_FRAGMENT &&
+ ctx->shader_info->info.ps.needs_sample_positions)
+ user_sgpr_info->need_ring_offsets = true;
+
+ /* 2 user sgprs will nearly always be allocated for scratch/rings */
+ if (ctx->options->supports_spill || user_sgpr_info->need_ring_offsets) {
+ user_sgpr_info->sgpr_count += 2;
+ }
+
+ switch (ctx->stage) {
+ case MESA_SHADER_COMPUTE:
+ user_sgpr_info->sgpr_count += ctx->shader_info->info.cs.grid_components_used;
+ break;
+ case MESA_SHADER_FRAGMENT:
+ user_sgpr_info->sgpr_count += ctx->shader_info->info.ps.needs_sample_positions;
+ break;
+ case MESA_SHADER_VERTEX:
+ if (!ctx->is_gs_copy_shader) {
+ user_sgpr_info->sgpr_count += ctx->shader_info->info.vs.has_vertex_buffers ? 2 : 0;
+ if (ctx->shader_info->info.vs.needs_draw_id) {
+ user_sgpr_info->sgpr_count += 3;
+ } else {
+ user_sgpr_info->sgpr_count += 2;
+ }
+ }
+ if (ctx->options->key.vs.as_ls)
+ user_sgpr_info->sgpr_count++;
+ break;
+ case MESA_SHADER_TESS_CTRL:
+ user_sgpr_info->sgpr_count += 4;
+ break;
+ case MESA_SHADER_TESS_EVAL:
+ user_sgpr_info->sgpr_count += 1;
+ break;
+ case MESA_SHADER_GEOMETRY:
+ user_sgpr_info->sgpr_count += 2;
+ break;
+ default:
+ break;
+ }
+
+ if (ctx->shader_info->info.needs_push_constants)
+ user_sgpr_info->sgpr_count += 2;
+
+ uint32_t remaining_sgprs = 16 - user_sgpr_info->sgpr_count;
+ if (remaining_sgprs / 2 < util_bitcount(ctx->shader_info->info.desc_set_used_mask)) {
+ user_sgpr_info->sgpr_count += 2;
+ user_sgpr_info->indirect_all_descriptor_sets = true;
+ } else {
+ user_sgpr_info->sgpr_count += util_bitcount(ctx->shader_info->info.desc_set_used_mask) * 2;
+ }
+}
static void create_function(struct nir_to_llvm_context *ctx)
{
unsigned i;
unsigned num_sets = ctx->options->layout ? ctx->options->layout->num_sets : 0;
unsigned user_sgpr_idx;
- bool need_push_constants;
- bool need_ring_offsets = false;
-
- /* until we sort out scratch/global buffers always assign ring offsets for gs/vs/es */
- if (ctx->stage == MESA_SHADER_GEOMETRY ||
- ctx->stage == MESA_SHADER_VERTEX ||
- ctx->is_gs_copy_shader)
- need_ring_offsets = true;
+ struct user_sgpr_info user_sgpr_info;
- need_push_constants = true;
- if (!ctx->options->layout)
- need_push_constants = false;
- else if (!ctx->options->layout->push_constant_size &&
- !ctx->options->layout->dynamic_offset_count)
- need_push_constants = false;
-
- if (need_ring_offsets && !ctx->options->supports_spill) {
- arg_types[arg_idx++] = const_array(ctx->v16i8, 8); /* address of rings */
+ allocate_user_sgprs(ctx, &user_sgpr_info);
+ if (user_sgpr_info.need_ring_offsets && !ctx->options->supports_spill) {
+ arg_types[arg_idx++] = const_array(ctx->v16i8, 16); /* address of rings */
}
/* 1 for each descriptor set */
- for (unsigned i = 0; i < num_sets; ++i) {
- if (ctx->options->layout->set[i].layout->shader_stages & (1 << ctx->stage)) {
- array_params_mask |= (1 << arg_idx);
- arg_types[arg_idx++] = const_array(ctx->i8, 1024 * 1024);
+ if (!user_sgpr_info.indirect_all_descriptor_sets) {
+ for (unsigned i = 0; i < num_sets; ++i) {
+ if (ctx->options->layout->set[i].layout->shader_stages & (1 << ctx->stage)) {
+ array_params_mask |= (1 << arg_idx);
+ arg_types[arg_idx++] = const_array(ctx->i8, 1024 * 1024);
+ }
}
+ } else {
+ array_params_mask |= (1 << arg_idx);
+ arg_types[arg_idx++] = const_array(const_array(ctx->i8, 1024 * 1024), 32);
}
- if (need_push_constants) {
+ if (ctx->shader_info->info.needs_push_constants) {
/* 1 for push constants and dynamic descriptors */
array_params_mask |= (1 << arg_idx);
arg_types[arg_idx++] = const_array(ctx->i8, 1024 * 1024);
switch (ctx->stage) {
case MESA_SHADER_COMPUTE:
- arg_types[arg_idx++] = LLVMVectorType(ctx->i32, 3); /* grid size */
+ if (ctx->shader_info->info.cs.grid_components_used)
+ arg_types[arg_idx++] = LLVMVectorType(ctx->i32, ctx->shader_info->info.cs.grid_components_used); /* grid size */
user_sgpr_count = arg_idx;
arg_types[arg_idx++] = LLVMVectorType(ctx->i32, 3);
arg_types[arg_idx++] = ctx->i32;
break;
case MESA_SHADER_VERTEX:
if (!ctx->is_gs_copy_shader) {
- arg_types[arg_idx++] = const_array(ctx->v16i8, 16); /* vertex buffers */
+ if (ctx->shader_info->info.vs.has_vertex_buffers)
+ arg_types[arg_idx++] = const_array(ctx->v16i8, 16); /* vertex buffers */
arg_types[arg_idx++] = ctx->i32; // base vertex
arg_types[arg_idx++] = ctx->i32; // start instance
- arg_types[arg_idx++] = ctx->i32; // draw index
+ if (ctx->shader_info->info.vs.needs_draw_id)
+ arg_types[arg_idx++] = ctx->i32; // draw index
}
user_sgpr_count = arg_idx;
if (ctx->options->key.vs.as_es)
arg_types[arg_idx++] = ctx->i32; //es2gs offset
+ else if (ctx->options->key.vs.as_ls) {
+ arg_types[arg_idx++] = ctx->i32; //ls out layout
+ user_sgpr_count++;
+ }
sgpr_count = arg_idx;
arg_types[arg_idx++] = ctx->i32; // vertex id
if (!ctx->is_gs_copy_shader) {
arg_types[arg_idx++] = ctx->i32; // instance id
}
break;
+ case MESA_SHADER_TESS_CTRL:
+ arg_types[arg_idx++] = ctx->i32; // tcs offchip layout
+ arg_types[arg_idx++] = ctx->i32; // tcs out offsets
+ arg_types[arg_idx++] = ctx->i32; // tcs out layout
+ arg_types[arg_idx++] = ctx->i32; // tcs in layout
+ user_sgpr_count = arg_idx;
+ arg_types[arg_idx++] = ctx->i32; // param oc lds
+ arg_types[arg_idx++] = ctx->i32; // tess factor offset
+ sgpr_count = arg_idx;
+ arg_types[arg_idx++] = ctx->i32; // patch id
+ arg_types[arg_idx++] = ctx->i32; // rel ids;
+ break;
+ case MESA_SHADER_TESS_EVAL:
+ arg_types[arg_idx++] = ctx->i32; // tcs offchip layout
+ user_sgpr_count = arg_idx;
+ if (ctx->options->key.tes.as_es) {
+ arg_types[arg_idx++] = ctx->i32; // OC LDS
+ arg_types[arg_idx++] = ctx->i32; //
+ arg_types[arg_idx++] = ctx->i32; // es2gs offset
+ } else {
+ arg_types[arg_idx++] = ctx->i32; //
+ arg_types[arg_idx++] = ctx->i32; // OC LDS
+ }
+ sgpr_count = arg_idx;
+ arg_types[arg_idx++] = ctx->f32; // tes_u
+ arg_types[arg_idx++] = ctx->f32; // tes_v
+ arg_types[arg_idx++] = ctx->i32; // tes rel patch id
+ arg_types[arg_idx++] = ctx->i32; // tes patch id
+ break;
case MESA_SHADER_GEOMETRY:
arg_types[arg_idx++] = ctx->i32; // gsvs stride
arg_types[arg_idx++] = ctx->i32; // gsvs num entires
arg_types[arg_idx++] = ctx->i32; // GS instance id
break;
case MESA_SHADER_FRAGMENT:
- arg_types[arg_idx++] = const_array(ctx->f32, 32); /* sample positions */
+ if (ctx->shader_info->info.ps.needs_sample_positions)
+ arg_types[arg_idx++] = ctx->i32; /* sample position offset */
user_sgpr_count = arg_idx;
arg_types[arg_idx++] = ctx->i32; /* prim mask */
sgpr_count = arg_idx;
ctx->main_function = create_llvm_function(
ctx->context, ctx->module, ctx->builder, NULL, 0, arg_types,
- arg_idx, array_params_mask, sgpr_count, ctx->options->unsafe_math);
+ arg_idx, array_params_mask, sgpr_count, ctx->max_workgroup_size,
+ ctx->options->unsafe_math);
set_llvm_calling_convention(ctx->main_function, ctx->stage);
ctx->shader_info->num_input_sgprs = 0;
arg_idx = 0;
user_sgpr_idx = 0;
- if (ctx->options->supports_spill || need_ring_offsets) {
+ if (ctx->options->supports_spill || user_sgpr_info.need_ring_offsets) {
set_userdata_location_shader(ctx, AC_UD_SCRATCH_RING_OFFSETS, user_sgpr_idx, 2);
user_sgpr_idx += 2;
if (ctx->options->supports_spill) {
LLVMPointerType(ctx->i8, CONST_ADDR_SPACE),
NULL, 0, AC_FUNC_ATTR_READNONE);
ctx->ring_offsets = LLVMBuildBitCast(ctx->builder, ctx->ring_offsets,
- const_array(ctx->v16i8, 8), "");
+ const_array(ctx->v16i8, 16), "");
} else
ctx->ring_offsets = LLVMGetParam(ctx->main_function, arg_idx++);
}
- for (unsigned i = 0; i < num_sets; ++i) {
- if (ctx->options->layout->set[i].layout->shader_stages & (1 << ctx->stage)) {
- set_userdata_location(&ctx->shader_info->user_sgprs_locs.descriptor_sets[i], user_sgpr_idx, 2);
- user_sgpr_idx += 2;
- ctx->descriptor_sets[i] =
- LLVMGetParam(ctx->main_function, arg_idx++);
- } else
- ctx->descriptor_sets[i] = NULL;
+ if (!user_sgpr_info.indirect_all_descriptor_sets) {
+ for (unsigned i = 0; i < num_sets; ++i) {
+ if (ctx->options->layout->set[i].layout->shader_stages & (1 << ctx->stage)) {
+ set_userdata_location(&ctx->shader_info->user_sgprs_locs.descriptor_sets[i], user_sgpr_idx, 2);
+ user_sgpr_idx += 2;
+ ctx->descriptor_sets[i] =
+ LLVMGetParam(ctx->main_function, arg_idx++);
+ } else
+ ctx->descriptor_sets[i] = NULL;
+ }
+ } else {
+ uint32_t desc_sgpr_idx = user_sgpr_idx;
+ LLVMValueRef desc_sets = LLVMGetParam(ctx->main_function, arg_idx++);
+ set_userdata_location_shader(ctx, AC_UD_INDIRECT_DESCRIPTOR_SETS, user_sgpr_idx, 2);
+ user_sgpr_idx += 2;
+
+ for (unsigned i = 0; i < num_sets; ++i) {
+ if (ctx->options->layout->set[i].layout->shader_stages & (1 << ctx->stage)) {
+ set_userdata_location_indirect(&ctx->shader_info->user_sgprs_locs.descriptor_sets[i], desc_sgpr_idx, 2, i * 8);
+ ctx->descriptor_sets[i] = ac_build_indexed_load_const(&ctx->ac, desc_sets, LLVMConstInt(ctx->i32, i, false));
+
+ } else
+ ctx->descriptor_sets[i] = NULL;
+ }
+ ctx->shader_info->need_indirect_descriptor_sets = true;
}
- if (need_push_constants) {
+ if (ctx->shader_info->info.needs_push_constants) {
ctx->push_constants = LLVMGetParam(ctx->main_function, arg_idx++);
set_userdata_location_shader(ctx, AC_UD_PUSH_CONSTANTS, user_sgpr_idx, 2);
user_sgpr_idx += 2;
switch (ctx->stage) {
case MESA_SHADER_COMPUTE:
- set_userdata_location_shader(ctx, AC_UD_CS_GRID_SIZE, user_sgpr_idx, 3);
- user_sgpr_idx += 3;
- ctx->num_work_groups =
- LLVMGetParam(ctx->main_function, arg_idx++);
+ if (ctx->shader_info->info.cs.grid_components_used) {
+ set_userdata_location_shader(ctx, AC_UD_CS_GRID_SIZE, user_sgpr_idx, ctx->shader_info->info.cs.grid_components_used);
+ user_sgpr_idx += ctx->shader_info->info.cs.grid_components_used;
+ ctx->num_work_groups =
+ LLVMGetParam(ctx->main_function, arg_idx++);
+ }
ctx->workgroup_ids =
LLVMGetParam(ctx->main_function, arg_idx++);
ctx->tg_size =
break;
case MESA_SHADER_VERTEX:
if (!ctx->is_gs_copy_shader) {
- set_userdata_location_shader(ctx, AC_UD_VS_VERTEX_BUFFERS, user_sgpr_idx, 2);
- user_sgpr_idx += 2;
- ctx->vertex_buffers = LLVMGetParam(ctx->main_function, arg_idx++);
- set_userdata_location_shader(ctx, AC_UD_VS_BASE_VERTEX_START_INSTANCE, user_sgpr_idx, 3);
- user_sgpr_idx += 3;
+ if (ctx->shader_info->info.vs.has_vertex_buffers) {
+ set_userdata_location_shader(ctx, AC_UD_VS_VERTEX_BUFFERS, user_sgpr_idx, 2);
+ user_sgpr_idx += 2;
+ ctx->vertex_buffers = LLVMGetParam(ctx->main_function, arg_idx++);
+ }
+ unsigned vs_num = 2;
+ if (ctx->shader_info->info.vs.needs_draw_id)
+ vs_num++;
+
+ set_userdata_location_shader(ctx, AC_UD_VS_BASE_VERTEX_START_INSTANCE, user_sgpr_idx, vs_num);
+ user_sgpr_idx += vs_num;
+
ctx->base_vertex = LLVMGetParam(ctx->main_function, arg_idx++);
ctx->start_instance = LLVMGetParam(ctx->main_function, arg_idx++);
- ctx->draw_index = LLVMGetParam(ctx->main_function, arg_idx++);
+ if (ctx->shader_info->info.vs.needs_draw_id)
+ ctx->draw_index = LLVMGetParam(ctx->main_function, arg_idx++);
}
if (ctx->options->key.vs.as_es)
ctx->es2gs_offset = LLVMGetParam(ctx->main_function, arg_idx++);
+ else if (ctx->options->key.vs.as_ls) {
+ set_userdata_location_shader(ctx, AC_UD_VS_LS_TCS_IN_LAYOUT, user_sgpr_idx, 1);
+ user_sgpr_idx += 1;
+ ctx->ls_out_layout = LLVMGetParam(ctx->main_function, arg_idx++);
+ }
ctx->vertex_id = LLVMGetParam(ctx->main_function, arg_idx++);
if (!ctx->is_gs_copy_shader) {
ctx->rel_auto_id = LLVMGetParam(ctx->main_function, arg_idx++);
ctx->vs_prim_id = LLVMGetParam(ctx->main_function, arg_idx++);
ctx->instance_id = LLVMGetParam(ctx->main_function, arg_idx++);
}
+ if (ctx->options->key.vs.as_ls)
+ declare_tess_lds(ctx);
+ break;
+ case MESA_SHADER_TESS_CTRL:
+ set_userdata_location_shader(ctx, AC_UD_TCS_OFFCHIP_LAYOUT, user_sgpr_idx, 4);
+ user_sgpr_idx += 4;
+ ctx->tcs_offchip_layout = LLVMGetParam(ctx->main_function, arg_idx++);
+ ctx->tcs_out_offsets = LLVMGetParam(ctx->main_function, arg_idx++);
+ ctx->tcs_out_layout = LLVMGetParam(ctx->main_function, arg_idx++);
+ ctx->tcs_in_layout = LLVMGetParam(ctx->main_function, arg_idx++);
+ ctx->oc_lds = LLVMGetParam(ctx->main_function, arg_idx++);
+ ctx->tess_factor_offset = LLVMGetParam(ctx->main_function, arg_idx++);
+ ctx->tcs_patch_id = LLVMGetParam(ctx->main_function, arg_idx++);
+ ctx->tcs_rel_ids = LLVMGetParam(ctx->main_function, arg_idx++);
+
+ declare_tess_lds(ctx);
+ break;
+ case MESA_SHADER_TESS_EVAL:
+ set_userdata_location_shader(ctx, AC_UD_TES_OFFCHIP_LAYOUT, user_sgpr_idx, 1);
+ user_sgpr_idx += 1;
+ ctx->tcs_offchip_layout = LLVMGetParam(ctx->main_function, arg_idx++);
+ if (ctx->options->key.tes.as_es) {
+ ctx->oc_lds = LLVMGetParam(ctx->main_function, arg_idx++);
+ arg_idx++;
+ ctx->es2gs_offset = LLVMGetParam(ctx->main_function, arg_idx++);
+ } else {
+ arg_idx++;
+ ctx->oc_lds = LLVMGetParam(ctx->main_function, arg_idx++);
+ }
+ ctx->tes_u = LLVMGetParam(ctx->main_function, arg_idx++);
+ ctx->tes_v = LLVMGetParam(ctx->main_function, arg_idx++);
+ ctx->tes_rel_patch_id = LLVMGetParam(ctx->main_function, arg_idx++);
+ ctx->tes_patch_id = LLVMGetParam(ctx->main_function, arg_idx++);
break;
case MESA_SHADER_GEOMETRY:
set_userdata_location_shader(ctx, AC_UD_GS_VS_RING_STRIDE_ENTRIES, user_sgpr_idx, 2);
ctx->gs_invocation_id = LLVMGetParam(ctx->main_function, arg_idx++);
break;
case MESA_SHADER_FRAGMENT:
- set_userdata_location_shader(ctx, AC_UD_PS_SAMPLE_POS, user_sgpr_idx, 2);
- user_sgpr_idx += 2;
- ctx->sample_positions = LLVMGetParam(ctx->main_function, arg_idx++);
+ if (ctx->shader_info->info.ps.needs_sample_positions) {
+ set_userdata_location_shader(ctx, AC_UD_PS_SAMPLE_POS_OFFSET, user_sgpr_idx, 1);
+ user_sgpr_idx += 1;
+ ctx->sample_pos_offset = LLVMGetParam(ctx->main_function, arg_idx++);
+ }
ctx->prim_mask = LLVMGetParam(ctx->main_function, arg_idx++);
ctx->persp_sample = LLVMGetParam(ctx->main_function, arg_idx++);
ctx->persp_center = LLVMGetParam(ctx->main_function, arg_idx++);
return LLVMBuildAnd(ctx->builder, src0, LLVMBuildBitCast(ctx->builder, LLVMConstReal(ctx->f32, 1.0), ctx->i32, ""), "");
}
+static LLVMValueRef emit_f2f16(struct nir_to_llvm_context *ctx,
+ LLVMValueRef src0)
+{
+ LLVMValueRef result;
+ LLVMValueRef cond;
+
+ src0 = to_float(ctx, src0);
+ result = LLVMBuildFPTrunc(ctx->builder, src0, ctx->f16, "");
+
+ /* TODO SI/CIK options here */
+ if (ctx->options->chip_class >= VI) {
+ LLVMValueRef args[2];
+ /* Check if the result is a denormal - and flush to 0 if so. */
+ args[0] = result;
+ args[1] = LLVMConstInt(ctx->i32, N_SUBNORMAL | P_SUBNORMAL, false);
+ cond = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.class.f16", ctx->i1, args, 2, AC_FUNC_ATTR_READNONE);
+ }
+
+ /* need to convert back up to f32 */
+ result = LLVMBuildFPExt(ctx->builder, result, ctx->f32, "");
+
+ if (ctx->options->chip_class >= VI)
+ result = LLVMBuildSelect(ctx->builder, cond, ctx->f32zero, result, "");
+
+ return result;
+}
+
static LLVMValueRef emit_umul_high(struct nir_to_llvm_context *ctx,
LLVMValueRef src0, LLVMValueRef src1)
{
case nir_op_fmax:
result = emit_intrin_2f_param(ctx, "llvm.maxnum",
to_float_type(ctx, def_type), src[0], src[1]);
+ if (instr->dest.dest.ssa.bit_size == 32)
+ result = emit_intrin_1f_param(ctx, "llvm.canonicalize",
+ to_float_type(ctx, def_type),
+ result);
break;
case nir_op_fmin:
result = emit_intrin_2f_param(ctx, "llvm.minnum",
to_float_type(ctx, def_type), src[0], src[1]);
+ if (instr->dest.dest.ssa.bit_size == 32)
+ result = emit_intrin_1f_param(ctx, "llvm.canonicalize",
+ to_float_type(ctx, def_type),
+ result);
break;
case nir_op_ffma:
result = emit_intrin_3f_param(ctx, "llvm.fma",
result = emit_b2f(ctx, src[0]);
break;
case nir_op_fquantize2f16:
- src[0] = to_float(ctx, src[0]);
- result = LLVMBuildFPTrunc(ctx->builder, src[0], ctx->f16, "");
- /* need to convert back up to f32 */
- result = LLVMBuildFPExt(ctx->builder, result, ctx->f32, "");
+ result = emit_f2f16(ctx, src[0]);
break;
case nir_op_umul_high:
result = emit_umul_high(ctx, src[0], src[1]);
static LLVMValueRef build_tex_intrinsic(struct nir_to_llvm_context *ctx,
nir_tex_instr *instr,
+ bool lod_is_zero,
struct ac_image_args *args)
{
if (instr->sampler_dim == GLSL_SAMPLER_DIM_BUF) {
args->bias = true;
break;
case nir_texop_txl:
- args->lod = true;
+ if (lod_is_zero)
+ args->level_zero = true;
+ else
+ args->lod = true;
break;
case nir_texop_txs:
case nir_texop_query_levels:
}
static void
-radv_get_deref_offset(struct nir_to_llvm_context *ctx, nir_deref *tail,
+radv_get_deref_offset(struct nir_to_llvm_context *ctx, nir_deref_var *deref,
bool vs_in, unsigned *vertex_index_out,
+ LLVMValueRef *vertex_index_ref,
unsigned *const_out, LLVMValueRef *indir_out)
{
unsigned const_offset = 0;
+ nir_deref *tail = &deref->deref;
LLVMValueRef offset = NULL;
- if (vertex_index_out != NULL) {
+ if (vertex_index_out != NULL || vertex_index_ref != NULL) {
tail = tail->child;
nir_deref_array *deref_array = nir_deref_as_array(tail);
- *vertex_index_out = deref_array->base_offset;
+ if (vertex_index_out)
+ *vertex_index_out = deref_array->base_offset;
+
+ if (vertex_index_ref) {
+ LLVMValueRef vtx = LLVMConstInt(ctx->i32, deref_array->base_offset, false);
+ if (deref_array->deref_array_type == nir_deref_array_type_indirect) {
+ vtx = LLVMBuildAdd(ctx->builder, vtx, get_src(ctx, deref_array->indirect), "");
+ }
+ *vertex_index_ref = vtx;
+ }
+ }
+
+ if (deref->var->data.compact) {
+ assert(tail->child->deref_type == nir_deref_type_array);
+ assert(glsl_type_is_scalar(glsl_without_array(deref->var->type)));
+ nir_deref_array *deref_array = nir_deref_as_array(tail->child);
+ /* We always lower indirect dereferences for "compact" array vars. */
+ assert(deref_array->deref_array_type == nir_deref_array_type_direct);
+
+ const_offset = deref_array->base_offset;
+ goto out;
}
while (tail->child != NULL) {
unreachable("unsupported deref type");
}
-
+out:
if (const_offset && offset)
offset = LLVMBuildAdd(ctx->builder, offset,
LLVMConstInt(ctx->i32, const_offset, 0),
*indir_out = offset;
}
+static LLVMValueRef
+lds_load(struct nir_to_llvm_context *ctx,
+ LLVMValueRef dw_addr)
+{
+ LLVMValueRef value;
+ value = ac_build_indexed_load(&ctx->ac, ctx->lds, dw_addr, false);
+ return value;
+}
+
+static void
+lds_store(struct nir_to_llvm_context *ctx,
+ LLVMValueRef dw_addr, LLVMValueRef value)
+{
+ value = LLVMBuildBitCast(ctx->builder, value, ctx->i32, "");
+ ac_build_indexed_store(&ctx->ac, ctx->lds,
+ dw_addr, value);
+}
+
+/* The offchip buffer layout for TCS->TES is
+ *
+ * - attribute 0 of patch 0 vertex 0
+ * - attribute 0 of patch 0 vertex 1
+ * - attribute 0 of patch 0 vertex 2
+ * ...
+ * - attribute 0 of patch 1 vertex 0
+ * - attribute 0 of patch 1 vertex 1
+ * ...
+ * - attribute 1 of patch 0 vertex 0
+ * - attribute 1 of patch 0 vertex 1
+ * ...
+ * - per patch attribute 0 of patch 0
+ * - per patch attribute 0 of patch 1
+ * ...
+ *
+ * Note that every attribute has 4 components.
+ */
+static LLVMValueRef get_tcs_tes_buffer_address(struct nir_to_llvm_context *ctx,
+ LLVMValueRef vertex_index,
+ LLVMValueRef param_index)
+{
+ LLVMValueRef base_addr, vertices_per_patch, num_patches, total_vertices;
+ LLVMValueRef param_stride, constant16;
+ LLVMValueRef rel_patch_id = get_rel_patch_id(ctx);
+
+ vertices_per_patch = unpack_param(ctx, ctx->tcs_offchip_layout, 9, 6);
+ num_patches = unpack_param(ctx, ctx->tcs_offchip_layout, 0, 9);
+ total_vertices = LLVMBuildMul(ctx->builder, vertices_per_patch,
+ num_patches, "");
+
+ constant16 = LLVMConstInt(ctx->i32, 16, false);
+ if (vertex_index) {
+ base_addr = LLVMBuildMul(ctx->builder, rel_patch_id,
+ vertices_per_patch, "");
+
+ base_addr = LLVMBuildAdd(ctx->builder, base_addr,
+ vertex_index, "");
+
+ param_stride = total_vertices;
+ } else {
+ base_addr = rel_patch_id;
+ param_stride = num_patches;
+ }
+
+ base_addr = LLVMBuildAdd(ctx->builder, base_addr,
+ LLVMBuildMul(ctx->builder, param_index,
+ param_stride, ""), "");
+
+ base_addr = LLVMBuildMul(ctx->builder, base_addr, constant16, "");
+
+ if (!vertex_index) {
+ LLVMValueRef patch_data_offset =
+ unpack_param(ctx, ctx->tcs_offchip_layout, 16, 16);
+
+ base_addr = LLVMBuildAdd(ctx->builder, base_addr,
+ patch_data_offset, "");
+ }
+ return base_addr;
+}
+
+static LLVMValueRef get_tcs_tes_buffer_address_params(struct nir_to_llvm_context *ctx,
+ unsigned param,
+ unsigned const_index,
+ bool is_compact,
+ LLVMValueRef vertex_index,
+ LLVMValueRef indir_index)
+{
+ LLVMValueRef param_index;
+
+ if (indir_index)
+ param_index = LLVMBuildAdd(ctx->builder, LLVMConstInt(ctx->i32, param, false),
+ indir_index, "");
+ else {
+ if (const_index && !is_compact)
+ param += const_index;
+ param_index = LLVMConstInt(ctx->i32, param, false);
+ }
+ return get_tcs_tes_buffer_address(ctx, vertex_index, param_index);
+}
+
+static void
+mark_tess_output(struct nir_to_llvm_context *ctx,
+ bool is_patch, uint32_t param)
+
+{
+ if (is_patch) {
+ ctx->tess_patch_outputs_written |= (1ull << param);
+ } else
+ ctx->tess_outputs_written |= (1ull << param);
+}
+
+static LLVMValueRef
+get_dw_address(struct nir_to_llvm_context *ctx,
+ LLVMValueRef dw_addr,
+ unsigned param,
+ unsigned const_index,
+ bool compact_const_index,
+ LLVMValueRef vertex_index,
+ LLVMValueRef stride,
+ LLVMValueRef indir_index)
+
+{
+
+ if (vertex_index) {
+ dw_addr = LLVMBuildAdd(ctx->builder, dw_addr,
+ LLVMBuildMul(ctx->builder,
+ vertex_index,
+ stride, ""), "");
+ }
+
+ if (indir_index)
+ dw_addr = LLVMBuildAdd(ctx->builder, dw_addr,
+ LLVMBuildMul(ctx->builder, indir_index,
+ LLVMConstInt(ctx->i32, 4, false), ""), "");
+ else if (const_index && !compact_const_index)
+ dw_addr = LLVMBuildAdd(ctx->builder, dw_addr,
+ LLVMConstInt(ctx->i32, const_index, false), "");
+
+ dw_addr = LLVMBuildAdd(ctx->builder, dw_addr,
+ LLVMConstInt(ctx->i32, param * 4, false), "");
+
+ if (const_index && compact_const_index)
+ dw_addr = LLVMBuildAdd(ctx->builder, dw_addr,
+ LLVMConstInt(ctx->i32, const_index, false), "");
+ return dw_addr;
+}
+
+static LLVMValueRef
+load_tcs_input(struct nir_to_llvm_context *ctx,
+ nir_intrinsic_instr *instr)
+{
+ LLVMValueRef dw_addr, stride;
+ unsigned const_index;
+ LLVMValueRef vertex_index;
+ LLVMValueRef indir_index;
+ unsigned param;
+ LLVMValueRef value[4], result;
+ const bool per_vertex = nir_is_per_vertex_io(instr->variables[0]->var, ctx->stage);
+ const bool is_compact = instr->variables[0]->var->data.compact;
+ param = shader_io_get_unique_index(instr->variables[0]->var->data.location);
+ radv_get_deref_offset(ctx, instr->variables[0],
+ false, NULL, per_vertex ? &vertex_index : NULL,
+ &const_index, &indir_index);
+
+ stride = unpack_param(ctx, ctx->tcs_in_layout, 13, 8);
+ dw_addr = get_tcs_in_current_patch_offset(ctx);
+ dw_addr = get_dw_address(ctx, dw_addr, param, const_index, is_compact, vertex_index, stride,
+ indir_index);
+
+ for (unsigned i = 0; i < instr->num_components; i++) {
+ value[i] = lds_load(ctx, dw_addr);
+ dw_addr = LLVMBuildAdd(ctx->builder, dw_addr,
+ ctx->i32one, "");
+ }
+ result = ac_build_gather_values(&ctx->ac, value, instr->num_components);
+ result = LLVMBuildBitCast(ctx->builder, result, get_def_type(ctx, &instr->dest.ssa), "");
+ return result;
+}
+
+static LLVMValueRef
+load_tcs_output(struct nir_to_llvm_context *ctx,
+ nir_intrinsic_instr *instr)
+{
+ LLVMValueRef dw_addr, stride;
+ LLVMValueRef value[4], result;
+ LLVMValueRef vertex_index = NULL;
+ LLVMValueRef indir_index = NULL;
+ unsigned const_index = 0;
+ unsigned param;
+ const bool per_vertex = nir_is_per_vertex_io(instr->variables[0]->var, ctx->stage);
+ const bool is_compact = instr->variables[0]->var->data.compact;
+ param = shader_io_get_unique_index(instr->variables[0]->var->data.location);
+ radv_get_deref_offset(ctx, instr->variables[0],
+ false, NULL, per_vertex ? &vertex_index : NULL,
+ &const_index, &indir_index);
+
+ if (!instr->variables[0]->var->data.patch) {
+ stride = unpack_param(ctx, ctx->tcs_out_layout, 13, 8);
+ dw_addr = get_tcs_out_current_patch_offset(ctx);
+ } else {
+ dw_addr = get_tcs_out_current_patch_data_offset(ctx);
+ }
+
+ dw_addr = get_dw_address(ctx, dw_addr, param, const_index, is_compact, vertex_index, stride,
+ indir_index);
+
+ for (unsigned i = 0; i < instr->num_components; i++) {
+ value[i] = lds_load(ctx, dw_addr);
+ dw_addr = LLVMBuildAdd(ctx->builder, dw_addr,
+ ctx->i32one, "");
+ }
+ result = ac_build_gather_values(&ctx->ac, value, instr->num_components);
+ result = LLVMBuildBitCast(ctx->builder, result, get_def_type(ctx, &instr->dest.ssa), "");
+ return result;
+}
+
+static void
+store_tcs_output(struct nir_to_llvm_context *ctx,
+ nir_intrinsic_instr *instr,
+ LLVMValueRef src,
+ unsigned writemask)
+{
+ LLVMValueRef stride, dw_addr;
+ LLVMValueRef buf_addr = NULL;
+ LLVMValueRef vertex_index = NULL;
+ LLVMValueRef indir_index = NULL;
+ unsigned const_index = 0;
+ unsigned param;
+ const bool per_vertex = nir_is_per_vertex_io(instr->variables[0]->var, ctx->stage);
+ const bool is_compact = instr->variables[0]->var->data.compact;
+
+ radv_get_deref_offset(ctx, instr->variables[0],
+ false, NULL, per_vertex ? &vertex_index : NULL,
+ &const_index, &indir_index);
+
+ param = shader_io_get_unique_index(instr->variables[0]->var->data.location);
+ if (instr->variables[0]->var->data.location == VARYING_SLOT_CLIP_DIST0 &&
+ is_compact && const_index > 3) {
+ const_index -= 3;
+ param++;
+ }
+
+ if (!instr->variables[0]->var->data.patch) {
+ stride = unpack_param(ctx, ctx->tcs_out_layout, 13, 8);
+ dw_addr = get_tcs_out_current_patch_offset(ctx);
+ } else {
+ dw_addr = get_tcs_out_current_patch_data_offset(ctx);
+ }
+
+ mark_tess_output(ctx, instr->variables[0]->var->data.patch, param);
+
+ dw_addr = get_dw_address(ctx, dw_addr, param, const_index, is_compact, vertex_index, stride,
+ indir_index);
+ buf_addr = get_tcs_tes_buffer_address_params(ctx, param, const_index, is_compact,
+ vertex_index, indir_index);
+
+ unsigned base = is_compact ? const_index : 0;
+ for (unsigned chan = 0; chan < 8; chan++) {
+ bool is_tess_factor = false;
+ if (!(writemask & (1 << chan)))
+ continue;
+ LLVMValueRef value = llvm_extract_elem(ctx, src, chan);
+
+ lds_store(ctx, dw_addr, value);
+
+ if (instr->variables[0]->var->data.location == VARYING_SLOT_TESS_LEVEL_INNER ||
+ instr->variables[0]->var->data.location == VARYING_SLOT_TESS_LEVEL_OUTER)
+ is_tess_factor = true;
+
+ if (!is_tess_factor && writemask != 0xF)
+ ac_build_buffer_store_dword(&ctx->ac, ctx->hs_ring_tess_offchip, value, 1,
+ buf_addr, ctx->oc_lds,
+ 4 * (base + chan), 1, 0, true, false);
+
+ dw_addr = LLVMBuildAdd(ctx->builder, dw_addr,
+ ctx->i32one, "");
+ }
+
+ if (writemask == 0xF) {
+ ac_build_buffer_store_dword(&ctx->ac, ctx->hs_ring_tess_offchip, src, 4,
+ buf_addr, ctx->oc_lds,
+ (base * 4), 1, 0, true, false);
+ }
+}
+
+static LLVMValueRef
+load_tes_input(struct nir_to_llvm_context *ctx,
+ nir_intrinsic_instr *instr)
+{
+ LLVMValueRef buf_addr;
+ LLVMValueRef result;
+ LLVMValueRef vertex_index = NULL;
+ LLVMValueRef indir_index = NULL;
+ unsigned const_index = 0;
+ unsigned param;
+ const bool per_vertex = nir_is_per_vertex_io(instr->variables[0]->var, ctx->stage);
+ const bool is_compact = instr->variables[0]->var->data.compact;
+
+ radv_get_deref_offset(ctx, instr->variables[0],
+ false, NULL, per_vertex ? &vertex_index : NULL,
+ &const_index, &indir_index);
+ param = shader_io_get_unique_index(instr->variables[0]->var->data.location);
+ if (instr->variables[0]->var->data.location == VARYING_SLOT_CLIP_DIST0 &&
+ is_compact && const_index > 3) {
+ const_index -= 3;
+ param++;
+ }
+ buf_addr = get_tcs_tes_buffer_address_params(ctx, param, const_index,
+ is_compact, vertex_index, indir_index);
+
+ result = ac_build_buffer_load(&ctx->ac, ctx->hs_ring_tess_offchip, instr->num_components, NULL,
+ buf_addr, ctx->oc_lds, is_compact ? (4 * const_index) : 0, 1, 0, true, false);
+ result = trim_vector(ctx, result, instr->num_components);
+ result = LLVMBuildBitCast(ctx->builder, result, get_def_type(ctx, &instr->dest.ssa), "");
+ return result;
+}
+
static LLVMValueRef
load_gs_input(struct nir_to_llvm_context *ctx,
nir_intrinsic_instr *instr)
unsigned param, vtx_offset_param;
LLVMValueRef value[4], result;
unsigned vertex_index;
- radv_get_deref_offset(ctx, &instr->variables[0]->deref,
- false, &vertex_index,
+ radv_get_deref_offset(ctx, instr->variables[0],
+ false, &vertex_index, NULL,
&const_index, &indir_index);
vtx_offset_param = vertex_index;
assert(vtx_offset_param < 6);
unsigned const_index;
bool vs_in = ctx->stage == MESA_SHADER_VERTEX &&
instr->variables[0]->var->data.mode == nir_var_shader_in;
- radv_get_deref_offset(ctx, &instr->variables[0]->deref, vs_in, NULL,
+ radv_get_deref_offset(ctx, instr->variables[0], vs_in, NULL, NULL,
&const_index, &indir_index);
if (instr->dest.ssa.bit_size == 64)
switch (instr->variables[0]->var->data.mode) {
case nir_var_shader_in:
+ if (ctx->stage == MESA_SHADER_TESS_CTRL)
+ return load_tcs_input(ctx, instr);
+ if (ctx->stage == MESA_SHADER_TESS_EVAL)
+ return load_tes_input(ctx, instr);
if (ctx->stage == MESA_SHADER_GEOMETRY) {
return load_gs_input(ctx, instr);
}
}
break;
case nir_var_shader_out:
+ if (ctx->stage == MESA_SHADER_TESS_CTRL)
+ return load_tcs_output(ctx, instr);
for (unsigned chan = 0; chan < ve; chan++) {
if (indir_index) {
unsigned count = glsl_count_attribute_slots(
int writemask = instr->const_index[0];
LLVMValueRef indir_index;
unsigned const_index;
- radv_get_deref_offset(ctx, &instr->variables[0]->deref, false,
- NULL, &const_index, &indir_index);
+ radv_get_deref_offset(ctx, instr->variables[0], false,
+ NULL, NULL, &const_index, &indir_index);
if (get_elem_bits(ctx, LLVMTypeOf(src)) == 64) {
int old_writemask = writemask;
switch (instr->variables[0]->var->data.mode) {
case nir_var_shader_out:
+
+ if (ctx->stage == MESA_SHADER_TESS_CTRL) {
+ store_tcs_output(ctx, instr, src, writemask);
+ return;
+ }
+
for (unsigned chan = 0; chan < 8; chan++) {
int stride = 4;
if (!(writemask & (1 << chan)))
static void emit_barrier(struct nir_to_llvm_context *ctx)
{
- // TODO tess
+ /* SI only (thanks to a hw bug workaround):
+ * The real barrier instruction isn’t needed, because an entire patch
+ * always fits into a single wave.
+ */
+ if (ctx->options->chip_class == SI &&
+ ctx->stage == MESA_SHADER_TESS_CTRL) {
+ emit_waitcnt(ctx, LGKM_CNT & VM_CNT);
+ return;
+ }
ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.s.barrier",
- ctx->voidt, NULL, 0, 0);
+ ctx->voidt, NULL, 0, AC_FUNC_ATTR_CONVERGENT);
}
static void emit_discard_if(struct nir_to_llvm_context *ctx,
static LLVMValueRef load_sample_position(struct nir_to_llvm_context *ctx,
LLVMValueRef sample_id)
{
- /* offset = sample_id * 8 (8 = 2 floats containing samplepos.xy) */
- LLVMValueRef offset0 = LLVMBuildMul(ctx->builder, sample_id, LLVMConstInt(ctx->i32, 8, false), "");
- LLVMValueRef offset1 = LLVMBuildAdd(ctx->builder, offset0, LLVMConstInt(ctx->i32, 4, false), "");
- LLVMValueRef result[2];
+ LLVMValueRef result;
+ LLVMValueRef ptr = ac_build_gep0(&ctx->ac, ctx->ring_offsets, LLVMConstInt(ctx->i32, RING_PS_SAMPLE_POSITIONS, false));
- result[0] = ac_build_indexed_load_const(&ctx->ac, ctx->sample_positions, offset0);
- result[1] = ac_build_indexed_load_const(&ctx->ac, ctx->sample_positions, offset1);
+ ptr = LLVMBuildBitCast(ctx->builder, ptr,
+ const_array(ctx->v2f32, 64), "");
- return ac_build_gather_values(&ctx->ac, result, 2);
+ sample_id = LLVMBuildAdd(ctx->builder, sample_id, ctx->sample_pos_offset, "");
+ result = ac_build_indexed_load(&ctx->ac, ptr, sample_id, false);
+
+ return result;
}
static LLVMValueRef load_sample_pos(struct nir_to_llvm_context *ctx)
location = INTERP_CENTROID;
break;
case nir_intrinsic_interp_var_at_sample:
- location = INTERP_SAMPLE;
- src0 = get_src(ctx, instr->src[0]);
- break;
case nir_intrinsic_interp_var_at_offset:
location = INTERP_CENTER;
src0 = get_src(ctx, instr->src[0]);
+ break;
default:
break;
}
ac_build_sendmsg(&ctx->ac, AC_SENDMSG_GS_OP_CUT | AC_SENDMSG_GS | (0 << 8), ctx->gs_wave_id);
}
+static LLVMValueRef
+visit_load_tess_coord(struct nir_to_llvm_context *ctx,
+ nir_intrinsic_instr *instr)
+{
+ LLVMValueRef coord[4] = {
+ ctx->tes_u,
+ ctx->tes_v,
+ ctx->f32zero,
+ ctx->f32zero,
+ };
+
+ if (ctx->tes_primitive_mode == GL_TRIANGLES)
+ coord[2] = LLVMBuildFSub(ctx->builder, ctx->f32one,
+ LLVMBuildFAdd(ctx->builder, coord[0], coord[1], ""), "");
+
+ LLVMValueRef result = ac_build_gather_values(&ctx->ac, coord, instr->num_components);
+ return LLVMBuildBitCast(ctx->builder, result,
+ get_def_type(ctx, &instr->dest.ssa), "");
+}
+
static void visit_intrinsic(struct nir_to_llvm_context *ctx,
nir_intrinsic_instr *instr)
{
result = ctx->draw_index;
break;
case nir_intrinsic_load_invocation_id:
- result = ctx->gs_invocation_id;
+ if (ctx->stage == MESA_SHADER_TESS_CTRL)
+ result = unpack_param(ctx, ctx->tcs_rel_ids, 8, 5);
+ else
+ result = ctx->gs_invocation_id;
break;
case nir_intrinsic_load_primitive_id:
if (ctx->stage == MESA_SHADER_GEOMETRY)
result = ctx->gs_prim_id;
+ else if (ctx->stage == MESA_SHADER_TESS_CTRL)
+ result = ctx->tcs_patch_id;
+ else if (ctx->stage == MESA_SHADER_TESS_EVAL)
+ result = ctx->tes_patch_id;
else
fprintf(stderr, "Unknown primitive id intrinsic: %d", ctx->stage);
break;
case nir_intrinsic_end_primitive:
visit_end_primitive(ctx, instr);
break;
+ case nir_intrinsic_load_tess_coord:
+ result = visit_load_tess_coord(ctx, instr);
+ break;
+ case nir_intrinsic_load_patch_vertices_in:
+ result = LLVMConstInt(ctx->i32, ctx->options->key.tcs.input_vertices, false);
+ break;
default:
fprintf(stderr, "Unknown intrinsic: ");
nir_print_instr(&instr->instr, stderr);
constant_index = child->base_offset;
}
- if (desc_type == DESC_SAMPLER && binding->immutable_samplers &&
+ if (desc_type == DESC_SAMPLER && binding->immutable_samplers_offset &&
(!index || binding->immutable_samplers_equal)) {
if (binding->immutable_samplers_equal)
constant_index = 0;
+ const uint32_t *samplers = radv_immutable_samplers(layout, binding);
+
LLVMValueRef constants[] = {
- LLVMConstInt(ctx->i32, binding->immutable_samplers[constant_index * 4 + 0], 0),
- LLVMConstInt(ctx->i32, binding->immutable_samplers[constant_index * 4 + 1], 0),
- LLVMConstInt(ctx->i32, binding->immutable_samplers[constant_index * 4 + 2], 0),
- LLVMConstInt(ctx->i32, binding->immutable_samplers[constant_index * 4 + 3], 0),
+ LLVMConstInt(ctx->i32, samplers[constant_index * 4 + 0], 0),
+ LLVMConstInt(ctx->i32, samplers[constant_index * 4 + 1], 0),
+ LLVMConstInt(ctx->i32, samplers[constant_index * 4 + 2], 0),
+ LLVMConstInt(ctx->i32, samplers[constant_index * 4 + 3], 0),
};
return ac_build_gather_values(&ctx->ac, constants, 4);
}
LLVMValueRef derivs[6];
unsigned chan, count = 0;
unsigned const_src = 0, num_deriv_comp = 0;
-
+ bool lod_is_zero = false;
tex_fetch_ptrs(ctx, instr, &res_ptr, &samp_ptr, &fmask_ptr);
for (unsigned i = 0; i < instr->num_srcs; i++) {
case nir_tex_src_bias:
bias = get_src(ctx, instr->src[i].src);
break;
- case nir_tex_src_lod:
+ case nir_tex_src_lod: {
+ nir_const_value *val = nir_src_as_const_value(instr->src[i].src);
+
+ if (val && val->i32[0] == 0)
+ lod_is_zero = true;
lod = get_src(ctx, instr->src[i].src);
break;
+ }
case nir_tex_src_ms_index:
sample_index = get_src(ctx, instr->src[i].src);
break;
}
for (unsigned i = 0; i < num_deriv_comp; i++) {
- derivs[i * 2] = to_float(ctx, llvm_extract_elem(ctx, ddx, i));
- derivs[i * 2 + 1] = to_float(ctx, llvm_extract_elem(ctx, ddy, i));
+ derivs[i] = to_float(ctx, llvm_extract_elem(ctx, ddx, i));
+ derivs[num_deriv_comp + i] = to_float(ctx, llvm_extract_elem(ctx, ddy, i));
}
}
if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE && coord) {
+ if (instr->is_array && instr->op != nir_texop_lod)
+ coords[3] = apply_round_slice(ctx, coords[3]);
for (chan = 0; chan < instr->coord_components; chan++)
coords[chan] = to_float(ctx, coords[chan]);
if (instr->coord_components == 3)
}
if (instr->coord_components > 2) {
/* This seems like a bit of a hack - but it passes Vulkan CTS with it */
- if (instr->sampler_dim != GLSL_SAMPLER_DIM_3D && instr->op != nir_texop_txf) {
+ if (instr->sampler_dim != GLSL_SAMPLER_DIM_3D &&
+ instr->sampler_dim != GLSL_SAMPLER_DIM_CUBE &&
+ instr->op != nir_texop_txf) {
coords[2] = apply_round_slice(ctx, coords[2]);
}
address[count++] = coords[2];
}
/* Pack LOD */
- if ((instr->op == nir_texop_txl || instr->op == nir_texop_txf) && lod) {
+ if (lod && ((instr->op == nir_texop_txl && !lod_is_zero) ||
+ instr->op == nir_texop_txf)) {
address[count++] = lod;
} else if (instr->op == nir_texop_txf_ms && sample_index) {
address[count++] = sample_index;
fmask_ptr, NULL,
txf_address, txf_count, 0xf);
- result = build_tex_intrinsic(ctx, instr, &txf_args);
+ result = build_tex_intrinsic(ctx, instr, false, &txf_args);
result = LLVMBuildExtractElement(ctx->builder, result, ctx->i32zero, "");
result = emit_int_cmp(ctx, LLVMIntEQ, result, ctx->i32zero);
set_tex_fetch_args(ctx, &args, instr, instr->op,
res_ptr, samp_ptr, address, count, dmask);
- result = build_tex_intrinsic(ctx, instr, &args);
+ result = build_tex_intrinsic(ctx, instr, lod_is_zero, &args);
if (instr->op == nir_texop_query_levels)
result = LLVMBuildExtractElement(ctx->builder, result, LLVMConstInt(ctx->i32, 3, false), "");
LLVMValueRef t_list_ptr = ctx->vertex_buffers;
LLVMValueRef t_offset;
LLVMValueRef t_list;
- LLVMValueRef args[3];
LLVMValueRef input;
LLVMValueRef buffer_index;
int index = variable->data.location - VERT_ATTRIB_GENERIC0;
t_offset = LLVMConstInt(ctx->i32, index + i, false);
t_list = ac_build_indexed_load_const(&ctx->ac, t_list_ptr, t_offset);
- args[0] = t_list;
- args[1] = LLVMConstInt(ctx->i32, 0, false);
- args[2] = buffer_index;
- input = ac_build_intrinsic(&ctx->ac,
- "llvm.SI.vs.load.input", ctx->v4f32, args, 3,
- AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_NOUNWIND |
- AC_FUNC_ATTR_LEGACY);
+
+ input = ac_build_buffer_load_format(&ctx->ac, t_list,
+ buffer_index,
+ LLVMConstInt(ctx->i32, 0, false),
+ true);
for (unsigned chan = 0; chan < 4; chan++) {
LLVMValueRef llvm_chan = LLVMConstInt(ctx->i32, chan, false);
uint64_t mask_attribs;
variable->data.driver_location = idx * 4;
+ /* tess ctrl has it's own load/store paths for outputs */
+ if (ctx->stage == MESA_SHADER_TESS_CTRL)
+ return;
+
mask_attribs = ((1ull << attrib_count) - 1) << idx;
if (ctx->stage == MESA_SHADER_VERTEX ||
+ ctx->stage == MESA_SHADER_TESS_EVAL ||
ctx->stage == MESA_SHADER_GEOMETRY) {
if (idx == VARYING_SLOT_CLIP_DIST0) {
int length = ctx->num_output_clips + ctx->num_output_culls;
ctx->shader_info->vs.outinfo.clip_dist_mask = (1 << ctx->num_output_clips) - 1;
ctx->shader_info->vs.outinfo.cull_dist_mask = (1 << ctx->num_output_culls) - 1;
}
+ if (ctx->stage == MESA_SHADER_TESS_EVAL) {
+ ctx->shader_info->tes.outinfo.clip_dist_mask = (1 << ctx->num_output_clips) - 1;
+ ctx->shader_info->tes.outinfo.cull_dist_mask = (1 << ctx->num_output_culls) - 1;
+ }
if (length > 4)
attrib_count = 2;
LLVMValueRef psize_value = NULL, layer_value = NULL, viewport_index_value = NULL;
int i;
- outinfo->prim_id_output = 0xffffffff;
- outinfo->layer_output = 0xffffffff;
+ memset(outinfo->vs_output_param_offset, AC_EXP_PARAM_UNDEFINED,
+ sizeof(outinfo->vs_output_param_offset));
+
if (ctx->output_mask & (1ull << VARYING_SLOT_CLIP_DIST0)) {
LLVMValueRef slots[8];
unsigned j;
} else if (i == VARYING_SLOT_LAYER) {
outinfo->writes_layer = true;
layer_value = values[0];
- outinfo->layer_output = param_count;
target = V_008DFC_SQ_EXP_PARAM + param_count;
+ outinfo->vs_output_param_offset[VARYING_SLOT_LAYER] = param_count;
param_count++;
} else if (i == VARYING_SLOT_VIEWPORT) {
outinfo->writes_viewport_index = true;
viewport_index_value = values[0];
continue;
} else if (i == VARYING_SLOT_PRIMITIVE_ID) {
- outinfo->prim_id_output = param_count;
target = V_008DFC_SQ_EXP_PARAM + param_count;
+ outinfo->vs_output_param_offset[VARYING_SLOT_PRIMITIVE_ID] = param_count;
param_count++;
} else if (i >= VARYING_SLOT_VAR0) {
outinfo->export_mask |= 1u << (i - VARYING_SLOT_VAR0);
target = V_008DFC_SQ_EXP_PARAM + param_count;
+ outinfo->vs_output_param_offset[i] = param_count;
param_count++;
}
}
static void
-si_export_mrt_color(struct nir_to_llvm_context *ctx,
- LLVMValueRef *color, unsigned param, bool is_last)
+handle_ls_outputs_post(struct nir_to_llvm_context *ctx)
{
+ LLVMValueRef vertex_id = ctx->rel_auto_id;
+ LLVMValueRef vertex_dw_stride = unpack_param(ctx, ctx->ls_out_layout, 13, 8);
+ LLVMValueRef base_dw_addr = LLVMBuildMul(ctx->builder, vertex_id,
+ vertex_dw_stride, "");
- struct ac_export_args args;
+ for (unsigned i = 0; i < RADEON_LLVM_MAX_OUTPUTS; ++i) {
+ LLVMValueRef *out_ptr = &ctx->outputs[i * 4];
+ int length = 4;
+
+ if (!(ctx->output_mask & (1ull << i)))
+ continue;
+
+ if (i == VARYING_SLOT_CLIP_DIST0)
+ length = ctx->num_output_clips + ctx->num_output_culls;
+ int param = shader_io_get_unique_index(i);
+ mark_tess_output(ctx, false, param);
+ if (length > 4)
+ mark_tess_output(ctx, false, param + 1);
+ LLVMValueRef dw_addr = LLVMBuildAdd(ctx->builder, base_dw_addr,
+ LLVMConstInt(ctx->i32, param * 4, false),
+ "");
+ for (unsigned j = 0; j < length; j++) {
+ lds_store(ctx, dw_addr,
+ LLVMBuildLoad(ctx->builder, out_ptr[j], ""));
+ dw_addr = LLVMBuildAdd(ctx->builder, dw_addr, ctx->i32one, "");
+ }
+ }
+}
+
+struct ac_build_if_state
+{
+ struct nir_to_llvm_context *ctx;
+ LLVMValueRef condition;
+ LLVMBasicBlockRef entry_block;
+ LLVMBasicBlockRef true_block;
+ LLVMBasicBlockRef false_block;
+ LLVMBasicBlockRef merge_block;
+};
+
+static LLVMBasicBlockRef
+ac_build_insert_new_block(struct nir_to_llvm_context *ctx, const char *name)
+{
+ LLVMBasicBlockRef current_block;
+ LLVMBasicBlockRef next_block;
+ LLVMBasicBlockRef new_block;
+
+ /* get current basic block */
+ current_block = LLVMGetInsertBlock(ctx->builder);
+
+ /* chqeck if there's another block after this one */
+ next_block = LLVMGetNextBasicBlock(current_block);
+ if (next_block) {
+ /* insert the new block before the next block */
+ new_block = LLVMInsertBasicBlockInContext(ctx->context, next_block, name);
+ }
+ else {
+ /* append new block after current block */
+ LLVMValueRef function = LLVMGetBasicBlockParent(current_block);
+ new_block = LLVMAppendBasicBlockInContext(ctx->context, function, name);
+ }
+ return new_block;
+}
+
+static void
+ac_nir_build_if(struct ac_build_if_state *ifthen,
+ struct nir_to_llvm_context *ctx,
+ LLVMValueRef condition)
+{
+ LLVMBasicBlockRef block = LLVMGetInsertBlock(ctx->builder);
+
+ memset(ifthen, 0, sizeof *ifthen);
+ ifthen->ctx = ctx;
+ ifthen->condition = condition;
+ ifthen->entry_block = block;
+
+ /* create endif/merge basic block for the phi functions */
+ ifthen->merge_block = ac_build_insert_new_block(ctx, "endif-block");
+
+ /* create/insert true_block before merge_block */
+ ifthen->true_block =
+ LLVMInsertBasicBlockInContext(ctx->context,
+ ifthen->merge_block,
+ "if-true-block");
+
+ /* successive code goes into the true block */
+ LLVMPositionBuilderAtEnd(ctx->builder, ifthen->true_block);
+}
+
+/**
+ * End a conditional.
+ */
+static void
+ac_nir_build_endif(struct ac_build_if_state *ifthen)
+{
+ LLVMBuilderRef builder = ifthen->ctx->builder;
+
+ /* Insert branch to the merge block from current block */
+ LLVMBuildBr(builder, ifthen->merge_block);
+
+ /*
+ * Now patch in the various branch instructions.
+ */
+
+ /* Insert the conditional branch instruction at the end of entry_block */
+ LLVMPositionBuilderAtEnd(builder, ifthen->entry_block);
+ if (ifthen->false_block) {
+ /* we have an else clause */
+ LLVMBuildCondBr(builder, ifthen->condition,
+ ifthen->true_block, ifthen->false_block);
+ }
+ else {
+ /* no else clause */
+ LLVMBuildCondBr(builder, ifthen->condition,
+ ifthen->true_block, ifthen->merge_block);
+ }
+
+ /* Resume building code at end of the ifthen->merge_block */
+ LLVMPositionBuilderAtEnd(builder, ifthen->merge_block);
+}
+
+static void
+write_tess_factors(struct nir_to_llvm_context *ctx)
+{
+ unsigned stride, outer_comps, inner_comps;
+ struct ac_build_if_state if_ctx, inner_if_ctx;
+ LLVMValueRef invocation_id = unpack_param(ctx, ctx->tcs_rel_ids, 8, 5);
+ LLVMValueRef rel_patch_id = unpack_param(ctx, ctx->tcs_rel_ids, 0, 8);
+ unsigned tess_inner_index, tess_outer_index;
+ LLVMValueRef lds_base, lds_inner, lds_outer, byteoffset, buffer;
+ LLVMValueRef out[6], vec0, vec1, tf_base, inner[4], outer[4];
+ int i;
+ emit_barrier(ctx);
+
+ switch (ctx->options->key.tcs.primitive_mode) {
+ case GL_ISOLINES:
+ stride = 2;
+ outer_comps = 2;
+ inner_comps = 0;
+ break;
+ case GL_TRIANGLES:
+ stride = 4;
+ outer_comps = 3;
+ inner_comps = 1;
+ break;
+ case GL_QUADS:
+ stride = 6;
+ outer_comps = 4;
+ inner_comps = 2;
+ break;
+ default:
+ return;
+ }
+
+ ac_nir_build_if(&if_ctx, ctx,
+ LLVMBuildICmp(ctx->builder, LLVMIntEQ,
+ invocation_id, ctx->i32zero, ""));
+
+ tess_inner_index = shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_INNER);
+ tess_outer_index = shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_OUTER);
+
+ mark_tess_output(ctx, true, tess_inner_index);
+ mark_tess_output(ctx, true, tess_outer_index);
+ lds_base = get_tcs_out_current_patch_data_offset(ctx);
+ lds_inner = LLVMBuildAdd(ctx->builder, lds_base,
+ LLVMConstInt(ctx->i32, tess_inner_index * 4, false), "");
+ lds_outer = LLVMBuildAdd(ctx->builder, lds_base,
+ LLVMConstInt(ctx->i32, tess_outer_index * 4, false), "");
+
+ for (i = 0; i < 4; i++) {
+ inner[i] = LLVMGetUndef(ctx->i32);
+ outer[i] = LLVMGetUndef(ctx->i32);
+ }
+
+ // LINES reverseal
+ if (ctx->options->key.tcs.primitive_mode == GL_ISOLINES) {
+ outer[0] = out[1] = lds_load(ctx, lds_outer);
+ lds_outer = LLVMBuildAdd(ctx->builder, lds_outer,
+ LLVMConstInt(ctx->i32, 1, false), "");
+ outer[1] = out[0] = lds_load(ctx, lds_outer);
+ } else {
+ for (i = 0; i < outer_comps; i++) {
+ outer[i] = out[i] =
+ lds_load(ctx, lds_outer);
+ lds_outer = LLVMBuildAdd(ctx->builder, lds_outer,
+ LLVMConstInt(ctx->i32, 1, false), "");
+ }
+ for (i = 0; i < inner_comps; i++) {
+ inner[i] = out[outer_comps+i] =
+ lds_load(ctx, lds_inner);
+ lds_inner = LLVMBuildAdd(ctx->builder, lds_inner,
+ LLVMConstInt(ctx->i32, 1, false), "");
+ }
+ }
+
+ /* Convert the outputs to vectors for stores. */
+ vec0 = ac_build_gather_values(&ctx->ac, out, MIN2(stride, 4));
+ vec1 = NULL;
+
+ if (stride > 4)
+ vec1 = ac_build_gather_values(&ctx->ac, out + 4, stride - 4);
+
+
+ buffer = ctx->hs_ring_tess_factor;
+ tf_base = ctx->tess_factor_offset;
+ byteoffset = LLVMBuildMul(ctx->builder, rel_patch_id,
+ LLVMConstInt(ctx->i32, 4 * stride, false), "");
+
+ ac_nir_build_if(&inner_if_ctx, ctx,
+ LLVMBuildICmp(ctx->builder, LLVMIntEQ,
+ rel_patch_id, ctx->i32zero, ""));
+
+ /* Store the dynamic HS control word. */
+ ac_build_buffer_store_dword(&ctx->ac, buffer,
+ LLVMConstInt(ctx->i32, 0x80000000, false),
+ 1, ctx->i32zero, tf_base,
+ 0, 1, 0, true, false);
+ ac_nir_build_endif(&inner_if_ctx);
+
+ /* Store the tessellation factors. */
+ ac_build_buffer_store_dword(&ctx->ac, buffer, vec0,
+ MIN2(stride, 4), byteoffset, tf_base,
+ 4, 1, 0, true, false);
+ if (vec1)
+ ac_build_buffer_store_dword(&ctx->ac, buffer, vec1,
+ stride - 4, byteoffset, tf_base,
+ 20, 1, 0, true, false);
+
+ //TODO store to offchip for TES to read - only if TES reads them
+ if (1) {
+ LLVMValueRef inner_vec, outer_vec, tf_outer_offset;
+ LLVMValueRef tf_inner_offset;
+ unsigned param_outer, param_inner;
+
+ param_outer = shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_OUTER);
+ tf_outer_offset = get_tcs_tes_buffer_address(ctx, NULL,
+ LLVMConstInt(ctx->i32, param_outer, 0));
+
+ outer_vec = ac_build_gather_values(&ctx->ac, outer,
+ util_next_power_of_two(outer_comps));
+
+ ac_build_buffer_store_dword(&ctx->ac, ctx->hs_ring_tess_offchip, outer_vec,
+ outer_comps, tf_outer_offset,
+ ctx->oc_lds, 0, 1, 0, true, false);
+ if (inner_comps) {
+ param_inner = shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_INNER);
+ tf_inner_offset = get_tcs_tes_buffer_address(ctx, NULL,
+ LLVMConstInt(ctx->i32, param_inner, 0));
+
+ inner_vec = inner_comps == 1 ? inner[0] :
+ ac_build_gather_values(&ctx->ac, inner, inner_comps);
+ ac_build_buffer_store_dword(&ctx->ac, ctx->hs_ring_tess_offchip, inner_vec,
+ inner_comps, tf_inner_offset,
+ ctx->oc_lds, 0, 1, 0, true, false);
+ }
+ }
+ ac_nir_build_endif(&if_ctx);
+}
+static void
+handle_tcs_outputs_post(struct nir_to_llvm_context *ctx)
+{
+ write_tess_factors(ctx);
+}
+
+static bool
+si_export_mrt_color(struct nir_to_llvm_context *ctx,
+ LLVMValueRef *color, unsigned param, bool is_last,
+ struct ac_export_args *args)
+{
/* Export */
si_llvm_init_export_args(ctx, color, param,
- &args);
+ args);
if (is_last) {
- args.valid_mask = 1; /* whether the EXEC mask is valid */
- args.done = 1; /* DONE bit */
- } else if (!args.enabled_channels)
- return; /* unnecessary NULL export */
+ args->valid_mask = 1; /* whether the EXEC mask is valid */
+ args->done = 1; /* DONE bit */
+ } else if (!args->enabled_channels)
+ return false; /* unnecessary NULL export */
- ac_build_export(&ctx->ac, &args);
+ return true;
}
static void
{
unsigned index = 0;
LLVMValueRef depth = NULL, stencil = NULL, samplemask = NULL;
+ struct ac_export_args color_args[8];
for (unsigned i = 0; i < RADEON_LLVM_MAX_OUTPUTS; ++i) {
LLVMValueRef values[4];
if (!ctx->shader_info->fs.writes_z && !ctx->shader_info->fs.writes_stencil && !ctx->shader_info->fs.writes_sample_mask)
last = ctx->output_mask <= ((1ull << (i + 1)) - 1);
- si_export_mrt_color(ctx, values, V_008DFC_SQ_EXP_MRT + index, last);
- index++;
+ bool ret = si_export_mrt_color(ctx, values, V_008DFC_SQ_EXP_MRT + (i - FRAG_RESULT_DATA0), last, &color_args[index]);
+ if (ret)
+ index++;
}
}
+ for (unsigned i = 0; i < index; i++)
+ ac_build_export(&ctx->ac, &color_args[i]);
if (depth || stencil || samplemask)
si_export_mrt_z(ctx, depth, stencil, samplemask);
- else if (!index)
- si_export_mrt_color(ctx, NULL, V_008DFC_SQ_EXP_NULL, true);
+ else if (!index) {
+ si_export_mrt_color(ctx, NULL, V_008DFC_SQ_EXP_NULL, true, &color_args[0]);
+ ac_build_export(&ctx->ac, &color_args[0]);
+ }
ctx->shader_info->fs.output_mask = index ? ((1ull << index) - 1) : 0;
}
{
switch (ctx->stage) {
case MESA_SHADER_VERTEX:
- if (ctx->options->key.vs.as_es)
+ if (ctx->options->key.vs.as_ls)
+ handle_ls_outputs_post(ctx);
+ else if (ctx->options->key.vs.as_es)
handle_es_outputs_post(ctx, &ctx->shader_info->vs.es_info);
else
handle_vs_outputs_post(ctx, &ctx->shader_info->vs.outinfo);
case MESA_SHADER_GEOMETRY:
emit_gs_epilogue(ctx);
break;
+ case MESA_SHADER_TESS_CTRL:
+ handle_tcs_outputs_post(ctx);
+ break;
+ case MESA_SHADER_TESS_EVAL:
+ if (ctx->options->key.tes.as_es)
+ handle_es_outputs_post(ctx, &ctx->shader_info->tes.es_info);
+ else
+ handle_vs_outputs_post(ctx, &ctx->shader_info->tes.outinfo);
+ break;
default:
break;
}
LLVMDisposePassManager(passmgr);
}
+static void
+ac_nir_eliminate_const_vs_outputs(struct nir_to_llvm_context *ctx)
+{
+ struct ac_vs_output_info *outinfo;
+
+ if (ctx->stage == MESA_SHADER_FRAGMENT ||
+ ctx->stage == MESA_SHADER_COMPUTE ||
+ ctx->stage == MESA_SHADER_TESS_CTRL ||
+ ctx->stage == MESA_SHADER_GEOMETRY)
+ return;
+
+ if (ctx->stage == MESA_SHADER_VERTEX) {
+ if (ctx->options->key.vs.as_ls ||
+ ctx->options->key.vs.as_es)
+ return;
+ outinfo = &ctx->shader_info->vs.outinfo;
+ }
+
+ if (ctx->stage == MESA_SHADER_TESS_EVAL) {
+ if (ctx->options->key.vs.as_es)
+ return;
+ outinfo = &ctx->shader_info->tes.outinfo;
+ }
+
+ ac_optimize_vs_outputs(&ctx->ac,
+ ctx->main_function,
+ outinfo->vs_output_param_offset,
+ VARYING_SLOT_MAX,
+ &outinfo->param_exports);
+}
+
static void
ac_setup_rings(struct nir_to_llvm_context *ctx)
{
- if (ctx->stage == MESA_SHADER_VERTEX && ctx->options->key.vs.as_es) {
- ctx->esgs_ring = ac_build_indexed_load_const(&ctx->ac, ctx->ring_offsets, ctx->i32one);
+ if ((ctx->stage == MESA_SHADER_VERTEX && ctx->options->key.vs.as_es) ||
+ (ctx->stage == MESA_SHADER_TESS_EVAL && ctx->options->key.tes.as_es)) {
+ ctx->esgs_ring = ac_build_indexed_load_const(&ctx->ac, ctx->ring_offsets, LLVMConstInt(ctx->i32, RING_ESGS_VS, false));
}
if (ctx->is_gs_copy_shader) {
- ctx->gsvs_ring = ac_build_indexed_load_const(&ctx->ac, ctx->ring_offsets, LLVMConstInt(ctx->i32, 3, false));
+ ctx->gsvs_ring = ac_build_indexed_load_const(&ctx->ac, ctx->ring_offsets, LLVMConstInt(ctx->i32, RING_GSVS_VS, false));
}
if (ctx->stage == MESA_SHADER_GEOMETRY) {
LLVMValueRef tmp;
- ctx->esgs_ring = ac_build_indexed_load_const(&ctx->ac, ctx->ring_offsets, LLVMConstInt(ctx->i32, 2, false));
- ctx->gsvs_ring = ac_build_indexed_load_const(&ctx->ac, ctx->ring_offsets, LLVMConstInt(ctx->i32, 4, false));
+ ctx->esgs_ring = ac_build_indexed_load_const(&ctx->ac, ctx->ring_offsets, LLVMConstInt(ctx->i32, RING_ESGS_GS, false));
+ ctx->gsvs_ring = ac_build_indexed_load_const(&ctx->ac, ctx->ring_offsets, LLVMConstInt(ctx->i32, RING_GSVS_GS, false));
ctx->gsvs_ring = LLVMBuildBitCast(ctx->builder, ctx->gsvs_ring, ctx->v4i32, "");
ctx->gsvs_ring = LLVMBuildBitCast(ctx->builder, ctx->gsvs_ring, ctx->v16i8, "");
}
+
+ if (ctx->stage == MESA_SHADER_TESS_CTRL ||
+ ctx->stage == MESA_SHADER_TESS_EVAL) {
+ ctx->hs_ring_tess_offchip = ac_build_indexed_load_const(&ctx->ac, ctx->ring_offsets, LLVMConstInt(ctx->i32, RING_HS_TESS_OFFCHIP, false));
+ ctx->hs_ring_tess_factor = ac_build_indexed_load_const(&ctx->ac, ctx->ring_offsets, LLVMConstInt(ctx->i32, RING_HS_TESS_FACTOR, false));
+ }
+}
+
+static unsigned
+ac_nir_get_max_workgroup_size(enum chip_class chip_class,
+ struct nir_shader *nir)
+{
+ switch (nir->stage) {
+ case MESA_SHADER_TESS_CTRL:
+ return chip_class >= CIK ? 128 : 64;
+ case MESA_SHADER_GEOMETRY:
+ return 64;
+ case MESA_SHADER_COMPUTE:
+ break;
+ default:
+ return 0;
+ }
+
+ unsigned max_workgroup_size = nir->info.cs.local_size[0] *
+ nir->info.cs.local_size[1] *
+ nir->info.cs.local_size[2];
+ return max_workgroup_size;
}
static
memset(shader_info, 0, sizeof(*shader_info));
+ ac_nir_shader_info_pass(nir, options, &shader_info->info);
+
LLVMSetTarget(ctx.module, options->supports_spill ? "amdgcn-mesa-mesa3d" : "amdgcn--");
LLVMTargetDataRef data_layout = LLVMCreateTargetDataLayout(tm);
ctx.builder = LLVMCreateBuilderInContext(ctx.context);
ctx.ac.builder = ctx.builder;
ctx.stage = nir->stage;
+ ctx.max_workgroup_size = ac_nir_get_max_workgroup_size(ctx.options->chip_class, nir);
for (i = 0; i < AC_UD_MAX_SETS; i++)
shader_info->user_sgprs_locs.descriptor_sets[i].sgpr_idx = -1;
} else if (nir->stage == MESA_SHADER_GEOMETRY) {
ctx.gs_next_vertex = ac_build_alloca(&ctx, ctx.i32, "gs_next_vertex");
- ctx.gs_max_out_vertices = nir->info->gs.vertices_out;
+ ctx.gs_max_out_vertices = nir->info.gs.vertices_out;
+ } else if (nir->stage == MESA_SHADER_TESS_EVAL) {
+ ctx.tes_primitive_mode = nir->info.tess.primitive_mode;
}
ac_setup_rings(&ctx);
if (nir->stage == MESA_SHADER_FRAGMENT)
handle_fs_inputs_pre(&ctx, nir);
- ctx.num_output_clips = nir->info->clip_distance_array_size;
- ctx.num_output_culls = nir->info->cull_distance_array_size;
+ ctx.num_output_clips = nir->info.clip_distance_array_size;
+ ctx.num_output_culls = nir->info.cull_distance_array_size;
nir_foreach_variable(variable, &nir->outputs)
handle_shader_output_decl(&ctx, variable);
LLVMBuildRetVoid(ctx.builder);
ac_llvm_finalize_module(&ctx);
+
+ ac_nir_eliminate_const_vs_outputs(&ctx);
free(ctx.locals);
ralloc_free(ctx.defs);
ralloc_free(ctx.phis);
unsigned addclip = ctx.num_output_clips + ctx.num_output_culls > 4;
shader_info->gs.gsvs_vertex_size = (util_bitcount64(ctx.output_mask) + addclip) * 16;
shader_info->gs.max_gsvs_emit_size = shader_info->gs.gsvs_vertex_size *
- nir->info->gs.vertices_out;
+ nir->info.gs.vertices_out;
+ } else if (nir->stage == MESA_SHADER_TESS_CTRL) {
+ shader_info->tcs.outputs_written = ctx.tess_outputs_written;
+ shader_info->tcs.patch_outputs_written = ctx.tess_patch_outputs_written;
+ } else if (nir->stage == MESA_SHADER_VERTEX && ctx.options->key.vs.as_ls) {
+ shader_info->vs.outputs_written = ctx.tess_outputs_written;
}
+
return ctx.module;
}
switch (nir->stage) {
case MESA_SHADER_COMPUTE:
for (int i = 0; i < 3; ++i)
- shader_info->cs.block_size[i] = nir->info->cs.local_size[i];
+ shader_info->cs.block_size[i] = nir->info.cs.local_size[i];
break;
case MESA_SHADER_FRAGMENT:
- shader_info->fs.early_fragment_test = nir->info->fs.early_fragment_tests;
+ shader_info->fs.early_fragment_test = nir->info.fs.early_fragment_tests;
break;
case MESA_SHADER_GEOMETRY:
- shader_info->gs.vertices_in = nir->info->gs.vertices_in;
- shader_info->gs.vertices_out = nir->info->gs.vertices_out;
- shader_info->gs.output_prim = nir->info->gs.output_primitive;
- shader_info->gs.invocations = nir->info->gs.invocations;
+ shader_info->gs.vertices_in = nir->info.gs.vertices_in;
+ shader_info->gs.vertices_out = nir->info.gs.vertices_out;
+ shader_info->gs.output_prim = nir->info.gs.output_primitive;
+ shader_info->gs.invocations = nir->info.gs.invocations;
+ break;
+ case MESA_SHADER_TESS_EVAL:
+ shader_info->tes.primitive_mode = nir->info.tess.primitive_mode;
+ shader_info->tes.spacing = nir->info.tess.spacing;
+ shader_info->tes.ccw = nir->info.tess.ccw;
+ shader_info->tes.point_mode = nir->info.tess.point_mode;
+ shader_info->tes.as_es = options->key.tes.as_es;
+ break;
+ case MESA_SHADER_TESS_CTRL:
+ shader_info->tcs.tcs_vertices_out = nir->info.tess.tcs_vertices_out;
break;
case MESA_SHADER_VERTEX:
shader_info->vs.as_es = options->key.vs.as_es;
+ shader_info->vs.as_ls = options->key.vs.as_ls;
+ /* in LS mode we need at least 1, invocation id needs 3, handled elsewhere */
+ if (options->key.vs.as_ls)
+ shader_info->vs.vgpr_comp_cnt = MAX2(1, shader_info->vs.vgpr_comp_cnt);
break;
default:
break;
create_function(&ctx);
- ctx.gs_max_out_vertices = geom_shader->info->gs.vertices_out;
+ ctx.gs_max_out_vertices = geom_shader->info.gs.vertices_out;
ac_setup_rings(&ctx);
- ctx.num_output_clips = geom_shader->info->clip_distance_array_size;
- ctx.num_output_culls = geom_shader->info->cull_distance_array_size;
+ ctx.num_output_clips = geom_shader->info.clip_distance_array_size;
+ ctx.num_output_culls = geom_shader->info.cull_distance_array_size;
nir_foreach_variable(variable, &geom_shader->outputs)
handle_shader_output_decl(&ctx, variable);