*/
-#include "main/imports.h"
+#include "main/errors.h"
+
#include "main/image.h"
#include "main/bufferobj.h"
#include "main/macros.h"
#include "st_debug.h"
#include "st_draw.h"
#include "st_program.h"
+#include "st_util.h"
#include "pipe/p_context.h"
#include "pipe/p_defines.h"
+#include "util/u_cpu_detect.h"
#include "util/u_inlines.h"
-#include "util/u_format.h"
+#include "util/format/u_format.h"
#include "util/u_prim.h"
#include "util/u_draw.h"
#include "util/u_upload_mgr.h"
#include "draw/draw_context.h"
#include "cso_cache/cso_context.h"
-
-/**
- * This is very similar to vbo_all_varyings_in_vbos() but we are
- * only interested in per-vertex data. See bug 38626.
- */
-static GLboolean
-all_varyings_in_vbos(const struct gl_vertex_array *arrays[])
-{
- GLuint i;
-
- for (i = 0; i < VERT_ATTRIB_MAX; i++)
- if (arrays[i]->StrideB &&
- !arrays[i]->InstanceDivisor &&
- !_mesa_is_bufferobj(arrays[i]->BufferObj))
- return GL_FALSE;
-
- return GL_TRUE;
-}
-
-
-/**
- * Basically, translate Mesa's index buffer information into
- * a pipe_index_buffer object.
- */
-static void
-setup_index_buffer(struct st_context *st,
- const struct _mesa_index_buffer *ib)
-{
- struct pipe_index_buffer ibuffer;
- struct gl_buffer_object *bufobj = ib->obj;
-
- ibuffer.index_size = vbo_sizeof_ib_type(ib->type);
-
- /* get/create the index buffer object */
- if (_mesa_is_bufferobj(bufobj)) {
- /* indices are in a real VBO */
- ibuffer.buffer = st_buffer_object(bufobj)->buffer;
- ibuffer.offset = pointer_to_offset(ib->ptr);
- ibuffer.user_buffer = NULL;
- }
- else {
- /* indices are in user space memory */
- ibuffer.buffer = NULL;
- ibuffer.offset = 0;
- ibuffer.user_buffer = ib->ptr;
- }
-
- cso_set_index_buffer(st->cso_context, &ibuffer);
-}
-
+#if defined(PIPE_OS_LINUX) && !defined(ANDROID)
+#include <sched.h>
+#define HAVE_SCHED_GETCPU 1
+#else
+#define sched_getcpu() 0
+#define HAVE_SCHED_GETCPU 0
+#endif
/**
* Set the restart index.
*/
static void
-setup_primitive_restart(struct gl_context *ctx,
- const struct _mesa_index_buffer *ib,
- struct pipe_draw_info *info)
+setup_primitive_restart(struct gl_context *ctx, struct pipe_draw_info *info)
{
if (ctx->Array._PrimitiveRestart) {
- info->restart_index = _mesa_primitive_restart_index(ctx, ib->type);
+ unsigned index_size = info->index_size;
+
+ info->restart_index = ctx->Array._RestartIndex[index_size - 1];
/* Enable primitive restart only when the restart index can have an
- * effect. This is required for correctness in radeonsi VI support.
+ * effect. This is required for correctness in radeonsi GFX8 support.
* Other hardware may also benefit from taking a faster, non-restart path
* when possible.
*/
- if ((ib->type == GL_UNSIGNED_INT) ||
- (ib->type == GL_UNSIGNED_SHORT && info->restart_index <= 0xffff) ||
- (ib->type == GL_UNSIGNED_BYTE && info->restart_index <= 0xff))
+ if (index_size == 4 || info->restart_index < (1 << (index_size * 8)))
info->primitive_restart = true;
}
}
return prim;
}
+static inline void
+prepare_draw(struct st_context *st, struct gl_context *ctx)
+{
+ /* Mesa core state should have been validated already */
+ assert(ctx->NewState == 0x0);
+
+ if (unlikely(!st->bitmap.cache.empty))
+ st_flush_bitmap_cache(st);
+
+ st_invalidate_readpix_cache(st);
+
+ /* Validate state. */
+ if ((st->dirty | ctx->NewDriverState) & ST_PIPELINE_RENDER_STATE_MASK ||
+ st->gfx_shaders_may_be_dirty) {
+ st_validate_state(st, ST_PIPELINE_RENDER);
+ }
+
+ struct pipe_context *pipe = st->pipe;
+
+ /* Pin threads regularly to the same Zen CCX that the main thread is
+ * running on. The main thread can move between CCXs.
+ */
+ if (unlikely(HAVE_SCHED_GETCPU && /* Linux */
+ /* AMD Zen */
+ util_cpu_caps.nr_cpus != util_cpu_caps.cores_per_L3 &&
+ /* no glthread */
+ ctx->CurrentClientDispatch != ctx->MarshalExec &&
+ /* driver support */
+ pipe->set_context_param &&
+ /* do it occasionally */
+ ++st->pin_thread_counter % 512 == 0)) {
+ int cpu = sched_getcpu();
+ if (cpu >= 0) {
+ unsigned L3_cache = cpu / util_cpu_caps.cores_per_L3;
+
+ pipe->set_context_param(pipe,
+ PIPE_CONTEXT_PARAM_PIN_THREADS_TO_L3_CACHE,
+ L3_cache);
+ }
+ }
+}
/**
* This function gets plugged into the VBO module and is called when
* we have something to render.
* Basically, translate the information into the format expected by gallium.
+ *
+ * Try to keep this logic in sync with st_feedback_draw_vbo.
*/
-void
+static void
st_draw_vbo(struct gl_context *ctx,
const struct _mesa_prim *prims,
GLuint nr_prims,
GLboolean index_bounds_valid,
GLuint min_index,
GLuint max_index,
+ GLuint num_instances,
+ GLuint base_instance,
struct gl_transform_feedback_object *tfb_vertcount,
- unsigned stream,
- struct gl_buffer_object *indirect)
+ unsigned stream)
{
struct st_context *st = st_context(ctx);
struct pipe_draw_info info;
- const struct gl_vertex_array **arrays = ctx->Array._DrawArrays;
unsigned i;
+ unsigned start = 0;
- /* Mesa core state should have been validated already */
- assert(ctx->NewState == 0x0);
+ prepare_draw(st, ctx);
- st_flush_bitmap_cache(st);
- st_invalidate_readpix_cache(st);
+ /* Initialize pipe_draw_info. */
+ info.primitive_restart = false;
+ info.vertices_per_patch = ctx->TessCtrlProgram.patch_vertices;
+ info.indirect = NULL;
+ info.count_from_stream_output = NULL;
+ info.restart_index = 0;
+ info.start_instance = base_instance;
+ info.instance_count = num_instances;
- /* Validate state. */
- if ((st->dirty | ctx->NewDriverState) & ST_PIPELINE_RENDER_STATE_MASK ||
- st->gfx_shaders_may_be_dirty) {
- st_validate_state(st, ST_PIPELINE_RENDER);
- }
+ if (ib) {
+ struct gl_buffer_object *bufobj = ib->obj;
- if (st->vertex_array_out_of_memory) {
- return;
- }
+ /* Get index bounds for user buffers. */
+ if (!index_bounds_valid && st->draw_needs_minmax_index) {
+ vbo_get_minmax_indices(ctx, prims, ib, &min_index, &max_index,
+ nr_prims);
+ }
- util_draw_init_info(&info);
+ info.index_size = 1 << ib->index_size_shift;
+ info.min_index = min_index;
+ info.max_index = max_index;
- if (ib) {
- /* Get index bounds for user buffers. */
- if (!index_bounds_valid)
- if (!all_varyings_in_vbos(arrays))
- vbo_get_minmax_indices(ctx, prims, ib, &min_index, &max_index,
- nr_prims);
+ if (bufobj) {
+ /* indices are in a real VBO */
+ info.has_user_indices = false;
+ info.index.resource = st_buffer_object(bufobj)->buffer;
- setup_index_buffer(st, ib);
+ /* Return if the bound element array buffer doesn't have any backing
+ * storage. (nothing to do)
+ */
+ if (!info.index.resource)
+ return;
- info.indexed = TRUE;
- if (min_index != ~0U && max_index != ~0U) {
- info.min_index = min_index;
- info.max_index = max_index;
+ start = pointer_to_offset(ib->ptr) >> ib->index_size_shift;
+ } else {
+ /* indices are in user space memory */
+ info.has_user_indices = true;
+ info.index.user = ib->ptr;
}
- /* The VBO module handles restart for the non-indexed GLDrawArrays
- * so we only set these fields for indexed drawing:
- */
- setup_primitive_restart(ctx, ib, &info);
+ setup_primitive_restart(ctx, &info);
}
else {
+ info.index_size = 0;
+ info.has_user_indices = false;
+
/* Transform feedback drawing is always non-indexed. */
/* Set info.count_from_stream_output. */
if (tfb_vertcount) {
}
}
- assert(!indirect);
-
/* do actual drawing */
for (i = 0; i < nr_prims; i++) {
- info.mode = translate_prim(ctx, prims[i].mode);
- info.start = prims[i].start;
info.count = prims[i].count;
- info.start_instance = prims[i].base_instance;
- info.instance_count = prims[i].num_instances;
- info.vertices_per_patch = ctx->TessCtrlProgram.patch_vertices;
+
+ /* Skip no-op draw calls. */
+ if (!info.count && !tfb_vertcount)
+ continue;
+
+ info.mode = translate_prim(ctx, prims[i].mode);
+ info.start = start + prims[i].start;
info.index_bias = prims[i].basevertex;
info.drawid = prims[i].draw_id;
if (!ib) {
}
if (ST_DEBUG & DEBUG_DRAW) {
- debug_printf("st/draw: mode %s start %u count %u indexed %d\n",
+ debug_printf("st/draw: mode %s start %u count %u index_size %d\n",
u_prim_name(info.mode),
info.start,
info.count,
- info.indexed);
+ info.index_size);
}
/* Don't call u_trim_pipe_prim. Drivers should do it if they need it. */
GLsizeiptr indirect_offset,
unsigned draw_count,
unsigned stride,
- struct gl_buffer_object *indirect_params,
- GLsizeiptr indirect_params_offset,
+ struct gl_buffer_object *indirect_draw_count,
+ GLsizeiptr indirect_draw_count_offset,
const struct _mesa_index_buffer *ib)
{
struct st_context *st = st_context(ctx);
struct pipe_draw_info info;
+ struct pipe_draw_indirect_info indirect;
- /* Mesa core state should have been validated already */
- assert(ctx->NewState == 0x0);
assert(stride);
+ prepare_draw(st, ctx);
- /* Validate state. */
- if ((st->dirty | ctx->NewDriverState) & ST_PIPELINE_RENDER_STATE_MASK ||
- st->gfx_shaders_may_be_dirty) {
- st_validate_state(st, ST_PIPELINE_RENDER);
- }
-
- if (st->vertex_array_out_of_memory) {
- return;
- }
-
+ memset(&indirect, 0, sizeof(indirect));
util_draw_init_info(&info);
+ info.start = 0; /* index offset / index size */
+ info.max_index = ~0u; /* so that u_vbuf can tell that it's unknown */
if (ib) {
- setup_index_buffer(st, ib);
+ struct gl_buffer_object *bufobj = ib->obj;
+
+ /* indices are always in a real VBO */
+ assert(bufobj);
- info.indexed = TRUE;
+ info.index_size = 1 << ib->index_size_shift;
+ info.index.resource = st_buffer_object(bufobj)->buffer;
+ info.start = pointer_to_offset(ib->ptr) >> ib->index_size_shift;
/* Primitive restart is not handled by the VBO module in this case. */
- setup_primitive_restart(ctx, ib, &info);
+ setup_primitive_restart(ctx, &info);
}
info.mode = translate_prim(ctx, mode);
info.vertices_per_patch = ctx->TessCtrlProgram.patch_vertices;
- info.indirect = st_buffer_object(indirect_data)->buffer;
- info.indirect_offset = indirect_offset;
+ info.indirect = &indirect;
+ indirect.buffer = st_buffer_object(indirect_data)->buffer;
+ indirect.offset = indirect_offset;
if (ST_DEBUG & DEBUG_DRAW) {
- debug_printf("st/draw indirect: mode %s drawcount %d indexed %d\n",
+ debug_printf("st/draw indirect: mode %s drawcount %d index_size %d\n",
u_prim_name(info.mode),
draw_count,
- info.indexed);
+ info.index_size);
}
if (!st->has_multi_draw_indirect) {
int i;
- assert(!indirect_params);
- info.indirect_count = 1;
+ assert(!indirect_draw_count);
+ indirect.draw_count = 1;
for (i = 0; i < draw_count; i++) {
info.drawid = i;
cso_draw_vbo(st->cso_context, &info);
- info.indirect_offset += stride;
+ indirect.offset += stride;
}
} else {
- info.indirect_count = draw_count;
- info.indirect_stride = stride;
- if (indirect_params) {
- info.indirect_params = st_buffer_object(indirect_params)->buffer;
- info.indirect_params_offset = indirect_params_offset;
+ indirect.draw_count = draw_count;
+ indirect.stride = stride;
+ if (indirect_draw_count) {
+ indirect.indirect_draw_count =
+ st_buffer_object(indirect_draw_count)->buffer;
+ indirect.indirect_draw_count_offset = indirect_draw_count_offset;
}
cso_draw_vbo(st->cso_context, &info);
}
void
-st_init_draw(struct st_context *st)
+st_init_draw_functions(struct dd_function_table *functions)
{
- struct gl_context *ctx = st->ctx;
-
- vbo_set_draw_func(ctx, st_draw_vbo);
- vbo_set_indirect_draw_func(ctx, st_indirect_draw_vbo);
+ functions->Draw = st_draw_vbo;
+ functions->DrawIndirect = st_indirect_draw_vbo;
}
u_upload_alloc(st->pipe->stream_uploader, 0,
4 * sizeof(struct st_util_vertex), 4,
- &vb.buffer_offset, &vb.buffer, (void **) &verts);
- if (!vb.buffer) {
+ &vb.buffer_offset, &vb.buffer.resource, (void **) &verts);
+ if (!vb.buffer.resource) {
return false;
}
u_upload_unmap(st->pipe->stream_uploader);
- /* At the time of writing, cso_get_aux_vertex_buffer_slot() always returns
- * zero. If that ever changes we need to audit the calls to that function
- * and make sure the slot number is used consistently everywhere.
- */
- assert(cso_get_aux_vertex_buffer_slot(st->cso_context) == 0);
-
- cso_set_vertex_buffers(st->cso_context,
- cso_get_aux_vertex_buffer_slot(st->cso_context),
- 1, &vb);
+ cso_set_vertex_buffers(st->cso_context, 0, 1, &vb);
if (num_instances > 1) {
cso_draw_arrays_instanced(st->cso_context, PIPE_PRIM_TRIANGLE_FAN, 0, 4,
cso_draw_arrays(st->cso_context, PIPE_PRIM_TRIANGLE_FAN, 0, 4);
}
- pipe_resource_reference(&vb.buffer, NULL);
+ pipe_resource_reference(&vb.buffer.resource, NULL);
return true;
}