| 1 |
/* |
| 2 |
* Copyright © 2025 Imagination Technologies Ltd. |
| 3 |
* |
| 4 |
* SPDX-License-Identifier: MIT |
| 5 |
*/ |
| 6 |
|
| 7 |
#include "pvr_mrt.h" |
| 8 |
|
| 9 |
#include "vk_log.h" |
| 10 |
|
| 11 |
#include "pvr_cmd_buffer.h" |
| 12 |
#include "pvr_csb.h" |
| 13 |
#include "pvr_device.h" |
| 14 |
#include "pvr_formats.h" |
| 15 |
#include "pvr_image.h" |
| 16 |
#include "pvr_mrt.h" |
| 17 |
#include "pvr_pass.h" |
| 18 |
#include "pvr_physical_device.h" |
| 19 |
|
| 20 |
#include "hwdef/rogue_hw_utils.h" |
| 21 |
#include "util/macros.h" |
| 22 |
#include "vulkan/vulkan_core.h" |
| 23 |
#include "vulkan/runtime/vk_graphics_state.h" |
| 24 |
#include "vulkan/runtime/vk_image.h" |
| 25 |
|
| 26 |
/* Which parts of the output registers/a tile buffer are currently allocated. */ |
| 27 |
struct pvr_mrt_alloc_mask { |
| 28 |
/* Bit array. A bit is set if the corresponding dword is allocated. */ |
| 29 |
BITSET_DECLARE(allocs, 8U); |
| 30 |
}; |
| 31 |
|
| 32 |
struct pvr_mrt_alloc_ctx { |
| 33 |
/* Which pixel output registers are allocated. */ |
| 34 |
struct pvr_mrt_alloc_mask output_reg; |
| 35 |
|
| 36 |
/* Range of allocated output registers. */ |
| 37 |
uint32_t output_regs_count; |
| 38 |
|
| 39 |
/* Number of tile buffers allocated. */ |
| 40 |
uint32_t tile_buffers_count; |
| 41 |
|
| 42 |
/* Which parts of each tile buffer are allocated. Length is |
| 43 |
* tile_buffers_count. |
| 44 |
*/ |
| 45 |
struct pvr_mrt_alloc_mask tile_buffers[PVR_MAX_TILE_BUFFER_COUNT]; |
| 46 |
}; |
| 47 |
|
| 48 |
static uint32_t pvr_get_accum_format_bitsize(VkFormat vk_format) |
| 49 |
{ |
| 50 |
return pvr_get_pbe_accum_format_size_in_bytes(vk_format) * 8; |
| 51 |
} |
| 52 |
|
| 53 |
/** |
| 54 |
* Check if there is space in a buffer for storing a render target of a |
| 55 |
* specified size. |
| 56 |
*/ |
| 57 |
static int32_t pvr_mrt_alloc_from_buffer(const struct pvr_device_info *dev_info, |
| 58 |
struct pvr_mrt_alloc_mask *buffer, |
| 59 |
uint32_t pixel_size) |
| 60 |
{ |
| 61 |
const uint32_t max_out_regs = rogue_get_max_output_regs_per_pixel(dev_info); |
| 62 |
uint32_t alignment = 1U; |
| 63 |
|
| 64 |
if (PVR_HAS_FEATURE(dev_info, pbe2_in_xe)) { |
| 65 |
/* For a 64-bit/128-bit source format: the start offset must be even. */ |
| 66 |
if (pixel_size == 2U || pixel_size == 4U) |
| 67 |
alignment = 2U; |
| 68 |
} |
| 69 |
|
| 70 |
assert(pixel_size <= max_out_regs); |
| 71 |
|
| 72 |
for (uint32_t i = 0U; i <= (max_out_regs - pixel_size); i += alignment) { |
| 73 |
if (!BITSET_TEST_COUNT(buffer->allocs, i, pixel_size)) { |
| 74 |
BITSET_SET_COUNT(buffer->allocs, i, pixel_size); |
| 75 |
return i; |
| 76 |
} |
| 77 |
} |
| 78 |
|
| 79 |
return -1; |
| 80 |
} |
| 81 |
|
| 82 |
void pvr_arch_init_mrt_desc(VkFormat format, struct usc_mrt_desc *desc) |
| 83 |
{ |
| 84 |
uint32_t pixel_size_in_chunks; |
| 85 |
uint32_t pixel_size_in_bits; |
| 86 |
|
| 87 |
/* TODO: Add support for packing multiple attachments into the same |
| 88 |
* register |
| 89 |
*/ |
| 90 |
const uint32_t part_bits = 0; |
| 91 |
if (vk_format_is_color(format) && |
| 92 |
pvr_arch_get_pbe_accum_format(format) == PVR_PBE_ACCUM_FORMAT_INVALID) { |
| 93 |
/* The VkFormat is not supported as a color attachment so `0`. |
| 94 |
* vulkan doesn't seem to restrict vkCreateRenderPass() to supported |
| 95 |
* formats only. |
| 96 |
*/ |
| 97 |
pixel_size_in_bits = 0; |
| 98 |
} else { |
| 99 |
/* TODO: handle IMG_PIXFMT_A8_UNORM |
| 100 |
* For alpha only formats alpha is still placed in channel 3, so channels |
| 101 |
* 0-2 need to be allocated but are left unused |
| 102 |
*/ |
| 103 |
pixel_size_in_bits = pvr_get_accum_format_bitsize(format); |
| 104 |
} |
| 105 |
|
| 106 |
desc->intermediate_size = DIV_ROUND_UP(pixel_size_in_bits, CHAR_BIT); |
| 107 |
|
| 108 |
pixel_size_in_chunks = DIV_ROUND_UP(pixel_size_in_bits, 32U); |
| 109 |
for (uint32_t j = 0U; j < pixel_size_in_chunks; j++) |
| 110 |
desc->valid_mask[j] = ~0; |
| 111 |
|
| 112 |
if (part_bits > 0U) |
| 113 |
desc->valid_mask[pixel_size_in_chunks] = BITFIELD_MASK(part_bits); |
| 114 |
} |
| 115 |
|
| 116 |
static VkResult pvr_alloc_mrt(const struct pvr_device_info *dev_info, |
| 117 |
struct pvr_mrt_alloc_ctx *alloc, |
| 118 |
struct usc_mrt_setup *setup, |
| 119 |
unsigned rt, |
| 120 |
VkFormat format) |
| 121 |
{ |
| 122 |
struct usc_mrt_resource *resource = &setup->mrt_resources[rt]; |
| 123 |
|
| 124 |
const uint32_t pixel_size = |
| 125 |
DIV_ROUND_UP(pvr_get_accum_format_bitsize(format), 32U); |
| 126 |
|
| 127 |
const int32_t output_reg = |
| 128 |
pvr_mrt_alloc_from_buffer(dev_info, &alloc->output_reg, pixel_size); |
| 129 |
|
| 130 |
if (output_reg != -1) { |
| 131 |
resource->type = USC_MRT_RESOURCE_TYPE_OUTPUT_REG; |
| 132 |
resource->reg.output_reg = output_reg; |
| 133 |
resource->reg.offset = 0; |
| 134 |
|
| 135 |
alloc->output_regs_count = |
| 136 |
MAX2(alloc->output_regs_count, resource->reg.output_reg + pixel_size); |
| 137 |
} else { |
| 138 |
resource->type = USC_MRT_RESOURCE_TYPE_MEMORY; |
| 139 |
|
| 140 |
unsigned tib = 0; |
| 141 |
for (; tib < alloc->tile_buffers_count; tib++) { |
| 142 |
struct pvr_mrt_alloc_mask *tib_alloc = &alloc->tile_buffers[tib]; |
| 143 |
|
| 144 |
const int32_t tile_buffer_offset = |
| 145 |
pvr_mrt_alloc_from_buffer(dev_info, tib_alloc, pixel_size); |
| 146 |
|
| 147 |
if (tile_buffer_offset != -1) { |
| 148 |
resource->mem.tile_buffer = tib; |
| 149 |
resource->mem.offset_dw = tile_buffer_offset; |
| 150 |
break; |
| 151 |
} |
| 152 |
} |
| 153 |
|
| 154 |
if (tib == alloc->tile_buffers_count) { |
| 155 |
if (alloc->tile_buffers_count == PVR_MAX_TILE_BUFFER_COUNT) |
| 156 |
return vk_error(NULL, VK_ERROR_TOO_MANY_OBJECTS); |
| 157 |
|
| 158 |
resource->mem.tile_buffer = alloc->tile_buffers_count; |
| 159 |
resource->mem.offset_dw = 0; |
| 160 |
} |
| 161 |
|
| 162 |
/* If needed a new tile buffer than those that were allocated, then wipe |
| 163 |
* it bump the global count. |
| 164 |
*/ |
| 165 |
if (resource->mem.tile_buffer >= alloc->tile_buffers_count) { |
| 166 |
memset( |
| 167 |
&alloc->tile_buffers[alloc->tile_buffers_count], |
| 168 |
0U, |
| 169 |
sizeof(alloc->tile_buffers[0U]) * |
| 170 |
(resource->mem.tile_buffer + 1U - alloc->tile_buffers_count)); |
| 171 |
alloc->tile_buffers_count = resource->mem.tile_buffer + 1U; |
| 172 |
|
| 173 |
/* Need to add the resource to the new tile buffer */ |
| 174 |
struct pvr_mrt_alloc_mask *tib_alloc = |
| 175 |
&alloc->tile_buffers[resource->mem.tile_buffer]; |
| 176 |
|
| 177 |
ASSERTED const int32_t tile_buffer_offset = |
| 178 |
pvr_mrt_alloc_from_buffer(dev_info, |
| 179 |
tib_alloc, |
| 180 |
pixel_size); |
| 181 |
assert(tile_buffer_offset == 0); |
| 182 |
} |
| 183 |
|
| 184 |
/* The hardware makes the bit depth of the on-chip storage and memory |
| 185 |
* storage the same so make sure the memory storage is large enough to |
| 186 |
* accommodate the largest render target. |
| 187 |
*/ |
| 188 |
alloc->output_regs_count = |
| 189 |
MAX2(alloc->output_regs_count, resource->mem.offset_dw + pixel_size); |
| 190 |
} |
| 191 |
|
| 192 |
pvr_arch_init_mrt_desc(format, &resource->mrt_desc); |
| 193 |
resource->intermediate_size = resource->mrt_desc.intermediate_size; |
| 194 |
|
| 195 |
setup->num_render_targets++; |
| 196 |
|
| 197 |
return VK_SUCCESS; |
| 198 |
} |
| 199 |
|
| 200 |
VkResult |
| 201 |
pvr_arch_mrt_setup_partial_init(struct pvr_device *const device, |
| 202 |
struct usc_mrt_setup *mrt_setup, |
| 203 |
uint32_t num_renger_targets, |
| 204 |
uint32_t num_output_regs, |
| 205 |
uint32_t num_tile_buffers) |
| 206 |
{ |
| 207 |
struct usc_mrt_resource *mrt_resources = NULL; |
| 208 |
|
| 209 |
if (num_renger_targets) { |
| 210 |
const uint32_t size = |
| 211 |
num_renger_targets * sizeof(*mrt_setup->mrt_resources); |
| 212 |
|
| 213 |
mrt_resources = vk_zalloc(&device->vk.alloc, |
| 214 |
size, |
| 215 |
8, |
| 216 |
VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); |
| 217 |
if (!mrt_resources) |
| 218 |
return VK_ERROR_OUT_OF_HOST_MEMORY; |
| 219 |
} |
| 220 |
|
| 221 |
*mrt_setup = (struct usc_mrt_setup){ |
| 222 |
.num_render_targets = num_renger_targets, |
| 223 |
.num_output_regs = num_output_regs, |
| 224 |
.num_tile_buffers = num_tile_buffers, |
| 225 |
.mrt_resources = mrt_resources, |
| 226 |
}; |
| 227 |
|
| 228 |
return VK_SUCCESS; |
| 229 |
} |
| 230 |
|
| 231 |
VkResult |
| 232 |
pvr_arch_init_usc_mrt_setup(struct pvr_device *device, |
| 233 |
uint32_t attachment_count, |
| 234 |
const VkFormat attachment_formats[attachment_count], |
| 235 |
struct usc_mrt_setup *setup) |
| 236 |
{ |
| 237 |
const struct pvr_device_info *dev_info = &device->pdevice->dev_info; |
| 238 |
struct pvr_mrt_alloc_ctx alloc = { 0 }; |
| 239 |
VkResult result; |
| 240 |
|
| 241 |
memset(setup, 0, sizeof(*setup)); |
| 242 |
|
| 243 |
if (!attachment_count) |
| 244 |
goto early_exit; |
| 245 |
|
| 246 |
setup->mrt_resources = |
| 247 |
vk_alloc(&device->vk.alloc, |
| 248 |
sizeof(*setup->mrt_resources) * attachment_count, |
| 249 |
8U, |
| 250 |
VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); |
| 251 |
if (!setup->mrt_resources) |
| 252 |
return vk_error(NULL, VK_ERROR_OUT_OF_HOST_MEMORY); |
| 253 |
|
| 254 |
for (unsigned i = 0; i < attachment_count; i++) { |
| 255 |
VkFormat att_format = attachment_formats[i]; |
| 256 |
assert(att_format != VK_FORMAT_UNDEFINED); |
| 257 |
|
| 258 |
result = pvr_alloc_mrt(dev_info, &alloc, setup, i, att_format); |
| 259 |
if (result != VK_SUCCESS) { |
| 260 |
result = vk_error(NULL, VK_ERROR_OUT_OF_DEVICE_MEMORY); |
| 261 |
goto fail; |
| 262 |
} |
| 263 |
} |
| 264 |
|
| 265 |
early_exit: |
| 266 |
setup->num_output_regs = util_next_power_of_two(alloc.output_regs_count); |
| 267 |
setup->num_tile_buffers = alloc.tile_buffers_count; |
| 268 |
return VK_SUCCESS; |
| 269 |
fail: |
| 270 |
vk_free(&device->vk.alloc, setup->mrt_resources); |
| 271 |
return result; |
| 272 |
} |
| 273 |
|
| 274 |
void pvr_arch_destroy_mrt_setup(const struct pvr_device *device, |
| 275 |
struct usc_mrt_setup *setup) |
| 276 |
{ |
| 277 |
if (!setup) |
| 278 |
return; |
| 279 |
|
| 280 |
vk_free(&device->vk.alloc, setup->mrt_resources); |
| 281 |
} |
| 282 |
|
| 283 |
static bool |
| 284 |
pvr_rendering_info_needs_load(const struct pvr_dynamic_render_info *dr_info) |
| 285 |
{ |
| 286 |
for (unsigned i = 0; i < dr_info->hw_render.color_init_count; i++) { |
| 287 |
const uint32_t index = dr_info->hw_render.color_init[i].index; |
| 288 |
if (index == VK_ATTACHMENT_UNUSED) |
| 289 |
continue; |
| 290 |
|
| 291 |
const VkAttachmentLoadOp op = dr_info->hw_render.color_init[i].op; |
| 292 |
if (op == VK_ATTACHMENT_LOAD_OP_LOAD || op == VK_ATTACHMENT_LOAD_OP_CLEAR) |
| 293 |
return true; |
| 294 |
} |
| 295 |
|
| 296 |
return false; |
| 297 |
} |
| 298 |
|
| 299 |
static VkResult pvr_mrt_load_op_init(struct pvr_device *device, |
| 300 |
const VkAllocationCallbacks *alloc, |
| 301 |
const struct pvr_render_pass_info *rp_info, |
| 302 |
struct pvr_load_op *load_op, |
| 303 |
uint32_t view_idx) |
| 304 |
{ |
| 305 |
const struct pvr_dynamic_render_info *dr_info = rp_info->dr_info; |
| 306 |
VkResult result; |
| 307 |
|
| 308 |
load_op->clears_loads_state.depth_clear_to_reg = PVR_NO_DEPTH_CLEAR_TO_REG; |
| 309 |
|
| 310 |
assert(dr_info->hw_render.color_init_count <= |
| 311 |
PVR_LOAD_OP_CLEARS_LOADS_MAX_RTS); |
| 312 |
for (unsigned i = 0; i < dr_info->hw_render.color_init_count; i++) { |
| 313 |
const struct pvr_renderpass_colorinit *color_init = |
| 314 |
&dr_info->hw_render.color_init[i]; |
| 315 |
const struct pvr_image *image; |
| 316 |
|
| 317 |
assert(color_init->index < rp_info->attachment_count); |
| 318 |
load_op->clears_loads_state.dest_vk_format[i] = |
| 319 |
rp_info->attachments[color_init->index]->vk.view_format; |
| 320 |
|
| 321 |
image = pvr_image_view_get_image(rp_info->attachments[color_init->index]); |
| 322 |
if (image->vk.samples > VK_SAMPLE_COUNT_1_BIT) |
| 323 |
load_op->clears_loads_state.unresolved_msaa_mask |= BITFIELD_BIT(i); |
| 324 |
|
| 325 |
switch (color_init->op) { |
| 326 |
case VK_ATTACHMENT_LOAD_OP_CLEAR: |
| 327 |
load_op->clears_loads_state.rt_clear_mask |= BITFIELD_BIT(i); |
| 328 |
break; |
| 329 |
case VK_ATTACHMENT_LOAD_OP_LOAD: |
| 330 |
load_op->clears_loads_state.rt_load_mask |= BITFIELD_BIT(i); |
| 331 |
break; |
| 332 |
case VK_ATTACHMENT_LOAD_OP_DONT_CARE: |
| 333 |
case VK_ATTACHMENT_LOAD_OP_NONE: |
| 334 |
break; |
| 335 |
default: |
| 336 |
UNREACHABLE("unsupported loadOp"); |
| 337 |
} |
| 338 |
} |
| 339 |
|
| 340 |
load_op->clears_loads_state.mrt_setup = &dr_info->hw_render.init_setup; |
| 341 |
|
| 342 |
result = pvr_arch_load_op_shader_generate(device, alloc, load_op); |
| 343 |
if (result != VK_SUCCESS) { |
| 344 |
vk_free2(&device->vk.alloc, alloc, load_op); |
| 345 |
return result; |
| 346 |
} |
| 347 |
|
| 348 |
load_op->view_indices[0] = view_idx; |
| 349 |
load_op->view_count = 1; |
| 350 |
|
| 351 |
load_op->is_hw_object = true; |
| 352 |
load_op->hw_render = &dr_info->hw_render; |
| 353 |
|
| 354 |
return VK_SUCCESS; |
| 355 |
} |
| 356 |
|
| 357 |
static void pvr_load_op_fini(struct pvr_load_op *load_op) |
| 358 |
{ |
| 359 |
pvr_bo_suballoc_free(load_op->pds_tex_state_prog.pvr_bo); |
| 360 |
pvr_bo_suballoc_free(load_op->pds_frag_prog.pvr_bo); |
| 361 |
pvr_bo_suballoc_free(load_op->usc_frag_prog_bo); |
| 362 |
} |
| 363 |
|
| 364 |
static void pvr_load_op_destroy(struct pvr_device *device, |
| 365 |
const VkAllocationCallbacks *allocator, |
| 366 |
struct pvr_load_op *load_op) |
| 367 |
{ |
| 368 |
pvr_load_op_fini(load_op); |
| 369 |
vk_free2(&device->vk.alloc, allocator, load_op); |
| 370 |
} |
| 371 |
|
| 372 |
void pvr_arch_mrt_load_op_state_cleanup(const struct pvr_device *device, |
| 373 |
const VkAllocationCallbacks *alloc, |
| 374 |
struct pvr_load_op_state *state) |
| 375 |
{ |
| 376 |
if (!state) |
| 377 |
return; |
| 378 |
|
| 379 |
while (state->load_op_count--) { |
| 380 |
const uint32_t load_op_idx = state->load_op_count; |
| 381 |
struct pvr_load_op *load_op = &state->load_ops[load_op_idx]; |
| 382 |
|
| 383 |
pvr_load_op_fini(load_op); |
| 384 |
} |
| 385 |
|
| 386 |
vk_free2(&device->vk.alloc, alloc, state); |
| 387 |
} |
| 388 |
|
| 389 |
static VkResult |
| 390 |
pvr_mrt_load_op_state_create(struct pvr_device *device, |
| 391 |
const VkAllocationCallbacks *alloc, |
| 392 |
const struct pvr_render_pass_info *rp_info, |
| 393 |
struct pvr_load_op_state **state) |
| 394 |
{ |
| 395 |
const struct pvr_dynamic_render_info *dr_info = rp_info->dr_info; |
| 396 |
const uint32_t view_count = util_bitcount(dr_info->hw_render.view_mask); |
| 397 |
struct pvr_load_op_state *load_op_state; |
| 398 |
struct pvr_load_op *load_ops; |
| 399 |
VkResult result; |
| 400 |
|
| 401 |
VK_MULTIALLOC(ma); |
| 402 |
vk_multialloc_add(&ma, &load_op_state, __typeof__(*load_op_state), 1); |
| 403 |
vk_multialloc_add(&ma, &load_ops, __typeof__(*load_ops), view_count); |
| 404 |
|
| 405 |
if (!vk_multialloc_zalloc(&ma, alloc, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE)) |
| 406 |
return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY); |
| 407 |
|
| 408 |
load_op_state->load_ops = load_ops; |
| 409 |
|
| 410 |
u_foreach_bit (view_idx, dr_info->hw_render.view_mask) { |
| 411 |
struct pvr_load_op *const load_op = |
| 412 |
&load_op_state->load_ops[load_op_state->load_op_count]; |
| 413 |
|
| 414 |
result = pvr_mrt_load_op_init(device, alloc, rp_info, load_op, view_idx); |
| 415 |
if (result != VK_SUCCESS) |
| 416 |
goto err_load_op_state_cleanup; |
| 417 |
|
| 418 |
load_op_state->load_op_count++; |
| 419 |
} |
| 420 |
|
| 421 |
*state = load_op_state; |
| 422 |
|
| 423 |
return VK_SUCCESS; |
| 424 |
|
| 425 |
err_load_op_state_cleanup: |
| 426 |
pvr_arch_mrt_load_op_state_cleanup(device, alloc, load_op_state); |
| 427 |
|
| 428 |
return result; |
| 429 |
} |
| 430 |
|
| 431 |
/* TODO: Can we gaurantee that if we have at least one render target there will |
| 432 |
* be a render target allocated as a REG? |
| 433 |
*/ |
| 434 |
static inline bool |
| 435 |
pvr_needs_output_register_writes(const struct usc_mrt_setup *setup) |
| 436 |
{ |
| 437 |
for (uint32_t i = 0; i < setup->num_render_targets; i++) { |
| 438 |
struct usc_mrt_resource *mrt_resource = &setup->mrt_resources[i]; |
| 439 |
|
| 440 |
if (mrt_resource->type == USC_MRT_RESOURCE_TYPE_OUTPUT_REG) |
| 441 |
return true; |
| 442 |
} |
| 443 |
|
| 444 |
return false; |
| 445 |
} |
| 446 |
|
| 447 |
static inline VkResult |
| 448 |
pvr_mrt_add_missing_output_register_write(struct usc_mrt_setup *setup, |
| 449 |
const VkAllocationCallbacks *alloc) |
| 450 |
{ |
| 451 |
const uint32_t last = setup->num_render_targets; |
| 452 |
struct usc_mrt_resource *mrt_resources; |
| 453 |
|
| 454 |
if (pvr_needs_output_register_writes(setup)) |
| 455 |
return VK_SUCCESS; |
| 456 |
|
| 457 |
setup->num_render_targets++; |
| 458 |
|
| 459 |
mrt_resources = |
| 460 |
vk_realloc(alloc, |
| 461 |
setup->mrt_resources, |
| 462 |
setup->num_render_targets * sizeof(*mrt_resources), |
| 463 |
8U, |
| 464 |
VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); |
| 465 |
|
| 466 |
if (!mrt_resources) |
| 467 |
return VK_ERROR_OUT_OF_HOST_MEMORY; |
| 468 |
|
| 469 |
setup->mrt_resources = mrt_resources; |
| 470 |
|
| 471 |
mrt_resources[last].type = USC_MRT_RESOURCE_TYPE_OUTPUT_REG; |
| 472 |
mrt_resources[last].reg.output_reg = 0U; |
| 473 |
mrt_resources[last].reg.offset = 0U; |
| 474 |
mrt_resources[last].intermediate_size = 4U; |
| 475 |
mrt_resources[last].mrt_desc.intermediate_size = 4U; |
| 476 |
mrt_resources[last].mrt_desc.priority = 0U; |
| 477 |
mrt_resources[last].mrt_desc.valid_mask[0U] = ~0; |
| 478 |
mrt_resources[last].mrt_desc.valid_mask[1U] = ~0; |
| 479 |
mrt_resources[last].mrt_desc.valid_mask[2U] = ~0; |
| 480 |
mrt_resources[last].mrt_desc.valid_mask[3U] = ~0; |
| 481 |
|
| 482 |
return VK_SUCCESS; |
| 483 |
} |
| 484 |
|
| 485 |
VkResult pvr_arch_mrt_load_ops_setup(struct pvr_cmd_buffer *cmd_buffer, |
| 486 |
const VkAllocationCallbacks *alloc, |
| 487 |
struct pvr_load_op_state **load_op_state) |
| 488 |
{ |
| 489 |
const struct pvr_cmd_buffer_state *state = &cmd_buffer->state; |
| 490 |
const struct pvr_dynamic_render_info *dr_info = |
| 491 |
state->render_pass_info.dr_info; |
| 492 |
struct pvr_device *device = cmd_buffer->device; |
| 493 |
VkResult result = VK_SUCCESS; |
| 494 |
|
| 495 |
if (dr_info->mrt_setup->num_tile_buffers) { |
| 496 |
result = pvr_device_tile_buffer_ensure_cap( |
| 497 |
device, |
| 498 |
dr_info->mrt_setup->num_tile_buffers); |
| 499 |
|
| 500 |
if (result != VK_SUCCESS) |
| 501 |
return result; |
| 502 |
} |
| 503 |
|
| 504 |
if (!pvr_rendering_info_needs_load(dr_info)) |
| 505 |
return VK_SUCCESS; |
| 506 |
|
| 507 |
result = |
| 508 |
pvr_mrt_add_missing_output_register_write(dr_info->mrt_setup, alloc); |
| 509 |
if (result != VK_SUCCESS) |
| 510 |
return result; |
| 511 |
|
| 512 |
result = pvr_mrt_load_op_state_create(device, |
| 513 |
alloc, |
| 514 |
&state->render_pass_info, |
| 515 |
load_op_state); |
| 516 |
|
| 517 |
return result; |
| 518 |
} |
| 519 |
|
| 520 |
VkResult pvr_arch_pds_unitex_state_program_create_and_upload( |
| 521 |
struct pvr_device *device, |
| 522 |
const VkAllocationCallbacks *allocator, |
| 523 |
uint32_t texture_kicks, |
| 524 |
uint32_t uniform_kicks, |
| 525 |
struct pvr_pds_upload *const pds_upload_out) |
| 526 |
{ |
| 527 |
struct pvr_pds_pixel_shader_sa_program program = { |
| 528 |
.num_texture_dma_kicks = texture_kicks, |
| 529 |
.num_uniform_dma_kicks = uniform_kicks, |
| 530 |
}; |
| 531 |
uint32_t staging_buffer_size; |
| 532 |
uint32_t *staging_buffer; |
| 533 |
VkResult result; |
| 534 |
|
| 535 |
pvr_pds_set_sizes_pixel_shader_uniform_texture_code(&program); |
| 536 |
|
| 537 |
staging_buffer_size = PVR_DW_TO_BYTES(program.code_size); |
| 538 |
|
| 539 |
staging_buffer = vk_alloc2(&device->vk.alloc, |
| 540 |
allocator, |
| 541 |
staging_buffer_size, |
| 542 |
8U, |
| 543 |
VK_SYSTEM_ALLOCATION_SCOPE_COMMAND); |
| 544 |
if (!staging_buffer) |
| 545 |
return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY); |
| 546 |
|
| 547 |
pvr_pds_generate_pixel_shader_sa_code_segment(&program, staging_buffer); |
| 548 |
|
| 549 |
/* FIXME: Figure out the define for alignment of 16. */ |
| 550 |
result = pvr_gpu_upload_pds(device, |
| 551 |
NULL, |
| 552 |
0U, |
| 553 |
0U, |
| 554 |
staging_buffer, |
| 555 |
program.code_size, |
| 556 |
16U, |
| 557 |
16U, |
| 558 |
pds_upload_out); |
| 559 |
if (result != VK_SUCCESS) { |
| 560 |
vk_free2(&device->vk.alloc, allocator, staging_buffer); |
| 561 |
return result; |
| 562 |
} |
| 563 |
|
| 564 |
vk_free2(&device->vk.alloc, allocator, staging_buffer); |
| 565 |
|
| 566 |
return VK_SUCCESS; |
| 567 |
} |
| 568 |
|
| 569 |
static VkResult pvr_pds_fragment_program_create_and_upload( |
| 570 |
struct pvr_device *device, |
| 571 |
const VkAllocationCallbacks *allocator, |
| 572 |
pco_shader *fs, |
| 573 |
struct pvr_suballoc_bo *shader_bo, |
| 574 |
struct pvr_pds_upload *pds_frag_prog, |
| 575 |
bool msaa) |
| 576 |
{ |
| 577 |
struct pvr_pds_kickusc_program program = { 0 }; |
| 578 |
pco_data *fs_data = pco_shader_data(fs); |
| 579 |
uint32_t staging_buffer_size; |
| 580 |
uint32_t *staging_buffer; |
| 581 |
VkResult result; |
| 582 |
|
| 583 |
const pvr_dev_addr_t exec_addr = |
| 584 |
PVR_DEV_ADDR_OFFSET(shader_bo->dev_addr, fs_data->common.entry_offset); |
| 585 |
|
| 586 |
/* Note this is not strictly required to be done before calculating the |
| 587 |
* staging_buffer_size in this particular case. It can also be done after |
| 588 |
* allocating the buffer. The size from pvr_pds_kick_usc() is constant. |
| 589 |
*/ |
| 590 |
pvr_pds_setup_doutu(&program.usc_task_control, |
| 591 |
exec_addr.addr, |
| 592 |
fs_data->common.temps, |
| 593 |
msaa ? ROGUE_PDSINST_DOUTU_SAMPLE_RATE_FULL |
| 594 |
: ROGUE_PDSINST_DOUTU_SAMPLE_RATE_INSTANCE, |
| 595 |
fs_data->fs.uses.phase_change); |
| 596 |
|
| 597 |
pvr_pds_kick_usc(&program, NULL, 0, false, PDS_GENERATE_SIZES); |
| 598 |
|
| 599 |
staging_buffer_size = PVR_DW_TO_BYTES(program.code_size + program.data_size); |
| 600 |
|
| 601 |
staging_buffer = vk_alloc2(&device->vk.alloc, |
| 602 |
allocator, |
| 603 |
staging_buffer_size, |
| 604 |
8, |
| 605 |
VK_SYSTEM_ALLOCATION_SCOPE_COMMAND); |
| 606 |
if (!staging_buffer) |
| 607 |
return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY); |
| 608 |
|
| 609 |
pvr_pds_kick_usc(&program, |
| 610 |
staging_buffer, |
| 611 |
0, |
| 612 |
false, |
| 613 |
PDS_GENERATE_CODEDATA_SEGMENTS); |
| 614 |
|
| 615 |
/* FIXME: Figure out the define for alignment of 16. */ |
| 616 |
result = pvr_gpu_upload_pds(device, |
| 617 |
&staging_buffer[0], |
| 618 |
program.data_size, |
| 619 |
16, |
| 620 |
&staging_buffer[program.data_size], |
| 621 |
program.code_size, |
| 622 |
16, |
| 623 |
16, |
| 624 |
pds_frag_prog); |
| 625 |
if (result != VK_SUCCESS) { |
| 626 |
vk_free2(&device->vk.alloc, allocator, staging_buffer); |
| 627 |
return result; |
| 628 |
} |
| 629 |
|
| 630 |
vk_free2(&device->vk.alloc, allocator, staging_buffer); |
| 631 |
|
| 632 |
return VK_SUCCESS; |
| 633 |
} |
| 634 |
|
| 635 |
VkResult |
| 636 |
pvr_arch_load_op_shader_generate(struct pvr_device *device, |
| 637 |
const VkAllocationCallbacks *allocator, |
| 638 |
struct pvr_load_op *load_op) |
| 639 |
{ |
| 640 |
const struct pvr_device_info *dev_info = &device->pdevice->dev_info; |
| 641 |
const uint32_t cache_line_size = pvr_get_slc_cache_line_size(dev_info); |
| 642 |
|
| 643 |
pco_shader *loadop = pvr_uscgen_loadop(device->pdevice->pco_ctx, load_op); |
| 644 |
|
| 645 |
VkResult result = pvr_gpu_upload_usc(device, |
| 646 |
pco_shader_binary_data(loadop), |
| 647 |
pco_shader_binary_size(loadop), |
| 648 |
cache_line_size, |
| 649 |
&load_op->usc_frag_prog_bo); |
| 650 |
|
| 651 |
if (result != VK_SUCCESS) { |
| 652 |
ralloc_free(loadop); |
| 653 |
return result; |
| 654 |
} |
| 655 |
|
| 656 |
const bool msaa = load_op->clears_loads_state.unresolved_msaa_mask & |
| 657 |
load_op->clears_loads_state.rt_load_mask; |
| 658 |
|
| 659 |
result = |
| 660 |
pvr_pds_fragment_program_create_and_upload(device, |
| 661 |
allocator, |
| 662 |
loadop, |
| 663 |
load_op->usc_frag_prog_bo, |
| 664 |
&load_op->pds_frag_prog, |
| 665 |
msaa); |
| 666 |
|
| 667 |
load_op->temps_count = pco_shader_data(loadop)->common.temps; |
| 668 |
ralloc_free(loadop); |
| 669 |
|
| 670 |
if (result != VK_SUCCESS) |
| 671 |
goto err_free_usc_frag_prog_bo; |
| 672 |
|
| 673 |
/* Manually hard coding `texture_kicks` to 1 since we'll pack everything into |
| 674 |
* one buffer to be DMAed. See `pvr_load_op_data_create_and_upload()`, where |
| 675 |
* we upload the buffer and upload the code section. |
| 676 |
*/ |
| 677 |
result = pvr_arch_pds_unitex_state_program_create_and_upload( |
| 678 |
device, |
| 679 |
allocator, |
| 680 |
1U, |
| 681 |
0U, |
| 682 |
&load_op->pds_tex_state_prog); |
| 683 |
if (result != VK_SUCCESS) |
| 684 |
goto err_free_pds_frag_prog; |
| 685 |
|
| 686 |
return VK_SUCCESS; |
| 687 |
|
| 688 |
err_free_pds_frag_prog: |
| 689 |
pvr_bo_suballoc_free(load_op->pds_frag_prog.pvr_bo); |
| 690 |
|
| 691 |
err_free_usc_frag_prog_bo: |
| 692 |
pvr_bo_suballoc_free(load_op->usc_frag_prog_bo); |
| 693 |
|
| 694 |
return result; |
| 695 |
} |