static bool
ilo_cp_detect_hang(struct ilo_cp *cp)
{
uint32_t active_lost, pending_lost;
bool guilty = false;
if (likely(!(ilo_debug & ILO_DEBUG_HANG)))
return false;
/* wait and get reset stats */
if (intel_bo_wait(cp->last_submitted_bo, -1) ||
intel_winsys_get_reset_stats(cp->winsys, cp->render_ctx,
&active_lost, &pending_lost))
return false;
if (cp->active_lost != active_lost) {
ilo_err("GPU hang caused by bo %p\n", cp->last_submitted_bo);
cp->active_lost = active_lost;
guilty = true;
}
if (cp->pending_lost != pending_lost) {
ilo_err("GPU hang detected\n");
cp->pending_lost = pending_lost;
}
return guilty;
}
/**
* Translate the TGSI tokens.
*/
static bool
vs_setup_tgsi(struct toy_compiler *tc, const struct tgsi_token *tokens,
struct toy_tgsi *tgsi)
{
if (ilo_debug & ILO_DEBUG_VS) {
ilo_printf("dumping vertex shader\n");
ilo_printf("\n");
tgsi_dump(tokens, 0);
ilo_printf("\n");
}
toy_compiler_translate_tgsi(tc, tokens, true, tgsi);
if (tc->fail) {
ilo_err("failed to translate VS TGSI tokens: %s\n", tc->reason);
return false;
}
if (ilo_debug & ILO_DEBUG_VS) {
ilo_printf("TGSI translator:\n");
toy_tgsi_dump(tgsi);
ilo_printf("\n");
toy_compiler_dump(tc);
ilo_printf("\n");
}
return true;
}
/**
* Compile the shader.
*/
static bool
gs_compile(struct gs_compile_context *gcc)
{
struct toy_compiler *tc = &gcc->tc;
struct ilo_shader *sh = gcc->shader;
get_num_prims_static(gcc);
if (gcc->is_static) {
tc_head(tc);
gs_init_vars(gcc);
gs_ff_sync(gcc, tdst_d(gcc->vars.tmp), tsrc_imm_d(gcc->static_data.total_prims));
gs_COPY1(tc, gcc->vars.urb_write_header, 0, tsrc_from(tdst_d(gcc->vars.tmp)), 0);
if (gcc->write_so)
gs_COPY4(tc, gcc->vars.so_index, 0, tsrc_from(tdst_d(gcc->vars.tmp)), 1);
tc_tail(tc);
}
else {
tc_fail(tc, "no control flow support");
return false;
}
if (!gcc->write_vue)
gs_discard(gcc);
gs_lower_virtual_opcodes(gcc);
toy_compiler_legalize_for_ra(tc);
toy_compiler_optimize(tc);
toy_compiler_allocate_registers(tc,
gcc->first_free_grf,
gcc->last_free_grf,
1);
toy_compiler_legalize_for_asm(tc);
if (tc->fail) {
ilo_err("failed to legalize GS instructions: %s\n", tc->reason);
return false;
}
if (ilo_debug & ILO_DEBUG_GS) {
ilo_printf("legalized instructions:\n");
toy_compiler_dump(tc);
ilo_printf("\n");
}
sh->kernel = toy_compiler_assemble(tc, &sh->kernel_size);
if (!sh->kernel)
return false;
if (ilo_debug & ILO_DEBUG_GS) {
ilo_printf("disassembly:\n");
toy_compiler_disassemble(tc->dev, sh->kernel, sh->kernel_size, false);
ilo_printf("\n");
}
return true;
}
static bool
tex_import_handle(struct ilo_texture *tex,
const struct winsys_handle *handle)
{
struct ilo_screen *is = ilo_screen(tex->base.screen);
const char *name = resource_get_bo_name(&tex->base);
enum intel_tiling_mode tiling;
unsigned long pitch;
tex->bo = intel_winsys_import_handle(is->winsys, name, handle,
tex->layout.bo_height, &tiling, &pitch);
if (!tex->bo)
return false;
if (!ilo_layout_update_for_imported_bo(&tex->layout, tiling, pitch)) {
ilo_err("imported handle has incompatible tiling/pitch\n");
intel_bo_unreference(tex->bo);
tex->bo = NULL;
return false;
}
return true;
}
static bool
gs_compile_passthrough(struct gs_compile_context *gcc)
{
struct toy_compiler *tc = &gcc->tc;
struct ilo_shader *sh = gcc->shader;
gcc->is_static = true;
gcc->static_data.total_vertices = gcc->in_vue_count;
gcc->static_data.total_prims = 1;
gcc->static_data.last_vertex[0] = 1 << (gcc->in_vue_count - 1);
gs_init_vars(gcc);
gs_ff_sync(gcc, tdst_d(gcc->vars.tmp), tsrc_imm_d(gcc->static_data.total_prims));
gs_COPY1(tc, gcc->vars.urb_write_header, 0, tsrc_from(tdst_d(gcc->vars.tmp)), 0);
if (gcc->write_so)
gs_COPY4(tc, gcc->vars.so_index, 0, tsrc_from(tdst_d(gcc->vars.tmp)), 1);
{
int vert, attr;
for (vert = 0; vert < gcc->out_vue_min_count; vert++) {
for (attr = 0; attr < gcc->shader->out.count; attr++) {
tc_MOV(tc, tdst_from(gcc->vars.tgsi_outs[attr]),
tsrc_offset(gcc->payload.vues[vert], attr / 2, (attr % 2) * 4));
}
gs_lower_opcode_emit(gcc, NULL);
}
gs_lower_opcode_endprim(gcc, NULL);
}
if (!gcc->write_vue)
gs_discard(gcc);
gs_lower_virtual_opcodes(gcc);
toy_compiler_legalize_for_ra(tc);
toy_compiler_optimize(tc);
toy_compiler_allocate_registers(tc,
gcc->first_free_grf,
gcc->last_free_grf,
1);
toy_compiler_legalize_for_asm(tc);
if (tc->fail) {
ilo_err("failed to translate GS TGSI tokens: %s\n", tc->reason);
return false;
}
if (ilo_debug & ILO_DEBUG_GS) {
int i;
ilo_printf("VUE count %d, VUE size %d\n",
gcc->in_vue_count, gcc->in_vue_size);
ilo_printf("%srasterizer discard\n",
(gcc->variant->u.gs.rasterizer_discard) ? "" : "no ");
for (i = 0; i < gcc->so_info->num_outputs; i++) {
ilo_printf("SO[%d] = OUT[%d]\n", i,
gcc->so_info->output[i].register_index);
}
ilo_printf("legalized instructions:\n");
toy_compiler_dump(tc);
ilo_printf("\n");
}
sh->kernel = toy_compiler_assemble(tc, &sh->kernel_size);
if (!sh->kernel) {
ilo_err("failed to compile GS: %s\n", tc->reason);
return false;
}
if (ilo_debug & ILO_DEBUG_GS) {
ilo_printf("disassembly:\n");
toy_compiler_disassemble(tc->dev, sh->kernel, sh->kernel_size, false);
ilo_printf("\n");
}
return true;
}
/**
* Compile the shader.
*/
static bool
vs_compile(struct vs_compile_context *vcc)
{
struct toy_compiler *tc = &vcc->tc;
struct ilo_shader *sh = vcc->shader;
vs_lower_virtual_opcodes(vcc);
toy_compiler_legalize_for_ra(tc);
toy_compiler_optimize(tc);
toy_compiler_allocate_registers(tc,
vcc->first_free_grf,
vcc->last_free_grf,
vcc->num_grf_per_vrf);
toy_compiler_legalize_for_asm(tc);
if (tc->fail) {
ilo_err("failed to legalize VS instructions: %s\n", tc->reason);
return false;
}
if (ilo_debug & ILO_DEBUG_VS) {
ilo_printf("legalized instructions:\n");
toy_compiler_dump(tc);
ilo_printf("\n");
}
if (true) {
sh->kernel = toy_compiler_assemble(tc, &sh->kernel_size);
}
else {
static const uint32_t microcode[] = {
/* fill in the microcode here */
0x0, 0x0, 0x0, 0x0,
};
const bool swap = true;
sh->kernel_size = sizeof(microcode);
sh->kernel = MALLOC(sh->kernel_size);
if (sh->kernel) {
const int num_dwords = sizeof(microcode) / 4;
const uint32_t *src = microcode;
uint32_t *dst = (uint32_t *) sh->kernel;
int i;
for (i = 0; i < num_dwords; i += 4) {
if (swap) {
dst[i + 0] = src[i + 3];
dst[i + 1] = src[i + 2];
dst[i + 2] = src[i + 1];
dst[i + 3] = src[i + 0];
}
else {
memcpy(dst, src, 16);
}
}
}
}
if (!sh->kernel) {
ilo_err("failed to compile VS: %s\n", tc->reason);
return false;
}
if (ilo_debug & ILO_DEBUG_VS) {
ilo_printf("disassembly:\n");
toy_compiler_disassemble(tc->dev, sh->kernel, sh->kernel_size, false);
ilo_printf("\n");
}
return true;
}
static bool
init_dev(struct ilo_dev_info *dev, const struct intel_winsys_info *info)
{
dev->devid = info->devid;
dev->has_llc = info->has_llc;
dev->has_gen7_sol_reset = info->has_gen7_sol_reset;
dev->has_address_swizzling = info->has_address_swizzling;
/*
* From the Sandy Bridge PRM, volume 4 part 2, page 18:
*
* "[DevSNB]: The GT1 product's URB provides 32KB of storage, arranged
* as 1024 256-bit rows. The GT2 product's URB provides 64KB of
* storage, arranged as 2048 256-bit rows. A row corresponds in size
* to an EU GRF register. Read/write access to the URB is generally
* supported on a row-granular basis."
*
* From the Ivy Bridge PRM, volume 4 part 2, page 17:
*
* "URB Size URB Rows URB Rows when SLM Enabled
* 128k 4096 2048
* 256k 8096 4096"
*/
if (IS_HASWELL(info->devid)) {
dev->gen = ILO_GEN(7.5);
if (IS_HSW_GT3(info->devid)) {
dev->gt = 3;
dev->urb_size = 512 * 1024;
}
else if (IS_HSW_GT2(info->devid)) {
dev->gt = 2;
dev->urb_size = 256 * 1024;
}
else {
dev->gt = 1;
dev->urb_size = 128 * 1024;
}
}
else if (IS_GEN7(info->devid)) {
dev->gen = ILO_GEN(7);
if (IS_IVB_GT2(info->devid)) {
dev->gt = 2;
dev->urb_size = 256 * 1024;
}
else {
dev->gt = 1;
dev->urb_size = 128 * 1024;
}
}
else if (IS_GEN6(info->devid)) {
dev->gen = ILO_GEN(6);
if (IS_SNB_GT2(info->devid)) {
dev->gt = 2;
dev->urb_size = 64 * 1024;
}
else {
dev->gt = 1;
dev->urb_size = 32 * 1024;
}
}
else {
ilo_err("unknown GPU generation\n");
return false;
}
return true;
}
/**
* Initialize the \p dev from \p winsys.
*/
bool
ilo_dev_init(struct ilo_dev *dev, struct intel_winsys *winsys)
{
const struct intel_winsys_info *info;
assert(ilo_is_zeroed(dev, sizeof(*dev)));
info = intel_winsys_get_info(winsys);
dev->winsys = winsys;
dev->devid = info->devid;
dev->aperture_total = info->aperture_total;
dev->aperture_mappable = info->aperture_mappable;
dev->has_llc = info->has_llc;
dev->has_address_swizzling = info->has_address_swizzling;
dev->has_logical_context = info->has_logical_context;
dev->has_ppgtt = info->has_ppgtt;
dev->has_timestamp = info->has_timestamp;
dev->has_gen7_sol_reset = info->has_gen7_sol_reset;
if (!dev->has_logical_context) {
ilo_err("missing hardware logical context support\n");
return false;
}
/*
* PIPE_CONTROL and MI_* use PPGTT writes on GEN7+ and privileged GGTT
* writes on GEN6.
*
* From the Sandy Bridge PRM, volume 1 part 3, page 101:
*
* "[DevSNB] When Per-Process GTT Enable is set, it is assumed that all
* code is in a secure environment, independent of address space.
* Under this condition, this bit only specifies the address space
* (GGTT or PPGTT). All commands are executed "as-is""
*
* We need PPGTT to be enabled on GEN6 too.
*/
if (!dev->has_ppgtt) {
/* experiments show that it does not really matter... */
ilo_warn("PPGTT disabled\n");
}
if (gen_is_bdw(info->devid) || gen_is_chv(info->devid)) {
dev->gen_opaque = ILO_GEN(8);
dev->gt = (gen_is_bdw(info->devid)) ? gen_get_bdw_gt(info->devid) : 1;
/* XXX random values */
if (dev->gt == 3) {
dev->eu_count = 48;
dev->thread_count = 336;
dev->urb_size = 384 * 1024;
} else if (dev->gt == 2) {
dev->eu_count = 24;
dev->thread_count = 168;
dev->urb_size = 384 * 1024;
} else {
dev->eu_count = 12;
dev->thread_count = 84;
dev->urb_size = 192 * 1024;
}
} else if (gen_is_hsw(info->devid)) {
/*
* From the Haswell PRM, volume 4, page 8:
*
* "Description GT3 GT2 GT1.5 GT1
* (...)
* EUs (Total) 40 20 12 10
* Threads (Total) 280 140 84 70
* (...)
* URB Size (max, within L3$) 512KB 256KB 256KB 128KB
*/
dev->gen_opaque = ILO_GEN(7.5);
dev->gt = gen_get_hsw_gt(info->devid);
if (dev->gt == 3) {
dev->eu_count = 40;
dev->thread_count = 280;
dev->urb_size = 512 * 1024;
} else if (dev->gt == 2) {
dev->eu_count = 20;
dev->thread_count = 140;
dev->urb_size = 256 * 1024;
} else {
dev->eu_count = 10;
dev->thread_count = 70;
dev->urb_size = 128 * 1024;
}
} else if (gen_is_ivb(info->devid) || gen_is_vlv(info->devid)) {
/*
* From the Ivy Bridge PRM, volume 1 part 1, page 18:
*
* "Device # of EUs #Threads/EU
* Ivy Bridge (GT2) 16 8
* Ivy Bridge (GT1) 6 6"
*
* From the Ivy Bridge PRM, volume 4 part 2, page 17:
*
* "URB Size URB Rows URB Rows when SLM Enabled
* 128k 4096 2048
* 256k 8096 4096"
*/
dev->gen_opaque = ILO_GEN(7);
dev->gt = (gen_is_ivb(info->devid)) ? gen_get_ivb_gt(info->devid) : 1;
if (dev->gt == 2) {
dev->eu_count = 16;
dev->thread_count = 128;
dev->urb_size = 256 * 1024;
} else {
dev->eu_count = 6;
dev->thread_count = 36;
dev->urb_size = 128 * 1024;
}
} else if (gen_is_snb(info->devid)) {
/*
* From the Sandy Bridge PRM, volume 1 part 1, page 22:
*
* "Device # of EUs #Threads/EU
* SNB GT2 12 5
* SNB GT1 6 4"
*
* From the Sandy Bridge PRM, volume 4 part 2, page 18:
*
* "[DevSNB]: The GT1 product's URB provides 32KB of storage,
* arranged as 1024 256-bit rows. The GT2 product's URB provides
* 64KB of storage, arranged as 2048 256-bit rows. A row
* corresponds in size to an EU GRF register. Read/write access to
* the URB is generally supported on a row-granular basis."
*/
dev->gen_opaque = ILO_GEN(6);
dev->gt = gen_get_snb_gt(info->devid);
if (dev->gt == 2) {
dev->eu_count = 12;
dev->thread_count = 60;
dev->urb_size = 64 * 1024;
} else {
dev->eu_count = 6;
dev->thread_count = 24;
dev->urb_size = 32 * 1024;
}
} else {
ilo_err("unknown GPU generation\n");
return false;
}
return true;
}