#include <nvif/class.h>
#include "util/u_format.h"
#include "util/u_format_s3tc.h"
+#include "util/u_screen.h"
#include "pipe/p_screen.h"
+#include "compiler/nir/nir.h"
#include "nouveau_vp3_video.h"
enum pipe_format format,
enum pipe_texture_target target,
unsigned sample_count,
+ unsigned storage_sample_count,
unsigned bindings)
{
const struct util_format_description *desc = util_format_description(format);
if (!(0x117 & (1 << sample_count))) /* 0, 1, 2, 4 or 8 */
return false;
+ if (MAX2(1, sample_count) != MAX2(1, storage_sample_count))
+ return false;
+
/* Short-circuit the rest of the logic -- this is used by the state tracker
* to determine valid MS levels in a no-attachments scenario.
*/
if (format == PIPE_FORMAT_NONE && bindings & PIPE_BIND_RENDER_TARGET)
return true;
- if (!util_format_is_supported(format, bindings))
- return false;
-
if ((bindings & PIPE_BIND_SAMPLER_VIEW) && (target != PIPE_BUFFER))
if (util_format_get_blocksizebits(format) == 3 * 32)
return false;
PIPE_BIND_SHARED);
if (bindings & PIPE_BIND_SHADER_IMAGE) {
- if (sample_count > 0 &&
- nouveau_screen(pscreen)->class_3d >= GM107_3D_CLASS) {
- /* MS images are currently unsupported on Maxwell because they have to
- * be handled explicitly. */
- return false;
- }
-
if (format == PIPE_FORMAT_B8G8R8A8_UNORM &&
nouveau_screen(pscreen)->class_3d < NVE4_3D_CLASS) {
/* This should work on Fermi, but for currently unknown reasons it
nvc0_screen_get_param(struct pipe_screen *pscreen, enum pipe_cap param)
{
const uint16_t class_3d = nouveau_screen(pscreen)->class_3d;
- struct nouveau_device *dev = nouveau_screen(pscreen)->device;
+ const struct nouveau_screen *screen = nouveau_screen(pscreen);
+ struct nouveau_device *dev = screen->device;
switch (param) {
/* non-boolean caps */
- case PIPE_CAP_MAX_TEXTURE_2D_LEVELS:
+ case PIPE_CAP_MAX_TEXTURE_2D_SIZE:
+ return 16384;
case PIPE_CAP_MAX_TEXTURE_CUBE_LEVELS:
return 15;
case PIPE_CAP_MAX_TEXTURE_3D_LEVELS:
case PIPE_CAP_GLSL_FEATURE_LEVEL:
return 430;
case PIPE_CAP_GLSL_FEATURE_LEVEL_COMPATIBILITY:
- return 140;
+ return 430;
case PIPE_CAP_MAX_RENDER_TARGETS:
return 8;
case PIPE_CAP_MAX_DUAL_SOURCE_RENDER_TARGETS:
return 1;
+ case PIPE_CAP_VIEWPORT_SUBPIXEL_BITS:
+ case PIPE_CAP_RASTERIZER_SUBPIXEL_BITS:
+ return 8;
case PIPE_CAP_MAX_STREAM_OUTPUT_BUFFERS:
return 4;
case PIPE_CAP_MAX_STREAM_OUTPUT_SEPARATE_COMPONENTS:
return 1024;
case PIPE_CAP_MAX_VERTEX_STREAMS:
return 4;
+ case PIPE_CAP_MAX_GS_INVOCATIONS:
+ return 32;
+ case PIPE_CAP_MAX_SHADER_BUFFER_SIZE:
+ return 1 << 27;
case PIPE_CAP_MAX_VERTEX_ATTRIB_STRIDE:
return 2048;
+ case PIPE_CAP_MAX_VERTEX_ELEMENT_SRC_OFFSET:
+ return 2047;
case PIPE_CAP_CONSTANT_BUFFER_OFFSET_ALIGNMENT:
return 256;
case PIPE_CAP_TEXTURE_BUFFER_OFFSET_ALIGNMENT:
return NVC0_MAX_WINDOW_RECTANGLES;
case PIPE_CAP_MAX_CONSERVATIVE_RASTER_SUBPIXEL_PRECISION_BIAS:
return class_3d >= GM200_3D_CLASS ? 8 : 0;
+ case PIPE_CAP_MAX_TEXTURE_UPLOAD_MEMORY_BUDGET:
+ return 64 * 1024 * 1024;
+ case PIPE_CAP_MAX_VARYINGS:
+ /* NOTE: These only count our slots for GENERIC varyings.
+ * The address space may be larger, but the actual hard limit seems to be
+ * less than what the address space layout permits, so don't add TEXCOORD,
+ * COLOR, etc. here.
+ */
+ return 0x1f0 / 16;
/* supported caps */
case PIPE_CAP_TEXTURE_MIRROR_CLAMP:
+ case PIPE_CAP_TEXTURE_MIRROR_CLAMP_TO_EDGE:
case PIPE_CAP_TEXTURE_SWIZZLE:
case PIPE_CAP_NPOT_TEXTURES:
case PIPE_CAP_MIXED_FRAMEBUFFER_SIZES:
case PIPE_CAP_CAN_BIND_CONST_BUFFER_AS_VERTEX:
case PIPE_CAP_ALLOW_MAPPED_BUFFERS_DURING_EXECUTION:
case PIPE_CAP_QUERY_SO_OVERFLOW:
+ case PIPE_CAP_DEST_SURFACE_SRGB_CONTROL:
return 1;
case PIPE_CAP_PREFER_BLIT_BASED_TEXTURE_TRANSFER:
return nouveau_screen(pscreen)->vram_domain & NOUVEAU_BO_VRAM ? 1 : 0;
- case PIPE_CAP_TGSI_FS_FBFETCH:
- return class_3d >= NVE4_3D_CLASS; /* needs testing on fermi */
+ case PIPE_CAP_FBFETCH:
+ return class_3d >= NVE4_3D_CLASS ? 1 : 0; /* needs testing on fermi */
+ case PIPE_CAP_SEAMLESS_CUBE_MAP_PER_TEXTURE:
+ case PIPE_CAP_TGSI_BALLOT:
+ return class_3d >= NVE4_3D_CLASS;
+ case PIPE_CAP_BINDLESS_TEXTURE:
+ return class_3d >= NVE4_3D_CLASS;
+ case PIPE_CAP_TGSI_ATOMFADD:
+ return class_3d < GM107_3D_CLASS; /* needs additional lowering */
case PIPE_CAP_POLYGON_MODE_FILL_RECTANGLE:
case PIPE_CAP_TGSI_VS_LAYER_VIEWPORT:
case PIPE_CAP_TGSI_TES_LAYER_VIEWPORT:
return class_3d >= GM200_3D_CLASS;
case PIPE_CAP_CONSERVATIVE_RASTER_PRE_SNAP_TRIANGLES:
return class_3d >= GP100_3D_CLASS;
- case PIPE_CAP_SEAMLESS_CUBE_MAP_PER_TEXTURE:
- case PIPE_CAP_TGSI_BALLOT:
- case PIPE_CAP_BINDLESS_TEXTURE:
- return class_3d >= NVE4_3D_CLASS;
+
+ /* caps has to be turned on with nir */
+ case PIPE_CAP_INT64_DIVMOD:
+ return screen->prefer_nir ? 1 : 0;
/* unsupported caps */
+ case PIPE_CAP_DEPTH_CLIP_DISABLE_SEPARATE:
case PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT:
case PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER:
case PIPE_CAP_SHADER_STENCIL_EXPORT:
case PIPE_CAP_PCI_BUS:
case PIPE_CAP_PCI_DEVICE:
case PIPE_CAP_PCI_FUNCTION:
- case PIPE_CAP_VIEWPORT_SUBPIXEL_BITS:
case PIPE_CAP_TGSI_CAN_READ_OUTPUTS:
case PIPE_CAP_NATIVE_FENCE_FD:
case PIPE_CAP_GLSL_OPTIMIZE_CONSERVATIVELY:
- case PIPE_CAP_INT64_DIVMOD:
case PIPE_CAP_SPARSE_BUFFER_PAGE_SIZE:
case PIPE_CAP_NIR_SAMPLERS_AS_DEREF:
case PIPE_CAP_MEMOBJ:
case PIPE_CAP_TGSI_ANY_REG_AS_ADDRESS:
case PIPE_CAP_TILE_RASTER_ORDER:
case PIPE_CAP_MAX_COMBINED_SHADER_OUTPUT_RESOURCES:
+ case PIPE_CAP_FRAMEBUFFER_MSAA_CONSTRAINTS:
case PIPE_CAP_SIGNED_VERTEX_BUFFER_OFFSET:
case PIPE_CAP_CONTEXT_PRIORITY_MASK:
case PIPE_CAP_FENCE_SIGNAL:
case PIPE_CAP_CONSTBUF0_FLAGS:
case PIPE_CAP_PACKED_UNIFORMS:
case PIPE_CAP_CONSERVATIVE_RASTER_PRE_SNAP_POINTS_LINES:
+ case PIPE_CAP_MAX_COMBINED_SHADER_BUFFERS:
+ case PIPE_CAP_MAX_COMBINED_HW_ATOMIC_COUNTERS:
+ case PIPE_CAP_MAX_COMBINED_HW_ATOMIC_COUNTER_BUFFERS:
+ case PIPE_CAP_SURFACE_SAMPLE_COUNT:
+ case PIPE_CAP_QUERY_PIPELINE_STATISTICS_SINGLE:
+ case PIPE_CAP_RGB_OVERRIDE_DST_ALPHA_BLEND:
+ case PIPE_CAP_GLSL_TESS_LEVELS_AS_INPUTS:
+ case PIPE_CAP_NIR_COMPACT_ARRAYS:
+ case PIPE_CAP_IMAGE_LOAD_FORMATTED:
return 0;
case PIPE_CAP_VENDOR_ID:
return dev->vram_size >> 20;
case PIPE_CAP_UMA:
return 0;
+ default:
+ debug_printf("%s: unhandled cap %d\n", __func__, param);
+ return u_pipe_screen_get_param_defaults(pscreen, param);
}
-
- NOUVEAU_ERR("unknown PIPE_CAP %d\n", param);
- return 0;
}
static int
enum pipe_shader_type shader,
enum pipe_shader_cap param)
{
- const uint16_t class_3d = nouveau_screen(pscreen)->class_3d;
+ const struct nouveau_screen *screen = nouveau_screen(pscreen);
+ const uint16_t class_3d = screen->class_3d;
switch (shader) {
case PIPE_SHADER_VERTEX:
switch (param) {
case PIPE_SHADER_CAP_PREFERRED_IR:
- return PIPE_SHADER_IR_TGSI;
+ return screen->prefer_nir ? PIPE_SHADER_IR_NIR : PIPE_SHADER_IR_TGSI;
case PIPE_SHADER_CAP_SUPPORTED_IRS:
- return 1 << PIPE_SHADER_IR_TGSI;
+ return 1 << PIPE_SHADER_IR_TGSI |
+ 1 << PIPE_SHADER_IR_NIR;
case PIPE_SHADER_CAP_MAX_INSTRUCTIONS:
case PIPE_SHADER_CAP_MAX_ALU_INSTRUCTIONS:
case PIPE_SHADER_CAP_MAX_TEX_INSTRUCTIONS:
case PIPE_SHADER_CAP_MAX_CONTROL_FLOW_DEPTH:
return 16;
case PIPE_SHADER_CAP_MAX_INPUTS:
- if (shader == PIPE_SHADER_VERTEX)
- return 32;
- /* NOTE: These only count our slots for GENERIC varyings.
- * The address space may be larger, but the actual hard limit seems to be
- * less than what the address space layout permits, so don't add TEXCOORD,
- * COLOR, etc. here.
- */
- if (shader == PIPE_SHADER_FRAGMENT)
- return 0x1f0 / 16;
- /* Actually this counts CLIPVERTEX, which occupies the last generic slot,
- * and excludes 0x60 per-patch inputs.
- */
return 0x200 / 16;
case PIPE_SHADER_CAP_MAX_OUTPUTS:
return 32;
case PIPE_SHADER_CAP_MAX_CONST_BUFFER_SIZE:
- return 65536;
+ return NVC0_MAX_CONSTBUF_SIZE;
case PIPE_SHADER_CAP_MAX_CONST_BUFFERS:
return NVC0_MAX_PIPE_CONSTBUFS;
case PIPE_SHADER_CAP_INDIRECT_OUTPUT_ADDR:
nouveau_heap_destroy(&screen->lib_code);
nouveau_heap_destroy(&screen->text_heap);
- FREE(screen->default_tsc);
FREE(screen->tic.entries);
nouveau_object_del(&screen->eng3d);
return 0;
}
+void
+nvc0_screen_bind_cb_3d(struct nvc0_screen *screen, bool *can_serialize,
+ int stage, int index, int size, uint64_t addr)
+{
+ assert(stage != 5);
+
+ struct nouveau_pushbuf *push = screen->base.pushbuf;
+
+ if (screen->base.class_3d >= GM107_3D_CLASS) {
+ struct nvc0_cb_binding *binding = &screen->cb_bindings[stage][index];
+
+ // TODO: Better figure out the conditions in which this is needed
+ bool serialize = binding->addr == addr && binding->size != size;
+ if (can_serialize)
+ serialize = serialize && *can_serialize;
+ if (serialize) {
+ IMMED_NVC0(push, NVC0_3D(SERIALIZE), 0);
+ if (can_serialize)
+ *can_serialize = false;
+ }
+
+ binding->addr = addr;
+ binding->size = size;
+ }
+
+ if (size >= 0) {
+ BEGIN_NVC0(push, NVC0_3D(CB_SIZE), 3);
+ PUSH_DATA (push, size);
+ PUSH_DATAh(push, addr);
+ PUSH_DATA (push, addr);
+ }
+ IMMED_NVC0(push, NVC0_3D(CB_BIND(stage)), (index << 4) | (size >= 0));
+}
+
+static const nir_shader_compiler_options nir_options = {
+ .lower_fdiv = false,
+ .lower_ffma = false,
+ .fuse_ffma = false, /* nir doesn't track mad vs fma */
+ .lower_flrp32 = true,
+ .lower_flrp64 = true,
+ .lower_fpow = false,
+ .lower_fsat = false,
+ .lower_fsqrt = false, // TODO: only before gm200
+ .lower_fmod = true,
+ .lower_bitfield_extract = false,
+ .lower_bitfield_extract_to_shifts = false,
+ .lower_bitfield_insert = false,
+ .lower_bitfield_insert_to_shifts = false,
+ .lower_bitfield_reverse = false,
+ .lower_bit_count = false,
+ .lower_ifind_msb = false,
+ .lower_find_lsb = false,
+ .lower_uadd_carry = true, // TODO
+ .lower_usub_borrow = true, // TODO
+ .lower_mul_high = false,
+ .lower_negate = false,
+ .lower_sub = false, // TODO
+ .lower_scmp = true, // TODO: not implemented yet
+ .lower_idiv = true,
+ .lower_isign = false, // TODO
+ .fdot_replicates = false, // TODO
+ .lower_ffloor = false, // TODO
+ .lower_ffract = true,
+ .lower_fceil = false, // TODO
+ .lower_ldexp = true,
+ .lower_pack_half_2x16 = true,
+ .lower_pack_unorm_2x16 = true,
+ .lower_pack_snorm_2x16 = true,
+ .lower_pack_unorm_4x8 = true,
+ .lower_pack_snorm_4x8 = true,
+ .lower_unpack_half_2x16 = true,
+ .lower_unpack_unorm_2x16 = true,
+ .lower_unpack_snorm_2x16 = true,
+ .lower_unpack_unorm_4x8 = true,
+ .lower_unpack_snorm_4x8 = true,
+ .lower_extract_byte = true,
+ .lower_extract_word = true,
+ .lower_all_io_to_temps = false,
+ .vertex_id_zero_based = false,
+ .lower_base_vertex = false,
+ .lower_helper_invocation = false,
+ .lower_cs_local_index_from_id = true,
+ .lower_cs_local_id_from_index = false,
+ .lower_device_index_to_zero = false, // TODO
+ .lower_wpos_pntc = false, // TODO
+ .lower_hadd = true, // TODO
+ .lower_add_sat = true, // TODO
+ .use_interpolated_input_intrinsics = true,
+ .lower_mul_2x32_64 = true, // TODO
+ .max_unroll_iterations = 32,
+ .lower_int64_options = nir_lower_divmod64, // TODO
+ .lower_doubles_options = nir_lower_dmod, // TODO
+};
+
+static const void *
+nvc0_screen_get_compiler_options(struct pipe_screen *pscreen,
+ enum pipe_shader_ir ir,
+ enum pipe_shader_type shader)
+{
+ if (ir == PIPE_SHADER_IR_NIR)
+ return &nir_options;
+ return NULL;
+}
+
#define FAIL_SCREEN_INIT(str, err) \
do { \
NOUVEAU_ERR(str, err); \
push->user_priv = screen;
push->rsvd_kick = 5;
+ /* TODO: could this be higher on Kepler+? how does reclocking vs no
+ * reclocking affect performance?
+ * TODO: could this be higher on Fermi?
+ */
+ if (dev->chipset >= 0xe0)
+ screen->base.transfer_pushbuf_threshold = 1024;
+
screen->base.vidmem_bindings |= PIPE_BIND_CONSTANT_BUFFER |
PIPE_BIND_SHADER_BUFFER |
PIPE_BIND_VERTEX_BUFFER | PIPE_BIND_INDEX_BUFFER |
pscreen->get_sample_pixel_grid = nvc0_screen_get_sample_pixel_grid;
pscreen->get_driver_query_info = nvc0_screen_get_driver_query_info;
pscreen->get_driver_query_group_info = nvc0_screen_get_driver_query_group_info;
+ /* nir stuff */
+ pscreen->get_compiler_options = nvc0_screen_get_compiler_options;
nvc0_screen_init_resource_functions(pscreen);
for (i = 0; i < NVC0_MAX_VIEWPORTS; i++) {
BEGIN_NVC0(push, NVC0_3D(SCISSOR_ENABLE(i)), 3);
PUSH_DATA (push, 1);
- PUSH_DATA (push, 8192 << 16);
- PUSH_DATA (push, 8192 << 16);
+ PUSH_DATA (push, 16384 << 16);
+ PUSH_DATA (push, 16384 << 16);
}
#define MK_MACRO(m, n) i = nvc0_graph_set_macro(screen, m, i, sizeof(n), n);
MK_MACRO(NVC0_3D_MACRO_DRAW_ELEMENTS_INDIRECT_COUNT, mme9097_draw_elts_indirect_count);
MK_MACRO(NVC0_3D_MACRO_QUERY_BUFFER_WRITE, mme9097_query_buffer_write);
MK_MACRO(NVC0_3D_MACRO_CONSERVATIVE_RASTER_STATE, mme9097_conservative_raster_state);
+ MK_MACRO(NVC0_3D_MACRO_COMPUTE_COUNTER, mme9097_compute_counter);
+ MK_MACRO(NVC0_3D_MACRO_COMPUTE_COUNTER_TO_QUERY, mme9097_compute_counter_to_query);
MK_MACRO(NVC0_CP_MACRO_LAUNCH_GRID_INDIRECT, mme90c0_launch_grid_indirect);
BEGIN_NVC0(push, NVC0_3D(RASTERIZE_ENABLE), 1);
/* XXX: Compute and 3D are somehow aliased on Fermi. */
for (i = 0; i < 5; ++i) {
+ unsigned j = 0;
+ for (j = 0; j < 16; j++)
+ screen->cb_bindings[i][j].size = -1;
+
/* TIC and TSC entries for each unit (nve4+ only) */
/* auxiliary constants (6 user clip planes, base instance id) */
- BEGIN_NVC0(push, NVC0_3D(CB_SIZE), 3);
- PUSH_DATA (push, NVC0_CB_AUX_SIZE);
- PUSH_DATAh(push, screen->uniform_bo->offset + NVC0_CB_AUX_INFO(i));
- PUSH_DATA (push, screen->uniform_bo->offset + NVC0_CB_AUX_INFO(i));
- BEGIN_NVC0(push, NVC0_3D(CB_BIND(i)), 1);
- PUSH_DATA (push, (15 << 4) | 1);
+ nvc0_screen_bind_cb_3d(screen, NULL, i, 15, NVC0_CB_AUX_SIZE,
+ screen->uniform_bo->offset + NVC0_CB_AUX_INFO(i));
if (screen->eng3d->oclass >= NVE4_3D_CLASS) {
unsigned j;
BEGIN_1IC0(push, NVC0_3D(CB_POS), 9);
if (!nvc0_blitter_create(screen))
goto fail;
- screen->default_tsc = CALLOC_STRUCT(nv50_tsc_entry);
- screen->default_tsc->tsc[0] = G80_TSC_0_SRGB_CONVERSION;
-
nouveau_fence_new(&screen->base, &screen->base.fence.current);
return &screen->base;