#include <llvm-c/Core.h>
#include <llvm-c/TargetMachine.h>
#include <llvm-c/Transforms/Scalar.h>
-#if HAVE_LLVM >= 0x0700
#include <llvm-c/Transforms/Utils.h>
-#endif
#include "sid.h"
-#include "gfx9d.h"
#include "ac_binary.h"
#include "ac_llvm_util.h"
#include "ac_llvm_build.h"
LLVMValueRef tes_u;
LLVMValueRef tes_v;
+ /* HW GS */
+ /* On gfx10:
+ * - bits 0..10: ordered_wave_id
+ * - bits 12..20: number of vertices in group
+ * - bits 22..30: number of primitives in group
+ */
+ LLVMValueRef gs_tg_info;
LLVMValueRef gs2vs_offset;
LLVMValueRef gs_wave_id;
LLVMValueRef gs_vtx_offset[6];
LLVMValueRef esgs_ring;
- LLVMValueRef gsvs_ring;
+ LLVMValueRef gsvs_ring[4];
LLVMValueRef hs_ring_tess_offchip;
LLVMValueRef hs_ring_tess_factor;
LLVMValueRef persp_sample, persp_center, persp_centroid;
LLVMValueRef linear_sample, linear_center, linear_centroid;
+ /* Streamout */
+ LLVMValueRef streamout_buffers;
+ LLVMValueRef streamout_write_idx;
+ LLVMValueRef streamout_config;
+ LLVMValueRef streamout_offset[4];
+
gl_shader_stage stage;
LLVMValueRef inputs[RADEON_LLVM_MAX_INPUTS * 4];
+ uint64_t float16_shaded_mask;
uint64_t input_mask;
uint64_t output_mask;
bool is_gs_copy_shader;
- LLVMValueRef gs_next_vertex;
+ LLVMValueRef gs_next_vertex[4];
unsigned gs_max_out_vertices;
unsigned tes_primitive_mode;
/* Make sure that the data fits in LDS. This assumes the shaders only
* use LDS for the inputs and outputs.
*/
- hardware_lds_size = ctx->options->chip_class >= CIK ? 65536 : 32768;
+ hardware_lds_size = 32768;
+
+ /* Looks like STONEY hangs if we use more than 32 KiB LDS in a single
+ * threadgroup, even though there is more than 32 KiB LDS.
+ *
+ * Test: dEQP-VK.tessellation.shader_input_output.barrier
+ */
+ if (ctx->options->chip_class >= GFX7 && ctx->options->family != CHIP_STONEY)
+ hardware_lds_size = 65536;
+
num_patches = MIN2(num_patches, hardware_lds_size / (input_patch_size + output_patch_size));
/* Make sure the output data fits in the offchip buffer */
num_patches = MIN2(num_patches, (ctx->options->tess_offchip_block_dw_size * 4) / output_patch_size);
*/
num_patches = MIN2(num_patches, 40);
- /* SI bug workaround - limit LS-HS threadgroups to only one wave. */
- if (ctx->options->chip_class == SI) {
+ /* GFX6 bug workaround - limit LS-HS threadgroups to only one wave. */
+ if (ctx->options->chip_class == GFX6) {
unsigned one_wave = 64 / MAX2(num_tcs_input_cp, num_tcs_output_cp);
num_patches = MIN2(num_patches, one_wave);
}
patch0_patch_data_offset);
}
-#define MAX_ARGS 23
+#define MAX_ARGS 64
struct arg_info {
LLVMTypeRef types[MAX_ARGS];
LLVMValueRef *assign[MAX_ARGS];
- unsigned array_params_mask;
uint8_t count;
uint8_t sgpr_count;
uint8_t num_sgprs_used;
}
}
-static inline void
-add_array_arg(struct arg_info *info, LLVMTypeRef type, LLVMValueRef *param_ptr)
-{
- info->array_params_mask |= (1 << info->count);
- add_arg(info, ARG_SGPR, type, param_ptr);
-}
-
static void assign_arguments(LLVMValueRef main_function,
struct arg_info *info)
{
LLVMSetFunctionCallConv(main_function, RADEON_LLVM_AMDGPU_CS);
for (unsigned i = 0; i < args->sgpr_count; ++i) {
+ LLVMValueRef P = LLVMGetParam(main_function, i);
+
ac_add_function_attr(ctx, main_function, i + 1, AC_FUNC_ATTR_INREG);
- if (args->array_params_mask & (1 << i)) {
- LLVMValueRef P = LLVMGetParam(main_function, i);
+ if (LLVMGetTypeKind(LLVMTypeOf(P)) == LLVMPointerTypeKind) {
ac_add_function_attr(ctx, main_function, i + 1, AC_FUNC_ATTR_NOALIAS);
ac_add_attr_dereferenceable(P, UINT64_MAX);
}
options->address32_hi);
}
- if (max_workgroup_size) {
- ac_llvm_add_target_dep_function_attr(main_function,
- "amdgpu-max-work-group-size",
- max_workgroup_size);
- }
+ ac_llvm_set_workgroup_size(main_function, max_workgroup_size);
+
if (options->unsafe_math) {
/* These were copied from some LLVM test. */
LLVMAddTargetDependentFunctionAttr(main_function,
static void
-set_loc(struct radv_userdata_info *ud_info, uint8_t *sgpr_idx, uint8_t num_sgprs,
- uint32_t indirect_offset)
+set_loc(struct radv_userdata_info *ud_info, uint8_t *sgpr_idx,
+ uint8_t num_sgprs)
{
ud_info->sgpr_idx = *sgpr_idx;
ud_info->num_sgprs = num_sgprs;
- ud_info->indirect = indirect_offset > 0;
- ud_info->indirect_offset = indirect_offset;
*sgpr_idx += num_sgprs;
}
&ctx->shader_info->user_sgprs_locs.shader_data[idx];
assert(ud_info);
- set_loc(ud_info, sgpr_idx, num_sgprs, 0);
+ set_loc(ud_info, sgpr_idx, num_sgprs);
}
static void
set_loc_shader_ptr(struct radv_shader_context *ctx, int idx, uint8_t *sgpr_idx)
{
- bool use_32bit_pointers = HAVE_32BIT_POINTERS &&
- idx != AC_UD_SCRATCH_RING_OFFSETS;
+ bool use_32bit_pointers = idx != AC_UD_SCRATCH_RING_OFFSETS;
set_loc_shader(ctx, idx, sgpr_idx, use_32bit_pointers ? 1 : 2);
}
static void
-set_loc_desc(struct radv_shader_context *ctx, int idx, uint8_t *sgpr_idx,
- uint32_t indirect_offset)
+set_loc_desc(struct radv_shader_context *ctx, int idx, uint8_t *sgpr_idx)
{
struct radv_userdata_locations *locs =
&ctx->shader_info->user_sgprs_locs;
struct radv_userdata_info *ud_info = &locs->descriptor_sets[idx];
assert(ud_info);
- set_loc(ud_info, sgpr_idx, HAVE_32BIT_POINTERS ? 1 : 2, indirect_offset);
- if (indirect_offset == 0)
- locs->descriptor_sets_enabled |= 1 << idx;
+ set_loc(ud_info, sgpr_idx, 1);
+
+ locs->descriptor_sets_enabled |= 1 << idx;
}
struct user_sgpr_info {
bool need_ring_offsets;
bool indirect_all_descriptor_sets;
+ uint8_t remaining_sgprs;
};
static bool needs_view_index_sgpr(struct radv_shader_context *ctx,
switch (stage) {
case MESA_SHADER_VERTEX:
if (ctx->shader_info->info.needs_multiview_view_index ||
- (!ctx->options->key.vs.as_es && !ctx->options->key.vs.as_ls && ctx->options->key.has_multiview_view_index))
+ (!ctx->options->key.vs.out.as_es && !ctx->options->key.vs.out.as_ls && ctx->options->key.has_multiview_view_index))
return true;
break;
case MESA_SHADER_TESS_EVAL:
- if (ctx->shader_info->info.needs_multiview_view_index || (!ctx->options->key.tes.as_es && ctx->options->key.has_multiview_view_index))
+ if (ctx->shader_info->info.needs_multiview_view_index || (!ctx->options->key.tes.out.as_es && ctx->options->key.has_multiview_view_index))
return true;
break;
case MESA_SHADER_GEOMETRY:
uint8_t count = 0;
if (ctx->shader_info->info.vs.has_vertex_buffers)
- count += HAVE_32BIT_POINTERS ? 1 : 2;
+ count++;
count += ctx->shader_info->info.vs.needs_draw_id ? 3 : 2;
return count;
}
+static void allocate_inline_push_consts(struct radv_shader_context *ctx,
+ struct user_sgpr_info *user_sgpr_info)
+{
+ uint8_t remaining_sgprs = user_sgpr_info->remaining_sgprs;
+
+ /* Only supported if shaders use push constants. */
+ if (ctx->shader_info->info.min_push_constant_used == UINT8_MAX)
+ return;
+
+ /* Only supported if shaders don't have indirect push constants. */
+ if (ctx->shader_info->info.has_indirect_push_constants)
+ return;
+
+ /* Only supported for 32-bit push constants. */
+ if (!ctx->shader_info->info.has_only_32bit_push_constants)
+ return;
+
+ uint8_t num_push_consts =
+ (ctx->shader_info->info.max_push_constant_used -
+ ctx->shader_info->info.min_push_constant_used) / 4;
+
+ /* Check if the number of user SGPRs is large enough. */
+ if (num_push_consts < remaining_sgprs) {
+ ctx->shader_info->info.num_inline_push_consts = num_push_consts;
+ } else {
+ ctx->shader_info->info.num_inline_push_consts = remaining_sgprs;
+ }
+
+ /* Clamp to the maximum number of allowed inlined push constants. */
+ if (ctx->shader_info->info.num_inline_push_consts > AC_MAX_INLINE_PUSH_CONSTS)
+ ctx->shader_info->info.num_inline_push_consts = AC_MAX_INLINE_PUSH_CONSTS;
+
+ if (ctx->shader_info->info.num_inline_push_consts == num_push_consts &&
+ !ctx->shader_info->info.loads_dynamic_offsets) {
+ /* Disable the default push constants path if all constants are
+ * inlined and if shaders don't use dynamic descriptors.
+ */
+ ctx->shader_info->info.loads_push_constants = false;
+ }
+
+ ctx->shader_info->info.base_inline_push_consts =
+ ctx->shader_info->info.min_push_constant_used / 4;
+}
+
static void allocate_user_sgprs(struct radv_shader_context *ctx,
gl_shader_stage stage,
bool has_previous_stage,
user_sgpr_count++;
if (ctx->shader_info->info.loads_push_constants)
- user_sgpr_count += HAVE_32BIT_POINTERS ? 1 : 2;
+ user_sgpr_count++;
+
+ if (ctx->streamout_buffers)
+ user_sgpr_count++;
- uint32_t available_sgprs = ctx->options->chip_class >= GFX9 ? 32 : 16;
+ uint32_t available_sgprs = ctx->options->chip_class >= GFX9 && stage != MESA_SHADER_COMPUTE ? 32 : 16;
uint32_t remaining_sgprs = available_sgprs - user_sgpr_count;
uint32_t num_desc_set =
util_bitcount(ctx->shader_info->info.desc_set_used_mask);
- if (remaining_sgprs / (HAVE_32BIT_POINTERS ? 1 : 2) < num_desc_set) {
+ if (remaining_sgprs < num_desc_set) {
user_sgpr_info->indirect_all_descriptor_sets = true;
+ user_sgpr_info->remaining_sgprs = remaining_sgprs - 1;
+ } else {
+ user_sgpr_info->remaining_sgprs = remaining_sgprs - num_desc_set;
}
+
+ allocate_inline_push_consts(ctx, user_sgpr_info);
}
static void
declare_global_input_sgprs(struct radv_shader_context *ctx,
- gl_shader_stage stage,
- bool has_previous_stage,
- gl_shader_stage previous_stage,
const struct user_sgpr_info *user_sgpr_info,
struct arg_info *args,
LLVMValueRef *desc_sets)
{
LLVMTypeRef type = ac_array_in_const32_addr_space(ctx->ac.i8);
- unsigned num_sets = ctx->options->layout ?
- ctx->options->layout->num_sets : 0;
- unsigned stage_mask = 1 << stage;
-
- if (has_previous_stage)
- stage_mask |= 1 << previous_stage;
/* 1 for each descriptor set */
if (!user_sgpr_info->indirect_all_descriptor_sets) {
- for (unsigned i = 0; i < num_sets; ++i) {
- if ((ctx->shader_info->info.desc_set_used_mask & (1 << i)) &&
- ctx->options->layout->set[i].layout->shader_stages & stage_mask) {
- add_array_arg(args, type,
- &ctx->descriptor_sets[i]);
- }
+ uint32_t mask = ctx->shader_info->info.desc_set_used_mask;
+
+ while (mask) {
+ int i = u_bit_scan(&mask);
+
+ add_arg(args, ARG_SGPR, type, &ctx->descriptor_sets[i]);
}
} else {
- add_array_arg(args, ac_array_in_const32_addr_space(type), desc_sets);
+ add_arg(args, ARG_SGPR, ac_array_in_const32_addr_space(type),
+ desc_sets);
}
if (ctx->shader_info->info.loads_push_constants) {
/* 1 for push constants and dynamic descriptors */
- add_array_arg(args, type, &ctx->abi.push_constants);
+ add_arg(args, ARG_SGPR, type, &ctx->abi.push_constants);
+ }
+
+ for (unsigned i = 0; i < ctx->shader_info->info.num_inline_push_consts; i++) {
+ add_arg(args, ARG_SGPR, ctx->ac.i32,
+ &ctx->abi.inline_push_consts[i]);
+ }
+ ctx->abi.num_inline_push_consts = ctx->shader_info->info.num_inline_push_consts;
+ ctx->abi.base_inline_push_consts = ctx->shader_info->info.base_inline_push_consts;
+
+ if (ctx->shader_info->info.so.num_outputs) {
+ add_arg(args, ARG_SGPR,
+ ac_array_in_const32_addr_space(ctx->ac.v4i32),
+ &ctx->streamout_buffers);
}
}
{
add_arg(args, ARG_VGPR, ctx->ac.i32, &ctx->abi.vertex_id);
if (!ctx->is_gs_copy_shader) {
- if (ctx->options->key.vs.as_ls) {
+ if (ctx->options->key.vs.out.as_ls) {
add_arg(args, ARG_VGPR, ctx->ac.i32, &ctx->rel_auto_id);
add_arg(args, ARG_VGPR, ctx->ac.i32, &ctx->abi.instance_id);
+ add_arg(args, ARG_VGPR, ctx->ac.i32, NULL); /* unused */
} else {
- add_arg(args, ARG_VGPR, ctx->ac.i32, &ctx->abi.instance_id);
- add_arg(args, ARG_VGPR, ctx->ac.i32, &ctx->vs_prim_id);
+ if (ctx->ac.chip_class >= GFX10) {
+ add_arg(args, ARG_VGPR, ctx->ac.i32, NULL); /* user vgpr */
+ add_arg(args, ARG_VGPR, ctx->ac.i32, NULL); /* user vgpr */
+ add_arg(args, ARG_VGPR, ctx->ac.i32, &ctx->abi.instance_id);
+ } else {
+ add_arg(args, ARG_VGPR, ctx->ac.i32, &ctx->abi.instance_id);
+ add_arg(args, ARG_VGPR, ctx->ac.i32, &ctx->vs_prim_id);
+ add_arg(args, ARG_VGPR, ctx->ac.i32, NULL); /* unused */
+ }
+ }
+ }
+}
+
+static void
+declare_streamout_sgprs(struct radv_shader_context *ctx, gl_shader_stage stage,
+ struct arg_info *args)
+{
+ int i;
+
+ if (ctx->ac.chip_class >= GFX10)
+ return;
+
+ /* Streamout SGPRs. */
+ if (ctx->shader_info->info.so.num_outputs) {
+ assert(stage == MESA_SHADER_VERTEX ||
+ stage == MESA_SHADER_TESS_EVAL);
+
+ if (stage != MESA_SHADER_TESS_EVAL) {
+ add_arg(args, ARG_SGPR, ctx->ac.i32, &ctx->streamout_config);
+ } else {
+ args->assign[args->count - 1] = &ctx->streamout_config;
+ args->types[args->count - 1] = ctx->ac.i32;
}
- add_arg(args, ARG_VGPR, ctx->ac.i32, NULL); /* unused */
+
+ add_arg(args, ARG_SGPR, ctx->ac.i32, &ctx->streamout_write_idx);
+ }
+
+ /* A streamout buffer offset is loaded if the stride is non-zero. */
+ for (i = 0; i < 4; i++) {
+ if (!ctx->shader_info->info.so.strides[i])
+ continue;
+
+ add_arg(args, ARG_SGPR, ctx->ac.i32, &ctx->streamout_offset[i]);
}
}
}
static void
-set_global_input_locs(struct radv_shader_context *ctx, gl_shader_stage stage,
- bool has_previous_stage, gl_shader_stage previous_stage,
+set_global_input_locs(struct radv_shader_context *ctx,
const struct user_sgpr_info *user_sgpr_info,
LLVMValueRef desc_sets, uint8_t *user_sgpr_idx)
{
- unsigned num_sets = ctx->options->layout ?
- ctx->options->layout->num_sets : 0;
- unsigned stage_mask = 1 << stage;
-
- if (has_previous_stage)
- stage_mask |= 1 << previous_stage;
+ uint32_t mask = ctx->shader_info->info.desc_set_used_mask;
if (!user_sgpr_info->indirect_all_descriptor_sets) {
- for (unsigned i = 0; i < num_sets; ++i) {
- if ((ctx->shader_info->info.desc_set_used_mask & (1 << i)) &&
- ctx->options->layout->set[i].layout->shader_stages & stage_mask) {
- set_loc_desc(ctx, i, user_sgpr_idx, 0);
- } else
- ctx->descriptor_sets[i] = NULL;
+ while (mask) {
+ int i = u_bit_scan(&mask);
+
+ set_loc_desc(ctx, i, user_sgpr_idx);
}
} else {
set_loc_shader_ptr(ctx, AC_UD_INDIRECT_DESCRIPTOR_SETS,
user_sgpr_idx);
- for (unsigned i = 0; i < num_sets; ++i) {
- if ((ctx->shader_info->info.desc_set_used_mask & (1 << i)) &&
- ctx->options->layout->set[i].layout->shader_stages & stage_mask) {
- set_loc_desc(ctx, i, user_sgpr_idx, i * 8);
- ctx->descriptor_sets[i] =
- ac_build_load_to_sgpr(&ctx->ac,
- desc_sets,
- LLVMConstInt(ctx->ac.i32, i, false));
+ while (mask) {
+ int i = u_bit_scan(&mask);
+
+ ctx->descriptor_sets[i] =
+ ac_build_load_to_sgpr(&ctx->ac, desc_sets,
+ LLVMConstInt(ctx->ac.i32, i, false));
- } else
- ctx->descriptor_sets[i] = NULL;
}
+
ctx->shader_info->need_indirect_descriptor_sets = true;
}
if (ctx->shader_info->info.loads_push_constants) {
set_loc_shader_ptr(ctx, AC_UD_PUSH_CONSTANTS, user_sgpr_idx);
}
+
+ if (ctx->shader_info->info.num_inline_push_consts) {
+ set_loc_shader(ctx, AC_UD_INLINE_PUSH_CONSTANTS, user_sgpr_idx,
+ ctx->shader_info->info.num_inline_push_consts);
+ }
+
+ if (ctx->streamout_buffers) {
+ set_loc_shader_ptr(ctx, AC_UD_STREAMOUT_BUFFERS,
+ user_sgpr_idx);
+ }
}
static void
LLVMSetFunctionCallConv(func, calling_conv);
}
+/* Returns whether the stage is a stage that can be directly before the GS */
+static bool is_pre_gs_stage(gl_shader_stage stage)
+{
+ return stage == MESA_SHADER_VERTEX || stage == MESA_SHADER_TESS_EVAL;
+}
+
static void create_function(struct radv_shader_context *ctx,
gl_shader_stage stage,
bool has_previous_stage,
&ctx->ring_offsets);
}
+ if (ctx->ac.chip_class >= GFX10) {
+ if (stage == MESA_SHADER_VERTEX && ctx->options->key.vs.out.as_ngg) {
+ /* On GFX10, VS is merged into GS for NGG. */
+ stage = MESA_SHADER_GEOMETRY;
+ has_previous_stage = true;
+ previous_stage = MESA_SHADER_VERTEX;
+ }
+ }
+
switch (stage) {
case MESA_SHADER_COMPUTE:
- declare_global_input_sgprs(ctx, stage, has_previous_stage,
- previous_stage, &user_sgpr_info,
- &args, &desc_sets);
+ declare_global_input_sgprs(ctx, &user_sgpr_info, &args,
+ &desc_sets);
if (ctx->shader_info->info.cs.uses_grid_size) {
add_arg(&args, ARG_SGPR, ctx->ac.v3i32,
&ctx->abi.local_invocation_ids);
break;
case MESA_SHADER_VERTEX:
- declare_global_input_sgprs(ctx, stage, has_previous_stage,
- previous_stage, &user_sgpr_info,
- &args, &desc_sets);
+ declare_global_input_sgprs(ctx, &user_sgpr_info, &args,
+ &desc_sets);
+
declare_vs_specific_input_sgprs(ctx, stage, has_previous_stage,
previous_stage, &args);
if (needs_view_index)
add_arg(&args, ARG_SGPR, ctx->ac.i32,
&ctx->abi.view_index);
- if (ctx->options->key.vs.as_es)
+ if (ctx->options->key.vs.out.as_es) {
add_arg(&args, ARG_SGPR, ctx->ac.i32,
&ctx->es2gs_offset);
+ } else if (ctx->options->key.vs.out.as_ls) {
+ /* no extra parameters */
+ } else {
+ declare_streamout_sgprs(ctx, stage, &args);
+ }
declare_vs_input_vgprs(ctx, &args);
break;
add_arg(&args, ARG_SGPR, ctx->ac.i32, NULL); // unknown
add_arg(&args, ARG_SGPR, ctx->ac.i32, NULL); // unknown
- declare_global_input_sgprs(ctx, stage,
- has_previous_stage,
- previous_stage,
- &user_sgpr_info, &args,
+ declare_global_input_sgprs(ctx, &user_sgpr_info, &args,
&desc_sets);
+
declare_vs_specific_input_sgprs(ctx, stage,
has_previous_stage,
previous_stage, &args);
declare_vs_input_vgprs(ctx, &args);
} else {
- declare_global_input_sgprs(ctx, stage,
- has_previous_stage,
- previous_stage,
- &user_sgpr_info, &args,
+ declare_global_input_sgprs(ctx, &user_sgpr_info, &args,
&desc_sets);
if (needs_view_index)
}
break;
case MESA_SHADER_TESS_EVAL:
- declare_global_input_sgprs(ctx, stage, has_previous_stage,
- previous_stage, &user_sgpr_info,
- &args, &desc_sets);
+ declare_global_input_sgprs(ctx, &user_sgpr_info, &args,
+ &desc_sets);
if (needs_view_index)
add_arg(&args, ARG_SGPR, ctx->ac.i32,
&ctx->abi.view_index);
- if (ctx->options->key.tes.as_es) {
+ if (ctx->options->key.tes.out.as_es) {
add_arg(&args, ARG_SGPR, ctx->ac.i32, &ctx->oc_lds);
add_arg(&args, ARG_SGPR, ctx->ac.i32, NULL);
add_arg(&args, ARG_SGPR, ctx->ac.i32,
&ctx->es2gs_offset);
} else {
add_arg(&args, ARG_SGPR, ctx->ac.i32, NULL);
+ declare_streamout_sgprs(ctx, stage, &args);
add_arg(&args, ARG_SGPR, ctx->ac.i32, &ctx->oc_lds);
}
declare_tes_input_vgprs(ctx, &args);
case MESA_SHADER_GEOMETRY:
if (has_previous_stage) {
// First 6 system regs
- add_arg(&args, ARG_SGPR, ctx->ac.i32,
- &ctx->gs2vs_offset);
+ if (ctx->options->key.vs.out.as_ngg) {
+ add_arg(&args, ARG_SGPR, ctx->ac.i32,
+ &ctx->gs_tg_info);
+ } else {
+ add_arg(&args, ARG_SGPR, ctx->ac.i32,
+ &ctx->gs2vs_offset);
+ }
+
add_arg(&args, ARG_SGPR, ctx->ac.i32,
&ctx->merged_wave_info);
add_arg(&args, ARG_SGPR, ctx->ac.i32, &ctx->oc_lds);
add_arg(&args, ARG_SGPR, ctx->ac.i32, NULL); // unknown
add_arg(&args, ARG_SGPR, ctx->ac.i32, NULL); // unknown
- declare_global_input_sgprs(ctx, stage,
- has_previous_stage,
- previous_stage,
- &user_sgpr_info, &args,
+ declare_global_input_sgprs(ctx, &user_sgpr_info, &args,
&desc_sets);
if (previous_stage != MESA_SHADER_TESS_EVAL) {
declare_tes_input_vgprs(ctx, &args);
}
} else {
- declare_global_input_sgprs(ctx, stage,
- has_previous_stage,
- previous_stage,
- &user_sgpr_info, &args,
+ declare_global_input_sgprs(ctx, &user_sgpr_info, &args,
&desc_sets);
if (needs_view_index)
}
break;
case MESA_SHADER_FRAGMENT:
- declare_global_input_sgprs(ctx, stage, has_previous_stage,
- previous_stage, &user_sgpr_info,
- &args, &desc_sets);
+ declare_global_input_sgprs(ctx, &user_sgpr_info, &args,
+ &desc_sets);
add_arg(&args, ARG_SGPR, ctx->ac.i32, &ctx->abi.prim_mask);
add_arg(&args, ARG_VGPR, ctx->ac.v2i32, &ctx->persp_sample);
&user_sgpr_idx);
if (ctx->options->supports_spill) {
ctx->ring_offsets = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.implicit.buffer.ptr",
- LLVMPointerType(ctx->ac.i8, AC_CONST_ADDR_SPACE),
+ LLVMPointerType(ctx->ac.i8, AC_ADDR_SPACE_CONST),
NULL, 0, AC_FUNC_ATTR_READNONE);
ctx->ring_offsets = LLVMBuildBitCast(ctx->ac.builder, ctx->ring_offsets,
ac_array_in_const_addr_space(ctx->ac.v4i32), "");
if (has_previous_stage)
user_sgpr_idx = 0;
- set_global_input_locs(ctx, stage, has_previous_stage, previous_stage,
- &user_sgpr_info, desc_sets, &user_sgpr_idx);
+ set_global_input_locs(ctx, &user_sgpr_info, desc_sets, &user_sgpr_idx);
switch (stage) {
case MESA_SHADER_COMPUTE:
}
if (stage == MESA_SHADER_TESS_CTRL ||
- (stage == MESA_SHADER_VERTEX && ctx->options->key.vs.as_ls) ||
+ (stage == MESA_SHADER_VERTEX && ctx->options->key.vs.out.as_ls) ||
/* GFX9 has the ESGS ring buffer in LDS. */
(stage == MESA_SHADER_GEOMETRY && has_previous_stage)) {
ac_declare_lds_as_pointer(&ctx->ac);
} else
stride = LLVMConstInt(ctx->ac.i32, layout->binding[binding].size, false);
- offset = ac_build_imad(&ctx->ac, index, stride,
- LLVMConstInt(ctx->ac.i32, base_offset, false));
+ offset = LLVMConstInt(ctx->ac.i32, base_offset, false);
- desc_ptr = ac_build_gep0(&ctx->ac, desc_ptr, offset);
+ if (layout->binding[binding].type != VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT) {
+ offset = ac_build_imad(&ctx->ac, index, stride, offset);
+ }
+
+ desc_ptr = LLVMBuildGEP(ctx->ac.builder, desc_ptr, &offset, 1, "");
desc_ptr = ac_cast_ptr(&ctx->ac, desc_ptr, ctx->ac.v4i32);
LLVMSetMetadata(desc_ptr, ctx->ac.uniform_md_kind, ctx->ac.empty_md);
+ if (layout->binding[binding].type == VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT) {
+ uint32_t desc_type = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) |
+ S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
+ S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) |
+ S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W) |
+ S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
+ S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32);
+
+ LLVMValueRef desc_components[4] = {
+ LLVMBuildPtrToInt(ctx->ac.builder, desc_ptr, ctx->ac.intptr, ""),
+ LLVMConstInt(ctx->ac.i32, S_008F04_BASE_ADDRESS_HI(ctx->options->address32_hi), false),
+ /* High limit to support variable sizes. */
+ LLVMConstInt(ctx->ac.i32, 0xffffffff, false),
+ LLVMConstInt(ctx->ac.i32, desc_type, false),
+ };
+
+ return ac_build_gather_values(&ctx->ac, desc_components, 4);
+ }
+
return desc_ptr;
}
{
struct radv_shader_context *ctx = radv_shader_context_from_abi(abi);
const unsigned location = var->data.location;
- const unsigned component = var->data.location_frac;
+ unsigned component = var->data.location_frac;
const bool is_patch = var->data.patch;
const bool is_compact = var->data.compact;
LLVMValueRef dw_addr;
}
param = shader_io_get_unique_index(location);
- if (location == VARYING_SLOT_CLIP_DIST0 &&
- is_compact && const_index > 3) {
- const_index -= 3;
- param++;
+ if ((location == VARYING_SLOT_CLIP_DIST0 || location == VARYING_SLOT_CLIP_DIST1) && is_compact) {
+ const_index += component;
+ component = 0;
+
+ if (const_index >= 4) {
+ const_index -= 4;
+ param++;
+ }
}
if (!is_patch) {
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);
+ 4 * (base + chan), ac_glc, false);
}
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);
+ (base * 4), ac_glc, false);
}
}
LLVMValueRef result;
unsigned param = shader_io_get_unique_index(location);
- if (location == VARYING_SLOT_CLIP_DIST0 && is_compact && const_index > 3) {
- const_index -= 3;
- param++;
+ if ((location == VARYING_SLOT_CLIP_DIST0 || location == VARYING_SLOT_CLIP_DIST1) && is_compact) {
+ const_index += component;
+ component = 0;
+ if (const_index >= 4) {
+ const_index -= 4;
+ param++;
+ }
}
buf_addr = get_tcs_tes_buffer_address_params(ctx, param, const_index,
buf_addr = LLVMBuildAdd(ctx->ac.builder, buf_addr, comp_offset, "");
result = ac_build_buffer_load(&ctx->ac, ctx->hs_ring_tess_offchip, num_components, NULL,
- buf_addr, ctx->oc_lds, is_compact ? (4 * const_index) : 0, 1, 0, true, false);
+ buf_addr, ctx->oc_lds, is_compact ? (4 * const_index) : 0, ac_glc, true, false);
result = ac_trim_vector(&ctx->ac, result, num_components);
return result;
}
ctx->esgs_ring, 1,
ctx->ac.i32_0,
vtx_offset, soffset,
- 0, 1, 0, true, false);
+ 0, ac_glc, true, false);
}
if (ac_get_type_size(type) == 2) {
case 8:
sample_pos_offset = 7;
break;
- case 16:
- sample_pos_offset = 15;
- break;
default:
break;
}
struct radv_shader_context *ctx = radv_shader_context_from_abi(abi);
LLVMValueRef result;
- LLVMValueRef ptr = ac_build_gep0(&ctx->ac, ctx->ring_offsets, LLVMConstInt(ctx->ac.i32, RING_PS_SAMPLE_POSITIONS, false));
+ LLVMValueRef index = LLVMConstInt(ctx->ac.i32, RING_PS_SAMPLE_POSITIONS, false);
+ LLVMValueRef ptr = LLVMBuildGEP(ctx->ac.builder, ctx->ring_offsets, &index, 1, "");
ptr = LLVMBuildBitCast(ctx->ac.builder, ptr,
ac_array_in_const_addr_space(ctx->ac.v2f32), "");
unsigned offset = 0;
struct radv_shader_context *ctx = radv_shader_context_from_abi(abi);
- assert(stream == 0);
-
/* Write vertex attribute values to GSVS ring */
gs_next_vertex = LLVMBuildLoad(ctx->ac.builder,
- ctx->gs_next_vertex,
+ ctx->gs_next_vertex[stream],
"");
/* If this thread has already emitted the declared maximum number of
for (unsigned i = 0; i < AC_LLVM_MAX_OUTPUTS; ++i) {
unsigned output_usage_mask =
ctx->shader_info->info.gs.output_usage_mask[i];
+ uint8_t output_stream =
+ ctx->shader_info->info.gs.output_streams[i];
LLVMValueRef *out_ptr = &addrs[i * 4];
- int length = 4;
+ int length = util_last_bit(output_usage_mask);
- if (!(ctx->output_mask & (1ull << i)))
+ if (!(ctx->output_mask & (1ull << i)) ||
+ output_stream != stream)
continue;
- if (i == VARYING_SLOT_CLIP_DIST0) {
- /* pack clip and cull into a single set of slots */
- length = util_last_bit(output_usage_mask);
- }
-
for (unsigned j = 0; j < length; j++) {
if (!(output_usage_mask & (1 << j)))
continue;
out_val = ac_to_integer(&ctx->ac, out_val);
out_val = LLVMBuildZExtOrBitCast(ctx->ac.builder, out_val, ctx->ac.i32, "");
- ac_build_buffer_store_dword(&ctx->ac, ctx->gsvs_ring,
+ ac_build_buffer_store_dword(&ctx->ac,
+ ctx->gsvs_ring[stream],
out_val, 1,
voffset, ctx->gs2vs_offset, 0,
- 1, 1, true, true);
+ ac_glc | ac_slc, true);
}
}
gs_next_vertex = LLVMBuildAdd(ctx->ac.builder, gs_next_vertex,
ctx->ac.i32_1, "");
- LLVMBuildStore(ctx->ac.builder, gs_next_vertex, ctx->gs_next_vertex);
+ LLVMBuildStore(ctx->ac.builder, gs_next_vertex, ctx->gs_next_vertex[stream]);
- ac_build_sendmsg(&ctx->ac, AC_SENDMSG_GS_OP_EMIT | AC_SENDMSG_GS | (0 << 8), ctx->gs_wave_id);
+ ac_build_sendmsg(&ctx->ac,
+ AC_SENDMSG_GS_OP_EMIT | AC_SENDMSG_GS | (stream << 8),
+ ctx->gs_wave_id);
}
static void
struct radv_shader_context *ctx = radv_shader_context_from_abi(abi);
LLVMValueRef result;
+ if (LLVMGetTypeKind(LLVMTypeOf(buffer_ptr)) != LLVMPointerTypeKind) {
+ /* Do not load the descriptor for inlined uniform blocks. */
+ return buffer_ptr;
+ }
+
LLVMSetMetadata(buffer_ptr, ctx->ac.uniform_md_kind, ctx->ac.empty_md);
result = LLVMBuildLoad(ctx->ac.builder, buffer_ptr, "");
break;
case AC_DESC_SAMPLER:
type = ctx->ac.v4i32;
- if (binding->type == VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER)
- offset += 64;
+ if (binding->type == VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER) {
+ offset += radv_combined_image_descriptor_sampler_offset(binding);
+ }
type_size = 16;
break;
type = ctx->ac.v4i32;
type_size = 16;
break;
+ case AC_DESC_PLANE_0:
+ case AC_DESC_PLANE_1:
+ case AC_DESC_PLANE_2:
+ type = ctx->ac.v8i32;
+ type_size = 32;
+ offset += 32 * (desc_type - AC_DESC_PLANE_0);
+ break;
default:
unreachable("invalid desc_type\n");
}
assert(stride % type_size == 0);
- if (!index)
- index = ctx->ac.i32_0;
+ LLVMValueRef adjusted_index = index;
+ if (!adjusted_index)
+ adjusted_index = ctx->ac.i32_0;
- index = LLVMBuildMul(builder, index, LLVMConstInt(ctx->ac.i32, stride / type_size, 0), "");
+ adjusted_index = LLVMBuildMul(builder, adjusted_index, LLVMConstInt(ctx->ac.i32, stride / type_size, 0), "");
- list = ac_build_gep0(&ctx->ac, list, LLVMConstInt(ctx->ac.i32, offset, 0));
+ LLVMValueRef val_offset = LLVMConstInt(ctx->ac.i32, offset, 0);
+ list = LLVMBuildGEP(builder, list, &val_offset, 1, "");
list = LLVMBuildPointerCast(builder, list,
ac_array_in_const32_addr_space(type), "");
- return ac_build_load_to_sgpr(&ctx->ac, list, index);
+ LLVMValueRef descriptor = ac_build_load_to_sgpr(&ctx->ac, list, adjusted_index);
+
+ /* 3 plane formats always have same size and format for plane 1 & 2, so
+ * use the tail from plane 1 so that we can store only the first 16 bytes
+ * of the last plane. */
+ if (desc_type == AC_DESC_PLANE_2) {
+ LLVMValueRef descriptor2 = radv_get_sampler_desc(abi, descriptor_set, base_index, constant_index, index, AC_DESC_PLANE_1,image, write, bindless);
+
+ LLVMValueRef components[8];
+ for (unsigned i = 0; i < 4; ++i)
+ components[i] = ac_llvm_extract_elem(&ctx->ac, descriptor, i);
+
+ for (unsigned i = 4; i < 8; ++i)
+ components[i] = ac_llvm_extract_elem(&ctx->ac, descriptor2, i);
+ descriptor = ac_build_gather_values(&ctx->ac, components, 8);
+ }
+
+ return descriptor;
}
/* For 2_10_10_10 formats the alpha is handled as unsigned by pre-vega HW.
LLVMValueRef c30 = LLVMConstInt(ctx->ac.i32, 30, 0);
+ alpha = LLVMBuildBitCast(ctx->ac.builder, alpha, ctx->ac.f32, "");
+
if (adjustment == RADV_ALPHA_ADJUST_SSCALED)
alpha = LLVMBuildFPToUI(ctx->ac.builder, alpha, ctx->ac.i32, "");
else
alpha = LLVMBuildSIToFP(ctx->ac.builder, alpha, ctx->ac.f32, "");
}
- return alpha;
+ return LLVMBuildBitCast(ctx->ac.builder, alpha, ctx->ac.i32, "");
+}
+
+static unsigned
+get_num_channels_from_data_format(unsigned data_format)
+{
+ switch (data_format) {
+ case V_008F0C_BUF_DATA_FORMAT_8:
+ case V_008F0C_BUF_DATA_FORMAT_16:
+ case V_008F0C_BUF_DATA_FORMAT_32:
+ return 1;
+ case V_008F0C_BUF_DATA_FORMAT_8_8:
+ case V_008F0C_BUF_DATA_FORMAT_16_16:
+ case V_008F0C_BUF_DATA_FORMAT_32_32:
+ return 2;
+ case V_008F0C_BUF_DATA_FORMAT_10_11_11:
+ case V_008F0C_BUF_DATA_FORMAT_11_11_10:
+ case V_008F0C_BUF_DATA_FORMAT_32_32_32:
+ return 3;
+ case V_008F0C_BUF_DATA_FORMAT_8_8_8_8:
+ case V_008F0C_BUF_DATA_FORMAT_10_10_10_2:
+ case V_008F0C_BUF_DATA_FORMAT_2_10_10_10:
+ case V_008F0C_BUF_DATA_FORMAT_16_16_16_16:
+ case V_008F0C_BUF_DATA_FORMAT_32_32_32_32:
+ return 4;
+ default:
+ break;
+ }
+
+ return 4;
+}
+
+static LLVMValueRef
+radv_fixup_vertex_input_fetches(struct radv_shader_context *ctx,
+ LLVMValueRef value,
+ unsigned num_channels,
+ bool is_float)
+{
+ LLVMValueRef zero = is_float ? ctx->ac.f32_0 : ctx->ac.i32_0;
+ LLVMValueRef one = is_float ? ctx->ac.f32_1 : ctx->ac.i32_1;
+ LLVMValueRef chan[4];
+
+ if (LLVMGetTypeKind(LLVMTypeOf(value)) == LLVMVectorTypeKind) {
+ unsigned vec_size = LLVMGetVectorSize(LLVMTypeOf(value));
+
+ if (num_channels == 4 && num_channels == vec_size)
+ return value;
+
+ num_channels = MIN2(num_channels, vec_size);
+
+ for (unsigned i = 0; i < num_channels; i++)
+ chan[i] = ac_llvm_extract_elem(&ctx->ac, value, i);
+ } else {
+ if (num_channels) {
+ assert(num_channels == 1);
+ chan[0] = value;
+ }
+ }
+
+ for (unsigned i = num_channels; i < 4; i++) {
+ chan[i] = i == 3 ? one : zero;
+ chan[i] = ac_to_integer(&ctx->ac, chan[i]);
+ }
+
+ return ac_build_gather_values(&ctx->ac, chan, 4);
}
static void
unsigned attrib_count = glsl_count_attribute_slots(variable->type, true);
uint8_t input_usage_mask =
ctx->shader_info->info.vs.input_usage_mask[variable->data.location];
- unsigned num_channels = util_last_bit(input_usage_mask);
+ unsigned num_input_channels = util_last_bit(input_usage_mask);
variable->data.driver_location = variable->data.location * 4;
for (unsigned i = 0; i < attrib_count; ++i) {
LLVMValueRef output[4];
unsigned attrib_index = variable->data.location + i - VERT_ATTRIB_GENERIC0;
+ unsigned attrib_format = ctx->options->key.vs.vertex_attribute_formats[attrib_index];
+ unsigned data_format = attrib_format & 0x0f;
+ unsigned num_format = (attrib_format >> 4) & 0x07;
+ bool is_float = num_format != V_008F0C_BUF_NUM_FORMAT_UINT &&
+ num_format != V_008F0C_BUF_NUM_FORMAT_SINT;
if (ctx->options->key.vs.instance_rate_inputs & (1u << attrib_index)) {
uint32_t divisor = ctx->options->key.vs.instance_rate_divisors[attrib_index];
buffer_index = LLVMBuildUDiv(ctx->ac.builder, buffer_index,
LLVMConstInt(ctx->ac.i32, divisor, 0), "");
}
-
- if (ctx->options->key.vs.as_ls) {
- ctx->shader_info->vs.vgpr_comp_cnt =
- MAX2(2, ctx->shader_info->vs.vgpr_comp_cnt);
- } else {
- ctx->shader_info->vs.vgpr_comp_cnt =
- MAX2(1, ctx->shader_info->vs.vgpr_comp_cnt);
- }
} else {
buffer_index = ctx->ac.i32_0;
}
} else
buffer_index = LLVMBuildAdd(ctx->ac.builder, ctx->abi.vertex_id,
ctx->abi.base_vertex, "");
- t_offset = LLVMConstInt(ctx->ac.i32, attrib_index, false);
- t_list = ac_build_load_to_sgpr(&ctx->ac, t_list_ptr, t_offset);
+ /* Adjust the number of channels to load based on the vertex
+ * attribute format.
+ */
+ unsigned num_format_channels = get_num_channels_from_data_format(data_format);
+ unsigned num_channels = MIN2(num_input_channels, num_format_channels);
+ unsigned attrib_binding = ctx->options->key.vs.vertex_attribute_bindings[attrib_index];
+ unsigned attrib_offset = ctx->options->key.vs.vertex_attribute_offsets[attrib_index];
+ unsigned attrib_stride = ctx->options->key.vs.vertex_attribute_strides[attrib_index];
+
+ if (ctx->options->key.vs.post_shuffle & (1 << attrib_index)) {
+ /* Always load, at least, 3 channels for formats that
+ * need to be shuffled because X<->Z.
+ */
+ num_channels = MAX2(num_channels, 3);
+ }
+
+ if (attrib_stride != 0 && attrib_offset > attrib_stride) {
+ LLVMValueRef buffer_offset =
+ LLVMConstInt(ctx->ac.i32,
+ attrib_offset / attrib_stride, false);
- input = ac_build_buffer_load_format(&ctx->ac, t_list,
+ buffer_index = LLVMBuildAdd(ctx->ac.builder,
buffer_index,
- ctx->ac.i32_0,
- num_channels, false, true);
+ buffer_offset, "");
- input = ac_build_expand_to_vec4(&ctx->ac, input, num_channels);
+ attrib_offset = attrib_offset % attrib_stride;
+ }
+
+ t_offset = LLVMConstInt(ctx->ac.i32, attrib_binding, false);
+ t_list = ac_build_load_to_sgpr(&ctx->ac, t_list_ptr, t_offset);
+
+ input = ac_build_struct_tbuffer_load(&ctx->ac, t_list,
+ buffer_index,
+ LLVMConstInt(ctx->ac.i32, attrib_offset, false),
+ ctx->ac.i32_0, ctx->ac.i32_0,
+ num_channels,
+ data_format, num_format, 0, true);
+
+ if (ctx->options->key.vs.post_shuffle & (1 << attrib_index)) {
+ LLVMValueRef c[4];
+ c[0] = ac_llvm_extract_elem(&ctx->ac, input, 2);
+ c[1] = ac_llvm_extract_elem(&ctx->ac, input, 1);
+ c[2] = ac_llvm_extract_elem(&ctx->ac, input, 0);
+ c[3] = ac_llvm_extract_elem(&ctx->ac, input, 3);
+
+ input = ac_build_gather_values(&ctx->ac, c, 4);
+ }
+
+ input = radv_fixup_vertex_input_fetches(ctx, input, num_channels,
+ is_float);
for (unsigned chan = 0; chan < 4; chan++) {
LLVMValueRef llvm_chan = LLVMConstInt(ctx->ac.i32, chan, false);
unsigned attr,
LLVMValueRef interp_param,
LLVMValueRef prim_mask,
+ bool float16,
LLVMValueRef result[4])
{
LLVMValueRef attr_number;
for (chan = 0; chan < 4; chan++) {
LLVMValueRef llvm_chan = LLVMConstInt(ctx->ac.i32, chan, false);
- if (interp) {
+ if (interp && float16) {
+ result[chan] = ac_build_fs_interp_f16(&ctx->ac,
+ llvm_chan,
+ attr_number,
+ prim_mask, i, j);
+ } else if (interp) {
result[chan] = ac_build_fs_interp(&ctx->ac,
llvm_chan,
attr_number,
attr_number,
prim_mask);
result[chan] = LLVMBuildBitCast(ctx->ac.builder, result[chan], ctx->ac.i32, "");
- result[chan] = LLVMBuildTruncOrBitCast(ctx->ac.builder, result[chan], LLVMTypeOf(interp_param), "");
+ result[chan] = LLVMBuildTruncOrBitCast(ctx->ac.builder, result[chan], float16 ? ctx->ac.i16 : ctx->ac.i32, "");
+ }
+ }
+}
+
+static void mark_16bit_fs_input(struct radv_shader_context *ctx,
+ const struct glsl_type *type,
+ int location)
+{
+ if (glsl_type_is_scalar(type) || glsl_type_is_vector(type) || glsl_type_is_matrix(type)) {
+ unsigned attrib_count = glsl_count_attribute_slots(type, false);
+ if (glsl_type_is_16bit(type)) {
+ ctx->float16_shaded_mask |= ((1ull << attrib_count) - 1) << location;
+ }
+ } else if (glsl_type_is_array(type)) {
+ unsigned stride = glsl_count_attribute_slots(glsl_get_array_element(type), false);
+ for (unsigned i = 0; i < glsl_get_length(type); ++i) {
+ mark_16bit_fs_input(ctx, glsl_get_array_element(type), location + i * stride);
+ }
+ } else {
+ assert(glsl_type_is_struct_or_ifc(type));
+ for (unsigned i = 0; i < glsl_get_length(type); i++) {
+ mark_16bit_fs_input(ctx, glsl_get_struct_field(type, i), location);
+ location += glsl_count_attribute_slots(glsl_get_struct_field(type, i), false);
}
}
}
uint64_t mask;
variable->data.driver_location = idx * 4;
+
+
+ if (variable->data.compact) {
+ unsigned component_count = variable->data.location_frac +
+ glsl_get_length(variable->type);
+ attrib_count = (component_count + 3) / 4;
+ } else
+ mark_16bit_fs_input(ctx, variable->type, idx);
+
mask = ((1ull << attrib_count) - 1) << variable->data.location;
- if (glsl_get_base_type(glsl_without_array(variable->type)) == GLSL_TYPE_FLOAT) {
+ if (glsl_get_base_type(glsl_without_array(variable->type)) == GLSL_TYPE_FLOAT ||
+ glsl_get_base_type(glsl_without_array(variable->type)) == GLSL_TYPE_FLOAT16 ||
+ glsl_get_base_type(glsl_without_array(variable->type)) == GLSL_TYPE_STRUCT) {
unsigned interp_type;
if (variable->data.sample)
interp_type = INTERP_SAMPLE;
interp = lookup_interp_param(&ctx->abi, variable->data.interpolation, interp_type);
}
- bool is_16bit = glsl_type_is_16bit(variable->type);
- LLVMTypeRef type = is_16bit ? ctx->ac.i16 : ctx->ac.i32;
if (interp == NULL)
- interp = LLVMGetUndef(type);
+ interp = LLVMGetUndef(ctx->ac.i32);
for (unsigned i = 0; i < attrib_count; ++i)
ctx->inputs[ac_llvm_reg_index_soa(idx + i, 0)] = interp;
- if (idx == VARYING_SLOT_CLIP_DIST0) {
- /* Do not account for the number of components inside the array
- * of clip/cull distances because this might wrongly set other
- * bits like primitive ID or layer.
- */
- mask = 1ull << VARYING_SLOT_CLIP_DIST0;
- }
-
ctx->input_mask |= mask;
}
unsigned index = 0;
- if (ctx->shader_info->info.ps.uses_input_attachments ||
- ctx->shader_info->info.needs_multiview_view_index) {
+ if (ctx->shader_info->info.needs_multiview_view_index ||
+ ctx->shader_info->info.ps.layer_input) {
ctx->input_mask |= 1ull << VARYING_SLOT_LAYER;
ctx->inputs[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER, 0)] = LLVMGetUndef(ctx->ac.i32);
}
if (i >= VARYING_SLOT_VAR0 || i == VARYING_SLOT_PNTC ||
i == VARYING_SLOT_PRIMITIVE_ID || i == VARYING_SLOT_LAYER) {
interp_param = *inputs;
- interp_fs_input(ctx, index, interp_param, ctx->abi.prim_mask,
+ bool float16 = (ctx->float16_shaded_mask >> i) & 1;
+ interp_fs_input(ctx, index, interp_param, ctx->abi.prim_mask, float16,
inputs);
if (LLVMIsUndef(interp_param))
ctx->shader_info->fs.flat_shaded_mask |= 1u << index;
+ if (float16)
+ ctx->shader_info->fs.float16_shaded_mask |= 1u << index;
+ if (i >= VARYING_SLOT_VAR0)
+ ctx->abi.fs_input_attr_indices[i - VARYING_SLOT_VAR0] = index;
++index;
} else if (i == VARYING_SLOT_CLIP_DIST0) {
int length = ctx->shader_info->info.ps.num_input_clips_culls;
interp_param = *inputs;
interp_fs_input(ctx, index, interp_param,
- ctx->abi.prim_mask, inputs);
+ ctx->abi.prim_mask, false, inputs);
++index;
}
} else if (i == VARYING_SLOT_POS) {
if (stage == MESA_SHADER_TESS_CTRL)
return;
+ if (variable->data.compact) {
+ unsigned component_count = variable->data.location_frac +
+ glsl_get_length(variable->type);
+ attrib_count = (component_count + 3) / 4;
+ }
+
mask_attribs = ((1ull << attrib_count) - 1) << idx;
if (stage == MESA_SHADER_VERTEX ||
stage == MESA_SHADER_TESS_EVAL ||
ctx->shader_info->tes.outinfo.cull_dist_mask = (1 << shader->info.cull_distance_array_size) - 1;
ctx->shader_info->tes.outinfo.cull_dist_mask <<= shader->info.clip_distance_array_size;
}
-
- mask_attribs = 1ull << idx;
}
}
return;
bool is_16bit = ac_get_type_size(LLVMTypeOf(values[0])) == 2;
- if (ctx->stage == MESA_SHADER_FRAGMENT && target >= V_008DFC_SQ_EXP_MRT) {
+ if (ctx->stage == MESA_SHADER_FRAGMENT) {
unsigned index = target - V_008DFC_SQ_EXP_MRT;
unsigned col_format = (ctx->options->key.fs.col_format >> (4 * index)) & 0xf;
bool is_int8 = (ctx->options->key.fs.is_int8 >> index) & 1;
break;
case V_028714_SPI_SHADER_32_AR:
- args->enabled_channels = 0x9;
- args->out[0] = values[0];
- args->out[3] = values[3];
+ if (ctx->ac.chip_class >= GFX10) {
+ args->enabled_channels = 0x3;
+ args->out[0] = values[0];
+ args->out[1] = values[3];
+ } else {
+ args->enabled_channels = 0x9;
+ args->out[0] = values[0];
+ args->out[3] = values[3];
+ }
break;
case V_028714_SPI_SHADER_FP16_ABGR:
if (is_16bit) {
for (unsigned chan = 0; chan < 4; chan++)
values[chan] = LLVMBuildZExt(ctx->ac.builder,
- values[chan],
+ ac_to_integer(&ctx->ac, values[chan]),
ctx->ac.i32, "");
}
break;
if (is_16bit) {
for (unsigned chan = 0; chan < 4; chan++)
values[chan] = LLVMBuildSExt(ctx->ac.builder,
- values[chan],
+ ac_to_integer(&ctx->ac, values[chan]),
ctx->ac.i32, "");
}
break;
} else
memcpy(&args->out[0], values, sizeof(values[0]) * 4);
- for (unsigned i = 0; i < 4; ++i) {
- if (!(args->enabled_channels & (1 << i)))
- continue;
-
+ for (unsigned i = 0; i < 4; ++i)
args->out[i] = ac_to_float(&ctx->ac, args->out[i]);
- }
}
static void
return LLVMBuildLoad(ctx->ac.builder, output, "");
}
+static void
+radv_emit_stream_output(struct radv_shader_context *ctx,
+ LLVMValueRef const *so_buffers,
+ LLVMValueRef const *so_write_offsets,
+ const struct radv_stream_output *output)
+{
+ unsigned num_comps = util_bitcount(output->component_mask);
+ unsigned loc = output->location;
+ unsigned buf = output->buffer;
+ unsigned offset = output->offset;
+ unsigned start;
+ LLVMValueRef out[4];
+
+ assert(num_comps && num_comps <= 4);
+ if (!num_comps || num_comps > 4)
+ return;
+
+ /* Get the first component. */
+ start = ffs(output->component_mask) - 1;
+
+ /* Load the output as int. */
+ for (int i = 0; i < num_comps; i++) {
+ out[i] = ac_to_integer(&ctx->ac,
+ radv_load_output(ctx, loc, start + i));
+ }
+
+ /* Pack the output. */
+ LLVMValueRef vdata = NULL;
+
+ switch (num_comps) {
+ case 1: /* as i32 */
+ vdata = out[0];
+ break;
+ case 2: /* as v2i32 */
+ case 3: /* as v4i32 (aligned to 4) */
+ out[3] = LLVMGetUndef(ctx->ac.i32);
+ /* fall through */
+ case 4: /* as v4i32 */
+ vdata = ac_build_gather_values(&ctx->ac, out,
+ !ac_has_vec3_support(ctx->ac.chip_class, false) ?
+ util_next_power_of_two(num_comps) :
+ num_comps);
+ break;
+ }
+
+ ac_build_buffer_store_dword(&ctx->ac, so_buffers[buf],
+ vdata, num_comps, so_write_offsets[buf],
+ ctx->ac.i32_0, offset,
+ ac_glc | ac_slc, false);
+}
+
+static void
+radv_emit_streamout(struct radv_shader_context *ctx, unsigned stream)
+{
+ struct ac_build_if_state if_ctx;
+ int i;
+
+ /* Get bits [22:16], i.e. (so_param >> 16) & 127; */
+ assert(ctx->streamout_config);
+ LLVMValueRef so_vtx_count =
+ ac_build_bfe(&ctx->ac, ctx->streamout_config,
+ LLVMConstInt(ctx->ac.i32, 16, false),
+ LLVMConstInt(ctx->ac.i32, 7, false), false);
+
+ LLVMValueRef tid = ac_get_thread_id(&ctx->ac);
+
+ /* can_emit = tid < so_vtx_count; */
+ LLVMValueRef can_emit = LLVMBuildICmp(ctx->ac.builder, LLVMIntULT,
+ tid, so_vtx_count, "");
+
+ /* Emit the streamout code conditionally. This actually avoids
+ * out-of-bounds buffer access. The hw tells us via the SGPR
+ * (so_vtx_count) which threads are allowed to emit streamout data.
+ */
+ ac_nir_build_if(&if_ctx, ctx, can_emit);
+ {
+ /* The buffer offset is computed as follows:
+ * ByteOffset = streamout_offset[buffer_id]*4 +
+ * (streamout_write_index + thread_id)*stride[buffer_id] +
+ * attrib_offset
+ */
+ LLVMValueRef so_write_index = ctx->streamout_write_idx;
+
+ /* Compute (streamout_write_index + thread_id). */
+ so_write_index =
+ LLVMBuildAdd(ctx->ac.builder, so_write_index, tid, "");
+
+ /* Load the descriptor and compute the write offset for each
+ * enabled buffer.
+ */
+ LLVMValueRef so_write_offset[4] = {};
+ LLVMValueRef so_buffers[4] = {};
+ LLVMValueRef buf_ptr = ctx->streamout_buffers;
+
+ for (i = 0; i < 4; i++) {
+ uint16_t stride = ctx->shader_info->info.so.strides[i];
+
+ if (!stride)
+ continue;
+
+ LLVMValueRef offset =
+ LLVMConstInt(ctx->ac.i32, i, false);
+
+ so_buffers[i] = ac_build_load_to_sgpr(&ctx->ac,
+ buf_ptr, offset);
+
+ LLVMValueRef so_offset = ctx->streamout_offset[i];
+
+ so_offset = LLVMBuildMul(ctx->ac.builder, so_offset,
+ LLVMConstInt(ctx->ac.i32, 4, false), "");
+
+ so_write_offset[i] =
+ ac_build_imad(&ctx->ac, so_write_index,
+ LLVMConstInt(ctx->ac.i32,
+ stride * 4, false),
+ so_offset);
+ }
+
+ /* Write streamout data. */
+ for (i = 0; i < ctx->shader_info->info.so.num_outputs; i++) {
+ struct radv_stream_output *output =
+ &ctx->shader_info->info.so.outputs[i];
+
+ if (stream != output->stream)
+ continue;
+
+ radv_emit_stream_output(ctx, so_buffers,
+ so_write_offset, output);
+ }
+ }
+ ac_nir_build_endif(&if_ctx);
+}
+
static void
handle_vs_outputs_post(struct radv_shader_context *ctx,
bool export_prim_id, bool export_layer_id,
+ bool export_clip_dists,
struct radv_vs_output_info *outinfo)
{
uint32_t param_count = 0;
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)) {
- unsigned output_usage_mask, length;
- LLVMValueRef slots[8];
- unsigned j;
-
- if (ctx->stage == MESA_SHADER_VERTEX &&
- !ctx->is_gs_copy_shader) {
- output_usage_mask =
- ctx->shader_info->info.vs.output_usage_mask[VARYING_SLOT_CLIP_DIST0];
- } else if (ctx->stage == MESA_SHADER_TESS_EVAL) {
- output_usage_mask =
- ctx->shader_info->info.tes.output_usage_mask[VARYING_SLOT_CLIP_DIST0];
- } else {
- assert(ctx->is_gs_copy_shader);
- output_usage_mask =
- ctx->shader_info->info.gs.output_usage_mask[VARYING_SLOT_CLIP_DIST0];
- }
+ for(unsigned location = VARYING_SLOT_CLIP_DIST0; location <= VARYING_SLOT_CLIP_DIST1; ++location) {
+ if (ctx->output_mask & (1ull << location)) {
+ unsigned output_usage_mask, length;
+ LLVMValueRef slots[4];
+ unsigned j;
+
+ if (ctx->stage == MESA_SHADER_VERTEX &&
+ !ctx->is_gs_copy_shader) {
+ output_usage_mask =
+ ctx->shader_info->info.vs.output_usage_mask[location];
+ } else if (ctx->stage == MESA_SHADER_TESS_EVAL) {
+ output_usage_mask =
+ ctx->shader_info->info.tes.output_usage_mask[location];
+ } else {
+ assert(ctx->is_gs_copy_shader);
+ output_usage_mask =
+ ctx->shader_info->info.gs.output_usage_mask[location];
+ }
- length = util_last_bit(output_usage_mask);
+ length = util_last_bit(output_usage_mask);
- i = VARYING_SLOT_CLIP_DIST0;
- for (j = 0; j < length; j++)
- slots[j] = ac_to_float(&ctx->ac, radv_load_output(ctx, i, j));
+ for (j = 0; j < length; j++)
+ slots[j] = ac_to_float(&ctx->ac, radv_load_output(ctx, location, j));
- for (i = length; i < 8; i++)
- slots[i] = LLVMGetUndef(ctx->ac.f32);
+ for (i = length; i < 4; i++)
+ slots[i] = LLVMGetUndef(ctx->ac.f32);
- if (length > 4) {
- target = V_008DFC_SQ_EXP_POS + 3;
- si_llvm_init_export_args(ctx, &slots[4], 0xf, target, &args);
+ target = V_008DFC_SQ_EXP_POS + 2 + (location - VARYING_SLOT_CLIP_DIST0);
+ si_llvm_init_export_args(ctx, &slots[0], 0xf, target, &args);
memcpy(&pos_args[target - V_008DFC_SQ_EXP_POS],
- &args, sizeof(args));
- }
-
- target = V_008DFC_SQ_EXP_POS + 2;
- si_llvm_init_export_args(ctx, &slots[0], 0xf, target, &args);
- memcpy(&pos_args[target - V_008DFC_SQ_EXP_POS],
- &args, sizeof(args));
+ &args, sizeof(args));
- /* Export the clip/cull distances values to the next stage. */
- radv_export_param(ctx, param_count, &slots[0], 0xf);
- outinfo->vs_output_param_offset[VARYING_SLOT_CLIP_DIST0] = param_count++;
- if (length > 4) {
- radv_export_param(ctx, param_count, &slots[4], 0xf);
- outinfo->vs_output_param_offset[VARYING_SLOT_CLIP_DIST1] = param_count++;
+ if (export_clip_dists) {
+ /* Export the clip/cull distances values to the next stage. */
+ radv_export_param(ctx, param_count, &slots[0], 0xf);
+ outinfo->vs_output_param_offset[location] = param_count++;
+ }
}
}
viewport_index_value = radv_load_output(ctx, VARYING_SLOT_VIEWPORT, 0);
}
+ if (ctx->shader_info->info.so.num_outputs &&
+ !ctx->is_gs_copy_shader) {
+ /* The GS copy shader emission already emits streamout. */
+ radv_emit_streamout(ctx, 0);
+ }
+
if (outinfo->writes_pointsize ||
outinfo->writes_layer ||
outinfo->writes_viewport_index) {
num_pos_exports++;
}
+ /* Navi10-14 skip POS0 exports if EXEC=0 and DONE=0, causing a hang.
+ * Setting valid_mask=1 prevents it and has no other effect.
+ */
+ if (ctx->ac.family == CHIP_NAVI10 ||
+ ctx->ac.family == CHIP_NAVI12 ||
+ ctx->ac.family == CHIP_NAVI14)
+ pos_args[0].valid_mask = 1;
+
pos_idx = 0;
for (i = 0; i < 4; i++) {
if (!pos_args[i].out[0])
LLVMValueRef values[4];
values[0] = ctx->vs_prim_id;
- ctx->shader_info->vs.vgpr_comp_cnt = MAX2(2,
- ctx->shader_info->vs.vgpr_comp_cnt);
for (unsigned j = 1; j < 4; j++)
values[j] = ctx->ac.f32_0;
LLVMValueRef lds_base = NULL;
for (unsigned i = 0; i < AC_LLVM_MAX_OUTPUTS; ++i) {
- unsigned output_usage_mask;
int param_index;
- int length = 4;
if (!(ctx->output_mask & (1ull << i)))
continue;
- if (ctx->stage == MESA_SHADER_VERTEX) {
- output_usage_mask =
- ctx->shader_info->info.vs.output_usage_mask[i];
- } else {
- assert(ctx->stage == MESA_SHADER_TESS_EVAL);
- output_usage_mask =
- ctx->shader_info->info.tes.output_usage_mask[i];
- }
-
- if (i == VARYING_SLOT_CLIP_DIST0)
- length = util_last_bit(output_usage_mask);
-
param_index = shader_io_get_unique_index(i);
- max_output_written = MAX2(param_index + (length > 4), max_output_written);
+ max_output_written = MAX2(param_index, max_output_written);
}
outinfo->esgs_itemsize = (max_output_written + 1) * 16;
LLVMValueRef *out_ptr = &ctx->abi.outputs[i * 4];
unsigned output_usage_mask;
int param_index;
- int length = 4;
if (!(ctx->output_mask & (1ull << i)))
continue;
ctx->shader_info->info.tes.output_usage_mask[i];
}
- if (i == VARYING_SLOT_CLIP_DIST0)
- length = util_last_bit(output_usage_mask);
-
param_index = shader_io_get_unique_index(i);
if (lds_base) {
"");
}
- for (j = 0; j < length; j++) {
+ for (j = 0; j < 4; j++) {
if (!(output_usage_mask & (1 << j)))
continue;
out_val, 1,
NULL, ctx->es2gs_offset,
(4 * param_index + j) * 4,
- 1, 1, true, true);
+ ac_glc | ac_slc, true);
}
}
}
vertex_dw_stride, "");
for (unsigned i = 0; i < AC_LLVM_MAX_OUTPUTS; ++i) {
- unsigned output_usage_mask =
- ctx->shader_info->info.vs.output_usage_mask[i];
LLVMValueRef *out_ptr = &ctx->abi.outputs[i * 4];
- int length = 4;
if (!(ctx->output_mask & (1ull << i)))
continue;
- if (i == VARYING_SLOT_CLIP_DIST0)
- length = util_last_bit(output_usage_mask);
-
int param = shader_io_get_unique_index(i);
LLVMValueRef dw_addr = LLVMBuildAdd(ctx->ac.builder, base_dw_addr,
LLVMConstInt(ctx->ac.i32, param * 4, false),
"");
- for (unsigned j = 0; j < length; j++) {
+ for (unsigned j = 0; j < 4; j++) {
LLVMValueRef value = LLVMBuildLoad(ctx->ac.builder, out_ptr[j], "");
value = ac_to_integer(&ctx->ac, value);
value = LLVMBuildZExtOrBitCast(ctx->ac.builder, value, ctx->ac.i32, "");
}
}
+static LLVMValueRef get_wave_id_in_tg(struct radv_shader_context *ctx)
+{
+ return ac_unpack_param(&ctx->ac, ctx->merged_wave_info, 24, 4);
+}
+
+static LLVMValueRef ngg_get_vtx_cnt(struct radv_shader_context *ctx)
+{
+ return ac_build_bfe(&ctx->ac, ctx->gs_tg_info,
+ LLVMConstInt(ctx->ac.i32, 12, false),
+ LLVMConstInt(ctx->ac.i32, 9, false),
+ false);
+}
+
+static LLVMValueRef ngg_get_prim_cnt(struct radv_shader_context *ctx)
+{
+ return ac_build_bfe(&ctx->ac, ctx->gs_tg_info,
+ LLVMConstInt(ctx->ac.i32, 22, false),
+ LLVMConstInt(ctx->ac.i32, 9, false),
+ false);
+}
+
+/* Send GS Alloc Req message from the first wave of the group to SPI.
+ * Message payload is:
+ * - bits 0..10: vertices in group
+ * - bits 12..22: primitives in group
+ */
+static void build_sendmsg_gs_alloc_req(struct radv_shader_context *ctx,
+ LLVMValueRef vtx_cnt,
+ LLVMValueRef prim_cnt)
+{
+ LLVMBuilderRef builder = ctx->ac.builder;
+ LLVMValueRef tmp;
+
+ tmp = LLVMBuildICmp(builder, LLVMIntEQ, get_wave_id_in_tg(ctx), ctx->ac.i32_0, "");
+ ac_build_ifcc(&ctx->ac, tmp, 5020);
+
+ tmp = LLVMBuildShl(builder, prim_cnt, LLVMConstInt(ctx->ac.i32, 12, false),"");
+ tmp = LLVMBuildOr(builder, tmp, vtx_cnt, "");
+ ac_build_sendmsg(&ctx->ac, AC_SENDMSG_GS_ALLOC_REQ, tmp);
+
+ ac_build_endif(&ctx->ac, 5020);
+}
+
+struct ngg_prim {
+ unsigned num_vertices;
+ LLVMValueRef isnull;
+ LLVMValueRef index[3];
+ LLVMValueRef edgeflag[3];
+};
+
+static void build_export_prim(struct radv_shader_context *ctx,
+ const struct ngg_prim *prim)
+{
+ LLVMBuilderRef builder = ctx->ac.builder;
+ struct ac_export_args args;
+ LLVMValueRef tmp;
+
+ tmp = LLVMBuildZExt(builder, prim->isnull, ctx->ac.i32, "");
+ args.out[0] = LLVMBuildShl(builder, tmp, LLVMConstInt(ctx->ac.i32, 31, false), "");
+
+ for (unsigned i = 0; i < prim->num_vertices; ++i) {
+ tmp = LLVMBuildShl(builder, prim->index[i],
+ LLVMConstInt(ctx->ac.i32, 10 * i, false), "");
+ args.out[0] = LLVMBuildOr(builder, args.out[0], tmp, "");
+ tmp = LLVMBuildZExt(builder, prim->edgeflag[i], ctx->ac.i32, "");
+ tmp = LLVMBuildShl(builder, tmp,
+ LLVMConstInt(ctx->ac.i32, 10 * i + 9, false), "");
+ args.out[0] = LLVMBuildOr(builder, args.out[0], tmp, "");
+ }
+
+ args.out[0] = LLVMBuildBitCast(builder, args.out[0], ctx->ac.f32, "");
+ args.out[1] = LLVMGetUndef(ctx->ac.f32);
+ args.out[2] = LLVMGetUndef(ctx->ac.f32);
+ args.out[3] = LLVMGetUndef(ctx->ac.f32);
+
+ args.target = V_008DFC_SQ_EXP_PRIM;
+ args.enabled_channels = 1;
+ args.done = true;
+ args.valid_mask = false;
+ args.compr = false;
+
+ ac_build_export(&ctx->ac, &args);
+}
+
+static void
+handle_ngg_outputs_post(struct radv_shader_context *ctx)
+{
+ LLVMBuilderRef builder = ctx->ac.builder;
+ struct ac_build_if_state if_state;
+ unsigned num_vertices = 3;
+ LLVMValueRef tmp;
+
+ assert(ctx->stage == MESA_SHADER_VERTEX && !ctx->is_gs_copy_shader);
+
+ LLVMValueRef prims_in_wave = ac_unpack_param(&ctx->ac, ctx->merged_wave_info, 8, 8);
+ LLVMValueRef vtx_in_wave = ac_unpack_param(&ctx->ac, ctx->merged_wave_info, 0, 8);
+ LLVMValueRef is_gs_thread = LLVMBuildICmp(builder, LLVMIntULT,
+ ac_get_thread_id(&ctx->ac), prims_in_wave, "");
+ LLVMValueRef is_es_thread = LLVMBuildICmp(builder, LLVMIntULT,
+ ac_get_thread_id(&ctx->ac), vtx_in_wave, "");
+ LLVMValueRef vtxindex[] = {
+ ac_unpack_param(&ctx->ac, ctx->gs_vtx_offset[0], 0, 16),
+ ac_unpack_param(&ctx->ac, ctx->gs_vtx_offset[0], 16, 16),
+ ac_unpack_param(&ctx->ac, ctx->gs_vtx_offset[2], 0, 16),
+ };
+
+ /* TODO: streamout */
+
+ /* TODO: VS primitive ID */
+ if (ctx->options->key.vs.out.export_prim_id)
+ assert(0);
+
+ /* TODO: primitive culling */
+
+ build_sendmsg_gs_alloc_req(ctx, ngg_get_vtx_cnt(ctx), ngg_get_prim_cnt(ctx));
+
+ /* TODO: streamout queries */
+ /* Export primitive data to the index buffer. Format is:
+ * - bits 0..8: index 0
+ * - bit 9: edge flag 0
+ * - bits 10..18: index 1
+ * - bit 19: edge flag 1
+ * - bits 20..28: index 2
+ * - bit 29: edge flag 2
+ * - bit 31: null primitive (skip)
+ *
+ * For the first version, we will always build up all three indices
+ * independent of the primitive type. The additional garbage data
+ * shouldn't hurt.
+ *
+ * TODO: culling depends on the primitive type, so can have some
+ * interaction here.
+ */
+ ac_nir_build_if(&if_state, ctx, is_gs_thread);
+ {
+ struct ngg_prim prim = {};
+
+ prim.num_vertices = num_vertices;
+ prim.isnull = ctx->ac.i1false;
+ memcpy(prim.index, vtxindex, sizeof(vtxindex[0]) * 3);
+
+ for (unsigned i = 0; i < num_vertices; ++i) {
+ tmp = LLVMBuildLShr(builder, ctx->abi.gs_invocation_id,
+ LLVMConstInt(ctx->ac.i32, 8 + i, false), "");
+ prim.edgeflag[i] = LLVMBuildTrunc(builder, tmp, ctx->ac.i1, "");
+ }
+
+ build_export_prim(ctx, &prim);
+ }
+ ac_nir_build_endif(&if_state);
+
+ /* Export per-vertex data (positions and parameters). */
+ ac_nir_build_if(&if_state, ctx, is_es_thread);
+ {
+ handle_vs_outputs_post(ctx, ctx->options->key.vs.out.export_prim_id,
+ ctx->options->key.vs.out.export_layer_id,
+ ctx->options->key.vs.out.export_clip_dists,
+ &ctx->shader_info->vs.outinfo);
+ }
+ ac_nir_build_endif(&if_state);
+}
+
static void
write_tess_factors(struct radv_shader_context *ctx)
{
LLVMConstInt(ctx->ac.i32, 4 * stride, false), "");
unsigned tf_offset = 0;
- if (ctx->options->chip_class <= VI) {
+ if (ctx->options->chip_class <= GFX8) {
ac_nir_build_if(&inner_if_ctx, ctx,
LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ,
rel_patch_id, ctx->ac.i32_0, ""));
ac_build_buffer_store_dword(&ctx->ac, buffer,
LLVMConstInt(ctx->ac.i32, 0x80000000, false),
1, ctx->ac.i32_0, tf_base,
- 0, 1, 0, true, false);
+ 0, ac_glc, false);
tf_offset += 4;
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,
- tf_offset, 1, 0, true, false);
+ tf_offset, ac_glc, false);
if (vec1)
ac_build_buffer_store_dword(&ctx->ac, buffer, vec1,
stride - 4, byteoffset, tf_base,
- 16 + tf_offset, 1, 0, true, false);
+ 16 + tf_offset, ac_glc, false);
//store to offchip for TES to read - only if TES reads them
if (ctx->options->key.tcs.tes_reads_tess_factors) {
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);
+ ctx->oc_lds, 0, ac_glc, 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,
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);
+ ctx->oc_lds, 0, ac_glc, false);
}
}
ac_nir_build_endif(&if_ctx);
switch (ctx->stage) {
case MESA_SHADER_VERTEX:
- if (ctx->options->key.vs.as_ls)
+ if (ctx->options->key.vs.out.as_ls)
handle_ls_outputs_post(ctx);
- else if (ctx->options->key.vs.as_es)
+ else if (ctx->options->key.vs.out.as_ngg)
+ break; /* handled outside of the shader body */
+ else if (ctx->options->key.vs.out.as_es)
handle_es_outputs_post(ctx, &ctx->shader_info->vs.es_info);
else
- handle_vs_outputs_post(ctx, ctx->options->key.vs.export_prim_id,
- ctx->options->key.vs.export_layer_id,
+ handle_vs_outputs_post(ctx, ctx->options->key.vs.out.export_prim_id,
+ ctx->options->key.vs.out.export_layer_id,
+ ctx->options->key.vs.out.export_clip_dists,
&ctx->shader_info->vs.outinfo);
break;
case MESA_SHADER_FRAGMENT:
handle_tcs_outputs_post(ctx);
break;
case MESA_SHADER_TESS_EVAL:
- if (ctx->options->key.tes.as_es)
+ if (ctx->options->key.tes.out.as_es)
handle_es_outputs_post(ctx, &ctx->shader_info->tes.es_info);
else
- handle_vs_outputs_post(ctx, ctx->options->key.tes.export_prim_id,
- ctx->options->key.tes.export_layer_id,
+ handle_vs_outputs_post(ctx, ctx->options->key.tes.out.export_prim_id,
+ ctx->options->key.tes.out.export_layer_id,
+ ctx->options->key.tes.out.export_clip_dists,
&ctx->shader_info->tes.outinfo);
break;
default:
case MESA_SHADER_GEOMETRY:
return;
case MESA_SHADER_VERTEX:
- if (ctx->options->key.vs.as_ls ||
- ctx->options->key.vs.as_es)
+ if (ctx->options->key.vs.out.as_ls ||
+ ctx->options->key.vs.out.as_es)
return;
outinfo = &ctx->shader_info->vs.outinfo;
break;
case MESA_SHADER_TESS_EVAL:
- if (ctx->options->key.vs.as_es)
+ if (ctx->options->key.vs.out.as_es)
return;
outinfo = &ctx->shader_info->tes.outinfo;
break;
static void
ac_setup_rings(struct radv_shader_context *ctx)
{
- if (ctx->options->chip_class <= VI &&
+ if (ctx->options->chip_class <= GFX8 &&
(ctx->stage == MESA_SHADER_GEOMETRY ||
- ctx->options->key.vs.as_es || ctx->options->key.tes.as_es)) {
+ ctx->options->key.vs.out.as_es || ctx->options->key.tes.out.as_es)) {
unsigned ring = ctx->stage == MESA_SHADER_GEOMETRY ? RING_ESGS_GS
: RING_ESGS_VS;
LLVMValueRef offset = LLVMConstInt(ctx->ac.i32, ring, false);
}
if (ctx->is_gs_copy_shader) {
- ctx->gsvs_ring = ac_build_load_to_sgpr(&ctx->ac, ctx->ring_offsets, LLVMConstInt(ctx->ac.i32, RING_GSVS_VS, false));
+ ctx->gsvs_ring[0] =
+ ac_build_load_to_sgpr(&ctx->ac, ctx->ring_offsets,
+ LLVMConstInt(ctx->ac.i32,
+ RING_GSVS_VS, false));
}
+
if (ctx->stage == MESA_SHADER_GEOMETRY) {
- LLVMValueRef tmp;
- uint32_t num_entries = 64;
- LLVMValueRef gsvs_ring_stride = LLVMConstInt(ctx->ac.i32, ctx->max_gsvs_emit_size, false);
- LLVMValueRef gsvs_ring_desc = LLVMConstInt(ctx->ac.i32, ctx->max_gsvs_emit_size << 16, false);
- ctx->gsvs_ring = ac_build_load_to_sgpr(&ctx->ac, ctx->ring_offsets, LLVMConstInt(ctx->ac.i32, RING_GSVS_GS, false));
+ /* The conceptual layout of the GSVS ring is
+ * v0c0 .. vLv0 v0c1 .. vLc1 ..
+ * but the real memory layout is swizzled across
+ * threads:
+ * t0v0c0 .. t15v0c0 t0v1c0 .. t15v1c0 ... t15vLcL
+ * t16v0c0 ..
+ * Override the buffer descriptor accordingly.
+ */
+ LLVMTypeRef v2i64 = LLVMVectorType(ctx->ac.i64, 2);
+ uint64_t stream_offset = 0;
+ unsigned num_records = 64;
+ LLVMValueRef base_ring;
+
+ base_ring =
+ ac_build_load_to_sgpr(&ctx->ac, ctx->ring_offsets,
+ LLVMConstInt(ctx->ac.i32,
+ RING_GSVS_GS, false));
- ctx->gsvs_ring = LLVMBuildBitCast(ctx->ac.builder, ctx->gsvs_ring, ctx->ac.v4i32, "");
+ for (unsigned stream = 0; stream < 4; stream++) {
+ unsigned num_components, stride;
+ LLVMValueRef ring, tmp;
- tmp = LLVMConstInt(ctx->ac.i32, num_entries, false);
- if (ctx->options->chip_class >= VI)
- tmp = LLVMBuildMul(ctx->ac.builder, gsvs_ring_stride, tmp, "");
- ctx->gsvs_ring = LLVMBuildInsertElement(ctx->ac.builder, ctx->gsvs_ring, tmp, LLVMConstInt(ctx->ac.i32, 2, false), "");
- tmp = LLVMBuildExtractElement(ctx->ac.builder, ctx->gsvs_ring, ctx->ac.i32_1, "");
- tmp = LLVMBuildOr(ctx->ac.builder, tmp, gsvs_ring_desc, "");
- ctx->gsvs_ring = LLVMBuildInsertElement(ctx->ac.builder, ctx->gsvs_ring, tmp, ctx->ac.i32_1, "");
+ num_components =
+ ctx->shader_info->info.gs.num_stream_output_components[stream];
+
+ if (!num_components)
+ continue;
+
+ stride = 4 * num_components * ctx->gs_max_out_vertices;
+
+ /* Limit on the stride field for <= GFX7. */
+ assert(stride < (1 << 14));
+
+ ring = LLVMBuildBitCast(ctx->ac.builder,
+ base_ring, v2i64, "");
+ tmp = LLVMBuildExtractElement(ctx->ac.builder,
+ ring, ctx->ac.i32_0, "");
+ tmp = LLVMBuildAdd(ctx->ac.builder, tmp,
+ LLVMConstInt(ctx->ac.i64,
+ stream_offset, 0), "");
+ ring = LLVMBuildInsertElement(ctx->ac.builder,
+ ring, tmp, ctx->ac.i32_0, "");
+
+ stream_offset += stride * 64;
+
+ ring = LLVMBuildBitCast(ctx->ac.builder, ring,
+ ctx->ac.v4i32, "");
+
+ tmp = LLVMBuildExtractElement(ctx->ac.builder, ring,
+ ctx->ac.i32_1, "");
+ tmp = LLVMBuildOr(ctx->ac.builder, tmp,
+ LLVMConstInt(ctx->ac.i32,
+ S_008F04_STRIDE(stride), false), "");
+ ring = LLVMBuildInsertElement(ctx->ac.builder, ring, tmp,
+ ctx->ac.i32_1, "");
+
+ ring = LLVMBuildInsertElement(ctx->ac.builder, ring,
+ LLVMConstInt(ctx->ac.i32,
+ num_records, false),
+ LLVMConstInt(ctx->ac.i32, 2, false), "");
+
+ ctx->gsvs_ring[stream] = ring;
+ }
}
if (ctx->stage == MESA_SHADER_TESS_CTRL ||
}
}
-static unsigned
-ac_nir_get_max_workgroup_size(enum chip_class chip_class,
- const struct nir_shader *nir)
+unsigned
+radv_nir_get_max_workgroup_size(enum chip_class chip_class,
+ const struct nir_shader *nir)
{
switch (nir->info.stage) {
case MESA_SHADER_TESS_CTRL:
- return chip_class >= CIK ? 128 : 64;
+ return chip_class >= GFX7 ? 128 : 64;
case MESA_SHADER_GEOMETRY:
return chip_class >= GFX9 ? 128 : 64;
case MESA_SHADER_COMPUTE:
memset(shader_info, 0, sizeof(*shader_info));
+ radv_nir_shader_info_init(&shader_info->info);
+
for(int i = 0; i < shader_count; ++i)
radv_nir_shader_info_pass(shaders[i], options, &shader_info->info);
ctx.max_workgroup_size = 0;
for (int i = 0; i < shader_count; ++i) {
ctx.max_workgroup_size = MAX2(ctx.max_workgroup_size,
- ac_nir_get_max_workgroup_size(ctx.options->chip_class,
+ radv_nir_get_max_workgroup_size(ctx.options->chip_class,
shaders[i]));
}
+ if (ctx.ac.chip_class >= GFX10) {
+ if (shaders[0]->info.stage == MESA_SHADER_VERTEX &&
+ options->key.vs.out.as_ngg) {
+ ctx.max_workgroup_size = 128;
+ }
+ }
+
create_function(&ctx, shaders[shader_count - 1]->info.stage, shader_count >= 2,
shader_count >= 2 ? shaders[shader_count - 2]->info.stage : MESA_SHADER_VERTEX);
ctx.abi.load_sampler_desc = radv_get_sampler_desc;
ctx.abi.load_resource = radv_load_resource;
ctx.abi.clamp_shadow_reference = false;
- ctx.abi.gfx9_stride_size_workaround = ctx.ac.chip_class == GFX9;
+ ctx.abi.gfx9_stride_size_workaround = ctx.ac.chip_class == GFX9 && HAVE_LLVM < 0x800;
+
+ /* Because the new raw/struct atomic intrinsics are buggy with LLVM 8,
+ * we fallback to the old intrinsics for atomic buffer image operations
+ * and thus we need to apply the indexing workaround...
+ */
+ ctx.abi.gfx9_stride_size_workaround_for_atomic = ctx.ac.chip_class == GFX9 && HAVE_LLVM < 0x900;
- if (shader_count >= 2)
+ bool is_ngg = is_pre_gs_stage(shaders[0]->info.stage) && ctx.options->key.vs.out.as_ngg;
+ if (shader_count >= 2 || is_ngg)
ac_init_exec_full_mask(&ctx.ac);
- if (ctx.ac.chip_class == GFX9 &&
+ if ((ctx.ac.family == CHIP_VEGA10 ||
+ ctx.ac.family == CHIP_RAVEN) &&
shaders[shader_count - 1]->info.stage == MESA_SHADER_TESS_CTRL)
ac_nir_fixup_ls_hs_input_vgprs(&ctx);
ctx.output_mask = 0;
if (shaders[i]->info.stage == MESA_SHADER_GEOMETRY) {
- ctx.gs_next_vertex = ac_build_alloca(&ctx.ac, ctx.ac.i32, "gs_next_vertex");
+ for (int i = 0; i < 4; i++) {
+ ctx.gs_next_vertex[i] =
+ ac_build_alloca(&ctx.ac, ctx.ac.i32, "");
+ }
ctx.gs_max_out_vertices = shaders[i]->info.gs.vertices_out;
ctx.abi.load_inputs = load_gs_input;
ctx.abi.emit_primitive = visit_end_primitive;
ctx.tcs_vertices_per_patch = shaders[i]->info.tess.tcs_vertices_out;
ctx.tcs_num_patches = ctx.options->key.tes.num_patches;
} else if (shaders[i]->info.stage == MESA_SHADER_VERTEX) {
- if (shader_info->info.vs.needs_instance_id) {
- if (ctx.options->key.vs.as_ls) {
- ctx.shader_info->vs.vgpr_comp_cnt =
- MAX2(2, ctx.shader_info->vs.vgpr_comp_cnt);
- } else {
- ctx.shader_info->vs.vgpr_comp_cnt =
- MAX2(1, ctx.shader_info->vs.vgpr_comp_cnt);
- }
- }
ctx.abi.load_base_vertex = radv_load_base_vertex;
} else if (shaders[i]->info.stage == MESA_SHADER_FRAGMENT) {
shader_info->fs.can_discard = shaders[i]->info.fs.uses_discard;
ac_setup_rings(&ctx);
LLVMBasicBlockRef merge_block;
- if (shader_count >= 2) {
+ if (shader_count >= 2 || is_ngg) {
LLVMValueRef fn = LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx.ac.builder));
LLVMBasicBlockRef then_block = LLVMAppendBasicBlockInContext(ctx.ac.context, fn, "");
merge_block = LLVMAppendBasicBlockInContext(ctx.ac.context, fn, "");
ac_nir_translate(&ctx.ac, &ctx.abi, shaders[i]);
- if (shader_count >= 2) {
+ if (shader_count >= 2 || is_ngg) {
LLVMBuildBr(ctx.ac.builder, merge_block);
LLVMPositionBuilderAtEnd(ctx.ac.builder, merge_block);
}
+ /* This needs to be outside the if wrapping the shader body, as sometimes
+ * the HW generates waves with 0 es/vs threads. */
+ if (is_pre_gs_stage(shaders[i]->info.stage) &&
+ ctx.options->key.vs.out.as_ngg &&
+ i == shader_count - 1) {
+ handle_ngg_outputs_post(&ctx);
+ }
+
if (shaders[i]->info.stage == MESA_SHADER_GEOMETRY) {
shader_info->gs.gsvs_vertex_size = ctx.gsvs_vertex_size;
shader_info->gs.max_gsvs_emit_size = ctx.max_gsvs_emit_size;
LLVMDisposeMessage(description);
}
-static unsigned ac_llvm_compile(LLVMModuleRef M,
- struct ac_shader_binary *binary,
- struct ac_llvm_compiler *ac_llvm)
+static unsigned radv_llvm_compile(LLVMModuleRef M,
+ char **pelf_buffer, size_t *pelf_size,
+ struct ac_llvm_compiler *ac_llvm)
{
unsigned retval = 0;
LLVMContextRef llvm_ctx;
&retval);
/* Compile IR*/
- if (!radv_compile_to_binary(ac_llvm, M, binary))
+ if (!radv_compile_to_elf(ac_llvm, M, pelf_buffer, pelf_size))
retval = 1;
return retval;
}
static void ac_compile_llvm_module(struct ac_llvm_compiler *ac_llvm,
LLVMModuleRef llvm_module,
- struct ac_shader_binary *binary,
- struct ac_shader_config *config,
+ struct radv_shader_binary **rbinary,
struct radv_shader_variant_info *shader_info,
gl_shader_stage stage,
const struct radv_nir_compiler_options *options)
{
+ char *elf_buffer = NULL;
+ size_t elf_size = 0;
+ char *llvm_ir_string = NULL;
if (options->dump_shader)
ac_dump_module(llvm_module);
- memset(binary, 0, sizeof(*binary));
-
if (options->record_llvm_ir) {
char *llvm_ir = LLVMPrintModuleToString(llvm_module);
- binary->llvm_ir_string = strdup(llvm_ir);
+ llvm_ir_string = strdup(llvm_ir);
LLVMDisposeMessage(llvm_ir);
}
- int v = ac_llvm_compile(llvm_module, binary, ac_llvm);
+ int v = radv_llvm_compile(llvm_module, &elf_buffer, &elf_size, ac_llvm);
if (v) {
fprintf(stderr, "compile failed\n");
}
- if (options->dump_shader)
- fprintf(stderr, "disasm:\n%s\n", binary->disasm_string);
-
- ac_shader_binary_read_config(binary, config, 0, options->supports_spill);
-
LLVMContextRef ctx = LLVMGetModuleContext(llvm_module);
LLVMDisposeModule(llvm_module);
LLVMContextDispose(ctx);
- if (stage == MESA_SHADER_FRAGMENT) {
- shader_info->num_input_vgprs = 0;
- if (G_0286CC_PERSP_SAMPLE_ENA(config->spi_ps_input_addr))
- shader_info->num_input_vgprs += 2;
- if (G_0286CC_PERSP_CENTER_ENA(config->spi_ps_input_addr))
- shader_info->num_input_vgprs += 2;
- if (G_0286CC_PERSP_CENTROID_ENA(config->spi_ps_input_addr))
- shader_info->num_input_vgprs += 2;
- if (G_0286CC_PERSP_PULL_MODEL_ENA(config->spi_ps_input_addr))
- shader_info->num_input_vgprs += 3;
- if (G_0286CC_LINEAR_SAMPLE_ENA(config->spi_ps_input_addr))
- shader_info->num_input_vgprs += 2;
- if (G_0286CC_LINEAR_CENTER_ENA(config->spi_ps_input_addr))
- shader_info->num_input_vgprs += 2;
- if (G_0286CC_LINEAR_CENTROID_ENA(config->spi_ps_input_addr))
- shader_info->num_input_vgprs += 2;
- if (G_0286CC_LINE_STIPPLE_TEX_ENA(config->spi_ps_input_addr))
- shader_info->num_input_vgprs += 1;
- if (G_0286CC_POS_X_FLOAT_ENA(config->spi_ps_input_addr))
- shader_info->num_input_vgprs += 1;
- if (G_0286CC_POS_Y_FLOAT_ENA(config->spi_ps_input_addr))
- shader_info->num_input_vgprs += 1;
- if (G_0286CC_POS_Z_FLOAT_ENA(config->spi_ps_input_addr))
- shader_info->num_input_vgprs += 1;
- if (G_0286CC_POS_W_FLOAT_ENA(config->spi_ps_input_addr))
- shader_info->num_input_vgprs += 1;
- if (G_0286CC_FRONT_FACE_ENA(config->spi_ps_input_addr))
- shader_info->num_input_vgprs += 1;
- if (G_0286CC_ANCILLARY_ENA(config->spi_ps_input_addr))
- shader_info->num_input_vgprs += 1;
- if (G_0286CC_SAMPLE_COVERAGE_ENA(config->spi_ps_input_addr))
- shader_info->num_input_vgprs += 1;
- if (G_0286CC_POS_FIXED_PT_ENA(config->spi_ps_input_addr))
- shader_info->num_input_vgprs += 1;
- }
- config->num_vgprs = MAX2(config->num_vgprs, shader_info->num_input_vgprs);
-
- /* +3 for scratch wave offset and VCC */
- config->num_sgprs = MAX2(config->num_sgprs,
- shader_info->num_input_sgprs + 3);
-
- /* Enable 64-bit and 16-bit denormals, because there is no performance
- * cost.
- *
- * If denormals are enabled, all floating-point output modifiers are
- * ignored.
- *
- * Don't enable denormals for 32-bit floats, because:
- * - Floating-point output modifiers would be ignored by the hw.
- * - Some opcodes don't support denormals, such as v_mad_f32. We would
- * have to stop using those.
- * - SI & CI would be very slow.
- */
- config->float_mode |= V_00B028_FP_64_DENORMS;
+ size_t llvm_ir_size = llvm_ir_string ? strlen(llvm_ir_string) : 0;
+ size_t alloc_size = sizeof(struct radv_shader_binary_rtld) + elf_size + llvm_ir_size + 1;
+ struct radv_shader_binary_rtld *rbin = calloc(1, alloc_size);
+ memcpy(rbin->data, elf_buffer, elf_size);
+ if (llvm_ir_string)
+ memcpy(rbin->data + elf_size, llvm_ir_string, llvm_ir_size + 1);
+
+ rbin->base.type = RADV_BINARY_TYPE_RTLD;
+ rbin->base.stage = stage;
+ rbin->base.total_size = alloc_size;
+ rbin->elf_size = elf_size;
+ rbin->llvm_ir_size = llvm_ir_size;
+ *rbinary = &rbin->base;
+
+ free(llvm_ir_string);
+ free(elf_buffer);
}
static void
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;
+ shader_info->tes.as_es = options->key.tes.out.as_es;
+ shader_info->tes.export_prim_id = options->key.tes.out.export_prim_id;
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 2, handled elsewhere */
- if (options->key.vs.as_ls)
- shader_info->vs.vgpr_comp_cnt = MAX2(1, shader_info->vs.vgpr_comp_cnt);
+ shader_info->vs.as_es = options->key.vs.out.as_es;
+ shader_info->vs.as_ls = options->key.vs.out.as_ls;
+ shader_info->vs.export_prim_id = options->key.vs.out.export_prim_id;
+ shader_info->is_ngg = options->key.vs.out.as_ngg;
break;
default:
break;
void
radv_compile_nir_shader(struct ac_llvm_compiler *ac_llvm,
- struct ac_shader_binary *binary,
- struct ac_shader_config *config,
+ struct radv_shader_binary **rbinary,
struct radv_shader_variant_info *shader_info,
struct nir_shader *const *nir,
int nir_count,
llvm_module = ac_translate_nir_to_llvm(ac_llvm, nir, nir_count, shader_info,
options);
- ac_compile_llvm_module(ac_llvm, llvm_module, binary, config, shader_info,
- nir[0]->info.stage, options);
+ ac_compile_llvm_module(ac_llvm, llvm_module, rbinary, shader_info,
+ nir[nir_count - 1]->info.stage, options);
for (int i = 0; i < nir_count; ++i)
ac_fill_shader_info(shader_info, nir[i], options);
LLVMValueRef vtx_offset =
LLVMBuildMul(ctx->ac.builder, ctx->abi.vertex_id,
LLVMConstInt(ctx->ac.i32, 4, false), "");
- unsigned offset = 0;
+ LLVMValueRef stream_id;
- for (unsigned i = 0; i < AC_LLVM_MAX_OUTPUTS; ++i) {
- unsigned output_usage_mask =
- ctx->shader_info->info.gs.output_usage_mask[i];
- int length = 4;
+ /* Fetch the vertex stream ID. */
+ if (ctx->shader_info->info.so.num_outputs) {
+ stream_id =
+ ac_unpack_param(&ctx->ac, ctx->streamout_config, 24, 2);
+ } else {
+ stream_id = ctx->ac.i32_0;
+ }
- if (!(ctx->output_mask & (1ull << i)))
+ LLVMBasicBlockRef end_bb;
+ LLVMValueRef switch_inst;
+
+ end_bb = LLVMAppendBasicBlockInContext(ctx->ac.context,
+ ctx->main_function, "end");
+ switch_inst = LLVMBuildSwitch(ctx->ac.builder, stream_id, end_bb, 4);
+
+ for (unsigned stream = 0; stream < 4; stream++) {
+ unsigned num_components =
+ ctx->shader_info->info.gs.num_stream_output_components[stream];
+ LLVMBasicBlockRef bb;
+ unsigned offset;
+
+ if (!num_components)
continue;
- if (i == VARYING_SLOT_CLIP_DIST0) {
- /* unpack clip and cull from a single set of slots */
- length = util_last_bit(output_usage_mask);
- }
+ if (stream > 0 && !ctx->shader_info->info.so.num_outputs)
+ continue;
- for (unsigned j = 0; j < length; j++) {
- LLVMValueRef value, soffset;
+ bb = LLVMInsertBasicBlockInContext(ctx->ac.context, end_bb, "out");
+ LLVMAddCase(switch_inst, LLVMConstInt(ctx->ac.i32, stream, 0), bb);
+ LLVMPositionBuilderAtEnd(ctx->ac.builder, bb);
- if (!(output_usage_mask & (1 << j)))
+ offset = 0;
+ for (unsigned i = 0; i < AC_LLVM_MAX_OUTPUTS; ++i) {
+ unsigned output_usage_mask =
+ ctx->shader_info->info.gs.output_usage_mask[i];
+ unsigned output_stream =
+ ctx->shader_info->info.gs.output_streams[i];
+ int length = util_last_bit(output_usage_mask);
+
+ if (!(ctx->output_mask & (1ull << i)) ||
+ output_stream != stream)
continue;
- soffset = LLVMConstInt(ctx->ac.i32,
- offset *
- ctx->gs_max_out_vertices * 16 * 4, false);
+ for (unsigned j = 0; j < length; j++) {
+ LLVMValueRef value, soffset;
- offset++;
+ if (!(output_usage_mask & (1 << j)))
+ continue;
+
+ soffset = LLVMConstInt(ctx->ac.i32,
+ offset *
+ ctx->gs_max_out_vertices * 16 * 4, false);
- value = ac_build_buffer_load(&ctx->ac, ctx->gsvs_ring,
- 1, ctx->ac.i32_0,
- vtx_offset, soffset,
- 0, 1, 1, true, false);
+ offset++;
+
+ value = ac_build_buffer_load(&ctx->ac,
+ ctx->gsvs_ring[0],
+ 1, ctx->ac.i32_0,
+ vtx_offset, soffset,
+ 0, ac_glc | ac_slc, true, false);
+
+ LLVMTypeRef type = LLVMGetAllocatedType(ctx->abi.outputs[ac_llvm_reg_index_soa(i, j)]);
+ if (ac_get_type_size(type) == 2) {
+ value = LLVMBuildBitCast(ctx->ac.builder, value, ctx->ac.i32, "");
+ value = LLVMBuildTrunc(ctx->ac.builder, value, ctx->ac.i16, "");
+ }
- LLVMTypeRef type = LLVMGetAllocatedType(ctx->abi.outputs[ac_llvm_reg_index_soa(i, j)]);
- if (ac_get_type_size(type) == 2) {
- value = LLVMBuildBitCast(ctx->ac.builder, value, ctx->ac.i32, "");
- value = LLVMBuildTrunc(ctx->ac.builder, value, ctx->ac.i16, "");
+ LLVMBuildStore(ctx->ac.builder,
+ ac_to_float(&ctx->ac, value), ctx->abi.outputs[ac_llvm_reg_index_soa(i, j)]);
}
+ }
+
+ if (ctx->shader_info->info.so.num_outputs)
+ radv_emit_streamout(ctx, stream);
- LLVMBuildStore(ctx->ac.builder,
- ac_to_float(&ctx->ac, value), ctx->abi.outputs[ac_llvm_reg_index_soa(i, j)]);
+ if (stream == 0) {
+ handle_vs_outputs_post(ctx, false, false, true,
+ &ctx->shader_info->vs.outinfo);
}
+
+ LLVMBuildBr(ctx->ac.builder, end_bb);
}
- handle_vs_outputs_post(ctx, false, false, &ctx->shader_info->vs.outinfo);
+
+ LLVMPositionBuilderAtEnd(ctx->ac.builder, end_bb);
}
void
radv_compile_gs_copy_shader(struct ac_llvm_compiler *ac_llvm,
struct nir_shader *geom_shader,
- struct ac_shader_binary *binary,
- struct ac_shader_config *config,
+ struct radv_shader_binary **rbinary,
struct radv_shader_variant_info *shader_info,
const struct radv_nir_compiler_options *options)
{
ac_llvm_finalize_module(&ctx, ac_llvm->passmgr, options);
- ac_compile_llvm_module(ac_llvm, ctx.ac.module, binary, config, shader_info,
+ ac_compile_llvm_module(ac_llvm, ctx.ac.module, rbinary, shader_info,
MESA_SHADER_VERTEX, options);
+ (*rbinary)->is_gs_copy_shader = true;
+
}