#include <llvm-c/TargetMachine.h>
#include "sid.h"
-#include "gfx9d.h"
#include "ac_binary.h"
#include "ac_llvm_util.h"
#include "ac_nir_to_llvm.h"
uint32_t lds_size;
};
+struct radv_ngg_state {
+ uint16_t ngg_emit_size; /* in dwords */
+ uint32_t hw_max_esverts;
+ uint32_t max_gsprims;
+ uint32_t max_out_verts;
+ uint32_t prim_amp_factor;
+ uint32_t vgt_esgs_ring_itemsize;
+ bool max_vert_out_per_gs_instance;
+};
+
+bool radv_pipeline_has_ngg(const struct radv_pipeline *pipeline)
+{
+ struct radv_shader_variant *variant = NULL;
+ if (pipeline->shaders[MESA_SHADER_GEOMETRY])
+ variant = pipeline->shaders[MESA_SHADER_GEOMETRY];
+ else if (pipeline->shaders[MESA_SHADER_TESS_EVAL])
+ variant = pipeline->shaders[MESA_SHADER_TESS_EVAL];
+ else if (pipeline->shaders[MESA_SHADER_VERTEX])
+ variant = pipeline->shaders[MESA_SHADER_VERTEX];
+ else
+ return false;
+ return variant->info.is_ngg;
+}
+
static void
radv_pipeline_destroy(struct radv_device *device,
struct radv_pipeline *pipeline,
col_format |= cf << (4 * i);
}
+ if (!(col_format & 0xf) && blend->need_src_alpha & (1 << 0)) {
+ /* When a subpass doesn't have any color attachments, write the
+ * alpha channel of MRT0 when alpha coverage is enabled because
+ * the depth attachment needs it.
+ */
+ col_format |= V_028714_SPI_SHADER_32_AR;
+ }
+
/* If the i-th target format is set, all previous target formats must
* be non-zero to avoid hangs.
*/
}
}
- blend->cb_shader_mask = ac_get_cb_shader_mask(col_format);
-
+ /* The output for dual source blending should have the same format as
+ * the first output.
+ */
if (blend->mrt0_is_dual_src)
col_format |= (col_format & 0xf) << 4;
+
+ blend->cb_shader_mask = ac_get_cb_shader_mask(col_format);
blend->spi_shader_col_format = col_format;
}
return false;
}
+/*
+ * Ordered so that for each i,
+ * radv_format_meta_fs_key(radv_fs_key_format_exemplars[i]) == i.
+ */
+const VkFormat radv_fs_key_format_exemplars[NUM_META_FS_KEYS] = {
+ VK_FORMAT_R32_SFLOAT,
+ VK_FORMAT_R32G32_SFLOAT,
+ VK_FORMAT_R8G8B8A8_UNORM,
+ VK_FORMAT_R16G16B16A16_UNORM,
+ VK_FORMAT_R16G16B16A16_SNORM,
+ VK_FORMAT_R16G16B16A16_UINT,
+ VK_FORMAT_R16G16B16A16_SINT,
+ VK_FORMAT_R32G32B32A32_SFLOAT,
+ VK_FORMAT_R8G8B8A8_UINT,
+ VK_FORMAT_R8G8B8A8_SINT,
+ VK_FORMAT_A2R10G10B10_UINT_PACK32,
+ VK_FORMAT_A2R10G10B10_SINT_PACK32,
+};
+
unsigned radv_format_meta_fs_key(VkFormat format)
{
- unsigned col_format = si_choose_spi_color_format(format, false, false) - 1;
+ unsigned col_format = si_choose_spi_color_format(format, false, false);
+
+ assert(col_format != V_028714_SPI_SHADER_32_AR);
+ if (col_format >= V_028714_SPI_SHADER_32_AR)
+ --col_format; /* Skip V_028714_SPI_SHADER_32_AR since there is no such VkFormat */
+
+ --col_format; /* Skip V_028714_SPI_SHADER_ZERO */
bool is_int8 = format_is_int8(format);
bool is_int10 = format_is_int10(format);
else
blend.cb_color_control |= S_028808_ROP3(V_028808_ROP3_COPY);
- blend.db_alpha_to_mask = S_028B70_ALPHA_TO_MASK_OFFSET0(2) |
- S_028B70_ALPHA_TO_MASK_OFFSET1(2) |
- S_028B70_ALPHA_TO_MASK_OFFSET2(2) |
- S_028B70_ALPHA_TO_MASK_OFFSET3(2);
+ blend.db_alpha_to_mask = S_028B70_ALPHA_TO_MASK_OFFSET0(3) |
+ S_028B70_ALPHA_TO_MASK_OFFSET1(1) |
+ S_028B70_ALPHA_TO_MASK_OFFSET2(0) |
+ S_028B70_ALPHA_TO_MASK_OFFSET3(2) |
+ S_028B70_OFFSET_ROUND(1);
if (vkms && vkms->alphaToCoverageEnable) {
blend.db_alpha_to_mask |= S_028B70_ALPHA_TO_MASK_ENABLE(1);
+ blend.need_src_alpha |= 0x1;
}
blend.cb_target_mask = 0;
};
if (pCreateInfo->pDepthStencilState &&
- subpass->depth_stencil_attachment.attachment != VK_ATTACHMENT_UNUSED) {
+ subpass->depth_stencil_attachment) {
const VkPipelineDepthStencilStateCreateInfo *vkds =
pCreateInfo->pDepthStencilState;
struct radv_render_pass_attachment *attachment =
- pass->attachments + subpass->depth_stencil_attachment.attachment;
+ pass->attachments + subpass->depth_stencil_attachment->attachment;
bool has_stencil = vk_format_is_stencil(attachment->format);
struct radv_dsa_order_invariance order_invariance[2];
struct radv_shader_variant *ps =
ps_iter_samples = ms->num_samples;
}
+ const struct VkPipelineRasterizationStateRasterizationOrderAMD *raster_order =
+ vk_find_struct_const(pCreateInfo->pRasterizationState->pNext, PIPELINE_RASTERIZATION_STATE_RASTERIZATION_ORDER_AMD);
+ if (raster_order && raster_order->rasterizationOrder == VK_RASTERIZATION_ORDER_RELAXED_AMD) {
+ /* Out-of-order rasterization is explicitly enabled by the
+ * application.
+ */
+ out_of_order_rast = true;
+ } else {
+ /* Determine if the driver can enable out-of-order
+ * rasterization internally.
+ */
+ out_of_order_rast =
+ radv_pipeline_out_of_order_rast(pipeline, blend, pCreateInfo);
+ }
+
ms->pa_sc_line_cntl = S_028BDC_DX10_DIAMOND_TEST_ENA(1);
ms->pa_sc_aa_config = 0;
ms->db_eqaa = S_028804_HIGH_QUALITY_INTERSECTIONS(1) |
- S_028804_STATIC_ANCHOR_ASSOCIATIONS(1);
+ S_028804_INCOHERENT_EQAA_READS(1) |
+ S_028804_INTERPOLATE_COMP_Z(1) |
+ S_028804_STATIC_ANCHOR_ASSOCIATIONS(1);
ms->pa_sc_mode_cntl_1 =
S_028A4C_WALK_FENCE_ENABLE(1) | //TODO linear dst fixes
S_028A4C_WALK_FENCE_SIZE(num_tile_pipes == 2 ? 2 : 3) |
+ S_028A4C_OUT_OF_ORDER_PRIMITIVE_ENABLE(out_of_order_rast) |
+ S_028A4C_OUT_OF_ORDER_WATER_MARK(0x7) |
/* always 1: */
S_028A4C_WALK_ALIGN8_PRIM_FITS_ST(1) |
S_028A4C_SUPERTILE_WALK_ORDER_ENABLE(1) |
S_028804_MASK_EXPORT_NUM_SAMPLES(log_samples) |
S_028804_ALPHA_TO_MASK_NUM_SAMPLES(log_samples);
ms->pa_sc_aa_config |= S_028BE0_MSAA_NUM_SAMPLES(log_samples) |
- S_028BE0_MAX_SAMPLE_DIST(radv_cayman_get_maxdist(log_samples)) |
+ S_028BE0_MAX_SAMPLE_DIST(radv_get_default_max_sample_dist(log_samples)) |
S_028BE0_MSAA_EXPOSED_SAMPLES(log_samples); /* CM_R_028BE0_PA_SC_AA_CONFIG */
ms->pa_sc_mode_cntl_1 |= S_028A4C_PS_ITER_SAMPLE(ps_iter_samples > 1);
if (ps_iter_samples > 1)
pipeline->graphics.spi_baryc_cntl |= S_0286E0_POS_FLOAT_LOCATION(2);
}
- const struct VkPipelineRasterizationStateRasterizationOrderAMD *raster_order =
- vk_find_struct_const(pCreateInfo->pRasterizationState->pNext, PIPELINE_RASTERIZATION_STATE_RASTERIZATION_ORDER_AMD);
- if (raster_order && raster_order->rasterizationOrder == VK_RASTERIZATION_ORDER_RELAXED_AMD) {
- /* Out-of-order rasterization is explicitly enabled by the
- * application.
- */
- out_of_order_rast = true;
- } else {
- /* Determine if the driver can enable out-of-order
- * rasterization internally.
- */
- out_of_order_rast =
- radv_pipeline_out_of_order_rast(pipeline, blend, pCreateInfo);
- }
-
- if (out_of_order_rast) {
- ms->pa_sc_mode_cntl_1 |= S_028A4C_OUT_OF_ORDER_PRIMITIVE_ENABLE(1) |
- S_028A4C_OUT_OF_ORDER_WATER_MARK(0x7);
- }
-
if (vkms && vkms->pSampleMask) {
mask = vkms->pSampleMask[0] & 0xffff;
}
}
}
-static unsigned si_map_swizzle(unsigned swizzle)
-{
- switch (swizzle) {
- case VK_SWIZZLE_Y:
- return V_008F0C_SQ_SEL_Y;
- case VK_SWIZZLE_Z:
- return V_008F0C_SQ_SEL_Z;
- case VK_SWIZZLE_W:
- return V_008F0C_SQ_SEL_W;
- case VK_SWIZZLE_0:
- return V_008F0C_SQ_SEL_0;
- case VK_SWIZZLE_1:
- return V_008F0C_SQ_SEL_1;
- default: /* VK_SWIZZLE_X */
- return V_008F0C_SQ_SEL_X;
- }
-}
-
-
static unsigned radv_dynamic_state_mask(VkDynamicState state)
{
switch(state) {
return RADV_DYNAMIC_STENCIL_REFERENCE;
case VK_DYNAMIC_STATE_DISCARD_RECTANGLE_EXT:
return RADV_DYNAMIC_DISCARD_RECTANGLE;
+ case VK_DYNAMIC_STATE_SAMPLE_LOCATIONS_EXT:
+ return RADV_DYNAMIC_SAMPLE_LOCATIONS;
default:
unreachable("Unhandled dynamic state");
}
if (!vk_find_struct_const(pCreateInfo->pNext, PIPELINE_DISCARD_RECTANGLE_STATE_CREATE_INFO_EXT))
states &= ~RADV_DYNAMIC_DISCARD_RECTANGLE;
+ if (!pCreateInfo->pMultisampleState ||
+ !vk_find_struct_const(pCreateInfo->pMultisampleState->pNext,
+ PIPELINE_SAMPLE_LOCATIONS_STATE_CREATE_INFO_EXT))
+ states &= ~RADV_DYNAMIC_SAMPLE_LOCATIONS;
+
/* TODO: blend constants & line width. */
return states;
* disabled or if the subpass of the render pass the pipeline is
* created against does not use any color attachments.
*/
- bool uses_color_att = false;
- for (unsigned i = 0; i < subpass->color_count; ++i) {
- if (subpass->color_attachments[i].attachment != VK_ATTACHMENT_UNUSED) {
- uses_color_att = true;
- break;
- }
- }
-
- if (uses_color_att && states & RADV_DYNAMIC_BLEND_CONSTANTS) {
+ if (subpass->has_color_att && states & RADV_DYNAMIC_BLEND_CONSTANTS) {
assert(pCreateInfo->pColorBlendState);
typed_memcpy(dynamic->blend_constants,
pCreateInfo->pColorBlendState->blendConstants, 4);
* disabled or if the subpass of the render pass the pipeline is created
* against does not use a depth/stencil attachment.
*/
- if (needed_states &&
- subpass->depth_stencil_attachment.attachment != VK_ATTACHMENT_UNUSED) {
+ if (needed_states && subpass->depth_stencil_attachment) {
assert(pCreateInfo->pDepthStencilState);
if (states & RADV_DYNAMIC_DEPTH_BOUNDS) {
const VkPipelineDiscardRectangleStateCreateInfoEXT *discard_rectangle_info =
vk_find_struct_const(pCreateInfo->pNext, PIPELINE_DISCARD_RECTANGLE_STATE_CREATE_INFO_EXT);
- if (states & RADV_DYNAMIC_DISCARD_RECTANGLE) {
+ if (needed_states & RADV_DYNAMIC_DISCARD_RECTANGLE) {
dynamic->discard_rectangle.count = discard_rectangle_info->discardRectangleCount;
- typed_memcpy(dynamic->discard_rectangle.rectangles,
- discard_rectangle_info->pDiscardRectangles,
- discard_rectangle_info->discardRectangleCount);
+ if (states & RADV_DYNAMIC_DISCARD_RECTANGLE) {
+ typed_memcpy(dynamic->discard_rectangle.rectangles,
+ discard_rectangle_info->pDiscardRectangles,
+ discard_rectangle_info->discardRectangleCount);
+ }
+ }
+
+ if (needed_states & RADV_DYNAMIC_SAMPLE_LOCATIONS) {
+ const VkPipelineSampleLocationsStateCreateInfoEXT *sample_location_info =
+ vk_find_struct_const(pCreateInfo->pMultisampleState->pNext,
+ PIPELINE_SAMPLE_LOCATIONS_STATE_CREATE_INFO_EXT);
+ /* If sampleLocationsEnable is VK_FALSE, the default sample
+ * locations are used and the values specified in
+ * sampleLocationsInfo are ignored.
+ */
+ if (sample_location_info->sampleLocationsEnable) {
+ const VkSampleLocationsInfoEXT *pSampleLocationsInfo =
+ &sample_location_info->sampleLocationsInfo;
+
+ assert(pSampleLocationsInfo->sampleLocationsCount <= MAX_SAMPLE_LOCATIONS);
+
+ dynamic->sample_location.per_pixel = pSampleLocationsInfo->sampleLocationsPerPixel;
+ dynamic->sample_location.grid_size = pSampleLocationsInfo->sampleLocationGridSize;
+ dynamic->sample_location.count = pSampleLocationsInfo->sampleLocationsCount;
+ typed_memcpy(&dynamic->sample_location.locations[0],
+ pSampleLocationsInfo->pSampleLocations,
+ pSampleLocationsInfo->sampleLocationsCount);
+ }
}
pipeline->dynamic_state.mask = states;
return gs;
}
+static void clamp_gsprims_to_esverts(unsigned *max_gsprims, unsigned max_esverts,
+ unsigned min_verts_per_prim, bool use_adjacency)
+{
+ unsigned max_reuse = max_esverts - min_verts_per_prim;
+ if (use_adjacency)
+ max_reuse /= 2;
+ *max_gsprims = MIN2(*max_gsprims, 1 + max_reuse);
+}
+
+static struct radv_ngg_state
+calculate_ngg_info(const VkGraphicsPipelineCreateInfo *pCreateInfo,
+ struct radv_pipeline *pipeline)
+{
+ struct radv_ngg_state ngg = {0};
+ struct radv_shader_variant_info *gs_info = &pipeline->shaders[MESA_SHADER_GEOMETRY]->info;
+ struct radv_es_output_info *es_info =
+ radv_pipeline_has_tess(pipeline) ? &gs_info->tes.es_info : &gs_info->vs.es_info;
+ unsigned gs_type = MESA_SHADER_VERTEX;
+ unsigned max_verts_per_prim = 3; // triangles
+ unsigned min_verts_per_prim =
+ gs_type == MESA_SHADER_GEOMETRY ? max_verts_per_prim : 1;
+ unsigned gs_num_invocations = 1;//MAX2(gs_info->gs.invocations, 1);
+ bool uses_adjacency;
+ switch(pCreateInfo->pInputAssemblyState->topology) {
+ case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY:
+ case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY:
+ case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY:
+ case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY:
+ uses_adjacency = true;
+ break;
+ default:
+ uses_adjacency = false;
+ break;
+ }
+
+ /* All these are in dwords: */
+ /* We can't allow using the whole LDS, because GS waves compete with
+ * other shader stages for LDS space.
+ *
+ * Streamout can increase the ESGS buffer size later on, so be more
+ * conservative with streamout and use 4K dwords. This may be suboptimal.
+ *
+ * Otherwise, use the limit of 7K dwords. The reason is that we need
+ * to leave some headroom for the max_esverts increase at the end.
+ *
+ * TODO: We should really take the shader's internal LDS use into
+ * account. The linker will fail if the size is greater than
+ * 8K dwords.
+ */
+ const unsigned max_lds_size = (0 /*gs_info->info.so.num_outputs*/ ? 4 : 7) * 1024 - 128;
+ const unsigned target_lds_size = max_lds_size;
+ unsigned esvert_lds_size = 0;
+ unsigned gsprim_lds_size = 0;
+
+ /* All these are per subgroup: */
+ bool max_vert_out_per_gs_instance = false;
+ unsigned max_esverts_base = 256;
+ unsigned max_gsprims_base = 128; /* default prim group size clamp */
+
+ /* Hardware has the following non-natural restrictions on the value
+ * of GE_CNTL.VERT_GRP_SIZE based on based on the primitive type of
+ * the draw:
+ * - at most 252 for any line input primitive type
+ * - at most 251 for any quad input primitive type
+ * - at most 251 for triangle strips with adjacency (this happens to
+ * be the natural limit for triangle *lists* with adjacency)
+ */
+ max_esverts_base = MIN2(max_esverts_base, 251 + max_verts_per_prim - 1);
+
+ if (gs_type == MESA_SHADER_GEOMETRY) {
+ unsigned max_out_verts_per_gsprim =
+ gs_info->gs.vertices_out * gs_num_invocations;
+
+ if (max_out_verts_per_gsprim <= 256) {
+ if (max_out_verts_per_gsprim) {
+ max_gsprims_base = MIN2(max_gsprims_base,
+ 256 / max_out_verts_per_gsprim);
+ }
+ } else {
+ /* Use special multi-cycling mode in which each GS
+ * instance gets its own subgroup. Does not work with
+ * tessellation. */
+ max_vert_out_per_gs_instance = true;
+ max_gsprims_base = 1;
+ max_out_verts_per_gsprim = gs_info->gs.vertices_out;
+ }
+
+ esvert_lds_size = es_info->esgs_itemsize / 4;
+ gsprim_lds_size = (gs_info->gs.gsvs_vertex_size / 4 + 1) * max_out_verts_per_gsprim;
+ } else {
+ /* TODO: This needs to be adjusted once LDS use for compaction
+ * after culling is implemented. */
+ /*
+ if (es_info->info.so.num_outputs)
+ esvert_lds_size = 4 * es_info->info.so.num_outputs + 1;
+ */
+ }
+
+ unsigned max_gsprims = max_gsprims_base;
+ unsigned max_esverts = max_esverts_base;
+
+ if (esvert_lds_size)
+ max_esverts = MIN2(max_esverts, target_lds_size / esvert_lds_size);
+ if (gsprim_lds_size)
+ max_gsprims = MIN2(max_gsprims, target_lds_size / gsprim_lds_size);
+
+ max_esverts = MIN2(max_esverts, max_gsprims * max_verts_per_prim);
+ clamp_gsprims_to_esverts(&max_gsprims, max_esverts, min_verts_per_prim, uses_adjacency);
+ assert(max_esverts >= max_verts_per_prim && max_gsprims >= 1);
+
+ if (esvert_lds_size || gsprim_lds_size) {
+ /* Now that we have a rough proportionality between esverts
+ * and gsprims based on the primitive type, scale both of them
+ * down simultaneously based on required LDS space.
+ *
+ * We could be smarter about this if we knew how much vertex
+ * reuse to expect.
+ */
+ unsigned lds_total = max_esverts * esvert_lds_size +
+ max_gsprims * gsprim_lds_size;
+ if (lds_total > target_lds_size) {
+ max_esverts = max_esverts * target_lds_size / lds_total;
+ max_gsprims = max_gsprims * target_lds_size / lds_total;
+
+ max_esverts = MIN2(max_esverts, max_gsprims * max_verts_per_prim);
+ clamp_gsprims_to_esverts(&max_gsprims, max_esverts,
+ min_verts_per_prim, uses_adjacency);
+ assert(max_esverts >= max_verts_per_prim && max_gsprims >= 1);
+ }
+ }
+
+ /* Round up towards full wave sizes for better ALU utilization. */
+ if (!max_vert_out_per_gs_instance) {
+ const unsigned wavesize = 64;
+ unsigned orig_max_esverts;
+ unsigned orig_max_gsprims;
+ do {
+ orig_max_esverts = max_esverts;
+ orig_max_gsprims = max_gsprims;
+
+ max_esverts = align(max_esverts, wavesize);
+ max_esverts = MIN2(max_esverts, max_esverts_base);
+ if (esvert_lds_size)
+ max_esverts = MIN2(max_esverts,
+ (max_lds_size - max_gsprims * gsprim_lds_size) /
+ esvert_lds_size);
+ max_esverts = MIN2(max_esverts, max_gsprims * max_verts_per_prim);
+
+ max_gsprims = align(max_gsprims, wavesize);
+ max_gsprims = MIN2(max_gsprims, max_gsprims_base);
+ if (gsprim_lds_size)
+ max_gsprims = MIN2(max_gsprims,
+ (max_lds_size - max_esverts * esvert_lds_size) /
+ gsprim_lds_size);
+ clamp_gsprims_to_esverts(&max_gsprims, max_esverts,
+ min_verts_per_prim, uses_adjacency);
+ assert(max_esverts >= max_verts_per_prim && max_gsprims >= 1);
+ } while (orig_max_esverts != max_esverts || orig_max_gsprims != max_gsprims);
+ }
+
+ /* Hardware restriction: minimum value of max_esverts */
+ max_esverts = MAX2(max_esverts, 23 + max_verts_per_prim);
+
+ unsigned max_out_vertices =
+ max_vert_out_per_gs_instance ? gs_info->gs.vertices_out :
+ gs_type == MESA_SHADER_GEOMETRY ?
+ max_gsprims * gs_num_invocations * gs_info->gs.vertices_out :
+ max_esverts;
+ assert(max_out_vertices <= 256);
+
+ unsigned prim_amp_factor = 1;
+ if (gs_type == MESA_SHADER_GEOMETRY) {
+ /* Number of output primitives per GS input primitive after
+ * GS instancing. */
+ prim_amp_factor = gs_info->gs.vertices_out;
+ }
+
+ /* The GE only checks against the maximum number of ES verts after
+ * allocating a full GS primitive. So we need to ensure that whenever
+ * this check passes, there is enough space for a full primitive without
+ * vertex reuse.
+ */
+ ngg.hw_max_esverts = max_esverts - max_verts_per_prim + 1;
+ ngg.max_gsprims = max_gsprims;
+ ngg.max_out_verts = max_out_vertices;
+ ngg.prim_amp_factor = prim_amp_factor;
+ ngg.max_vert_out_per_gs_instance = max_vert_out_per_gs_instance;
+ ngg.ngg_emit_size = max_gsprims * gsprim_lds_size;
+ ngg.vgt_esgs_ring_itemsize = 1;
+
+ pipeline->graphics.esgs_ring_size = 4 * max_esverts * esvert_lds_size;
+
+ assert(ngg.hw_max_esverts >= 24); /* HW limitation */
+
+ return ngg;
+}
+
static void
calculate_gs_ring_sizes(struct radv_pipeline *pipeline, const struct radv_gs_state *gs)
{
unsigned num_se = device->physical_device->rad_info.max_se;
unsigned wave_size = 64;
unsigned max_gs_waves = 32 * num_se; /* max 32 per SE on GCN */
- unsigned gs_vertex_reuse = 16 * num_se; /* GS_VERTEX_REUSE register (per SE) */
+ /* On GFX6-GFX7, the value comes from VGT_GS_VERTEX_REUSE = 16.
+ * On GFX8+, the value comes from VGT_VERTEX_REUSE_BLOCK_CNTL = 30 (+2).
+ */
+ unsigned gs_vertex_reuse =
+ (device->physical_device->rad_info.chip_class >= GFX8 ? 32 : 16) * num_se;
unsigned alignment = 256 * num_se;
/* The maximum size is 63.999 MB per SE. */
unsigned max_size = ((unsigned)(63.999 * 1024 * 1024) & ~255) * num_se;
unsigned esgs_ring_size = max_gs_waves * 2 * wave_size *
gs->vgt_esgs_ring_itemsize * 4 * gs_info->gs.vertices_in;
unsigned gsvs_ring_size = max_gs_waves * 2 * wave_size *
- gs_info->gs.max_gsvs_emit_size * 1; // no streams in VK (gs->max_gs_stream + 1);
+ gs_info->gs.max_gsvs_emit_size;
min_esgs_ring_size = align(min_esgs_ring_size, alignment);
esgs_ring_size = align(esgs_ring_size, alignment);
gsvs_ring_size = align(gsvs_ring_size, alignment);
- if (pipeline->device->physical_device->rad_info.chip_class <= VI)
+ if (pipeline->device->physical_device->rad_info.chip_class <= GFX8)
pipeline->graphics.esgs_ring_size = CLAMP(esgs_ring_size, min_esgs_ring_size, max_size);
pipeline->graphics.gsvs_ring_size = MIN2(gsvs_ring_size, max_size);
* It applies to workgroup sizes of more than one wavefront.
*/
if (device->physical_device->rad_info.family == CHIP_BONAIRE ||
- device->physical_device->rad_info.family == CHIP_KABINI ||
- device->physical_device->rad_info.family == CHIP_MULLINS)
+ device->physical_device->rad_info.family == CHIP_KABINI)
*lds_size = MAX2(*lds_size, 8);
}
lds_size = pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.tcs.lds_size;
- if (pipeline->device->physical_device->rad_info.chip_class >= CIK) {
+ if (pipeline->device->physical_device->rad_info.chip_class >= GFX7) {
assert(lds_size <= 65536);
lds_size = align(lds_size, 512) / 512;
} else {
}
bool ccw = tes->info.tes.ccw;
- const VkPipelineTessellationDomainOriginStateCreateInfoKHR *domain_origin_state =
+ const VkPipelineTessellationDomainOriginStateCreateInfo *domain_origin_state =
vk_find_struct_const(pCreateInfo->pTessellationState,
- PIPELINE_TESSELLATION_DOMAIN_ORIGIN_STATE_CREATE_INFO_KHR);
+ PIPELINE_TESSELLATION_DOMAIN_ORIGIN_STATE_CREATE_INFO);
- if (domain_origin_state && domain_origin_state->domainOrigin != VK_TESSELLATION_DOMAIN_ORIGIN_UPPER_LEFT_KHR)
+ if (domain_origin_state && domain_origin_state->domainOrigin != VK_TESSELLATION_DOMAIN_ORIGIN_UPPER_LEFT)
ccw = !ccw;
if (tes->info.tes.point_mode)
ordered_shaders[shader_count++] = shaders[MESA_SHADER_VERTEX];
}
+ if (shader_count > 1) {
+ unsigned first = ordered_shaders[shader_count - 1]->info.stage;
+ unsigned last = ordered_shaders[0]->info.stage;
+
+ if (ordered_shaders[0]->info.stage == MESA_SHADER_FRAGMENT &&
+ ordered_shaders[1]->info.has_transform_feedback_varyings)
+ nir_link_xfb_varyings(ordered_shaders[1], ordered_shaders[0]);
+
+ for (int i = 0; i < shader_count; ++i) {
+ nir_variable_mode mask = 0;
+
+ if (ordered_shaders[i]->info.stage != first)
+ mask = mask | nir_var_shader_in;
+
+ if (ordered_shaders[i]->info.stage != last)
+ mask = mask | nir_var_shader_out;
+
+ nir_lower_io_to_scalar_early(ordered_shaders[i], mask);
+ radv_optimize_nir(ordered_shaders[i], false, false);
+ }
+ }
+
for (int i = 1; i < shader_count; ++i) {
nir_lower_io_arrays_to_elements(ordered_shaders[i],
ordered_shaders[i - 1]);
+ if (nir_link_opt_varyings(ordered_shaders[i],
+ ordered_shaders[i - 1]))
+ radv_optimize_nir(ordered_shaders[i - 1], false, false);
+
nir_remove_dead_variables(ordered_shaders[i],
nir_var_shader_out);
nir_remove_dead_variables(ordered_shaders[i - 1],
ac_lower_indirect_derefs(ordered_shaders[i],
pipeline->device->physical_device->rad_info.chip_class);
}
- radv_optimize_nir(ordered_shaders[i], false);
+ radv_optimize_nir(ordered_shaders[i], false, false);
if (nir_lower_global_vars_to_local(ordered_shaders[i - 1])) {
ac_lower_indirect_derefs(ordered_shaders[i - 1],
pipeline->device->physical_device->rad_info.chip_class);
}
- radv_optimize_nir(ordered_shaders[i - 1], false);
+ radv_optimize_nir(ordered_shaders[i - 1], false, false);
}
}
}
+static uint32_t
+radv_get_attrib_stride(const VkPipelineVertexInputStateCreateInfo *input_state,
+ uint32_t attrib_binding)
+{
+ for (uint32_t i = 0; i < input_state->vertexBindingDescriptionCount; i++) {
+ const VkVertexInputBindingDescription *input_binding =
+ &input_state->pVertexBindingDescriptions[i];
+
+ if (input_binding->binding == attrib_binding)
+ return input_binding->stride;
+ }
+
+ return 0;
+}
static struct radv_pipeline_key
radv_generate_graphics_pipeline_key(struct radv_pipeline *pipeline,
}
for (unsigned i = 0; i < input_state->vertexAttributeDescriptionCount; ++i) {
- unsigned location = input_state->pVertexAttributeDescriptions[i].location;
- unsigned binding = input_state->pVertexAttributeDescriptions[i].binding;
+ const VkVertexInputAttributeDescription *desc =
+ &input_state->pVertexAttributeDescriptions[i];
+ const struct vk_format_description *format_desc;
+ unsigned location = desc->location;
+ unsigned binding = desc->binding;
+ unsigned num_format, data_format;
+ int first_non_void;
+
if (binding_input_rate & (1u << binding)) {
key.instance_rate_inputs |= 1u << location;
key.instance_rate_divisors[location] = instance_rate_divisors[binding];
}
- if (pipeline->device->physical_device->rad_info.chip_class <= VI &&
+ format_desc = vk_format_description(desc->format);
+ first_non_void = vk_format_get_first_non_void_channel(desc->format);
+
+ num_format = radv_translate_buffer_numformat(format_desc, first_non_void);
+ data_format = radv_translate_buffer_dataformat(format_desc, first_non_void);
+
+ key.vertex_attribute_formats[location] = data_format | (num_format << 4);
+ key.vertex_attribute_bindings[location] = desc->binding;
+ key.vertex_attribute_offsets[location] = desc->offset;
+ key.vertex_attribute_strides[location] = radv_get_attrib_stride(input_state, desc->binding);
+
+ if (pipeline->device->physical_device->rad_info.chip_class <= GFX8 &&
pipeline->device->physical_device->rad_info.family != CHIP_STONEY) {
VkFormat format = input_state->pVertexAttributeDescriptions[i].format;
uint64_t adjust;
}
key.vertex_alpha_adjust |= adjust << (2 * location);
}
+
+ switch (desc->format) {
+ case VK_FORMAT_B8G8R8A8_UNORM:
+ case VK_FORMAT_B8G8R8A8_SNORM:
+ case VK_FORMAT_B8G8R8A8_USCALED:
+ case VK_FORMAT_B8G8R8A8_SSCALED:
+ case VK_FORMAT_B8G8R8A8_UINT:
+ case VK_FORMAT_B8G8R8A8_SINT:
+ case VK_FORMAT_B8G8R8A8_SRGB:
+ case VK_FORMAT_A2R10G10B10_UNORM_PACK32:
+ case VK_FORMAT_A2R10G10B10_SNORM_PACK32:
+ case VK_FORMAT_A2R10G10B10_USCALED_PACK32:
+ case VK_FORMAT_A2R10G10B10_SSCALED_PACK32:
+ case VK_FORMAT_A2R10G10B10_UINT_PACK32:
+ case VK_FORMAT_A2R10G10B10_SINT_PACK32:
+ key.vertex_post_shuffle |= 1 << location;
+ break;
+ default:
+ break;
+ }
}
if (pCreateInfo->pTessellationState)
pCreateInfo->pMultisampleState->rasterizationSamples > 1) {
uint32_t num_samples = pCreateInfo->pMultisampleState->rasterizationSamples;
uint32_t ps_iter_samples = radv_pipeline_get_ps_iter_samples(pCreateInfo->pMultisampleState);
- key.log2_num_samples = util_logbase2(num_samples);
+ key.num_samples = num_samples;
key.log2_ps_iter_samples = util_logbase2(ps_iter_samples);
}
key.col_format = blend->spi_shader_col_format;
- if (pipeline->device->physical_device->rad_info.chip_class < VI)
+ if (pipeline->device->physical_device->rad_info.chip_class < GFX8)
radv_pipeline_compute_get_int_clamp(pCreateInfo, &key.is_int8, &key.is_int10);
return key;
}
static void
-radv_fill_shader_keys(struct radv_shader_variant_key *keys,
+radv_fill_shader_keys(struct radv_device *device,
+ struct radv_shader_variant_key *keys,
const struct radv_pipeline_key *key,
nir_shader **nir)
{
keys[MESA_SHADER_VERTEX].vs.instance_rate_inputs = key->instance_rate_inputs;
keys[MESA_SHADER_VERTEX].vs.alpha_adjust = key->vertex_alpha_adjust;
- for (unsigned i = 0; i < MAX_VERTEX_ATTRIBS; ++i)
+ keys[MESA_SHADER_VERTEX].vs.post_shuffle = key->vertex_post_shuffle;
+ for (unsigned i = 0; i < MAX_VERTEX_ATTRIBS; ++i) {
keys[MESA_SHADER_VERTEX].vs.instance_rate_divisors[i] = key->instance_rate_divisors[i];
+ keys[MESA_SHADER_VERTEX].vs.vertex_attribute_formats[i] = key->vertex_attribute_formats[i];
+ keys[MESA_SHADER_VERTEX].vs.vertex_attribute_bindings[i] = key->vertex_attribute_bindings[i];
+ keys[MESA_SHADER_VERTEX].vs.vertex_attribute_offsets[i] = key->vertex_attribute_offsets[i];
+ keys[MESA_SHADER_VERTEX].vs.vertex_attribute_strides[i] = key->vertex_attribute_strides[i];
+ }
if (nir[MESA_SHADER_TESS_CTRL]) {
- keys[MESA_SHADER_VERTEX].vs.as_ls = true;
+ keys[MESA_SHADER_VERTEX].vs.out.as_ls = true;
keys[MESA_SHADER_TESS_CTRL].tcs.num_inputs = 0;
keys[MESA_SHADER_TESS_CTRL].tcs.input_vertices = key->tess_input_vertices;
keys[MESA_SHADER_TESS_CTRL].tcs.primitive_mode = nir[MESA_SHADER_TESS_EVAL]->info.tess.primitive_mode;
if (nir[MESA_SHADER_GEOMETRY]) {
if (nir[MESA_SHADER_TESS_CTRL])
- keys[MESA_SHADER_TESS_EVAL].tes.as_es = true;
+ keys[MESA_SHADER_TESS_EVAL].tes.out.as_es = true;
else
- keys[MESA_SHADER_VERTEX].vs.as_es = true;
+ keys[MESA_SHADER_VERTEX].vs.out.as_es = true;
+ }
+
+ if (device->physical_device->rad_info.chip_class >= GFX10) {
+ keys[MESA_SHADER_VERTEX].vs.out.as_ngg = true;
}
for(int i = 0; i < MESA_SHADER_STAGES; ++i)
keys[MESA_SHADER_FRAGMENT].fs.is_int8 = key->is_int8;
keys[MESA_SHADER_FRAGMENT].fs.is_int10 = key->is_int10;
keys[MESA_SHADER_FRAGMENT].fs.log2_ps_iter_samples = key->log2_ps_iter_samples;
- keys[MESA_SHADER_FRAGMENT].fs.log2_num_samples = key->log2_num_samples;
+ keys[MESA_SHADER_FRAGMENT].fs.num_samples = key->num_samples;
}
static void
tes_info->tess.point_mode |= tcs_info->tess.point_mode;
}
+static
+void radv_init_feedback(const VkPipelineCreationFeedbackCreateInfoEXT *ext)
+{
+ if (!ext)
+ return;
+
+ if (ext->pPipelineCreationFeedback) {
+ ext->pPipelineCreationFeedback->flags = 0;
+ ext->pPipelineCreationFeedback->duration = 0;
+ }
+
+ for (unsigned i = 0; i < ext->pipelineStageCreationFeedbackCount; ++i) {
+ ext->pPipelineStageCreationFeedbacks[i].flags = 0;
+ ext->pPipelineStageCreationFeedbacks[i].duration = 0;
+ }
+}
+
+static
+void radv_start_feedback(VkPipelineCreationFeedbackEXT *feedback)
+{
+ if (!feedback)
+ return;
+
+ feedback->duration -= radv_get_current_time();
+ feedback ->flags = VK_PIPELINE_CREATION_FEEDBACK_VALID_BIT_EXT;
+}
+
+static
+void radv_stop_feedback(VkPipelineCreationFeedbackEXT *feedback, bool cache_hit)
+{
+ if (!feedback)
+ return;
+
+ feedback->duration += radv_get_current_time();
+ feedback ->flags = VK_PIPELINE_CREATION_FEEDBACK_VALID_BIT_EXT |
+ (cache_hit ? VK_PIPELINE_CREATION_FEEDBACK_APPLICATION_PIPELINE_CACHE_HIT_BIT_EXT : 0);
+}
+
static
void radv_create_shaders(struct radv_pipeline *pipeline,
struct radv_device *device,
struct radv_pipeline_cache *cache,
- struct radv_pipeline_key key,
+ const struct radv_pipeline_key *key,
const VkPipelineShaderStageCreateInfo **pStages,
- const VkPipelineCreateFlags flags)
+ const VkPipelineCreateFlags flags,
+ VkPipelineCreationFeedbackEXT *pipeline_feedback,
+ VkPipelineCreationFeedbackEXT **stage_feedbacks)
{
struct radv_shader_module fs_m = {0};
struct radv_shader_module *modules[MESA_SHADER_STAGES] = { 0, };
nir_shader *nir[MESA_SHADER_STAGES] = {0};
- void *codes[MESA_SHADER_STAGES] = {0};
- unsigned code_sizes[MESA_SHADER_STAGES] = {0};
- struct radv_shader_variant_key keys[MESA_SHADER_STAGES] = {{{{0}}}};
+ struct radv_shader_binary *binaries[MESA_SHADER_STAGES] = {NULL};
+ struct radv_shader_variant_key keys[MESA_SHADER_STAGES] = {{{{{0}}}}};
unsigned char hash[20], gs_copy_hash[20];
+ radv_start_feedback(pipeline_feedback);
+
for (unsigned i = 0; i < MESA_SHADER_STAGES; ++i) {
if (pStages[i]) {
modules[i] = radv_shader_module_from_handle(pStages[i]->module);
}
}
- radv_hash_shaders(hash, pStages, pipeline->layout, &key, get_hash_flags(device));
+ radv_hash_shaders(hash, pStages, pipeline->layout, key, get_hash_flags(device));
memcpy(gs_copy_hash, hash, 20);
gs_copy_hash[0] ^= 1;
+ bool found_in_application_cache = true;
if (modules[MESA_SHADER_GEOMETRY]) {
struct radv_shader_variant *variants[MESA_SHADER_STAGES] = {0};
- radv_create_shader_variants_from_pipeline_cache(device, cache, gs_copy_hash, variants);
+ radv_create_shader_variants_from_pipeline_cache(device, cache, gs_copy_hash, variants,
+ &found_in_application_cache);
pipeline->gs_copy_shader = variants[MESA_SHADER_GEOMETRY];
}
- if (radv_create_shader_variants_from_pipeline_cache(device, cache, hash, pipeline->shaders) &&
+ if (radv_create_shader_variants_from_pipeline_cache(device, cache, hash, pipeline->shaders,
+ &found_in_application_cache) &&
(!modules[MESA_SHADER_GEOMETRY] || pipeline->gs_copy_shader)) {
+ radv_stop_feedback(pipeline_feedback, found_in_application_cache);
return;
}
modules[MESA_SHADER_FRAGMENT] = &fs_m;
}
- /* Determine first and last stage. */
- unsigned first = MESA_SHADER_STAGES;
- unsigned last = 0;
- for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
- if (!pStages[i])
- continue;
- if (first == MESA_SHADER_STAGES)
- first = i;
- last = i;
- }
-
for (unsigned i = 0; i < MESA_SHADER_STAGES; ++i) {
const VkPipelineShaderStageCreateInfo *stage = pStages[i];
if (!modules[i])
continue;
+ radv_start_feedback(stage_feedbacks[i]);
+
nir[i] = radv_shader_compile_to_nir(device, modules[i],
stage ? stage->pName : "main", i,
stage ? stage->pSpecializationInfo : NULL,
- flags);
+ flags, pipeline->layout);
/* We don't want to alter meta shaders IR directly so clone it
* first.
nir[i] = nir_shader_clone(NULL, nir[i]);
}
- if (first != last) {
- nir_variable_mode mask = 0;
-
- if (i != first)
- mask = mask | nir_var_shader_in;
-
- if (i != last)
- mask = mask | nir_var_shader_out;
-
- if (!(flags & VK_PIPELINE_CREATE_DISABLE_OPTIMIZATION_BIT)) {
- nir_lower_io_to_scalar_early(nir[i], mask);
- radv_optimize_nir(nir[i], false);
- }
- }
+ radv_stop_feedback(stage_feedbacks[i], false);
}
if (nir[MESA_SHADER_TESS_CTRL]) {
- nir_lower_tes_patch_vertices(nir[MESA_SHADER_TESS_EVAL], nir[MESA_SHADER_TESS_CTRL]->info.tess.tcs_vertices_out);
+ nir_lower_patch_vertices(nir[MESA_SHADER_TESS_EVAL], nir[MESA_SHADER_TESS_CTRL]->info.tess.tcs_vertices_out, NULL);
merge_tess_info(&nir[MESA_SHADER_TESS_EVAL]->info, &nir[MESA_SHADER_TESS_CTRL]->info);
}
radv_link_shaders(pipeline, nir);
for (int i = 0; i < MESA_SHADER_STAGES; ++i) {
+ if (nir[i]) {
+ NIR_PASS_V(nir[i], nir_lower_bool_to_int32);
+ NIR_PASS_V(nir[i], nir_lower_non_uniform_access,
+ nir_lower_non_uniform_ubo_access |
+ nir_lower_non_uniform_ssbo_access |
+ nir_lower_non_uniform_texture_access |
+ nir_lower_non_uniform_image_access);
+ }
+
if (radv_can_dump_shader(device, modules[i], false))
nir_print_shader(nir[i], stderr);
}
- radv_fill_shader_keys(keys, &key, nir);
+ radv_fill_shader_keys(device, keys, key, nir);
if (nir[MESA_SHADER_FRAGMENT]) {
if (!pipeline->shaders[MESA_SHADER_FRAGMENT]) {
+ radv_start_feedback(stage_feedbacks[MESA_SHADER_FRAGMENT]);
+
pipeline->shaders[MESA_SHADER_FRAGMENT] =
- radv_shader_variant_create(device, modules[MESA_SHADER_FRAGMENT], &nir[MESA_SHADER_FRAGMENT], 1,
+ radv_shader_variant_compile(device, modules[MESA_SHADER_FRAGMENT], &nir[MESA_SHADER_FRAGMENT], 1,
pipeline->layout, keys + MESA_SHADER_FRAGMENT,
- &codes[MESA_SHADER_FRAGMENT], &code_sizes[MESA_SHADER_FRAGMENT]);
+ &binaries[MESA_SHADER_FRAGMENT]);
+
+ radv_stop_feedback(stage_feedbacks[MESA_SHADER_FRAGMENT], false);
}
/* TODO: These are no longer used as keys we should refactor this */
- keys[MESA_SHADER_VERTEX].vs.export_prim_id =
+ keys[MESA_SHADER_VERTEX].vs.out.export_prim_id =
pipeline->shaders[MESA_SHADER_FRAGMENT]->info.info.ps.prim_id_input;
- keys[MESA_SHADER_VERTEX].vs.export_layer_id =
+ keys[MESA_SHADER_VERTEX].vs.out.export_layer_id =
pipeline->shaders[MESA_SHADER_FRAGMENT]->info.info.ps.layer_input;
- keys[MESA_SHADER_TESS_EVAL].tes.export_prim_id =
+ keys[MESA_SHADER_VERTEX].vs.out.export_clip_dists =
+ !!pipeline->shaders[MESA_SHADER_FRAGMENT]->info.info.ps.num_input_clips_culls;
+ keys[MESA_SHADER_TESS_EVAL].tes.out.export_prim_id =
pipeline->shaders[MESA_SHADER_FRAGMENT]->info.info.ps.prim_id_input;
- keys[MESA_SHADER_TESS_EVAL].tes.export_layer_id =
+ keys[MESA_SHADER_TESS_EVAL].tes.out.export_layer_id =
pipeline->shaders[MESA_SHADER_FRAGMENT]->info.info.ps.layer_input;
+ keys[MESA_SHADER_TESS_EVAL].tes.out.export_clip_dists =
+ !!pipeline->shaders[MESA_SHADER_FRAGMENT]->info.info.ps.num_input_clips_culls;
}
if (device->physical_device->rad_info.chip_class >= GFX9 && modules[MESA_SHADER_TESS_CTRL]) {
struct nir_shader *combined_nir[] = {nir[MESA_SHADER_VERTEX], nir[MESA_SHADER_TESS_CTRL]};
struct radv_shader_variant_key key = keys[MESA_SHADER_TESS_CTRL];
key.tcs.vs_key = keys[MESA_SHADER_VERTEX].vs;
- pipeline->shaders[MESA_SHADER_TESS_CTRL] = radv_shader_variant_create(device, modules[MESA_SHADER_TESS_CTRL], combined_nir, 2,
+
+ radv_start_feedback(stage_feedbacks[MESA_SHADER_TESS_CTRL]);
+
+ pipeline->shaders[MESA_SHADER_TESS_CTRL] = radv_shader_variant_compile(device, modules[MESA_SHADER_TESS_CTRL], combined_nir, 2,
pipeline->layout,
- &key, &codes[MESA_SHADER_TESS_CTRL],
- &code_sizes[MESA_SHADER_TESS_CTRL]);
+ &key, &binaries[MESA_SHADER_TESS_CTRL]);
+
+ radv_stop_feedback(stage_feedbacks[MESA_SHADER_TESS_CTRL], false);
}
modules[MESA_SHADER_VERTEX] = NULL;
keys[MESA_SHADER_TESS_EVAL].tes.num_patches = pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.tcs.num_patches;
gl_shader_stage pre_stage = modules[MESA_SHADER_TESS_EVAL] ? MESA_SHADER_TESS_EVAL : MESA_SHADER_VERTEX;
if (!pipeline->shaders[MESA_SHADER_GEOMETRY]) {
struct nir_shader *combined_nir[] = {nir[pre_stage], nir[MESA_SHADER_GEOMETRY]};
- pipeline->shaders[MESA_SHADER_GEOMETRY] = radv_shader_variant_create(device, modules[MESA_SHADER_GEOMETRY], combined_nir, 2,
+
+ radv_start_feedback(stage_feedbacks[MESA_SHADER_GEOMETRY]);
+
+ pipeline->shaders[MESA_SHADER_GEOMETRY] = radv_shader_variant_compile(device, modules[MESA_SHADER_GEOMETRY], combined_nir, 2,
pipeline->layout,
- &keys[pre_stage] , &codes[MESA_SHADER_GEOMETRY],
- &code_sizes[MESA_SHADER_GEOMETRY]);
+ &keys[pre_stage] , &binaries[MESA_SHADER_GEOMETRY]);
+
+ radv_stop_feedback(stage_feedbacks[MESA_SHADER_GEOMETRY], false);
}
modules[pre_stage] = NULL;
}
keys[MESA_SHADER_TESS_EVAL].tes.num_patches = pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.tcs.num_patches;
keys[MESA_SHADER_TESS_EVAL].tes.tcs_num_outputs = util_last_bit64(pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.info.tcs.outputs_written);
}
- pipeline->shaders[i] = radv_shader_variant_create(device, modules[i], &nir[i], 1,
+
+ radv_start_feedback(stage_feedbacks[i]);
+
+ pipeline->shaders[i] = radv_shader_variant_compile(device, modules[i], &nir[i], 1,
pipeline->layout,
- keys + i, &codes[i],
- &code_sizes[i]);
+ keys + i, &binaries[i]);
+
+ radv_stop_feedback(stage_feedbacks[i], false);
}
}
if(modules[MESA_SHADER_GEOMETRY]) {
- void *gs_copy_code = NULL;
- unsigned gs_copy_code_size = 0;
+ struct radv_shader_binary *gs_copy_binary = NULL;
if (!pipeline->gs_copy_shader) {
pipeline->gs_copy_shader = radv_create_gs_copy_shader(
- device, nir[MESA_SHADER_GEOMETRY], &gs_copy_code,
- &gs_copy_code_size,
+ device, nir[MESA_SHADER_GEOMETRY], &gs_copy_binary,
keys[MESA_SHADER_GEOMETRY].has_multiview_view_index);
}
if (pipeline->gs_copy_shader) {
- void *code[MESA_SHADER_STAGES] = {0};
- unsigned code_size[MESA_SHADER_STAGES] = {0};
+ struct radv_shader_binary *binaries[MESA_SHADER_STAGES] = {NULL};
struct radv_shader_variant *variants[MESA_SHADER_STAGES] = {0};
- code[MESA_SHADER_GEOMETRY] = gs_copy_code;
- code_size[MESA_SHADER_GEOMETRY] = gs_copy_code_size;
+ binaries[MESA_SHADER_GEOMETRY] = gs_copy_binary;
variants[MESA_SHADER_GEOMETRY] = pipeline->gs_copy_shader;
radv_pipeline_cache_insert_shaders(device, cache,
gs_copy_hash,
variants,
- (const void**)code,
- code_size);
+ binaries);
}
- free(gs_copy_code);
+ free(gs_copy_binary);
}
radv_pipeline_cache_insert_shaders(device, cache, hash, pipeline->shaders,
- (const void**)codes, code_sizes);
+ binaries);
for (int i = 0; i < MESA_SHADER_STAGES; ++i) {
- free(codes[i]);
- if (modules[i]) {
+ free(binaries[i]);
+ if (nir[i]) {
if (!pipeline->device->keep_shader_info)
ralloc_free(nir[i]);
if (fs_m.nir)
ralloc_free(fs_m.nir);
+
+ radv_stop_feedback(pipeline_feedback, false);
}
static uint32_t
{
bool has_gs = radv_pipeline_has_gs(pipeline);
bool has_tess = radv_pipeline_has_tess(pipeline);
+ bool has_ngg = radv_pipeline_has_ngg(pipeline);
+
switch (stage) {
case MESA_SHADER_FRAGMENT:
return R_00B030_SPI_SHADER_USER_DATA_PS_0;
case MESA_SHADER_VERTEX:
- if (chip_class >= GFX9) {
- return has_tess ? R_00B430_SPI_SHADER_USER_DATA_LS_0 :
- has_gs ? R_00B330_SPI_SHADER_USER_DATA_ES_0 :
- R_00B130_SPI_SHADER_USER_DATA_VS_0;
+ if (has_tess) {
+ if (chip_class >= GFX10) {
+ return R_00B430_SPI_SHADER_USER_DATA_HS_0;
+ } else if (chip_class == GFX9) {
+ return R_00B430_SPI_SHADER_USER_DATA_LS_0;
+ } else {
+ return R_00B530_SPI_SHADER_USER_DATA_LS_0;
+ }
+
}
- if (has_tess)
- return R_00B530_SPI_SHADER_USER_DATA_LS_0;
- else
- return has_gs ? R_00B330_SPI_SHADER_USER_DATA_ES_0 : R_00B130_SPI_SHADER_USER_DATA_VS_0;
+
+ if (has_gs) {
+ if (chip_class >= GFX10) {
+ return R_00B230_SPI_SHADER_USER_DATA_GS_0;
+ } else {
+ return R_00B330_SPI_SHADER_USER_DATA_ES_0;
+ }
+ }
+
+ if (has_ngg)
+ return R_00B230_SPI_SHADER_USER_DATA_GS_0;
+
+ return R_00B130_SPI_SHADER_USER_DATA_VS_0;
case MESA_SHADER_GEOMETRY:
- return chip_class >= GFX9 ? R_00B330_SPI_SHADER_USER_DATA_ES_0 :
+ return chip_class == GFX9 ? R_00B330_SPI_SHADER_USER_DATA_ES_0 :
R_00B230_SPI_SHADER_USER_DATA_GS_0;
case MESA_SHADER_COMPUTE:
return R_00B900_COMPUTE_USER_DATA_0;
case MESA_SHADER_TESS_CTRL:
- return chip_class >= GFX9 ? R_00B430_SPI_SHADER_USER_DATA_LS_0 :
+ return chip_class == GFX9 ? R_00B430_SPI_SHADER_USER_DATA_LS_0 :
R_00B430_SPI_SHADER_USER_DATA_HS_0;
case MESA_SHADER_TESS_EVAL:
- if (chip_class >= GFX9) {
- return has_gs ? R_00B330_SPI_SHADER_USER_DATA_ES_0 :
- R_00B130_SPI_SHADER_USER_DATA_VS_0;
- }
- if (has_gs)
- return R_00B330_SPI_SHADER_USER_DATA_ES_0;
- else
+ if (has_gs) {
+ return chip_class >= GFX10 ? R_00B230_SPI_SHADER_USER_DATA_GS_0 :
+ R_00B330_SPI_SHADER_USER_DATA_ES_0;
+ } else {
return R_00B130_SPI_SHADER_USER_DATA_VS_0;
+ }
default:
unreachable("unknown shader");
}
pipeline->device->physical_device->rad_info.max_se);
unsigned log_num_se = util_logbase2_ceil(pipeline->device->physical_device->rad_info.max_se);
- unsigned total_samples = 1u << G_028BE0_MSAA_NUM_SAMPLES(pipeline->graphics.ms.pa_sc_mode_cntl_1);
+ unsigned total_samples = 1u << G_028BE0_MSAA_NUM_SAMPLES(pipeline->graphics.ms.pa_sc_aa_config);
unsigned ps_iter_samples = 1u << G_028804_PS_ITER_SAMPLES(pipeline->graphics.ms.db_eqaa);
unsigned effective_samples = total_samples;
unsigned color_bytes_per_pixel = 0;
}
const struct radv_bin_size_entry *color_entry = color_size_table[log_num_rb_per_se][log_num_se];
- while(color_entry->bpp <= color_bytes_per_pixel)
+ while(color_entry[1].bpp <= color_bytes_per_pixel)
++color_entry;
extent = color_entry->extent;
- if (subpass->depth_stencil_attachment.attachment != VK_ATTACHMENT_UNUSED) {
- struct radv_render_pass_attachment *attachment = pass->attachments + subpass->depth_stencil_attachment.attachment;
+ if (subpass->depth_stencil_attachment) {
+ struct radv_render_pass_attachment *attachment = pass->attachments + subpass->depth_stencil_attachment->attachment;
/* Coefficients taken from AMDVLK */
unsigned depth_coeff = vk_format_is_depth(attachment->format) ? 5 : 0;
unsigned ds_bytes_per_pixel = 4 * (depth_coeff + stencil_coeff) * total_samples;
const struct radv_bin_size_entry *ds_entry = ds_size_table[log_num_rb_per_se][log_num_se];
- while(ds_entry->bpp <= ds_bytes_per_pixel)
+ while(ds_entry[1].bpp <= ds_bytes_per_pixel)
++ds_entry;
extent.width = MIN2(extent.width, ds_entry->extent.width);
}
static void
-radv_pipeline_generate_binning_state(struct radeon_winsys_cs *cs,
+radv_pipeline_generate_binning_state(struct radeon_cmdbuf *ctx_cs,
struct radv_pipeline *pipeline,
const VkGraphicsPipelineCreateInfo *pCreateInfo)
{
VkExtent2D bin_size = radv_compute_bin_size(pipeline, pCreateInfo);
- unsigned context_states_per_bin; /* allowed range: [1, 6] */
- unsigned persistent_states_per_bin; /* allowed range: [1, 32] */
- unsigned fpovs_per_batch; /* allowed range: [0, 255], 0 = unlimited */
-
- switch (pipeline->device->physical_device->rad_info.family) {
- case CHIP_VEGA10:
- case CHIP_VEGA12:
- context_states_per_bin = 1;
- persistent_states_per_bin = 1;
- fpovs_per_batch = 63;
- break;
- case CHIP_RAVEN:
- context_states_per_bin = 6;
- persistent_states_per_bin = 32;
- fpovs_per_batch = 63;
- break;
- default:
- unreachable("unhandled family while determining binning state.");
- }
-
if (pipeline->device->pbb_allowed && bin_size.width && bin_size.height) {
+ unsigned context_states_per_bin; /* allowed range: [1, 6] */
+ unsigned persistent_states_per_bin; /* allowed range: [1, 32] */
+ unsigned fpovs_per_batch; /* allowed range: [0, 255], 0 = unlimited */
+
+ switch (pipeline->device->physical_device->rad_info.family) {
+ case CHIP_VEGA10:
+ case CHIP_VEGA12:
+ case CHIP_VEGA20:
+ context_states_per_bin = 1;
+ persistent_states_per_bin = 1;
+ fpovs_per_batch = 63;
+ break;
+ case CHIP_RAVEN:
+ case CHIP_RAVEN2:
+ context_states_per_bin = 6;
+ persistent_states_per_bin = 32;
+ fpovs_per_batch = 63;
+ break;
+ default:
+ unreachable("unhandled family while determining binning state.");
+ }
+
pa_sc_binner_cntl_0 =
S_028C44_BINNING_MODE(V_028C44_BINNING_ALLOWED) |
S_028C44_BIN_SIZE_X(bin_size.width == 16) |
S_028C44_OPTIMAL_BIN_SELECTION(1);
}
- radeon_set_context_reg(cs, R_028C44_PA_SC_BINNER_CNTL_0,
+ radeon_set_context_reg(ctx_cs, R_028C44_PA_SC_BINNER_CNTL_0,
pa_sc_binner_cntl_0);
- radeon_set_context_reg(cs, R_028060_DB_DFSM_CONTROL,
- db_dfsm_control);
+
+ if (pipeline->device->physical_device->rad_info.chip_class >= GFX10) {
+ radeon_set_context_reg(ctx_cs, R_028038_DB_DFSM_CONTROL,
+ db_dfsm_control);
+ } else {
+ radeon_set_context_reg(ctx_cs, R_028060_DB_DFSM_CONTROL,
+ db_dfsm_control);
+ }
}
static void
-radv_pipeline_generate_depth_stencil_state(struct radeon_winsys_cs *cs,
+radv_pipeline_generate_depth_stencil_state(struct radeon_cmdbuf *ctx_cs,
struct radv_pipeline *pipeline,
const VkGraphicsPipelineCreateInfo *pCreateInfo,
const struct radv_graphics_pipeline_create_info *extra)
const VkPipelineDepthStencilStateCreateInfo *vkds = pCreateInfo->pDepthStencilState;
RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass);
struct radv_subpass *subpass = pass->subpasses + pCreateInfo->subpass;
- struct radv_shader_variant *ps = pipeline->shaders[MESA_SHADER_FRAGMENT];
struct radv_render_pass_attachment *attachment = NULL;
uint32_t db_depth_control = 0, db_stencil_control = 0;
uint32_t db_render_control = 0, db_render_override2 = 0;
uint32_t db_render_override = 0;
- if (subpass->depth_stencil_attachment.attachment != VK_ATTACHMENT_UNUSED)
- attachment = pass->attachments + subpass->depth_stencil_attachment.attachment;
+ if (subpass->depth_stencil_attachment)
+ attachment = pass->attachments + subpass->depth_stencil_attachment->attachment;
bool has_depth_attachment = attachment && vk_format_is_depth(attachment->format);
bool has_stencil_attachment = attachment && vk_format_is_stencil(attachment->format);
db_render_override |= S_02800C_FORCE_HIS_ENABLE0(V_02800C_FORCE_DISABLE) |
S_02800C_FORCE_HIS_ENABLE1(V_02800C_FORCE_DISABLE);
- if (pipeline->device->enabled_extensions.EXT_depth_range_unrestricted &&
- !pCreateInfo->pRasterizationState->depthClampEnable &&
- ps->info.info.ps.writes_z) {
+ if (!pCreateInfo->pRasterizationState->depthClampEnable) {
/* From VK_EXT_depth_range_unrestricted spec:
*
* "The behavior described in Primitive Clipping still applies.
db_render_override |= S_02800C_DISABLE_VIEWPORT_CLAMP(1);
}
- radeon_set_context_reg(cs, R_028800_DB_DEPTH_CONTROL, db_depth_control);
- radeon_set_context_reg(cs, R_02842C_DB_STENCIL_CONTROL, db_stencil_control);
+ radeon_set_context_reg(ctx_cs, R_028800_DB_DEPTH_CONTROL, db_depth_control);
+ radeon_set_context_reg(ctx_cs, R_02842C_DB_STENCIL_CONTROL, db_stencil_control);
- radeon_set_context_reg(cs, R_028000_DB_RENDER_CONTROL, db_render_control);
- radeon_set_context_reg(cs, R_02800C_DB_RENDER_OVERRIDE, db_render_override);
- radeon_set_context_reg(cs, R_028010_DB_RENDER_OVERRIDE2, db_render_override2);
+ radeon_set_context_reg(ctx_cs, R_028000_DB_RENDER_CONTROL, db_render_control);
+ radeon_set_context_reg(ctx_cs, R_02800C_DB_RENDER_OVERRIDE, db_render_override);
+ radeon_set_context_reg(ctx_cs, R_028010_DB_RENDER_OVERRIDE2, db_render_override2);
}
static void
-radv_pipeline_generate_blend_state(struct radeon_winsys_cs *cs,
+radv_pipeline_generate_blend_state(struct radeon_cmdbuf *ctx_cs,
struct radv_pipeline *pipeline,
const struct radv_blend_state *blend)
{
- radeon_set_context_reg_seq(cs, R_028780_CB_BLEND0_CONTROL, 8);
- radeon_emit_array(cs, blend->cb_blend_control,
+ radeon_set_context_reg_seq(ctx_cs, R_028780_CB_BLEND0_CONTROL, 8);
+ radeon_emit_array(ctx_cs, blend->cb_blend_control,
8);
- radeon_set_context_reg(cs, R_028808_CB_COLOR_CONTROL, blend->cb_color_control);
- radeon_set_context_reg(cs, R_028B70_DB_ALPHA_TO_MASK, blend->db_alpha_to_mask);
+ radeon_set_context_reg(ctx_cs, R_028808_CB_COLOR_CONTROL, blend->cb_color_control);
+ radeon_set_context_reg(ctx_cs, R_028B70_DB_ALPHA_TO_MASK, blend->db_alpha_to_mask);
if (pipeline->device->physical_device->has_rbplus) {
- radeon_set_context_reg_seq(cs, R_028760_SX_MRT0_BLEND_OPT, 8);
- radeon_emit_array(cs, blend->sx_mrt_blend_opt, 8);
+ radeon_set_context_reg_seq(ctx_cs, R_028760_SX_MRT0_BLEND_OPT, 8);
+ radeon_emit_array(ctx_cs, blend->sx_mrt_blend_opt, 8);
}
- radeon_set_context_reg(cs, R_028714_SPI_SHADER_COL_FORMAT, blend->spi_shader_col_format);
+ radeon_set_context_reg(ctx_cs, R_028714_SPI_SHADER_COL_FORMAT, blend->spi_shader_col_format);
- radeon_set_context_reg(cs, R_028238_CB_TARGET_MASK, blend->cb_target_mask);
- radeon_set_context_reg(cs, R_02823C_CB_SHADER_MASK, blend->cb_shader_mask);
+ radeon_set_context_reg(ctx_cs, R_028238_CB_TARGET_MASK, blend->cb_target_mask);
+ radeon_set_context_reg(ctx_cs, R_02823C_CB_SHADER_MASK, blend->cb_shader_mask);
pipeline->graphics.col_format = blend->spi_shader_col_format;
pipeline->graphics.cb_target_mask = blend->cb_target_mask;
}
+static const VkConservativeRasterizationModeEXT
+radv_get_conservative_raster_mode(const VkPipelineRasterizationStateCreateInfo *pCreateInfo)
+{
+ const VkPipelineRasterizationConservativeStateCreateInfoEXT *conservative_raster =
+ vk_find_struct_const(pCreateInfo->pNext, PIPELINE_RASTERIZATION_CONSERVATIVE_STATE_CREATE_INFO_EXT);
+
+ if (!conservative_raster)
+ return VK_CONSERVATIVE_RASTERIZATION_MODE_DISABLED_EXT;
+ return conservative_raster->conservativeRasterizationMode;
+}
static void
-radv_pipeline_generate_raster_state(struct radeon_winsys_cs *cs,
+radv_pipeline_generate_raster_state(struct radeon_cmdbuf *ctx_cs,
+ struct radv_pipeline *pipeline,
const VkGraphicsPipelineCreateInfo *pCreateInfo)
{
const VkPipelineRasterizationStateCreateInfo *vkraster = pCreateInfo->pRasterizationState;
+ const VkConservativeRasterizationModeEXT mode =
+ radv_get_conservative_raster_mode(vkraster);
+ uint32_t pa_sc_conservative_rast = S_028C4C_NULL_SQUAD_AA_MASK_ENABLE(1);
+ bool depth_clip_disable = vkraster->depthClampEnable;
+
+ const VkPipelineRasterizationDepthClipStateCreateInfoEXT *depth_clip_state =
+ vk_find_struct_const(vkraster->pNext, PIPELINE_RASTERIZATION_DEPTH_CLIP_STATE_CREATE_INFO_EXT);
+ if (depth_clip_state) {
+ depth_clip_disable = !depth_clip_state->depthClipEnable;
+ }
- radeon_set_context_reg(cs, R_028810_PA_CL_CLIP_CNTL,
- S_028810_PS_UCP_MODE(3) |
+ radeon_set_context_reg(ctx_cs, R_028810_PA_CL_CLIP_CNTL,
S_028810_DX_CLIP_SPACE_DEF(1) | // vulkan uses DX conventions.
- S_028810_ZCLIP_NEAR_DISABLE(vkraster->depthClampEnable ? 1 : 0) |
- S_028810_ZCLIP_FAR_DISABLE(vkraster->depthClampEnable ? 1 : 0) |
+ S_028810_ZCLIP_NEAR_DISABLE(depth_clip_disable ? 1 : 0) |
+ S_028810_ZCLIP_FAR_DISABLE(depth_clip_disable ? 1 : 0) |
S_028810_DX_RASTERIZATION_KILL(vkraster->rasterizerDiscardEnable ? 1 : 0) |
S_028810_DX_LINEAR_ATTR_CLIP_ENA(1));
- radeon_set_context_reg(cs, R_0286D4_SPI_INTERP_CONTROL_0,
+ radeon_set_context_reg(ctx_cs, R_0286D4_SPI_INTERP_CONTROL_0,
S_0286D4_FLAT_SHADE_ENA(1) |
S_0286D4_PNT_SPRITE_ENA(1) |
S_0286D4_PNT_SPRITE_OVRD_X(V_0286D4_SPI_PNT_SPRITE_SEL_S) |
S_0286D4_PNT_SPRITE_OVRD_W(V_0286D4_SPI_PNT_SPRITE_SEL_1) |
S_0286D4_PNT_SPRITE_TOP_1(0)); /* vulkan is top to bottom - 1.0 at bottom */
- radeon_set_context_reg(cs, R_028BE4_PA_SU_VTX_CNTL,
+ radeon_set_context_reg(ctx_cs, R_028BE4_PA_SU_VTX_CNTL,
S_028BE4_PIX_CENTER(1) | // TODO verify
S_028BE4_ROUND_MODE(V_028BE4_X_ROUND_TO_EVEN) |
S_028BE4_QUANT_MODE(V_028BE4_X_16_8_FIXED_POINT_1_256TH));
- radeon_set_context_reg(cs, R_028814_PA_SU_SC_MODE_CNTL,
+ radeon_set_context_reg(ctx_cs, R_028814_PA_SU_SC_MODE_CNTL,
S_028814_FACE(vkraster->frontFace) |
S_028814_CULL_FRONT(!!(vkraster->cullMode & VK_CULL_MODE_FRONT_BIT)) |
S_028814_CULL_BACK(!!(vkraster->cullMode & VK_CULL_MODE_BACK_BIT)) |
S_028814_POLY_OFFSET_FRONT_ENABLE(vkraster->depthBiasEnable ? 1 : 0) |
S_028814_POLY_OFFSET_BACK_ENABLE(vkraster->depthBiasEnable ? 1 : 0) |
S_028814_POLY_OFFSET_PARA_ENABLE(vkraster->depthBiasEnable ? 1 : 0));
+
+ /* Conservative rasterization. */
+ if (mode != VK_CONSERVATIVE_RASTERIZATION_MODE_DISABLED_EXT) {
+ struct radv_multisample_state *ms = &pipeline->graphics.ms;
+
+ ms->pa_sc_aa_config |= S_028BE0_AA_MASK_CENTROID_DTMN(1);
+ ms->db_eqaa |= S_028804_ENABLE_POSTZ_OVERRASTERIZATION(1) |
+ S_028804_OVERRASTERIZATION_AMOUNT(4);
+
+ pa_sc_conservative_rast = S_028C4C_PREZ_AA_MASK_ENABLE(1) |
+ S_028C4C_POSTZ_AA_MASK_ENABLE(1) |
+ S_028C4C_CENTROID_SAMPLE_OVERRIDE(1);
+
+ if (mode == VK_CONSERVATIVE_RASTERIZATION_MODE_OVERESTIMATE_EXT) {
+ pa_sc_conservative_rast |=
+ S_028C4C_OVER_RAST_ENABLE(1) |
+ S_028C4C_OVER_RAST_SAMPLE_SELECT(0) |
+ S_028C4C_UNDER_RAST_ENABLE(0) |
+ S_028C4C_UNDER_RAST_SAMPLE_SELECT(1) |
+ S_028C4C_PBB_UNCERTAINTY_REGION_ENABLE(1);
+ } else {
+ assert(mode == VK_CONSERVATIVE_RASTERIZATION_MODE_UNDERESTIMATE_EXT);
+ pa_sc_conservative_rast |=
+ S_028C4C_OVER_RAST_ENABLE(0) |
+ S_028C4C_OVER_RAST_SAMPLE_SELECT(1) |
+ S_028C4C_UNDER_RAST_ENABLE(1) |
+ S_028C4C_UNDER_RAST_SAMPLE_SELECT(0) |
+ S_028C4C_PBB_UNCERTAINTY_REGION_ENABLE(0);
+ }
+ }
+
+ radeon_set_context_reg(ctx_cs, R_028C4C_PA_SC_CONSERVATIVE_RASTERIZATION_CNTL,
+ pa_sc_conservative_rast);
}
static void
-radv_pipeline_generate_multisample_state(struct radeon_winsys_cs *cs,
+radv_pipeline_generate_multisample_state(struct radeon_cmdbuf *ctx_cs,
struct radv_pipeline *pipeline)
{
struct radv_multisample_state *ms = &pipeline->graphics.ms;
- radeon_set_context_reg_seq(cs, R_028C38_PA_SC_AA_MASK_X0Y0_X1Y0, 2);
- radeon_emit(cs, ms->pa_sc_aa_mask[0]);
- radeon_emit(cs, ms->pa_sc_aa_mask[1]);
-
- radeon_set_context_reg(cs, R_028804_DB_EQAA, ms->db_eqaa);
- radeon_set_context_reg(cs, R_028A4C_PA_SC_MODE_CNTL_1, ms->pa_sc_mode_cntl_1);
-
- if (pipeline->shaders[MESA_SHADER_FRAGMENT]->info.info.ps.needs_sample_positions) {
- uint32_t offset;
- struct radv_userdata_info *loc = radv_lookup_user_sgpr(pipeline, MESA_SHADER_FRAGMENT, AC_UD_PS_SAMPLE_POS_OFFSET);
- uint32_t base_reg = pipeline->user_data_0[MESA_SHADER_FRAGMENT];
- if (loc->sgpr_idx == -1)
- return;
- assert(loc->num_sgprs == 1);
- assert(!loc->indirect);
- switch (pipeline->graphics.ms.num_samples) {
- default:
- offset = 0;
- break;
- case 2:
- offset = 1;
- break;
- case 4:
- offset = 3;
- break;
- case 8:
- offset = 7;
- break;
- case 16:
- offset = 15;
- break;
- }
+ radeon_set_context_reg_seq(ctx_cs, R_028C38_PA_SC_AA_MASK_X0Y0_X1Y0, 2);
+ radeon_emit(ctx_cs, ms->pa_sc_aa_mask[0]);
+ radeon_emit(ctx_cs, ms->pa_sc_aa_mask[1]);
- radeon_set_sh_reg(cs, base_reg + loc->sgpr_idx * 4, offset);
- }
+ radeon_set_context_reg(ctx_cs, R_028804_DB_EQAA, ms->db_eqaa);
+ radeon_set_context_reg(ctx_cs, R_028A4C_PA_SC_MODE_CNTL_1, ms->pa_sc_mode_cntl_1);
+
+ /* The exclusion bits can be set to improve rasterization efficiency
+ * if no sample lies on the pixel boundary (-8 sample offset). It's
+ * currently always TRUE because the driver doesn't support 16 samples.
+ */
+ bool exclusion = pipeline->device->physical_device->rad_info.chip_class >= GFX7;
+ radeon_set_context_reg(ctx_cs, R_02882C_PA_SU_PRIM_FILTER_CNTL,
+ S_02882C_XMAX_RIGHT_EXCLUSION(exclusion) |
+ S_02882C_YMAX_BOTTOM_EXCLUSION(exclusion));
}
static void
-radv_pipeline_generate_vgt_gs_mode(struct radeon_winsys_cs *cs,
- const struct radv_pipeline *pipeline)
+radv_pipeline_generate_vgt_gs_mode(struct radeon_cmdbuf *ctx_cs,
+ struct radv_pipeline *pipeline)
{
const struct radv_vs_output_info *outinfo = get_vs_output_info(pipeline);
-
- uint32_t vgt_primitiveid_en = false;
+ unsigned vgt_primitiveid_en = 0;
uint32_t vgt_gs_mode = 0;
if (radv_pipeline_has_gs(pipeline)) {
vgt_gs_mode = ac_vgt_gs_mode(gs->info.gs.vertices_out,
pipeline->device->physical_device->rad_info.chip_class);
+ } else if (radv_pipeline_has_ngg(pipeline)) {
+ const struct radv_shader_variant *vs =
+ pipeline->shaders[MESA_SHADER_VERTEX];
+ bool enable_prim_id =
+ outinfo->export_prim_id || vs->info.info.uses_prim_id;
+
+ vgt_primitiveid_en |= S_028A84_PRIMITIVEID_EN(enable_prim_id) |
+ S_028A84_NGG_DISABLE_PROVOK_REUSE(enable_prim_id);
} else if (outinfo->export_prim_id) {
vgt_gs_mode = S_028A40_MODE(V_028A40_GS_SCENARIO_A);
- vgt_primitiveid_en = true;
+ vgt_primitiveid_en |= S_028A84_PRIMITIVEID_EN(1);
}
- radeon_set_context_reg(cs, R_028A84_VGT_PRIMITIVEID_EN, vgt_primitiveid_en);
- radeon_set_context_reg(cs, R_028A40_VGT_GS_MODE, vgt_gs_mode);
+ radeon_set_context_reg(ctx_cs, R_028A84_VGT_PRIMITIVEID_EN, vgt_primitiveid_en);
+ radeon_set_context_reg(ctx_cs, R_028A40_VGT_GS_MODE, vgt_gs_mode);
}
static void
-radv_pipeline_generate_hw_vs(struct radeon_winsys_cs *cs,
+radv_pipeline_generate_hw_vs(struct radeon_cmdbuf *ctx_cs,
+ struct radeon_cmdbuf *cs,
struct radv_pipeline *pipeline,
struct radv_shader_variant *shader)
{
radeon_set_sh_reg_seq(cs, R_00B120_SPI_SHADER_PGM_LO_VS, 4);
radeon_emit(cs, va >> 8);
radeon_emit(cs, S_00B124_MEM_BASE(va >> 40));
- radeon_emit(cs, shader->rsrc1);
- radeon_emit(cs, shader->rsrc2);
+ radeon_emit(cs, shader->config.rsrc1);
+ radeon_emit(cs, shader->config.rsrc2);
const struct radv_vs_output_info *outinfo = get_vs_output_info(pipeline);
unsigned clip_dist_mask, cull_dist_mask, total_mask;
outinfo->writes_layer ||
outinfo->writes_viewport_index;
- radeon_set_context_reg(cs, R_0286C4_SPI_VS_OUT_CONFIG,
+ radeon_set_context_reg(ctx_cs, R_0286C4_SPI_VS_OUT_CONFIG,
S_0286C4_VS_EXPORT_COUNT(MAX2(1, outinfo->param_exports) - 1));
- radeon_set_context_reg(cs, R_02870C_SPI_SHADER_POS_FORMAT,
+ radeon_set_context_reg(ctx_cs, R_02870C_SPI_SHADER_POS_FORMAT,
S_02870C_POS0_EXPORT_FORMAT(V_02870C_SPI_SHADER_4COMP) |
S_02870C_POS1_EXPORT_FORMAT(outinfo->pos_exports > 1 ?
V_02870C_SPI_SHADER_4COMP :
V_02870C_SPI_SHADER_4COMP :
V_02870C_SPI_SHADER_NONE));
- radeon_set_context_reg(cs, R_028818_PA_CL_VTE_CNTL,
+ radeon_set_context_reg(ctx_cs, R_028818_PA_CL_VTE_CNTL,
S_028818_VTX_W0_FMT(1) |
S_028818_VPORT_X_SCALE_ENA(1) | S_028818_VPORT_X_OFFSET_ENA(1) |
S_028818_VPORT_Y_SCALE_ENA(1) | S_028818_VPORT_Y_OFFSET_ENA(1) |
S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1));
- radeon_set_context_reg(cs, R_02881C_PA_CL_VS_OUT_CNTL,
+ radeon_set_context_reg(ctx_cs, R_02881C_PA_CL_VS_OUT_CNTL,
S_02881C_USE_VTX_POINT_SIZE(outinfo->writes_pointsize) |
S_02881C_USE_VTX_RENDER_TARGET_INDX(outinfo->writes_layer) |
S_02881C_USE_VTX_VIEWPORT_INDX(outinfo->writes_viewport_index) |
cull_dist_mask << 8 |
clip_dist_mask);
- if (pipeline->device->physical_device->rad_info.chip_class <= VI)
- radeon_set_context_reg(cs, R_028AB4_VGT_REUSE_OFF,
+ if (pipeline->device->physical_device->rad_info.chip_class <= GFX8)
+ radeon_set_context_reg(ctx_cs, R_028AB4_VGT_REUSE_OFF,
outinfo->writes_viewport_index);
}
static void
-radv_pipeline_generate_hw_es(struct radeon_winsys_cs *cs,
+radv_pipeline_generate_hw_es(struct radeon_cmdbuf *cs,
struct radv_pipeline *pipeline,
struct radv_shader_variant *shader)
{
radeon_set_sh_reg_seq(cs, R_00B320_SPI_SHADER_PGM_LO_ES, 4);
radeon_emit(cs, va >> 8);
radeon_emit(cs, S_00B324_MEM_BASE(va >> 40));
- radeon_emit(cs, shader->rsrc1);
- radeon_emit(cs, shader->rsrc2);
+ radeon_emit(cs, shader->config.rsrc1);
+ radeon_emit(cs, shader->config.rsrc2);
}
static void
-radv_pipeline_generate_hw_ls(struct radeon_winsys_cs *cs,
+radv_pipeline_generate_hw_ls(struct radeon_cmdbuf *cs,
struct radv_pipeline *pipeline,
struct radv_shader_variant *shader,
const struct radv_tessellation_state *tess)
{
uint64_t va = radv_buffer_get_va(shader->bo) + shader->bo_offset;
- uint32_t rsrc2 = shader->rsrc2;
+ uint32_t rsrc2 = shader->config.rsrc2;
radeon_set_sh_reg_seq(cs, R_00B520_SPI_SHADER_PGM_LO_LS, 2);
radeon_emit(cs, va >> 8);
radeon_emit(cs, S_00B524_MEM_BASE(va >> 40));
rsrc2 |= S_00B52C_LDS_SIZE(tess->lds_size);
- if (pipeline->device->physical_device->rad_info.chip_class == CIK &&
+ if (pipeline->device->physical_device->rad_info.chip_class == GFX7 &&
pipeline->device->physical_device->rad_info.family != CHIP_HAWAII)
radeon_set_sh_reg(cs, R_00B52C_SPI_SHADER_PGM_RSRC2_LS, rsrc2);
radeon_set_sh_reg_seq(cs, R_00B528_SPI_SHADER_PGM_RSRC1_LS, 2);
- radeon_emit(cs, shader->rsrc1);
+ radeon_emit(cs, shader->config.rsrc1);
radeon_emit(cs, rsrc2);
}
static void
-radv_pipeline_generate_hw_hs(struct radeon_winsys_cs *cs,
+radv_pipeline_generate_hw_ngg(struct radeon_cmdbuf *ctx_cs,
+ struct radeon_cmdbuf *cs,
+ struct radv_pipeline *pipeline,
+ struct radv_shader_variant *shader,
+ const struct radv_ngg_state *ngg_state)
+{
+ uint64_t va = radv_buffer_get_va(shader->bo) + shader->bo_offset;
+
+ radeon_set_sh_reg_seq(cs, R_00B320_SPI_SHADER_PGM_LO_ES, 2);
+ radeon_emit(cs, va >> 8);
+ radeon_emit(cs, va >> 40);
+ radeon_set_sh_reg_seq(cs, R_00B228_SPI_SHADER_PGM_RSRC1_GS, 2);
+ radeon_emit(cs, shader->config.rsrc1);
+ radeon_emit(cs, shader->config.rsrc2);
+
+ const struct radv_vs_output_info *outinfo = get_vs_output_info(pipeline);
+ unsigned clip_dist_mask, cull_dist_mask, total_mask;
+ clip_dist_mask = outinfo->clip_dist_mask;
+ cull_dist_mask = outinfo->cull_dist_mask;
+ total_mask = clip_dist_mask | cull_dist_mask;
+ bool misc_vec_ena = outinfo->writes_pointsize ||
+ outinfo->writes_layer ||
+ outinfo->writes_viewport_index;
+ bool break_wave_at_eoi = false;
+
+ radeon_set_context_reg(ctx_cs, R_0286C4_SPI_VS_OUT_CONFIG,
+ S_0286C4_VS_EXPORT_COUNT(MAX2(1, outinfo->param_exports) - 1));
+ radeon_set_context_reg(ctx_cs, R_028708_SPI_SHADER_IDX_FORMAT,
+ S_028708_IDX0_EXPORT_FORMAT(V_028708_SPI_SHADER_1COMP));
+ radeon_set_context_reg(ctx_cs, R_02870C_SPI_SHADER_POS_FORMAT,
+ S_02870C_POS0_EXPORT_FORMAT(V_02870C_SPI_SHADER_4COMP) |
+ S_02870C_POS1_EXPORT_FORMAT(outinfo->pos_exports > 1 ?
+ V_02870C_SPI_SHADER_4COMP :
+ V_02870C_SPI_SHADER_NONE) |
+ S_02870C_POS2_EXPORT_FORMAT(outinfo->pos_exports > 2 ?
+ V_02870C_SPI_SHADER_4COMP :
+ V_02870C_SPI_SHADER_NONE) |
+ S_02870C_POS3_EXPORT_FORMAT(outinfo->pos_exports > 3 ?
+ V_02870C_SPI_SHADER_4COMP :
+ V_02870C_SPI_SHADER_NONE));
+
+ radeon_set_context_reg(ctx_cs, R_028818_PA_CL_VTE_CNTL,
+ S_028818_VTX_W0_FMT(1) |
+ S_028818_VPORT_X_SCALE_ENA(1) | S_028818_VPORT_X_OFFSET_ENA(1) |
+ S_028818_VPORT_Y_SCALE_ENA(1) | S_028818_VPORT_Y_OFFSET_ENA(1) |
+ S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1));
+ radeon_set_context_reg(ctx_cs, R_02881C_PA_CL_VS_OUT_CNTL,
+ S_02881C_USE_VTX_POINT_SIZE(outinfo->writes_pointsize) |
+ S_02881C_USE_VTX_RENDER_TARGET_INDX(outinfo->writes_layer) |
+ S_02881C_USE_VTX_VIEWPORT_INDX(outinfo->writes_viewport_index) |
+ S_02881C_VS_OUT_MISC_VEC_ENA(misc_vec_ena) |
+ S_02881C_VS_OUT_MISC_SIDE_BUS_ENA(misc_vec_ena) |
+ S_02881C_VS_OUT_CCDIST0_VEC_ENA((total_mask & 0x0f) != 0) |
+ S_02881C_VS_OUT_CCDIST1_VEC_ENA((total_mask & 0xf0) != 0) |
+ cull_dist_mask << 8 |
+ clip_dist_mask);
+
+ /* TODO: Correctly set REUSE_OFF */
+ radeon_set_context_reg(ctx_cs, R_028AB4_VGT_REUSE_OFF,
+ S_028AB4_REUSE_OFF(0));
+ radeon_set_context_reg(ctx_cs, R_028AAC_VGT_ESGS_RING_ITEMSIZE,
+ ngg_state->vgt_esgs_ring_itemsize);
+
+ /* NGG specific registers. */
+ struct radv_shader_variant *gs = pipeline->shaders[MESA_SHADER_GEOMETRY];
+ uint32_t gs_num_invocations = gs ? gs->info.gs.invocations : 1;
+
+ radeon_set_context_reg(ctx_cs, R_028A44_VGT_GS_ONCHIP_CNTL,
+ S_028A44_ES_VERTS_PER_SUBGRP(ngg_state->hw_max_esverts) |
+ S_028A44_GS_PRIMS_PER_SUBGRP(ngg_state->max_gsprims) |
+ S_028A44_GS_INST_PRIMS_IN_SUBGRP(ngg_state->max_gsprims * gs_num_invocations));
+ radeon_set_context_reg(ctx_cs, R_0287FC_GE_MAX_OUTPUT_PER_SUBGROUP,
+ S_0287FC_MAX_VERTS_PER_SUBGROUP(ngg_state->max_out_verts));
+ radeon_set_context_reg(ctx_cs, R_028B4C_GE_NGG_SUBGRP_CNTL,
+ S_028B4C_PRIM_AMP_FACTOR(ngg_state->prim_amp_factor) |
+ S_028B4C_THDS_PER_SUBGRP(0)); /* for fast launch */
+ radeon_set_context_reg(ctx_cs, R_028B90_VGT_GS_INSTANCE_CNT,
+ S_028B90_CNT(gs_num_invocations) |
+ S_028B90_ENABLE(gs_num_invocations > 1) |
+ S_028B90_EN_MAX_VERT_OUT_PER_GS_INSTANCE(ngg_state->max_vert_out_per_gs_instance));
+
+ /* User edge flags are set by the pos exports. If user edge flags are
+ * not used, we must use hw-generated edge flags and pass them via
+ * the prim export to prevent drawing lines on internal edges of
+ * decomposed primitives (such as quads) with polygon mode = lines.
+ *
+ * TODO: We should combine hw-generated edge flags with user edge
+ * flags in the shader.
+ */
+ radeon_set_context_reg(ctx_cs, R_028838_PA_CL_NGG_CNTL,
+ S_028838_INDEX_BUF_EDGE_FLAG_ENA(!radv_pipeline_has_tess(pipeline) &&
+ !radv_pipeline_has_gs(pipeline)));
+
+ radeon_set_context_reg(ctx_cs, R_03096C_GE_CNTL,
+ S_03096C_PRIM_GRP_SIZE(ngg_state->max_gsprims) |
+ S_03096C_VERT_GRP_SIZE(ngg_state->hw_max_esverts) |
+ S_03096C_BREAK_WAVE_AT_EOI(break_wave_at_eoi));
+}
+
+static void
+radv_pipeline_generate_hw_hs(struct radeon_cmdbuf *cs,
struct radv_pipeline *pipeline,
struct radv_shader_variant *shader,
const struct radv_tessellation_state *tess)
uint64_t va = radv_buffer_get_va(shader->bo) + shader->bo_offset;
if (pipeline->device->physical_device->rad_info.chip_class >= GFX9) {
- radeon_set_sh_reg_seq(cs, R_00B410_SPI_SHADER_PGM_LO_LS, 2);
- radeon_emit(cs, va >> 8);
- radeon_emit(cs, S_00B414_MEM_BASE(va >> 40));
+ unsigned hs_rsrc2 = shader->config.rsrc2;
+
+ if (pipeline->device->physical_device->rad_info.chip_class >= GFX10) {
+ hs_rsrc2 |= S_00B42C_LDS_SIZE_GFX10(tess->lds_size);
+ } else {
+ hs_rsrc2 |= S_00B42C_LDS_SIZE_GFX9(tess->lds_size);
+ }
+
+ if (pipeline->device->physical_device->rad_info.chip_class >= GFX10) {
+ radeon_set_sh_reg_seq(cs, R_00B520_SPI_SHADER_PGM_LO_LS, 2);
+ radeon_emit(cs, va >> 8);
+ radeon_emit(cs, S_00B524_MEM_BASE(va >> 40));
+ } else {
+ radeon_set_sh_reg_seq(cs, R_00B410_SPI_SHADER_PGM_LO_LS, 2);
+ radeon_emit(cs, va >> 8);
+ radeon_emit(cs, S_00B414_MEM_BASE(va >> 40));
+ }
radeon_set_sh_reg_seq(cs, R_00B428_SPI_SHADER_PGM_RSRC1_HS, 2);
- radeon_emit(cs, shader->rsrc1);
- radeon_emit(cs, shader->rsrc2 |
- S_00B42C_LDS_SIZE(tess->lds_size));
+ radeon_emit(cs, shader->config.rsrc1);
+ radeon_emit(cs, hs_rsrc2);
} else {
radeon_set_sh_reg_seq(cs, R_00B420_SPI_SHADER_PGM_LO_HS, 4);
radeon_emit(cs, va >> 8);
radeon_emit(cs, S_00B424_MEM_BASE(va >> 40));
- radeon_emit(cs, shader->rsrc1);
- radeon_emit(cs, shader->rsrc2);
+ radeon_emit(cs, shader->config.rsrc1);
+ radeon_emit(cs, shader->config.rsrc2);
}
}
static void
-radv_pipeline_generate_vertex_shader(struct radeon_winsys_cs *cs,
+radv_pipeline_generate_vertex_shader(struct radeon_cmdbuf *ctx_cs,
+ struct radeon_cmdbuf *cs,
struct radv_pipeline *pipeline,
- const struct radv_tessellation_state *tess)
+ const struct radv_tessellation_state *tess,
+ const struct radv_ngg_state *ngg)
{
struct radv_shader_variant *vs;
radv_pipeline_generate_hw_ls(cs, pipeline, vs, tess);
else if (vs->info.vs.as_es)
radv_pipeline_generate_hw_es(cs, pipeline, vs);
+ else if (vs->info.is_ngg)
+ radv_pipeline_generate_hw_ngg(ctx_cs, cs, pipeline, vs, ngg);
else
- radv_pipeline_generate_hw_vs(cs, pipeline, vs);
+ radv_pipeline_generate_hw_vs(ctx_cs, cs, pipeline, vs);
}
static void
-radv_pipeline_generate_tess_shaders(struct radeon_winsys_cs *cs,
+radv_pipeline_generate_tess_shaders(struct radeon_cmdbuf *ctx_cs,
+ struct radeon_cmdbuf *cs,
struct radv_pipeline *pipeline,
const struct radv_tessellation_state *tess)
{
if (tes->info.tes.as_es)
radv_pipeline_generate_hw_es(cs, pipeline, tes);
else
- radv_pipeline_generate_hw_vs(cs, pipeline, tes);
+ radv_pipeline_generate_hw_vs(ctx_cs, cs, pipeline, tes);
}
radv_pipeline_generate_hw_hs(cs, pipeline, tcs, tess);
- radeon_set_context_reg(cs, R_028B6C_VGT_TF_PARAM,
+ radeon_set_context_reg(ctx_cs, R_028B6C_VGT_TF_PARAM,
tess->tf_param);
- if (pipeline->device->physical_device->rad_info.chip_class >= CIK)
- radeon_set_context_reg_idx(cs, R_028B58_VGT_LS_HS_CONFIG, 2,
+ if (pipeline->device->physical_device->rad_info.chip_class >= GFX7)
+ radeon_set_context_reg_idx(ctx_cs, R_028B58_VGT_LS_HS_CONFIG, 2,
tess->ls_hs_config);
else
- radeon_set_context_reg(cs, R_028B58_VGT_LS_HS_CONFIG,
+ radeon_set_context_reg(ctx_cs, R_028B58_VGT_LS_HS_CONFIG,
tess->ls_hs_config);
}
static void
-radv_pipeline_generate_geometry_shader(struct radeon_winsys_cs *cs,
+radv_pipeline_generate_geometry_shader(struct radeon_cmdbuf *ctx_cs,
+ struct radeon_cmdbuf *cs,
struct radv_pipeline *pipeline,
const struct radv_gs_state *gs_state)
{
struct radv_shader_variant *gs;
+ unsigned gs_max_out_vertices;
+ uint8_t *num_components;
+ uint8_t max_stream;
+ unsigned offset;
uint64_t va;
gs = pipeline->shaders[MESA_SHADER_GEOMETRY];
if (!gs)
return;
- uint32_t gsvs_itemsize = gs->info.gs.max_gsvs_emit_size >> 2;
+ gs_max_out_vertices = gs->info.gs.vertices_out;
+ max_stream = gs->info.info.gs.max_stream;
+ num_components = gs->info.info.gs.num_stream_output_components;
- radeon_set_context_reg_seq(cs, R_028A60_VGT_GSVS_RING_OFFSET_1, 3);
- radeon_emit(cs, gsvs_itemsize);
- radeon_emit(cs, gsvs_itemsize);
- radeon_emit(cs, gsvs_itemsize);
+ offset = num_components[0] * gs_max_out_vertices;
- radeon_set_context_reg(cs, R_028AB0_VGT_GSVS_RING_ITEMSIZE, gsvs_itemsize);
+ radeon_set_context_reg_seq(ctx_cs, R_028A60_VGT_GSVS_RING_OFFSET_1, 3);
+ radeon_emit(ctx_cs, offset);
+ if (max_stream >= 1)
+ offset += num_components[1] * gs_max_out_vertices;
+ radeon_emit(ctx_cs, offset);
+ if (max_stream >= 2)
+ offset += num_components[2] * gs_max_out_vertices;
+ radeon_emit(ctx_cs, offset);
+ if (max_stream >= 3)
+ offset += num_components[3] * gs_max_out_vertices;
+ radeon_set_context_reg(ctx_cs, R_028AB0_VGT_GSVS_RING_ITEMSIZE, offset);
- radeon_set_context_reg(cs, R_028B38_VGT_GS_MAX_VERT_OUT, gs->info.gs.vertices_out);
+ radeon_set_context_reg(ctx_cs, R_028B38_VGT_GS_MAX_VERT_OUT, gs->info.gs.vertices_out);
- uint32_t gs_vert_itemsize = gs->info.gs.gsvs_vertex_size;
- radeon_set_context_reg_seq(cs, R_028B5C_VGT_GS_VERT_ITEMSIZE, 4);
- radeon_emit(cs, gs_vert_itemsize >> 2);
- radeon_emit(cs, 0);
- radeon_emit(cs, 0);
- radeon_emit(cs, 0);
+ radeon_set_context_reg_seq(ctx_cs, R_028B5C_VGT_GS_VERT_ITEMSIZE, 4);
+ radeon_emit(ctx_cs, num_components[0]);
+ radeon_emit(ctx_cs, (max_stream >= 1) ? num_components[1] : 0);
+ radeon_emit(ctx_cs, (max_stream >= 2) ? num_components[2] : 0);
+ radeon_emit(ctx_cs, (max_stream >= 3) ? num_components[3] : 0);
uint32_t gs_num_invocations = gs->info.gs.invocations;
- radeon_set_context_reg(cs, R_028B90_VGT_GS_INSTANCE_CNT,
+ radeon_set_context_reg(ctx_cs, R_028B90_VGT_GS_INSTANCE_CNT,
S_028B90_CNT(MIN2(gs_num_invocations, 127)) |
S_028B90_ENABLE(gs_num_invocations > 0));
- radeon_set_context_reg(cs, R_028AAC_VGT_ESGS_RING_ITEMSIZE,
+ radeon_set_context_reg(ctx_cs, R_028AAC_VGT_ESGS_RING_ITEMSIZE,
gs_state->vgt_esgs_ring_itemsize);
va = radv_buffer_get_va(gs->bo) + gs->bo_offset;
if (pipeline->device->physical_device->rad_info.chip_class >= GFX9) {
- radeon_set_sh_reg_seq(cs, R_00B210_SPI_SHADER_PGM_LO_ES, 2);
- radeon_emit(cs, va >> 8);
- radeon_emit(cs, S_00B214_MEM_BASE(va >> 40));
+ if (pipeline->device->physical_device->rad_info.chip_class >= GFX10) {
+ radeon_set_sh_reg_seq(cs, R_00B320_SPI_SHADER_PGM_LO_ES, 2);
+ radeon_emit(cs, va >> 8);
+ radeon_emit(cs, S_00B324_MEM_BASE(va >> 40));
+ } else {
+ radeon_set_sh_reg_seq(cs, R_00B210_SPI_SHADER_PGM_LO_ES, 2);
+ radeon_emit(cs, va >> 8);
+ radeon_emit(cs, S_00B214_MEM_BASE(va >> 40));
+ }
radeon_set_sh_reg_seq(cs, R_00B228_SPI_SHADER_PGM_RSRC1_GS, 2);
- radeon_emit(cs, gs->rsrc1);
- radeon_emit(cs, gs->rsrc2 | S_00B22C_LDS_SIZE(gs_state->lds_size));
+ radeon_emit(cs, gs->config.rsrc1);
+ radeon_emit(cs, gs->config.rsrc2 | S_00B22C_LDS_SIZE(gs_state->lds_size));
- radeon_set_context_reg(cs, R_028A44_VGT_GS_ONCHIP_CNTL, gs_state->vgt_gs_onchip_cntl);
- radeon_set_context_reg(cs, R_028A94_VGT_GS_MAX_PRIMS_PER_SUBGROUP, gs_state->vgt_gs_max_prims_per_subgroup);
+ radeon_set_context_reg(ctx_cs, R_028A44_VGT_GS_ONCHIP_CNTL, gs_state->vgt_gs_onchip_cntl);
+ radeon_set_context_reg(ctx_cs, R_028A94_VGT_GS_MAX_PRIMS_PER_SUBGROUP, gs_state->vgt_gs_max_prims_per_subgroup);
} else {
radeon_set_sh_reg_seq(cs, R_00B220_SPI_SHADER_PGM_LO_GS, 4);
radeon_emit(cs, va >> 8);
radeon_emit(cs, S_00B224_MEM_BASE(va >> 40));
- radeon_emit(cs, gs->rsrc1);
- radeon_emit(cs, gs->rsrc2);
+ radeon_emit(cs, gs->config.rsrc1);
+ radeon_emit(cs, gs->config.rsrc2);
}
- radv_pipeline_generate_hw_vs(cs, pipeline, pipeline->gs_copy_shader);
+ radv_pipeline_generate_hw_vs(ctx_cs, cs, pipeline, pipeline->gs_copy_shader);
}
-static uint32_t offset_to_ps_input(uint32_t offset, bool flat_shade)
+static uint32_t offset_to_ps_input(uint32_t offset, bool flat_shade, bool float16)
{
uint32_t ps_input_cntl;
if (offset <= AC_EXP_PARAM_OFFSET_31) {
ps_input_cntl = S_028644_OFFSET(offset);
if (flat_shade)
ps_input_cntl |= S_028644_FLAT_SHADE(1);
+ if (float16) {
+ ps_input_cntl |= S_028644_FP16_INTERP_MODE(1) |
+ S_028644_ATTR0_VALID(1);
+ }
} else {
/* The input is a DEFAULT_VAL constant. */
assert(offset >= AC_EXP_PARAM_DEFAULT_VAL_0000 &&
}
static void
-radv_pipeline_generate_ps_inputs(struct radeon_winsys_cs *cs,
- struct radv_pipeline *pipeline)
+radv_pipeline_generate_ps_inputs(struct radeon_cmdbuf *ctx_cs,
+ struct radv_pipeline *pipeline)
{
struct radv_shader_variant *ps = pipeline->shaders[MESA_SHADER_FRAGMENT];
const struct radv_vs_output_info *outinfo = get_vs_output_info(pipeline);
if (ps->info.info.ps.prim_id_input) {
unsigned vs_offset = outinfo->vs_output_param_offset[VARYING_SLOT_PRIMITIVE_ID];
if (vs_offset != AC_EXP_PARAM_UNDEFINED) {
- ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, true);
+ ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, true, false);
++ps_offset;
}
}
if (ps->info.info.ps.layer_input ||
- ps->info.info.ps.uses_input_attachments ||
ps->info.info.needs_multiview_view_index) {
unsigned vs_offset = outinfo->vs_output_param_offset[VARYING_SLOT_LAYER];
if (vs_offset != AC_EXP_PARAM_UNDEFINED)
- ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, true);
+ ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, true, false);
else
- ps_input_cntl[ps_offset] = offset_to_ps_input(AC_EXP_PARAM_DEFAULT_VAL_0000, true);
+ ps_input_cntl[ps_offset] = offset_to_ps_input(AC_EXP_PARAM_DEFAULT_VAL_0000, true, false);
++ps_offset;
}
ps_offset++;
}
+ if (ps->info.info.ps.num_input_clips_culls) {
+ unsigned vs_offset;
+
+ vs_offset = outinfo->vs_output_param_offset[VARYING_SLOT_CLIP_DIST0];
+ if (vs_offset != AC_EXP_PARAM_UNDEFINED) {
+ ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, false, false);
+ ++ps_offset;
+ }
+
+ vs_offset = outinfo->vs_output_param_offset[VARYING_SLOT_CLIP_DIST1];
+ if (vs_offset != AC_EXP_PARAM_UNDEFINED &&
+ ps->info.info.ps.num_input_clips_culls > 4) {
+ ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, false, false);
+ ++ps_offset;
+ }
+ }
+
for (unsigned i = 0; i < 32 && (1u << i) <= ps->info.fs.input_mask; ++i) {
unsigned vs_offset;
bool flat_shade;
+ bool float16;
if (!(ps->info.fs.input_mask & (1u << i)))
continue;
}
flat_shade = !!(ps->info.fs.flat_shaded_mask & (1u << ps_offset));
+ float16 = !!(ps->info.fs.float16_shaded_mask & (1u << ps_offset));
- ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, flat_shade);
+ ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, flat_shade, float16);
++ps_offset;
}
if (ps_offset) {
- radeon_set_context_reg_seq(cs, R_028644_SPI_PS_INPUT_CNTL_0, ps_offset);
+ radeon_set_context_reg_seq(ctx_cs, R_028644_SPI_PS_INPUT_CNTL_0, ps_offset);
for (unsigned i = 0; i < ps_offset; i++) {
- radeon_emit(cs, ps_input_cntl[i]);
+ radeon_emit(ctx_cs, ps_input_cntl[i]);
}
}
}
static uint32_t
radv_compute_db_shader_control(const struct radv_device *device,
+ const struct radv_pipeline *pipeline,
const struct radv_shader_variant *ps)
{
unsigned z_order;
bool disable_rbplus = device->physical_device->has_rbplus &&
!device->physical_device->rbplus_allowed;
+ /* It shouldn't be needed to export gl_SampleMask when MSAA is disabled
+ * but this appears to break Project Cars (DXVK). See
+ * https://bugs.freedesktop.org/show_bug.cgi?id=109401
+ */
+ bool mask_export_enable = ps->info.info.ps.writes_sample_mask;
+
return S_02880C_Z_EXPORT_ENABLE(ps->info.info.ps.writes_z) |
S_02880C_STENCIL_TEST_VAL_EXPORT_ENABLE(ps->info.info.ps.writes_stencil) |
S_02880C_KILL_ENABLE(!!ps->info.fs.can_discard) |
- S_02880C_MASK_EXPORT_ENABLE(ps->info.info.ps.writes_sample_mask) |
+ S_02880C_MASK_EXPORT_ENABLE(mask_export_enable) |
S_02880C_Z_ORDER(z_order) |
S_02880C_DEPTH_BEFORE_SHADER(ps->info.fs.early_fragment_test) |
S_02880C_EXEC_ON_HIER_FAIL(ps->info.info.ps.writes_memory) |
}
static void
-radv_pipeline_generate_fragment_shader(struct radeon_winsys_cs *cs,
+radv_pipeline_generate_fragment_shader(struct radeon_cmdbuf *ctx_cs,
+ struct radeon_cmdbuf *cs,
struct radv_pipeline *pipeline)
{
struct radv_shader_variant *ps;
radeon_set_sh_reg_seq(cs, R_00B020_SPI_SHADER_PGM_LO_PS, 4);
radeon_emit(cs, va >> 8);
radeon_emit(cs, S_00B024_MEM_BASE(va >> 40));
- radeon_emit(cs, ps->rsrc1);
- radeon_emit(cs, ps->rsrc2);
+ radeon_emit(cs, ps->config.rsrc1);
+ radeon_emit(cs, ps->config.rsrc2);
- radeon_set_context_reg(cs, R_02880C_DB_SHADER_CONTROL,
- radv_compute_db_shader_control(pipeline->device, ps));
+ radeon_set_context_reg(ctx_cs, R_02880C_DB_SHADER_CONTROL,
+ radv_compute_db_shader_control(pipeline->device,
+ pipeline, ps));
- radeon_set_context_reg(cs, R_0286CC_SPI_PS_INPUT_ENA,
+ radeon_set_context_reg(ctx_cs, R_0286CC_SPI_PS_INPUT_ENA,
ps->config.spi_ps_input_ena);
- radeon_set_context_reg(cs, R_0286D0_SPI_PS_INPUT_ADDR,
+ radeon_set_context_reg(ctx_cs, R_0286D0_SPI_PS_INPUT_ADDR,
ps->config.spi_ps_input_addr);
- radeon_set_context_reg(cs, R_0286D8_SPI_PS_IN_CONTROL,
+ radeon_set_context_reg(ctx_cs, R_0286D8_SPI_PS_IN_CONTROL,
S_0286D8_NUM_INTERP(ps->info.fs.num_interp));
- radeon_set_context_reg(cs, R_0286E0_SPI_BARYC_CNTL, pipeline->graphics.spi_baryc_cntl);
+ radeon_set_context_reg(ctx_cs, R_0286E0_SPI_BARYC_CNTL, pipeline->graphics.spi_baryc_cntl);
- radeon_set_context_reg(cs, R_028710_SPI_SHADER_Z_FORMAT,
+ radeon_set_context_reg(ctx_cs, R_028710_SPI_SHADER_Z_FORMAT,
ac_get_spi_shader_z_format(ps->info.info.ps.writes_z,
ps->info.info.ps.writes_stencil,
ps->info.info.ps.writes_sample_mask));
}
static void
-radv_pipeline_generate_vgt_vertex_reuse(struct radeon_winsys_cs *cs,
+radv_pipeline_generate_vgt_vertex_reuse(struct radeon_cmdbuf *ctx_cs,
struct radv_pipeline *pipeline)
{
- if (pipeline->device->physical_device->rad_info.family < CHIP_POLARIS10)
+ if (pipeline->device->physical_device->rad_info.family < CHIP_POLARIS10 ||
+ pipeline->device->physical_device->rad_info.chip_class >= GFX10)
return;
unsigned vtx_reuse_depth = 30;
radv_get_shader(pipeline, MESA_SHADER_TESS_EVAL)->info.tes.spacing == TESS_SPACING_FRACTIONAL_ODD) {
vtx_reuse_depth = 14;
}
- radeon_set_context_reg(cs, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL,
+ radeon_set_context_reg(ctx_cs, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL,
S_028C58_VTX_REUSE_DEPTH(vtx_reuse_depth));
}
stages |= S_028B54_ES_EN(V_028B54_ES_STAGE_DS) |
S_028B54_GS_EN(1) |
S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER);
+ else if (radv_pipeline_has_ngg(pipeline))
+ stages |= S_028B54_ES_EN(V_028B54_ES_STAGE_DS) |
+ S_028B54_PRIMGEN_EN(1);
else
stages |= S_028B54_VS_EN(V_028B54_VS_STAGE_DS);
- } else if (radv_pipeline_has_gs(pipeline))
+ } else if (radv_pipeline_has_gs(pipeline)) {
stages |= S_028B54_ES_EN(V_028B54_ES_STAGE_REAL) |
S_028B54_GS_EN(1) |
S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER);
+ } else if (radv_pipeline_has_ngg(pipeline)) {
+ stages |= S_028B54_ES_EN(V_028B54_ES_STAGE_REAL) |
+ S_028B54_PRIMGEN_EN(1);
+ }
if (pipeline->device->physical_device->rad_info.chip_class >= GFX9)
stages |= S_028B54_MAX_PRIMGRP_IN_WAVE(2);
return mask;
}
+static void
+gfx10_pipeline_generate_ge_cntl(struct radeon_cmdbuf *ctx_cs,
+ struct radv_pipeline *pipeline,
+ const struct radv_tessellation_state *tess,
+ const struct radv_gs_state *gs_state)
+{
+ bool break_wave_at_eoi = false;
+ unsigned primgroup_size;
+ unsigned vertgroup_size;
+
+ if (radv_pipeline_has_tess(pipeline)) {
+ primgroup_size = tess->num_patches; /* must be a multiple of NUM_PATCHES */
+ vertgroup_size = 0;
+ } else if (radv_pipeline_has_gs(pipeline)) {
+ unsigned vgt_gs_onchip_cntl = gs_state->vgt_gs_onchip_cntl;
+ primgroup_size = G_028A44_GS_PRIMS_PER_SUBGRP(vgt_gs_onchip_cntl);
+ vertgroup_size = G_028A44_ES_VERTS_PER_SUBGRP(vgt_gs_onchip_cntl);
+ } else {
+ primgroup_size = 128; /* recommended without a GS and tess */
+ vertgroup_size = 0;
+ }
+
+ if (radv_pipeline_has_tess(pipeline)) {
+ if (pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.info.uses_prim_id ||
+ radv_get_shader(pipeline, MESA_SHADER_TESS_EVAL)->info.info.uses_prim_id)
+ break_wave_at_eoi = true;
+ }
+
+ radeon_set_uconfig_reg(ctx_cs, R_03096C_GE_CNTL,
+ S_03096C_PRIM_GRP_SIZE(primgroup_size) |
+ S_03096C_VERT_GRP_SIZE(vertgroup_size) |
+ S_03096C_PACKET_TO_ONE_PA(0) /* line stipple */ |
+ S_03096C_BREAK_WAVE_AT_EOI(break_wave_at_eoi));
+}
+
static void
radv_pipeline_generate_pm4(struct radv_pipeline *pipeline,
const VkGraphicsPipelineCreateInfo *pCreateInfo,
const struct radv_blend_state *blend,
const struct radv_tessellation_state *tess,
const struct radv_gs_state *gs,
+ const struct radv_ngg_state *ngg,
unsigned prim, unsigned gs_out)
{
- pipeline->cs.buf = malloc(4 * 256);
- pipeline->cs.max_dw = 256;
-
- radv_pipeline_generate_depth_stencil_state(&pipeline->cs, pipeline, pCreateInfo, extra);
- radv_pipeline_generate_blend_state(&pipeline->cs, pipeline, blend);
- radv_pipeline_generate_raster_state(&pipeline->cs, pCreateInfo);
- radv_pipeline_generate_multisample_state(&pipeline->cs, pipeline);
- radv_pipeline_generate_vgt_gs_mode(&pipeline->cs, pipeline);
- radv_pipeline_generate_vertex_shader(&pipeline->cs, pipeline, tess);
- radv_pipeline_generate_tess_shaders(&pipeline->cs, pipeline, tess);
- radv_pipeline_generate_geometry_shader(&pipeline->cs, pipeline, gs);
- radv_pipeline_generate_fragment_shader(&pipeline->cs, pipeline);
- radv_pipeline_generate_ps_inputs(&pipeline->cs, pipeline);
- radv_pipeline_generate_vgt_vertex_reuse(&pipeline->cs, pipeline);
- radv_pipeline_generate_binning_state(&pipeline->cs, pipeline, pCreateInfo);
-
- radeon_set_context_reg(&pipeline->cs, R_0286E8_SPI_TMPRING_SIZE,
+ struct radeon_cmdbuf *ctx_cs = &pipeline->ctx_cs;
+ struct radeon_cmdbuf *cs = &pipeline->cs;
+
+ cs->max_dw = 64;
+ ctx_cs->max_dw = 256;
+ cs->buf = malloc(4 * (cs->max_dw + ctx_cs->max_dw));
+ ctx_cs->buf = cs->buf + cs->max_dw;
+
+ radv_pipeline_generate_depth_stencil_state(ctx_cs, pipeline, pCreateInfo, extra);
+ radv_pipeline_generate_blend_state(ctx_cs, pipeline, blend);
+ radv_pipeline_generate_raster_state(ctx_cs, pipeline, pCreateInfo);
+ radv_pipeline_generate_multisample_state(ctx_cs, pipeline);
+ radv_pipeline_generate_vgt_gs_mode(ctx_cs, pipeline);
+ radv_pipeline_generate_vertex_shader(ctx_cs, cs, pipeline, tess, ngg);
+ radv_pipeline_generate_tess_shaders(ctx_cs, cs, pipeline, tess);
+ radv_pipeline_generate_geometry_shader(ctx_cs, cs, pipeline, gs);
+ radv_pipeline_generate_fragment_shader(ctx_cs, cs, pipeline);
+ radv_pipeline_generate_ps_inputs(ctx_cs, pipeline);
+ radv_pipeline_generate_vgt_vertex_reuse(ctx_cs, pipeline);
+ radv_pipeline_generate_binning_state(ctx_cs, pipeline, pCreateInfo);
+
+ if (pipeline->device->physical_device->rad_info.chip_class >= GFX10 && !radv_pipeline_has_ngg(pipeline))
+ gfx10_pipeline_generate_ge_cntl(ctx_cs, pipeline, tess, gs);
+
+ radeon_set_context_reg(ctx_cs, R_0286E8_SPI_TMPRING_SIZE,
S_0286E8_WAVES(pipeline->max_waves) |
S_0286E8_WAVESIZE(pipeline->scratch_bytes_per_wave >> 10));
- radeon_set_context_reg(&pipeline->cs, R_028B54_VGT_SHADER_STAGES_EN, radv_compute_vgt_shader_stages_en(pipeline));
+ radeon_set_context_reg(ctx_cs, R_028B54_VGT_SHADER_STAGES_EN, radv_compute_vgt_shader_stages_en(pipeline));
- if (pipeline->device->physical_device->rad_info.chip_class >= CIK) {
- radeon_set_uconfig_reg_idx(&pipeline->cs, R_030908_VGT_PRIMITIVE_TYPE, 1, prim);
+ if (pipeline->device->physical_device->rad_info.chip_class >= GFX7) {
+ radeon_set_uconfig_reg_idx(pipeline->device->physical_device,
+ cs, R_030908_VGT_PRIMITIVE_TYPE, 1, prim);
} else {
- radeon_set_config_reg(&pipeline->cs, R_008958_VGT_PRIMITIVE_TYPE, prim);
+ radeon_set_config_reg(cs, R_008958_VGT_PRIMITIVE_TYPE, prim);
}
- radeon_set_context_reg(&pipeline->cs, R_028A6C_VGT_GS_OUT_PRIM_TYPE, gs_out);
+ radeon_set_context_reg(ctx_cs, R_028A6C_VGT_GS_OUT_PRIM_TYPE, gs_out);
- radeon_set_context_reg(&pipeline->cs, R_02820C_PA_SC_CLIPRECT_RULE, radv_compute_cliprect_rule(pCreateInfo));
+ radeon_set_context_reg(ctx_cs, R_02820C_PA_SC_CLIPRECT_RULE, radv_compute_cliprect_rule(pCreateInfo));
- assert(pipeline->cs.cdw <= pipeline->cs.max_dw);
+ pipeline->ctx_cs_hash = _mesa_hash_data(ctx_cs->buf, ctx_cs->cdw * 4);
+
+ assert(ctx_cs->cdw <= ctx_cs->max_dw);
+ assert(cs->cdw <= cs->max_dw);
}
static struct radv_ia_multi_vgt_param_helpers
else
ia_multi_vgt_param.primgroup_size = 128; /* recommended without a GS */
- ia_multi_vgt_param.partial_es_wave = false;
- if (pipeline->device->has_distributed_tess) {
- if (radv_pipeline_has_gs(pipeline)) {
- if (device->physical_device->rad_info.chip_class <= VI)
- ia_multi_vgt_param.partial_es_wave = true;
- }
- }
/* GS requirement. */
- if (radv_pipeline_has_gs(pipeline) && device->physical_device->rad_info.chip_class <= VI)
+ ia_multi_vgt_param.partial_es_wave = false;
+ if (radv_pipeline_has_gs(pipeline) && device->physical_device->rad_info.chip_class <= GFX8)
if (SI_GS_PER_ES / ia_multi_vgt_param.primgroup_size >= pipeline->device->gs_table_depth - 3)
ia_multi_vgt_param.partial_es_wave = true;
ia_multi_vgt_param.wd_switch_on_eop = false;
- if (device->physical_device->rad_info.chip_class >= CIK) {
+ if (device->physical_device->rad_info.chip_class >= GFX7) {
/* WD_SWITCH_ON_EOP has no effect on GPUs with less than
* 4 shader engines. Set 1 to pass the assertion below.
* The other cases are hardware requirements. */
(pipeline->graphics.prim_restart_enable &&
(device->physical_device->rad_info.family < CHIP_POLARIS10 ||
(prim != V_008958_DI_PT_POINTLIST &&
- prim != V_008958_DI_PT_LINESTRIP &&
- prim != V_008958_DI_PT_TRISTRIP))))
+ prim != V_008958_DI_PT_LINESTRIP))))
ia_multi_vgt_param.wd_switch_on_eop = true;
}
/* Needed for 028B6C_DISTRIBUTION_MODE != 0 */
if (device->has_distributed_tess) {
if (radv_pipeline_has_gs(pipeline)) {
- if (device->physical_device->rad_info.family == CHIP_TONGA ||
- device->physical_device->rad_info.family == CHIP_FIJI ||
- device->physical_device->rad_info.family == CHIP_POLARIS10 ||
- device->physical_device->rad_info.family == CHIP_POLARIS11 ||
- device->physical_device->rad_info.family == CHIP_POLARIS12 ||
- device->physical_device->rad_info.family == CHIP_VEGAM)
- ia_multi_vgt_param.partial_vs_wave = true;
+ if (device->physical_device->rad_info.chip_class <= GFX8)
+ ia_multi_vgt_param.partial_es_wave = true;
} else {
ia_multi_vgt_param.partial_vs_wave = true;
}
}
}
+ /* Workaround for a VGT hang when strip primitive types are used with
+ * primitive restart.
+ */
+ if (pipeline->graphics.prim_restart_enable &&
+ (prim == V_008958_DI_PT_LINESTRIP ||
+ prim == V_008958_DI_PT_TRISTRIP ||
+ prim == V_008958_DI_PT_LINESTRIP_ADJ ||
+ prim == V_008958_DI_PT_TRISTRIP_ADJ)) {
+ ia_multi_vgt_param.partial_vs_wave = true;
+ }
+
+ if (radv_pipeline_has_gs(pipeline)) {
+ /* On these chips there is the possibility of a hang if the
+ * pipeline uses a GS and partial_vs_wave is not set.
+ *
+ * This mostly does not hit 4-SE chips, as those typically set
+ * ia_switch_on_eoi and then partial_vs_wave is set for pipelines
+ * with GS due to another workaround.
+ *
+ * Reproducer: https://bugs.freedesktop.org/show_bug.cgi?id=109242
+ */
+ if (device->physical_device->rad_info.family == CHIP_TONGA ||
+ device->physical_device->rad_info.family == CHIP_FIJI ||
+ device->physical_device->rad_info.family == CHIP_POLARIS10 ||
+ device->physical_device->rad_info.family == CHIP_POLARIS11 ||
+ device->physical_device->rad_info.family == CHIP_POLARIS12 ||
+ device->physical_device->rad_info.family == CHIP_VEGAM) {
+ ia_multi_vgt_param.partial_vs_wave = true;
+ }
+ }
+
ia_multi_vgt_param.base =
S_028AA8_PRIMGROUP_SIZE(ia_multi_vgt_param.primgroup_size - 1) |
/* The following field was moved to VGT_SHADER_STAGES_EN in GFX9. */
- S_028AA8_MAX_PRIMGRP_IN_WAVE(device->physical_device->rad_info.chip_class == VI ? 2 : 0) |
+ S_028AA8_MAX_PRIMGRP_IN_WAVE(device->physical_device->rad_info.chip_class == GFX8 ? 2 : 0) |
S_030960_EN_INST_OPT_BASIC(device->physical_device->rad_info.chip_class >= GFX9) |
S_030960_EN_INST_OPT_ADV(device->physical_device->rad_info.chip_class >= GFX9);
&vi_info->pVertexAttributeDescriptions[i];
unsigned loc = desc->location;
const struct vk_format_description *format_desc;
- int first_non_void;
- uint32_t num_format, data_format;
- format_desc = vk_format_description(desc->format);
- first_non_void = vk_format_get_first_non_void_channel(desc->format);
- num_format = radv_translate_buffer_numformat(format_desc, first_non_void);
- data_format = radv_translate_buffer_dataformat(format_desc, first_non_void);
+ format_desc = vk_format_description(desc->format);
- velems->rsrc_word3[loc] = S_008F0C_DST_SEL_X(si_map_swizzle(format_desc->swizzle[0])) |
- S_008F0C_DST_SEL_Y(si_map_swizzle(format_desc->swizzle[1])) |
- S_008F0C_DST_SEL_Z(si_map_swizzle(format_desc->swizzle[2])) |
- S_008F0C_DST_SEL_W(si_map_swizzle(format_desc->swizzle[3])) |
- S_008F0C_NUM_FORMAT(num_format) |
- S_008F0C_DATA_FORMAT(data_format);
velems->format_size[loc] = format_desc->block.bits / 8;
- velems->offset[loc] = desc->offset;
- velems->binding[loc] = desc->binding;
- velems->count = MAX2(velems->count, loc + 1);
}
for (uint32_t i = 0; i < vi_info->vertexBindingDescriptionCount; i++) {
&vi_info->pVertexBindingDescriptions[i];
pipeline->binding_stride[desc->binding] = desc->stride;
+ pipeline->num_vertex_bindings =
+ MAX2(pipeline->num_vertex_bindings, desc->binding + 1);
+ }
+}
+
+static struct radv_shader_variant *
+radv_pipeline_get_streamout_shader(struct radv_pipeline *pipeline)
+{
+ int i;
+
+ for (i = MESA_SHADER_GEOMETRY; i >= MESA_SHADER_VERTEX; i--) {
+ struct radv_shader_variant *shader =
+ radv_get_shader(pipeline, i);
+
+ if (shader && shader->info.info.so.num_outputs > 0)
+ return shader;
}
+
+ return NULL;
}
static VkResult
struct radv_device *device,
struct radv_pipeline_cache *cache,
const VkGraphicsPipelineCreateInfo *pCreateInfo,
- const struct radv_graphics_pipeline_create_info *extra,
- const VkAllocationCallbacks *alloc)
+ const struct radv_graphics_pipeline_create_info *extra)
{
VkResult result;
bool has_view_index = false;
struct radv_subpass *subpass = pass->subpasses + pCreateInfo->subpass;
if (subpass->view_mask)
has_view_index = true;
- if (alloc == NULL)
- alloc = &device->alloc;
pipeline->device = device;
pipeline->layout = radv_pipeline_layout_from_handle(pCreateInfo->layout);
struct radv_blend_state blend = radv_pipeline_init_blend_state(pipeline, pCreateInfo, extra);
+ const VkPipelineCreationFeedbackCreateInfoEXT *creation_feedback =
+ vk_find_struct_const(pCreateInfo->pNext, PIPELINE_CREATION_FEEDBACK_CREATE_INFO_EXT);
+ radv_init_feedback(creation_feedback);
+
+ VkPipelineCreationFeedbackEXT *pipeline_feedback = creation_feedback ? creation_feedback->pPipelineCreationFeedback : NULL;
+
const VkPipelineShaderStageCreateInfo *pStages[MESA_SHADER_STAGES] = { 0, };
+ VkPipelineCreationFeedbackEXT *stage_feedbacks[MESA_SHADER_STAGES] = { 0 };
for (uint32_t i = 0; i < pCreateInfo->stageCount; i++) {
gl_shader_stage stage = ffs(pCreateInfo->pStages[i].stage) - 1;
pStages[stage] = &pCreateInfo->pStages[i];
+ if(creation_feedback)
+ stage_feedbacks[stage] = &creation_feedback->pPipelineStageCreationFeedbacks[i];
}
- radv_create_shaders(pipeline, device, cache,
- radv_generate_graphics_pipeline_key(pipeline, pCreateInfo, &blend, has_view_index),
- pStages, pCreateInfo->flags);
+ struct radv_pipeline_key key = radv_generate_graphics_pipeline_key(pipeline, pCreateInfo, &blend, has_view_index);
+ radv_create_shaders(pipeline, device, cache, &key, pStages, pCreateInfo->flags, pipeline_feedback, stage_feedbacks);
pipeline->graphics.spi_baryc_cntl = S_0286E0_FRONT_FACE_ALL_BITS(1);
radv_pipeline_init_multisample_state(pipeline, &blend, pCreateInfo);
if (radv_pipeline_has_gs(pipeline)) {
gs_out = si_conv_gl_prim_to_gs_out(pipeline->shaders[MESA_SHADER_GEOMETRY]->info.gs.output_prim);
pipeline->graphics.can_use_guardband = gs_out == V_028A6C_OUTPRIM_TYPE_TRISTRIP;
+ } else if (radv_pipeline_has_tess(pipeline)) {
+ gs_out = si_conv_gl_prim_to_gs_out(pipeline->shaders[MESA_SHADER_TESS_EVAL]->info.tes.primitive_mode);
+ pipeline->graphics.can_use_guardband = gs_out == V_028A6C_OUTPRIM_TYPE_TRISTRIP;
} else {
gs_out = si_conv_prim_to_gs_out(pCreateInfo->pInputAssemblyState->topology);
}
prim = V_008958_DI_PT_RECTLIST;
gs_out = V_028A6C_OUTPRIM_TYPE_TRISTRIP;
pipeline->graphics.can_use_guardband = true;
+ if (radv_pipeline_has_ngg(pipeline))
+ gs_out = V_028A6C_VGT_OUT_RECT_V0;
}
pipeline->graphics.prim_restart_enable = !!pCreateInfo->pInputAssemblyState->primitiveRestartEnable;
/* prim vertex count will need TESS changes */
}
}
+ struct radv_ngg_state ngg = {0};
struct radv_gs_state gs = {0};
- if (radv_pipeline_has_gs(pipeline)) {
+
+ if (radv_pipeline_has_ngg(pipeline)) {
+ ngg = calculate_ngg_info(pCreateInfo, pipeline);
+ } else if (radv_pipeline_has_gs(pipeline)) {
gs = calculate_gs_info(pCreateInfo, pipeline);
calculate_gs_ring_sizes(pipeline, &gs);
}
pipeline->graphics.vtx_emit_num = 2;
}
+ /* Find the last vertex shader stage that eventually uses streamout. */
+ pipeline->streamout_shader = radv_pipeline_get_streamout_shader(pipeline);
+
result = radv_pipeline_scratch_init(device, pipeline);
- radv_pipeline_generate_pm4(pipeline, pCreateInfo, extra, &blend, &tess, &gs, prim, gs_out);
+ radv_pipeline_generate_pm4(pipeline, pCreateInfo, extra, &blend, &tess, &gs, &ngg, prim, gs_out);
return result;
}
return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
result = radv_pipeline_init(pipeline, device, cache,
- pCreateInfo, extra, pAllocator);
+ pCreateInfo, extra);
if (result != VK_SUCCESS) {
radv_pipeline_destroy(device, pipeline, pAllocator);
return result;
radeon_emit(&pipeline->cs, S_00B834_DATA(va >> 40));
radeon_set_sh_reg_seq(&pipeline->cs, R_00B848_COMPUTE_PGM_RSRC1, 2);
- radeon_emit(&pipeline->cs, compute_shader->rsrc1);
- radeon_emit(&pipeline->cs, compute_shader->rsrc2);
+ radeon_emit(&pipeline->cs, compute_shader->config.rsrc1);
+ radeon_emit(&pipeline->cs, compute_shader->config.rsrc2);
radeon_set_sh_reg(&pipeline->cs, R_00B860_COMPUTE_TMPRING_SIZE,
S_00B860_WAVES(pipeline->max_waves) |
compute_resource_limits =
S_00B854_SIMD_DEST_CNTL(waves_per_threadgroup % 4 == 0);
- if (device->physical_device->rad_info.chip_class >= CIK) {
+ if (device->physical_device->rad_info.chip_class >= GFX7) {
unsigned num_cu_per_se =
device->physical_device->rad_info.num_good_compute_units /
device->physical_device->rad_info.max_se;
RADV_FROM_HANDLE(radv_device, device, _device);
RADV_FROM_HANDLE(radv_pipeline_cache, cache, _cache);
const VkPipelineShaderStageCreateInfo *pStages[MESA_SHADER_STAGES] = { 0, };
+ VkPipelineCreationFeedbackEXT *stage_feedbacks[MESA_SHADER_STAGES] = { 0 };
struct radv_pipeline *pipeline;
VkResult result;
pipeline->layout = radv_pipeline_layout_from_handle(pCreateInfo->layout);
assert(pipeline->layout);
+ const VkPipelineCreationFeedbackCreateInfoEXT *creation_feedback =
+ vk_find_struct_const(pCreateInfo->pNext, PIPELINE_CREATION_FEEDBACK_CREATE_INFO_EXT);
+ radv_init_feedback(creation_feedback);
+
+ VkPipelineCreationFeedbackEXT *pipeline_feedback = creation_feedback ? creation_feedback->pPipelineCreationFeedback : NULL;
+ if (creation_feedback)
+ stage_feedbacks[MESA_SHADER_COMPUTE] = &creation_feedback->pPipelineStageCreationFeedbacks[0];
+
pStages[MESA_SHADER_COMPUTE] = &pCreateInfo->stage;
- radv_create_shaders(pipeline, device, cache, (struct radv_pipeline_key) {0}, pStages, pCreateInfo->flags);
+ radv_create_shaders(pipeline, device, cache, &(struct radv_pipeline_key) {0}, pStages, pCreateInfo->flags, pipeline_feedback, stage_feedbacks);
pipeline->user_data_0[MESA_SHADER_COMPUTE] = radv_pipeline_stage_to_user_data_0(pipeline, MESA_SHADER_COMPUTE, device->physical_device->rad_info.chip_class);
pipeline->need_indirect_descriptor_sets |= pipeline->shaders[MESA_SHADER_COMPUTE]->info.need_indirect_descriptor_sets;