#include <sys/poll.h>
#include <errno.h>
+#include "pan_bo.h"
#include "pan_context.h"
-#include "pan_format.h"
+#include "pan_minmax_cache.h"
+#include "panfrost-quirks.h"
#include "util/macros.h"
-#include "util/u_format.h"
+#include "util/format/u_format.h"
#include "util/u_inlines.h"
#include "util/u_upload_mgr.h"
#include "util/u_memory.h"
#include "util/u_vbuf.h"
#include "util/half_float.h"
#include "util/u_helpers.h"
-#include "util/u_format.h"
+#include "util/format/u_format.h"
#include "util/u_prim.h"
#include "util/u_prim_restart.h"
#include "indices/u_primconvert.h"
#include "pan_blending.h"
#include "pan_blend_shaders.h"
#include "pan_util.h"
+#include "pandecode/decode.h"
-/* Framebuffer descriptor */
-
-static struct midgard_tiler_descriptor
+struct midgard_tiler_descriptor
panfrost_emit_midg_tiler(struct panfrost_batch *batch, unsigned vertex_count)
{
- struct panfrost_context *ctx = batch->ctx;
- struct midgard_tiler_descriptor t = {};
+ struct panfrost_screen *screen = pan_screen(batch->ctx->base.screen);
+ bool hierarchy = !(screen->quirks & MIDGARD_NO_HIER_TILING);
+ struct midgard_tiler_descriptor t = {0};
unsigned height = batch->key.height;
unsigned width = batch->key.width;
t.hierarchy_mask =
- panfrost_choose_hierarchy_mask(width, height, vertex_count);
+ panfrost_choose_hierarchy_mask(width, height, vertex_count, hierarchy);
/* Compute the polygon header size and use that to offset the body */
unsigned header_size = panfrost_tiler_header_size(
- width, height, t.hierarchy_mask);
+ width, height, t.hierarchy_mask, hierarchy);
t.polygon_list_size = panfrost_tiler_full_size(
- width, height, t.hierarchy_mask);
+ width, height, t.hierarchy_mask, hierarchy);
/* Sanity check */
- if (t.hierarchy_mask) {
+ if (vertex_count) {
+ struct panfrost_bo *tiler_heap;
+
+ tiler_heap = panfrost_batch_get_tiler_heap(batch);
t.polygon_list = panfrost_batch_get_polygon_list(batch,
header_size +
t.polygon_list_size);
/* Allow the entire tiler heap */
- t.heap_start = ctx->tiler_heap->gpu;
- t.heap_end = ctx->tiler_heap->gpu + ctx->tiler_heap->size;
+ t.heap_start = tiler_heap->gpu;
+ t.heap_end = tiler_heap->gpu + tiler_heap->size;
} else {
+ struct panfrost_bo *tiler_dummy;
+
+ tiler_dummy = panfrost_batch_get_tiler_dummy(batch);
+ header_size = MALI_TILER_MINIMUM_HEADER_SIZE;
+
/* The tiler is disabled, so don't allow the tiler heap */
- t.heap_start = ctx->tiler_heap->gpu;
+ t.heap_start = tiler_dummy->gpu;
t.heap_end = t.heap_start;
/* Use a dummy polygon list */
- t.polygon_list = ctx->tiler_dummy->gpu;
+ t.polygon_list = tiler_dummy->gpu;
/* Disable the tiler */
- t.hierarchy_mask |= MALI_TILER_DISABLED;
+ if (hierarchy)
+ t.hierarchy_mask |= MALI_TILER_DISABLED;
+ else {
+ t.hierarchy_mask = MALI_TILER_USER;
+ t.polygon_list_size = MALI_TILER_MINIMUM_HEADER_SIZE + 4;
+
+ /* We don't have a WRITE_VALUE job, so write the polygon list manually */
+ uint32_t *polygon_list_body = (uint32_t *) (tiler_dummy->cpu + header_size);
+ polygon_list_body[0] = 0xa0000000; /* TODO: Just that? */
+ }
}
t.polygon_list_body =
return t;
}
-struct mali_single_framebuffer
-panfrost_emit_sfbd(struct panfrost_batch *batch, unsigned vertex_count)
-{
- struct panfrost_context *ctx = batch->ctx;
- unsigned width = batch->key.width;
- unsigned height = batch->key.height;
-
- struct mali_single_framebuffer framebuffer = {
- .width = MALI_POSITIVE(width),
- .height = MALI_POSITIVE(height),
- .unknown2 = 0x1f,
- .format = 0x30000000,
- .clear_flags = 0x1000,
- .unknown_address_0 = ctx->scratchpad->gpu,
- .tiler = panfrost_emit_midg_tiler(batch, vertex_count),
- };
-
- return framebuffer;
-}
-
-struct bifrost_framebuffer
-panfrost_emit_mfbd(struct panfrost_batch *batch, unsigned vertex_count)
-{
- struct panfrost_context *ctx = batch->ctx;
- unsigned width = batch->key.width;
- unsigned height = batch->key.height;
-
- struct bifrost_framebuffer framebuffer = {
- .unk0 = 0x1e5, /* 1e4 if no spill */
- .width1 = MALI_POSITIVE(width),
- .height1 = MALI_POSITIVE(height),
- .width2 = MALI_POSITIVE(width),
- .height2 = MALI_POSITIVE(height),
-
- .unk1 = 0x1080,
-
- .rt_count_1 = MALI_POSITIVE(batch->key.nr_cbufs),
- .rt_count_2 = 4,
-
- .unknown2 = 0x1f,
-
- .scratchpad = ctx->scratchpad->gpu,
- .tiler = panfrost_emit_midg_tiler(batch, vertex_count)
- };
-
- return framebuffer;
-}
-
static void
panfrost_clear(
struct pipe_context *pipe,
double depth, unsigned stencil)
{
struct panfrost_context *ctx = pan_context(pipe);
- struct panfrost_batch *batch = panfrost_get_batch_for_fbo(ctx);
- panfrost_batch_clear(batch, buffers, color, depth, stencil);
-}
-
-static mali_ptr
-panfrost_attach_vt_mfbd(struct panfrost_batch *batch)
-{
- struct bifrost_framebuffer mfbd = panfrost_emit_mfbd(batch, ~0);
+ /* TODO: panfrost_get_fresh_batch_for_fbo() instantiates a new batch if
+ * the existing batch targeting this FBO has draws. We could probably
+ * avoid that by replacing plain clears by quad-draws with a specific
+ * color/depth/stencil value, thus avoiding the generation of extra
+ * fragment jobs.
+ */
+ struct panfrost_batch *batch = panfrost_get_fresh_batch_for_fbo(ctx);
- return panfrost_upload_transient(batch, &mfbd, sizeof(mfbd)) | MALI_MFBD;
+ panfrost_batch_add_fbo_bos(batch);
+ panfrost_batch_clear(batch, buffers, color, depth, stencil);
}
-static mali_ptr
-panfrost_attach_vt_sfbd(struct panfrost_batch *batch)
-{
- struct mali_single_framebuffer sfbd = panfrost_emit_sfbd(batch, ~0);
-
- return panfrost_upload_transient(batch, &sfbd, sizeof(sfbd)) | MALI_SFBD;
-}
+/* TODO: Bifrost requires just a mali_shared_memory, without the rest of the
+ * framebuffer */
static void
panfrost_attach_vt_framebuffer(struct panfrost_context *ctx)
{
- /* Skip the attach if we can */
-
- if (ctx->payloads[PIPE_SHADER_VERTEX].postfix.framebuffer) {
- assert(ctx->payloads[PIPE_SHADER_FRAGMENT].postfix.framebuffer);
- return;
- }
-
struct panfrost_screen *screen = pan_screen(ctx->base.screen);
struct panfrost_batch *batch = panfrost_get_batch_for_fbo(ctx);
- if (!batch->framebuffer)
- batch->framebuffer = screen->require_sfbd ?
- panfrost_attach_vt_sfbd(batch) :
- panfrost_attach_vt_mfbd(batch);
+ /* If we haven't, reserve space for the framebuffer */
+
+ if (!batch->framebuffer.gpu) {
+ unsigned size = (screen->quirks & MIDGARD_SFBD) ?
+ sizeof(struct mali_single_framebuffer) :
+ sizeof(struct mali_framebuffer);
+
+ batch->framebuffer = panfrost_allocate_transient(batch, size);
+
+ /* Tag the pointer */
+ if (!(screen->quirks & MIDGARD_SFBD))
+ batch->framebuffer.gpu |= MALI_MFBD;
+ }
for (unsigned i = 0; i < PIPE_SHADER_TYPES; ++i)
- ctx->payloads[i].postfix.framebuffer = batch->framebuffer;
+ ctx->payloads[i].postfix.shared_memory = batch->framebuffer.gpu;
}
/* Reset per-frame context, called on context initialisation as well as after
panfrost_invalidate_frame(struct panfrost_context *ctx)
{
for (unsigned i = 0; i < PIPE_SHADER_TYPES; ++i)
- ctx->payloads[i].postfix.framebuffer = 0;
-
- if (ctx->rasterizer)
- ctx->dirty |= PAN_DIRTY_RASTERIZER;
-
- /* XXX */
- ctx->dirty |= PAN_DIRTY_SAMPLERS | PAN_DIRTY_TEXTURES;
+ ctx->payloads[i].postfix.shared_memory = 0;
/* TODO: When does this need to be handled? */
ctx->active_queries = true;
memcpy(&ctx->payloads[PIPE_SHADER_COMPUTE], &payload, sizeof(payload));
}
-static void
-panfrost_emit_tiler_payload(struct panfrost_context *ctx)
-{
- struct midgard_payload_vertex_tiler payload = {
- .prefix = {
- .zero1 = 0xffff, /* Why is this only seen on test-quad-textured? */
- },
- };
-
- memcpy(&ctx->payloads[PIPE_SHADER_FRAGMENT], &payload, sizeof(payload));
-}
-
static unsigned
translate_tex_wrap(enum pipe_tex_wrap w)
{
case PIPE_TEX_WRAP_REPEAT:
return MALI_WRAP_REPEAT;
- /* TODO: lower GL_CLAMP? */
case PIPE_TEX_WRAP_CLAMP:
+ return MALI_WRAP_CLAMP;
+
case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
return MALI_WRAP_CLAMP_TO_EDGE;
case PIPE_TEX_WRAP_MIRROR_REPEAT:
return MALI_WRAP_MIRRORED_REPEAT;
+ case PIPE_TEX_WRAP_MIRROR_CLAMP:
+ return MALI_WRAP_MIRRORED_CLAMP;
+
+ case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
+ return MALI_WRAP_MIRRORED_CLAMP_TO_EDGE;
+
+ case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
+ return MALI_WRAP_MIRRORED_CLAMP_TO_BORDER;
+
default:
unreachable("Invalid wrap");
}
}
}
-static unsigned
-panfrost_translate_alt_compare_func(enum pipe_compare_func in)
-{
- switch (in) {
- case PIPE_FUNC_NEVER:
- return MALI_ALT_FUNC_NEVER;
-
- case PIPE_FUNC_LESS:
- return MALI_ALT_FUNC_LESS;
-
- case PIPE_FUNC_EQUAL:
- return MALI_ALT_FUNC_EQUAL;
-
- case PIPE_FUNC_LEQUAL:
- return MALI_ALT_FUNC_LEQUAL;
-
- case PIPE_FUNC_GREATER:
- return MALI_ALT_FUNC_GREATER;
-
- case PIPE_FUNC_NOTEQUAL:
- return MALI_ALT_FUNC_NOTEQUAL;
-
- case PIPE_FUNC_GEQUAL:
- return MALI_ALT_FUNC_GEQUAL;
-
- case PIPE_FUNC_ALWAYS:
- return MALI_ALT_FUNC_ALWAYS;
-
- default:
- unreachable("Invalid alt func");
- }
-}
-
static unsigned
panfrost_translate_stencil_op(enum pipe_stencil_op in)
{
static void
panfrost_default_shader_backend(struct panfrost_context *ctx)
{
+ struct panfrost_screen *screen = pan_screen(ctx->base.screen);
struct mali_shader_meta shader = {
.alpha_coverage = ~MALI_ALPHA_COVERAGE(0.000000),
.unknown2_4 = MALI_NO_MSAA | 0x4e0,
};
- /* unknown2_4 has 0x10 bit set on T6XX. We don't know why this is
+ /* unknown2_4 has 0x10 bit set on T6XX and T720. We don't know why this is
* required (independent of 32-bit/64-bit descriptors), or why it's not
* used on later GPU revisions. Otherwise, all shader jobs fault on
* these earlier chips (perhaps this is a chicken bit of some kind).
* More investigation is needed. */
- if (ctx->is_t6xx) {
+ if (screen->quirks & MIDGARD_SFBD)
shader.unknown2_4 |= 0x10;
- }
struct pipe_stencil_state default_stencil = {
.enabled = 0,
memcpy(&ctx->fragment_shader_core, &shader, sizeof(shader));
}
-/* Generates a vertex/tiler job. This is, in some sense, the heart of the
- * graphics command stream. It should be called once per draw, accordding to
- * presentations. Set is_tiler for "tiler" jobs (fragment shader jobs, but in
- * Mali parlance, "fragment" refers to framebuffer writeout). Clear it for
- * vertex jobs. */
-
-struct panfrost_transfer
-panfrost_vertex_tiler_job(struct panfrost_context *ctx, bool is_tiler)
-{
- struct panfrost_batch *batch = panfrost_get_batch_for_fbo(ctx);
- struct mali_job_descriptor_header job = {
- .job_type = is_tiler ? JOB_TYPE_TILER : JOB_TYPE_VERTEX,
- .job_descriptor_size = 1,
- };
-
- struct midgard_payload_vertex_tiler *payload = is_tiler ? &ctx->payloads[PIPE_SHADER_FRAGMENT] : &ctx->payloads[PIPE_SHADER_VERTEX];
-
- struct panfrost_transfer transfer = panfrost_allocate_transient(batch, sizeof(job) + sizeof(*payload));
- memcpy(transfer.cpu, &job, sizeof(job));
- memcpy(transfer.cpu + sizeof(job), payload, sizeof(*payload));
- return transfer;
-}
-
-mali_ptr
-panfrost_vertex_buffer_address(struct panfrost_context *ctx, unsigned i)
-{
- struct pipe_vertex_buffer *buf = &ctx->vertex_buffers[i];
- struct panfrost_resource *rsrc = (struct panfrost_resource *) (buf->buffer.resource);
-
- return rsrc->bo->gpu + buf->buffer_offset;
-}
-
static bool
panfrost_writes_point_size(struct panfrost_context *ctx)
{
struct panfrost_batch *batch = panfrost_get_batch_for_fbo(ctx);
struct panfrost_vertex_state *so = ctx->vertex;
- size_t sz = sizeof(struct mali_attr_meta) * so->num_elements;
+ size_t sz = sizeof(struct mali_attr_meta) * PAN_MAX_ATTRIBUTE;
struct panfrost_transfer transfer = panfrost_allocate_transient(batch, sz);
struct mali_attr_meta *target = (struct mali_attr_meta *) transfer.cpu;
for (unsigned i = 0; i < so->num_elements; ++i) {
unsigned vbi = so->pipe[i].vertex_buffer_index;
struct pipe_vertex_buffer *buf = &ctx->vertex_buffers[vbi];
- mali_ptr addr = panfrost_vertex_buffer_address(ctx, vbi);
+ struct panfrost_resource *rsrc = (struct panfrost_resource *) (buf->buffer.resource);
+ mali_ptr addr = rsrc->bo->gpu + buf->buffer_offset;
- /* Adjust by the masked off bits of the offset */
- target[i].src_offset += (addr & 63);
+ /* Adjust by the masked off bits of the offset. Make sure we
+ * read src_offset from so->hw (which is not GPU visible)
+ * rather than target (which is) due to caching effects */
+
+ unsigned src_offset = so->hw[i].src_offset;
+ src_offset += (addr & 63);
/* Also, somewhat obscurely per-instance data needs to be
* offset in response to a delayed start in an indexed draw */
- if (so->pipe[i].instance_divisor && ctx->instance_count > 1 && start) {
- target[i].src_offset -= buf->stride * start;
- }
-
+ if (so->pipe[i].instance_divisor && ctx->instance_count > 1 && start)
+ src_offset -= buf->stride * start;
+ target[i].src_offset = src_offset;
}
+ /* Let's also include vertex builtins */
+
+ struct mali_attr_meta builtin = {
+ .format = MALI_R32UI,
+ .swizzle = panfrost_get_default_swizzle(1)
+ };
+
+ /* See mali_attr_meta specification for the magic number */
+
+ builtin.index = so->vertexid_index;
+ memcpy(&target[PAN_VERTEX_ID], &builtin, 4);
+
+ builtin.index = so->vertexid_index + 1;
+ memcpy(&target[PAN_INSTANCE_ID], &builtin, 4);
+
ctx->payloads[PIPE_SHADER_VERTEX].postfix.attribute_meta = transfer.gpu;
}
for (int t = 0; t <= PIPE_SHADER_FRAGMENT; ++t) {
mali_ptr upload = 0;
- if (ctx->sampler_count[t] && ctx->sampler_view_count[t]) {
+ if (ctx->sampler_count[t]) {
size_t transfer_size = desc_size * ctx->sampler_count[t];
struct panfrost_transfer transfer =
}
}
-static enum mali_texture_layout
-panfrost_layout_for_texture(struct panfrost_resource *rsrc)
-{
- /* TODO: other linear depth textures */
- bool is_depth = rsrc->base.format == PIPE_FORMAT_Z32_UNORM;
-
- switch (rsrc->layout) {
- case PAN_AFBC:
- return MALI_TEXTURE_AFBC;
- case PAN_TILED:
- assert(!is_depth);
- return MALI_TEXTURE_TILED;
- case PAN_LINEAR:
- return is_depth ? MALI_TEXTURE_TILED : MALI_TEXTURE_LINEAR;
- default:
- unreachable("Invalid texture layout");
- }
-}
-
static mali_ptr
panfrost_upload_tex(
struct panfrost_context *ctx,
+ enum pipe_shader_type st,
struct panfrost_sampler_view *view)
{
if (!view)
struct pipe_sampler_view *pview = &view->base;
struct panfrost_resource *rsrc = pan_resource(pview->texture);
- /* Do we interleave an explicit stride with every element? */
-
- bool has_manual_stride = view->manual_stride;
-
- /* For easy access */
-
- bool is_buffer = pview->target == PIPE_BUFFER;
- unsigned first_level = is_buffer ? 0 : pview->u.tex.first_level;
- unsigned last_level = is_buffer ? 0 : pview->u.tex.last_level;
- unsigned first_layer = is_buffer ? 0 : pview->u.tex.first_layer;
- unsigned last_layer = is_buffer ? 0 : pview->u.tex.last_layer;
-
- /* Lower-bit is set when sampling from colour AFBC */
- bool is_afbc = rsrc->layout == PAN_AFBC;
- bool is_zs = rsrc->base.bind & PIPE_BIND_DEPTH_STENCIL;
- unsigned afbc_bit = (is_afbc && !is_zs) ? 1 : 0;
-
/* Add the BO to the job so it's retained until the job is done. */
struct panfrost_batch *batch = panfrost_get_batch_for_fbo(ctx);
- panfrost_batch_add_bo(batch, rsrc->bo);
-
- /* Add the usage flags in, since they can change across the CSO
- * lifetime due to layout switches */
- view->hw.format.layout = panfrost_layout_for_texture(rsrc);
- view->hw.format.manual_stride = has_manual_stride;
+ panfrost_batch_add_bo(batch, rsrc->bo,
+ PAN_BO_ACCESS_SHARED | PAN_BO_ACCESS_READ |
+ panfrost_bo_access_for_stage(st));
- /* Inject the addresses in, interleaving mip levels, cube faces, and
- * strides in that order */
+ panfrost_batch_add_bo(batch, view->bo,
+ PAN_BO_ACCESS_SHARED | PAN_BO_ACCESS_READ |
+ panfrost_bo_access_for_stage(st));
- unsigned idx = 0;
-
- for (unsigned l = first_level; l <= last_level; ++l) {
- for (unsigned f = first_layer; f <= last_layer; ++f) {
-
- view->hw.payload[idx++] =
- panfrost_get_texture_address(rsrc, l, f) + afbc_bit;
-
- if (has_manual_stride) {
- view->hw.payload[idx++] =
- rsrc->slices[l].stride;
- }
- }
- }
-
- return panfrost_upload_transient(batch, &view->hw,
- sizeof(struct mali_texture_descriptor));
+ return view->bo->gpu;
}
static void
for (int i = 0; i < ctx->sampler_view_count[t]; ++i)
trampolines[i] =
- panfrost_upload_tex(ctx, ctx->sampler_views[t][i]);
+ panfrost_upload_tex(ctx, t, ctx->sampler_views[t][i]);
trampoline = panfrost_upload_transient(batch, trampolines, sizeof(uint64_t) * ctx->sampler_view_count[t]);
}
struct panfrost_batch *batch = panfrost_get_batch_for_fbo(ctx);
struct panfrost_bo *bo = pan_resource(sb.buffer)->bo;
- panfrost_batch_add_bo(batch, bo);
+ panfrost_batch_add_bo(batch, bo,
+ PAN_BO_ACCESS_SHARED | PAN_BO_ACCESS_RW |
+ panfrost_bo_access_for_stage(st));
/* Upload address and size as sysval */
uniform->du[0] = bo->gpu + sb.buffer_offset;
uniform->u[2] = sb.buffer_size;
}
+static void
+panfrost_upload_sampler_sysval(
+ struct panfrost_context *ctx,
+ enum pipe_shader_type st,
+ unsigned sampler_index,
+ struct sysval_uniform *uniform)
+{
+ struct pipe_sampler_state *sampl =
+ &ctx->samplers[st][sampler_index]->base;
+
+ uniform->f[0] = sampl->min_lod;
+ uniform->f[1] = sampl->max_lod;
+ uniform->f[2] = sampl->lod_bias;
+
+ /* Even without any errata, Midgard represents "no mipmapping" as
+ * fixing the LOD with the clamps; keep behaviour consistent. c.f.
+ * panfrost_create_sampler_state which also explains our choice of
+ * epsilon value (again to keep behaviour consistent) */
+
+ if (sampl->min_mip_filter == PIPE_TEX_MIPFILTER_NONE)
+ uniform->f[1] = uniform->f[0] + (1.0/256.0);
+}
+
static void panfrost_upload_num_work_groups_sysval(struct panfrost_context *ctx,
struct sysval_uniform *uniform)
{
case PAN_SYSVAL_NUM_WORK_GROUPS:
panfrost_upload_num_work_groups_sysval(ctx, &uniforms[i]);
break;
-
+ case PAN_SYSVAL_SAMPLER:
+ panfrost_upload_sampler_sysval(ctx, st, PAN_SYSVAL_ID(sysval),
+ &uniforms[i]);
+ break;
default:
assert(0);
}
static mali_ptr
panfrost_map_constant_buffer_gpu(
struct panfrost_context *ctx,
+ enum pipe_shader_type st,
struct panfrost_constant_buffer *buf,
unsigned index)
{
struct panfrost_batch *batch = panfrost_get_batch_for_fbo(ctx);
if (rsrc) {
- panfrost_batch_add_bo(batch, rsrc->bo);
- return rsrc->bo->gpu;
+ panfrost_batch_add_bo(batch, rsrc->bo,
+ PAN_BO_ACCESS_SHARED |
+ PAN_BO_ACCESS_READ |
+ panfrost_bo_access_for_stage(st));
+
+ /* Alignment gauranteed by PIPE_CAP_CONSTANT_BUFFER_OFFSET_ALIGNMENT */
+ return rsrc->bo->gpu + cb->buffer_offset;
} else if (cb->user_buffer) {
- return panfrost_upload_transient(batch, cb->user_buffer, cb->buffer_size);
+ return panfrost_upload_transient(batch, cb->user_buffer + cb->buffer_offset, cb->buffer_size);
} else {
unreachable("No constant buffer");
}
return 32 - __builtin_clz(mask);
}
-/* Fixes up a shader state with current state, returning a GPU address to the
- * patched shader */
+/* Fixes up a shader state with current state */
-static mali_ptr
-panfrost_patch_shader_state(
- struct panfrost_context *ctx,
- struct panfrost_shader_state *ss,
- enum pipe_shader_type stage,
- bool should_upload)
+static void
+panfrost_patch_shader_state(struct panfrost_context *ctx,
+ enum pipe_shader_type stage)
{
+ struct panfrost_shader_variants *all = ctx->shader[stage];
+
+ if (!all) {
+ ctx->payloads[stage].postfix.shader = 0;
+ return;
+ }
+
+ struct panfrost_shader_state *ss = &all->variants[all->active_variant];
+
ss->tripipe->texture_count = ctx->sampler_view_count[stage];
ss->tripipe->sampler_count = ctx->sampler_count[stage];
- ss->tripipe->midgard1.flags = 0x220;
+ ss->tripipe->midgard1.flags_lo = 0x220;
unsigned ubo_count = panfrost_ubo_count(ctx, stage);
ss->tripipe->midgard1.uniform_buffer_count = ubo_count;
- /* We can't reuse over frames; that's not safe. The descriptor must be
- * transient uploaded */
-
- if (should_upload) {
- struct panfrost_batch *batch = panfrost_get_batch_for_fbo(ctx);
-
- return panfrost_upload_transient(batch, ss->tripipe,
- sizeof(struct mali_shader_meta));
- }
-
- /* If we don't need an upload, don't bother */
- return 0;
-
-}
-
-static void
-panfrost_patch_shader_state_compute(
- struct panfrost_context *ctx,
- enum pipe_shader_type stage,
- bool should_upload)
-{
- struct panfrost_shader_variants *all = ctx->shader[stage];
-
- if (!all) {
- ctx->payloads[stage].postfix._shader_upper = 0;
- return;
- }
+ struct panfrost_batch *batch = panfrost_get_batch_for_fbo(ctx);
- struct panfrost_shader_state *s = &all->variants[all->active_variant];
+ /* Add the shader BO to the batch. */
+ panfrost_batch_add_bo(batch, ss->bo,
+ PAN_BO_ACCESS_PRIVATE |
+ PAN_BO_ACCESS_READ |
+ panfrost_bo_access_for_stage(stage));
- ctx->payloads[stage].postfix._shader_upper =
- panfrost_patch_shader_state(ctx, s, stage, should_upload) >> 4;
+ ctx->payloads[stage].postfix.shader = panfrost_upload_transient(batch,
+ ss->tripipe,
+ sizeof(struct mali_shader_meta));
}
/* Go through dirty flags and actualise them in the cmdstream. */
struct panfrost_batch *batch = panfrost_get_batch_for_fbo(ctx);
struct panfrost_screen *screen = pan_screen(ctx->base.screen);
+ panfrost_batch_add_fbo_bos(batch);
panfrost_attach_vt_framebuffer(ctx);
if (with_vertex_data) {
panfrost_emit_varying_descriptor(ctx, total_count);
}
- bool msaa = ctx->rasterizer->base.multisample;
- if (ctx->dirty & PAN_DIRTY_RASTERIZER) {
+ if (ctx->rasterizer) {
+ bool msaa = ctx->rasterizer->base.multisample;
ctx->payloads[PIPE_SHADER_FRAGMENT].gl_enables = ctx->rasterizer->tiler_gl_enables;
/* TODO: Sample size */
panfrost_batch_set_requirements(batch);
if (ctx->occlusion_query) {
- ctx->payloads[PIPE_SHADER_FRAGMENT].gl_enables |= MALI_OCCLUSION_QUERY | MALI_OCCLUSION_PRECISE;
- ctx->payloads[PIPE_SHADER_FRAGMENT].postfix.occlusion_counter = ctx->occlusion_query->transfer.gpu;
+ ctx->payloads[PIPE_SHADER_FRAGMENT].gl_enables |= MALI_OCCLUSION_QUERY;
+ ctx->payloads[PIPE_SHADER_FRAGMENT].postfix.occlusion_counter = ctx->occlusion_query->bo->gpu;
}
- panfrost_patch_shader_state_compute(ctx, PIPE_SHADER_VERTEX, true);
- panfrost_patch_shader_state_compute(ctx, PIPE_SHADER_COMPUTE, true);
+ panfrost_patch_shader_state(ctx, PIPE_SHADER_VERTEX);
+ panfrost_patch_shader_state(ctx, PIPE_SHADER_COMPUTE);
- if (ctx->dirty & (PAN_DIRTY_RASTERIZER | PAN_DIRTY_VS)) {
+ if (ctx->shader[PIPE_SHADER_VERTEX] && ctx->shader[PIPE_SHADER_FRAGMENT]) {
/* Check if we need to link the gl_PointSize varying */
if (!panfrost_writes_point_size(ctx)) {
/* If the size is constant, write it out. Otherwise,
* don't touch primitive_size (since we would clobber
* the pointer there) */
- ctx->payloads[PIPE_SHADER_FRAGMENT].primitive_size.constant = ctx->rasterizer->base.line_width;
+ bool points = ctx->payloads[PIPE_SHADER_FRAGMENT].prefix.draw_mode == MALI_POINTS;
+
+ ctx->payloads[PIPE_SHADER_FRAGMENT].primitive_size.constant = points ?
+ ctx->rasterizer->base.point_size :
+ ctx->rasterizer->base.line_width;
}
}
- /* TODO: Maybe dirty track FS, maybe not. For now, it's transient. */
- if (ctx->shader[PIPE_SHADER_FRAGMENT])
- ctx->dirty |= PAN_DIRTY_FS;
-
- if (ctx->dirty & PAN_DIRTY_FS) {
- assert(ctx->shader[PIPE_SHADER_FRAGMENT]);
+ if (ctx->shader[PIPE_SHADER_FRAGMENT]) {
struct panfrost_shader_state *variant = &ctx->shader[PIPE_SHADER_FRAGMENT]->variants[ctx->shader[PIPE_SHADER_FRAGMENT]->active_variant];
- panfrost_patch_shader_state(ctx, variant, PIPE_SHADER_FRAGMENT, false);
-
- panfrost_batch_add_bo(batch, variant->bo);
+ panfrost_patch_shader_state(ctx, PIPE_SHADER_FRAGMENT);
#define COPY(name) ctx->fragment_shader_core.name = variant->tripipe->name
COPY(midgard1.uniform_count);
COPY(midgard1.uniform_buffer_count);
COPY(midgard1.work_count);
- COPY(midgard1.flags);
- COPY(midgard1.unknown2);
+ COPY(midgard1.flags_lo);
+ COPY(midgard1.flags_hi);
#undef COPY
unsigned rt_count = MAX2(ctx->pipe_framebuffer.nr_cbufs, 1);
struct panfrost_blend_final blend[PIPE_MAX_COLOR_BUFS];
+ unsigned shader_offset = 0;
+ struct panfrost_bo *shader_bo = NULL;
- for (unsigned c = 0; c < rt_count; ++c)
- blend[c] = panfrost_get_blend_for_context(ctx, c);
+ for (unsigned c = 0; c < rt_count; ++c) {
+ blend[c] = panfrost_get_blend_for_context(ctx, c, &shader_bo, &shader_offset);
+ }
/* If there is a blend shader, work registers are shared. XXX: opt */
ctx->fragment_shader_core.midgard1.work_count = 16;
}
- /* Set late due to depending on render state */
- unsigned flags = ctx->fragment_shader_core.midgard1.flags;
-
/* Depending on whether it's legal to in the given shader, we
* try to enable early-z testing (or forward-pixel kill?) */
- if (!variant->can_discard)
- flags |= MALI_EARLY_Z;
+ SET_BIT(ctx->fragment_shader_core.midgard1.flags_lo, MALI_EARLY_Z,
+ !variant->can_discard && !variant->writes_depth);
+
+ /* Add the writes Z/S flags if needed. */
+ SET_BIT(ctx->fragment_shader_core.midgard1.flags_lo,
+ MALI_WRITES_Z, variant->writes_depth);
+ SET_BIT(ctx->fragment_shader_core.midgard1.flags_hi,
+ MALI_WRITES_S, variant->writes_stencil);
/* Any time texturing is used, derivatives are implicitly
* calculated, so we need to enable helper invocations */
- if (variant->helper_invocations)
- flags |= MALI_HELPER_INVOCATIONS;
-
- ctx->fragment_shader_core.midgard1.flags = flags;
+ SET_BIT(ctx->fragment_shader_core.midgard1.flags_lo, MALI_HELPER_INVOCATIONS, variant->helper_invocations);
/* Assign the stencil refs late */
* thing?" by Peter Harris
*/
- if (variant->can_discard) {
- ctx->fragment_shader_core.unknown2_3 |= MALI_CAN_DISCARD;
- ctx->fragment_shader_core.midgard1.flags |= 0x400;
- }
+ SET_BIT(ctx->fragment_shader_core.unknown2_3, MALI_CAN_DISCARD, variant->can_discard);
+ SET_BIT(ctx->fragment_shader_core.midgard1.flags_lo, 0x400, variant->can_discard);
/* Even on MFBD, the shader descriptor gets blend shaders. It's
* *also* copied to the blend_meta appended (by convention),
* but this is the field actually read by the hardware. (Or
- * maybe both are read...?) */
+ * maybe both are read...?). Specify the last RTi with a blend
+ * shader. */
- if (blend[0].is_shader) {
- ctx->fragment_shader_core.blend.shader =
- blend[0].shader.bo->gpu | blend[0].shader.first_tag;
- } else {
- ctx->fragment_shader_core.blend.shader = 0;
+ ctx->fragment_shader_core.blend.shader = 0;
+
+ for (signed rt = (rt_count - 1); rt >= 0; --rt) {
+ if (blend[rt].is_shader) {
+ ctx->fragment_shader_core.blend.shader =
+ blend[rt].shader.gpu | blend[rt].shader.first_tag;
+ break;
+ }
}
- if (screen->require_sfbd) {
+ if (screen->quirks & MIDGARD_SFBD) {
/* When only a single render target platform is used, the blend
* information is inside the shader meta itself. We
* additionally need to signal CAN_DISCARD for nontrivial blend
* modes (so we're able to read back the destination buffer) */
+ SET_BIT(ctx->fragment_shader_core.unknown2_3, MALI_HAS_BLEND_SHADER, blend[0].is_shader);
+
if (!blend[0].is_shader) {
ctx->fragment_shader_core.blend.equation =
*blend[0].equation.equation;
blend[0].equation.constant;
}
- if (!blend[0].no_blending) {
- ctx->fragment_shader_core.unknown2_3 |= MALI_CAN_DISCARD;
- }
+ SET_BIT(ctx->fragment_shader_core.unknown2_3, MALI_CAN_DISCARD, !blend[0].no_blending);
}
size_t size = sizeof(struct mali_shader_meta) + (sizeof(struct midgard_blend_rt) * rt_count);
struct panfrost_transfer transfer = panfrost_allocate_transient(batch, size);
memcpy(transfer.cpu, &ctx->fragment_shader_core, sizeof(struct mali_shader_meta));
- ctx->payloads[PIPE_SHADER_FRAGMENT].postfix._shader_upper = (transfer.gpu) >> 4;
+ ctx->payloads[PIPE_SHADER_FRAGMENT].postfix.shader = transfer.gpu;
- if (!screen->require_sfbd) {
+ if (!(screen->quirks & MIDGARD_SFBD)) {
/* Additional blend descriptor tacked on for jobs using MFBD */
struct midgard_blend_rt rts[4];
for (unsigned i = 0; i < rt_count; ++i) {
- unsigned blend_count = 0x200;
-
- if (blend[i].is_shader) {
- /* For a blend shader, the bottom nibble corresponds to
- * the number of work registers used, which signals the
- * -existence- of a blend shader */
-
- assert(blend[i].shader.work_count >= 2);
- blend_count |= MIN2(blend[i].shader.work_count, 3);
- } else {
- /* Otherwise, the bottom bit simply specifies if
- * blending (anything other than REPLACE) is enabled */
-
- if (!blend[i].no_blending)
- blend_count |= 0x1;
- }
-
+ rts[i].flags = 0x200;
bool is_srgb =
(ctx->pipe_framebuffer.nr_cbufs > i) &&
(ctx->pipe_framebuffer.cbufs[i]) &&
util_format_is_srgb(ctx->pipe_framebuffer.cbufs[i]->format);
- rts[i].flags = blend_count;
-
- if (is_srgb)
- rts[i].flags |= MALI_BLEND_SRGB;
-
- if (!ctx->blend->base.dither)
- rts[i].flags |= MALI_BLEND_NO_DITHER;
-
- /* TODO: sRGB in blend shaders is currently
- * unimplemented. Contact me (Alyssa) if you're
- * interested in working on this. We have
- * native Midgard ops for helping here, but
- * they're not well-understood yet. */
-
- assert(!(is_srgb && blend[i].is_shader));
+ SET_BIT(rts[i].flags, MALI_BLEND_MRT_SHADER, blend[i].is_shader);
+ SET_BIT(rts[i].flags, MALI_BLEND_LOAD_TIB, !blend[i].no_blending);
+ SET_BIT(rts[i].flags, MALI_BLEND_SRGB, is_srgb);
+ SET_BIT(rts[i].flags, MALI_BLEND_NO_DITHER, !ctx->blend->base.dither);
if (blend[i].is_shader) {
- rts[i].blend.shader = blend[i].shader.bo->gpu | blend[i].shader.first_tag;
+ rts[i].blend.shader = blend[i].shader.gpu | blend[i].shader.first_tag;
} else {
rts[i].blend.equation = *blend[i].equation.equation;
rts[i].blend.constant = blend[i].equation.constant;
if (ctx->vertex)
panfrost_stage_attributes(ctx);
- if (ctx->dirty & PAN_DIRTY_SAMPLERS)
- panfrost_upload_sampler_descriptors(ctx);
-
- if (ctx->dirty & PAN_DIRTY_TEXTURES)
- panfrost_upload_texture_descriptors(ctx);
+ panfrost_upload_sampler_descriptors(ctx);
+ panfrost_upload_texture_descriptors(ctx);
const struct pipe_viewport_state *vp = &ctx->pipe_viewport;
struct panfrost_shader_state *ss = &all->variants[all->active_variant];
- panfrost_batch_add_bo(batch, ss->bo);
-
/* Uniforms are implicitly UBO #0 */
bool has_uniforms = buf->enabled_mask & (1 << 0);
panfrost_upload_sysvals(ctx, transfer.cpu, ss, i);
/* Upload uniforms */
- if (has_uniforms) {
+ if (has_uniforms && uniform_size) {
const void *cpu = panfrost_map_constant_buffer_cpu(buf, 0);
memcpy(transfer.cpu + sys_size, cpu, uniform_size);
}
unsigned ubo_count = panfrost_ubo_count(ctx, i);
assert(ubo_count >= 1);
- size_t sz = sizeof(struct mali_uniform_buffer_meta) * ubo_count;
- struct mali_uniform_buffer_meta ubos[PAN_MAX_CONST_BUFFERS];
+ size_t sz = sizeof(uint64_t) * ubo_count;
+ uint64_t ubos[PAN_MAX_CONST_BUFFERS];
/* Upload uniforms as a UBO */
- ubos[0].size = MALI_POSITIVE((2 + uniform_count));
- ubos[0].ptr = transfer.gpu >> 2;
+ ubos[0] = MALI_MAKE_UBO(2 + uniform_count, transfer.gpu);
/* The rest are honest-to-goodness UBOs */
if (!enabled || empty) {
/* Stub out disabled UBOs to catch accesses */
-
- ubos[ubo].size = 0;
- ubos[ubo].ptr = 0xDEAD0000;
+ ubos[ubo] = MALI_MAKE_UBO(0, 0xDEAD0000);
continue;
}
- mali_ptr gpu = panfrost_map_constant_buffer_gpu(ctx, buf, ubo);
+ mali_ptr gpu = panfrost_map_constant_buffer_gpu(ctx, i, buf, ubo);
unsigned bytes_per_field = 16;
unsigned aligned = ALIGN_POT(usz, bytes_per_field);
- unsigned fields = aligned / bytes_per_field;
-
- ubos[ubo].size = MALI_POSITIVE(fields);
- ubos[ubo].ptr = gpu >> 2;
+ ubos[ubo] = MALI_MAKE_UBO(aligned / bytes_per_field, gpu);
}
mali_ptr ubufs = panfrost_upload_transient(batch, ubos, sz);
panfrost_upload_transient(batch,
&view,
sizeof(struct mali_viewport));
-
- ctx->dirty = 0;
}
/* Corresponds to exactly one draw, but does not submit anything */
bool rasterizer_discard = ctx->rasterizer
&& ctx->rasterizer->base.rasterizer_discard;
- struct panfrost_transfer vertex = panfrost_vertex_tiler_job(ctx, false);
- struct panfrost_transfer tiler;
- if (!rasterizer_discard)
- tiler = panfrost_vertex_tiler_job(ctx, true);
+ struct midgard_payload_vertex_tiler *vertex_payload = &ctx->payloads[PIPE_SHADER_VERTEX];
+ struct midgard_payload_vertex_tiler *tiler_payload = &ctx->payloads[PIPE_SHADER_FRAGMENT];
struct panfrost_batch *batch = panfrost_get_batch_for_fbo(ctx);
+ bool wallpapering = ctx->wallpaper_batch && batch->tiler_dep;
- if (rasterizer_discard)
- panfrost_scoreboard_queue_vertex_job(batch, vertex, FALSE);
- else if (ctx->wallpaper_batch)
- panfrost_scoreboard_queue_fused_job_prepend(batch, vertex, tiler);
- else
- panfrost_scoreboard_queue_fused_job(batch, vertex, tiler);
+ if (wallpapering) {
+ /* Inject in reverse order, with "predicted" job indices. THIS IS A HACK XXX */
+ panfrost_new_job(batch, JOB_TYPE_TILER, false, batch->job_index + 2, tiler_payload, sizeof(*tiler_payload), true);
+ panfrost_new_job(batch, JOB_TYPE_VERTEX, false, 0, vertex_payload, sizeof(*vertex_payload), true);
+ } else {
+ unsigned vertex = panfrost_new_job(batch, JOB_TYPE_VERTEX, false, 0, vertex_payload, sizeof(*vertex_payload), false);
+
+ if (!rasterizer_discard)
+ panfrost_new_job(batch, JOB_TYPE_TILER, false, vertex, tiler_payload, sizeof(*tiler_payload), false);
+ }
+
+ for (unsigned i = 0; i < PIPE_SHADER_TYPES; ++i) {
+ struct panfrost_shader_variants *all = ctx->shader[i];
+
+ if (!all)
+ continue;
+
+ struct panfrost_shader_state *ss = &all->variants[all->active_variant];
+ batch->stack_size = MAX2(batch->stack_size, ss->stack_size);
+ }
}
/* The entire frame is in memory -- send it off to the kernel! */
unsigned flags)
{
struct panfrost_context *ctx = pan_context(pipe);
- struct panfrost_batch *batch = panfrost_get_batch_for_fbo(ctx);
+ struct util_dynarray fences;
- /* Submit the frame itself */
- panfrost_batch_submit(batch);
+ /* We must collect the fences before the flush is done, otherwise we'll
+ * lose track of them.
+ */
+ if (fence) {
+ util_dynarray_init(&fences, NULL);
+ hash_table_foreach(ctx->batches, hentry) {
+ struct panfrost_batch *batch = hentry->data;
+
+ panfrost_batch_fence_reference(batch->out_sync);
+ util_dynarray_append(&fences,
+ struct panfrost_batch_fence *,
+ batch->out_sync);
+ }
+ }
+
+ /* Submit all pending jobs */
+ panfrost_flush_all_batches(ctx, false);
if (fence) {
- struct panfrost_fence *f = panfrost_fence_create(ctx);
+ struct panfrost_fence *f = panfrost_fence_create(ctx, &fences);
pipe->screen->fence_reference(pipe->screen, fence, NULL);
*fence = (struct pipe_fence_handle *)f;
+
+ util_dynarray_foreach(&fences, struct panfrost_batch_fence *, fence)
+ panfrost_batch_fence_unreference(*fence);
+
+ util_dynarray_fini(&fences);
}
+
+ if (pan_debug & PAN_DBG_TRACE)
+ pandecode_next_frame();
}
#define DEFINE_CASE(c) case PIPE_PRIM_##c: return MALI_##c;
}
/* Gets a GPU address for the associated index buffer. Only gauranteed to be
- * good for the duration of the draw (transient), could last longer */
+ * good for the duration of the draw (transient), could last longer. Also get
+ * the bounds on the index buffer for the range accessed by the draw. We do
+ * these operations together because there are natural optimizations which
+ * require them to be together. */
static mali_ptr
-panfrost_get_index_buffer_mapped(struct panfrost_context *ctx, const struct pipe_draw_info *info)
+panfrost_get_index_buffer_bounded(struct panfrost_context *ctx, const struct pipe_draw_info *info, unsigned *min_index, unsigned *max_index)
{
struct panfrost_resource *rsrc = (struct panfrost_resource *) (info->index.resource);
off_t offset = info->start * info->index_size;
struct panfrost_batch *batch = panfrost_get_batch_for_fbo(ctx);
+ mali_ptr out = 0;
+
+ bool needs_indices = true;
+
+ if (info->max_index != ~0u) {
+ *min_index = info->min_index;
+ *max_index = info->max_index;
+ needs_indices = false;
+ }
if (!info->has_user_indices) {
/* Only resources can be directly mapped */
- panfrost_batch_add_bo(batch, rsrc->bo);
- return rsrc->bo->gpu + offset;
+ panfrost_batch_add_bo(batch, rsrc->bo,
+ PAN_BO_ACCESS_SHARED |
+ PAN_BO_ACCESS_READ |
+ PAN_BO_ACCESS_VERTEX_TILER);
+ out = rsrc->bo->gpu + offset;
+
+ /* Check the cache */
+ needs_indices = !panfrost_minmax_cache_get(rsrc->index_cache, info->start, info->count,
+ min_index, max_index);
} else {
/* Otherwise, we need to upload to transient memory */
const uint8_t *ibuf8 = (const uint8_t *) info->index.user;
- return panfrost_upload_transient(batch, ibuf8 + offset, info->count * info->index_size);
+ out = panfrost_upload_transient(batch, ibuf8 + offset, info->count * info->index_size);
+ }
+
+ if (needs_indices) {
+ /* Fallback */
+ u_vbuf_get_minmax_index(&ctx->base, info, min_index, max_index);
+
+ if (!info->has_user_indices) {
+ panfrost_minmax_cache_add(rsrc->index_cache, info->start, info->count,
+ *min_index, *max_index);
+ }
}
+
+
+ return out;
}
static bool
if (info->primitive_restart)
draw_flags |= MALI_DRAW_PRIMITIVE_RESTART_FIXED_INDEX;
- /* For higher amounts of vertices (greater than what fits in a 16-bit
- * short), the other value is needed, otherwise there will be bizarre
- * rendering artefacts. It's not clear what these values mean yet. This
- * change is also needed for instancing and sometimes points (perhaps
- * related to dynamically setting gl_PointSize) */
+ /* These doesn't make much sense */
- bool is_points = mode == PIPE_PRIM_POINTS;
- bool many_verts = ctx->vertex_count > 0xFFFF;
- bool instanced = ctx->instance_count > 1;
+ draw_flags |= 0x3000;
- draw_flags |= (is_points || many_verts || instanced) ? 0x3000 : 0x18000;
-
- /* This doesn't make much sense */
- if (mode == PIPE_PRIM_LINE_STRIP) {
- draw_flags |= 0x800;
- }
+ if (ctx->rasterizer && ctx->rasterizer->base.flatshade_first)
+ draw_flags |= MALI_DRAW_FLATSHADE_FIRST;
panfrost_statistics_record(ctx, info);
if (info->index_size) {
- /* Calculate the min/max index used so we can figure out how
- * many times to invoke the vertex shader */
-
- /* Fetch / calculate index bounds */
unsigned min_index = 0, max_index = 0;
-
- if (info->max_index == ~0u) {
- u_vbuf_get_minmax_index(pipe, info, &min_index, &max_index);
- } else {
- min_index = info->min_index;
- max_index = info->max_index;
- }
+ ctx->payloads[PIPE_SHADER_FRAGMENT].prefix.indices =
+ panfrost_get_index_buffer_bounded(ctx, info, &min_index, &max_index);
/* Use the corresponding values */
vertex_count = max_index - min_index + 1;
ctx->payloads[PIPE_SHADER_FRAGMENT].prefix.offset_bias_correction = -min_index;
ctx->payloads[PIPE_SHADER_FRAGMENT].prefix.index_count = MALI_POSITIVE(info->count);
- //assert(!info->restart_index); /* TODO: Research */
-
draw_flags |= panfrost_translate_index_size(info->index_size);
- ctx->payloads[PIPE_SHADER_FRAGMENT].prefix.indices = panfrost_get_index_buffer_mapped(ctx, info);
} else {
/* Index count == vertex count, if no indexing is applied, as
* if it is internally indexed in the expected order */
ctx->payloads[PIPE_SHADER_FRAGMENT].prefix.index_count = MALI_POSITIVE(ctx->vertex_count);
/* Reverse index state */
- ctx->payloads[PIPE_SHADER_FRAGMENT].prefix.indices = (u64) NULL;
+ ctx->payloads[PIPE_SHADER_FRAGMENT].prefix.indices = (mali_ptr) 0;
}
/* Dispatch "compute jobs" for the vertex/tiler pair as (1,
/* Encode the padded vertex count */
if (info->instance_count > 1) {
- /* Triangles have non-even vertex counts so they change how
- * padding works internally */
-
- bool is_triangle =
- mode == PIPE_PRIM_TRIANGLES ||
- mode == PIPE_PRIM_TRIANGLE_STRIP ||
- mode == PIPE_PRIM_TRIANGLE_FAN;
+ ctx->padded_count = panfrost_padded_vertex_count(vertex_count);
- struct pan_shift_odd so =
- panfrost_padded_vertex_count(vertex_count, !is_triangle);
+ unsigned shift = __builtin_ctz(ctx->padded_count);
+ unsigned k = ctx->padded_count >> (shift + 1);
- ctx->payloads[PIPE_SHADER_VERTEX].instance_shift = so.shift;
- ctx->payloads[PIPE_SHADER_FRAGMENT].instance_shift = so.shift;
+ ctx->payloads[PIPE_SHADER_VERTEX].instance_shift = shift;
+ ctx->payloads[PIPE_SHADER_FRAGMENT].instance_shift = shift;
- ctx->payloads[PIPE_SHADER_VERTEX].instance_odd = so.odd;
- ctx->payloads[PIPE_SHADER_FRAGMENT].instance_odd = so.odd;
-
- ctx->padded_count = pan_expand_shift_odd(so);
+ ctx->payloads[PIPE_SHADER_VERTEX].instance_odd = k;
+ ctx->payloads[PIPE_SHADER_FRAGMENT].instance_odd = k;
} else {
- ctx->padded_count = ctx->vertex_count;
+ ctx->padded_count = vertex_count;
/* Reset instancing state */
ctx->payloads[PIPE_SHADER_VERTEX].instance_shift = 0;
{
struct panfrost_context *ctx = pan_context(pctx);
- /* TODO: Why can't rasterizer be NULL ever? Other drivers are fine.. */
+ ctx->rasterizer = hwcso;
+
if (!hwcso)
return;
- ctx->rasterizer = hwcso;
- ctx->dirty |= PAN_DIRTY_RASTERIZER;
-
- ctx->fragment_shader_core.depth_units = ctx->rasterizer->base.offset_units;
+ ctx->fragment_shader_core.depth_units = ctx->rasterizer->base.offset_units * 2.0f;
ctx->fragment_shader_core.depth_factor = ctx->rasterizer->base.offset_scale;
/* Gauranteed with the core GL call, so don't expose ARB_polygon_offset */
void *hwcso)
{
struct panfrost_context *ctx = pan_context(pctx);
-
ctx->vertex = hwcso;
- ctx->dirty |= PAN_DIRTY_VERTEX;
}
static void *
panfrost_create_shader_state(
struct pipe_context *pctx,
- const struct pipe_shader_state *cso)
+ const struct pipe_shader_state *cso,
+ enum pipe_shader_type stage)
{
struct panfrost_shader_variants *so = CALLOC_STRUCT(panfrost_shader_variants);
so->base = *cso;
if (cso->type == PIPE_SHADER_IR_TGSI)
so->base.tokens = tgsi_dup_tokens(so->base.tokens);
+ /* Precompile for shader-db if we need to */
+ if (unlikely((pan_debug & PAN_DBG_PRECOMPILE) && cso->type == PIPE_SHADER_IR_NIR)) {
+ struct panfrost_context *ctx = pan_context(pctx);
+
+ struct mali_shader_meta meta;
+ struct panfrost_shader_state state;
+ uint64_t outputs_written;
+
+ panfrost_shader_compile(ctx, &meta,
+ PIPE_SHADER_IR_NIR,
+ so->base.ir.nir,
+ tgsi_processor_to_shader_stage(stage), &state,
+ &outputs_written);
+ }
+
return so;
}
for (unsigned i = 0; i < cso->variant_count; ++i) {
struct panfrost_shader_state *shader_state = &cso->variants[i];
- panfrost_bo_unreference(pctx->screen, shader_state->bo);
+ panfrost_bo_unreference(shader_state->bo);
shader_state->bo = NULL;
}
+ free(cso->variants);
free(so);
}
.wrap_s = translate_tex_wrap(cso->wrap_s),
.wrap_t = translate_tex_wrap(cso->wrap_t),
.wrap_r = translate_tex_wrap(cso->wrap_r),
- .compare_func = panfrost_translate_alt_compare_func(cso->compare_func),
+ .compare_func = panfrost_flip_compare_func(
+ panfrost_translate_compare_func(
+ cso->compare_func)),
.border_color = {
cso->border_color.f[0],
cso->border_color.f[1],
cso->border_color.f[2],
cso->border_color.f[3]
},
- .min_lod = FIXED_16(cso->min_lod),
- .max_lod = FIXED_16(cso->max_lod),
+ .min_lod = FIXED_16(cso->min_lod, false), /* clamp at 0 */
+ .max_lod = FIXED_16(cso->max_lod, false),
+ .lod_bias = FIXED_16(cso->lod_bias, true), /* can be negative */
.seamless_cube_map = cso->seamless_cube_map,
};
* essentially -- remember these are fixed point numbers, so
* epsilon=1/256) */
- if (cso->min_mip_filter == PIPE_TEX_MIPFILTER_NONE)
+ if (cso->min_mip_filter == PIPE_TEX_MIPFILTER_NONE) {
sampler_descriptor.max_lod = sampler_descriptor.min_lod;
- /* Enforce that there is something in the middle by adding epsilon*/
+ /* Enforce that there is something in the middle by adding epsilon*/
- if (sampler_descriptor.min_lod == sampler_descriptor.max_lod)
- sampler_descriptor.max_lod++;
+ if (sampler_descriptor.min_lod == sampler_descriptor.max_lod)
+ sampler_descriptor.max_lod++;
- /* Sanity check */
- assert(sampler_descriptor.max_lod > sampler_descriptor.min_lod);
+ /* Sanity check */
+ assert(sampler_descriptor.max_lod > sampler_descriptor.min_lod);
+ }
so->hw = sampler_descriptor;
/* XXX: Should upload, not just copy? */
ctx->sampler_count[shader] = num_sampler;
memcpy(ctx->samplers[shader], sampler, num_sampler * sizeof (void *));
-
- ctx->dirty |= PAN_DIRTY_SAMPLERS;
}
static bool
uint64_t outputs_written)
{
uint64_t so_outputs = 0;
- uint8_t reverse_map[64] = {};
+ uint8_t reverse_map[64] = {0};
unsigned slot = 0;
while (outputs_written)
enum pipe_shader_type type)
{
struct panfrost_context *ctx = pan_context(pctx);
-
ctx->shader[type] = hwcso;
- if (type == PIPE_SHADER_FRAGMENT)
- ctx->dirty |= PAN_DIRTY_FS;
- else
- ctx->dirty |= PAN_DIRTY_VS;
-
if (!hwcso) return;
/* Match the appropriate variant */
if (variant == -1) {
/* No variant matched, so create a new one */
variant = variants->variant_count++;
- assert(variants->variant_count < MAX_SHADER_VARIANTS);
+
+ if (variants->variant_count > variants->variant_space) {
+ unsigned old_space = variants->variant_space;
+
+ variants->variant_space *= 2;
+ if (variants->variant_space == 0)
+ variants->variant_space = 1;
+
+ /* Arbitrary limit to stop runaway programs from
+ * creating an unbounded number of shader variants. */
+ assert(variants->variant_space < 1024);
+
+ unsigned msize = sizeof(struct panfrost_shader_state);
+ variants->variants = realloc(variants->variants,
+ variants->variant_space * msize);
+
+ memset(&variants->variants[old_space], 0,
+ (variants->variant_space - old_space) * msize);
+ }
struct panfrost_shader_state *v =
&variants->variants[variant];
}
}
+static void *
+panfrost_create_vs_state(struct pipe_context *pctx, const struct pipe_shader_state *hwcso)
+{
+ return panfrost_create_shader_state(pctx, hwcso, PIPE_SHADER_VERTEX);
+}
+
+static void *
+panfrost_create_fs_state(struct pipe_context *pctx, const struct pipe_shader_state *hwcso)
+{
+ return panfrost_create_shader_state(pctx, hwcso, PIPE_SHADER_FRAGMENT);
+}
+
static void
panfrost_bind_vs_state(struct pipe_context *pctx, void *hwcso)
{
{
struct panfrost_context *ctx = pan_context(pctx);
ctx->stencil_ref = *ref;
-
- /* Shader core dirty */
- ctx->dirty |= PAN_DIRTY_FS;
}
static enum mali_texture_type
struct pipe_resource *texture,
const struct pipe_sampler_view *template)
{
+ struct panfrost_screen *screen = pan_screen(pctx->screen);
struct panfrost_sampler_view *so = rzalloc(pctx, struct panfrost_sampler_view);
- int bytes_per_pixel = util_format_get_blocksize(texture->format);
pipe_reference(NULL, &texture->reference);
so->base.reference.count = 1;
so->base.context = pctx;
- /* sampler_views correspond to texture descriptors, minus the texture
- * (data) itself. So, we serialise the descriptor here and cache it for
- * later. */
-
- const struct util_format_description *desc = util_format_description(prsrc->base.format);
-
unsigned char user_swizzle[4] = {
template->swizzle_r,
template->swizzle_g,
template->swizzle_a
};
- enum mali_format format = panfrost_find_format(desc);
-
- /* Check if we need to set a custom stride by computing the "expected"
- * stride and comparing it to what the BO actually wants. Only applies
- * to linear textures, since tiled/compressed textures have strict
- * alignment requirements for their strides as it is */
-
- unsigned first_level = template->u.tex.first_level;
- unsigned last_level = template->u.tex.last_level;
-
- if (prsrc->layout == PAN_LINEAR) {
- for (unsigned l = first_level; l <= last_level; ++l) {
- unsigned actual_stride = prsrc->slices[l].stride;
- unsigned width = u_minify(texture->width0, l);
- unsigned comp_stride = width * bytes_per_pixel;
-
- if (comp_stride != actual_stride) {
- so->manual_stride = true;
- break;
- }
- }
- }
-
/* In the hardware, array_size refers specifically to array textures,
* whereas in Gallium, it also covers cubemaps */
array_size /= 6;
}
- struct mali_texture_descriptor texture_descriptor = {
- .width = MALI_POSITIVE(u_minify(texture->width0, first_level)),
- .height = MALI_POSITIVE(u_minify(texture->height0, first_level)),
- .depth = MALI_POSITIVE(u_minify(texture->depth0, first_level)),
- .array_size = MALI_POSITIVE(array_size),
-
- .format = {
- .swizzle = panfrost_translate_swizzle_4(desc->swizzle),
- .format = format,
- .srgb = desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB,
- .type = panfrost_translate_texture_type(template->target),
- .unknown2 = 0x1,
- },
-
- .swizzle = panfrost_translate_swizzle_4(user_swizzle)
- };
-
- texture_descriptor.levels = last_level - first_level;
-
- so->hw = texture_descriptor;
+ enum mali_texture_type type =
+ panfrost_translate_texture_type(template->target);
+
+ unsigned size = panfrost_estimate_texture_size(
+ template->u.tex.first_level,
+ template->u.tex.last_level,
+ template->u.tex.first_layer,
+ template->u.tex.last_layer,
+ type, prsrc->layout);
+
+ so->bo = panfrost_bo_create(screen, size, 0);
+
+ panfrost_new_texture(
+ so->bo->cpu,
+ texture->width0, texture->height0,
+ texture->depth0, array_size,
+ texture->format,
+ type, prsrc->layout,
+ template->u.tex.first_level,
+ template->u.tex.last_level,
+ template->u.tex.first_layer,
+ template->u.tex.last_layer,
+ prsrc->cubemap_stride,
+ panfrost_translate_swizzle_4(user_swizzle),
+ prsrc->bo->gpu,
+ prsrc->slices);
return (struct pipe_sampler_view *) so;
}
struct pipe_sampler_view **views)
{
struct panfrost_context *ctx = pan_context(pctx);
+ unsigned new_nr = 0;
+ unsigned i;
assert(start_slot == 0);
- unsigned new_nr = 0;
- for (unsigned i = 0; i < num_views; ++i) {
+ for (i = 0; i < num_views; ++i) {
if (views[i])
new_nr = i + 1;
+ pipe_sampler_view_reference((struct pipe_sampler_view **)&ctx->sampler_views[shader][i],
+ views[i]);
}
+ for (; i < ctx->sampler_view_count[shader]; i++) {
+ pipe_sampler_view_reference((struct pipe_sampler_view **)&ctx->sampler_views[shader][i],
+ NULL);
+ }
ctx->sampler_view_count[shader] = new_nr;
- memcpy(ctx->sampler_views[shader], views, num_views * sizeof (void *));
-
- ctx->dirty |= PAN_DIRTY_TEXTURES;
}
static void
panfrost_sampler_view_destroy(
struct pipe_context *pctx,
- struct pipe_sampler_view *view)
+ struct pipe_sampler_view *pview)
{
- pipe_resource_reference(&view->texture, NULL);
+ struct panfrost_sampler_view *view = (struct panfrost_sampler_view *) pview;
+
+ pipe_resource_reference(&pview->texture, NULL);
+ panfrost_bo_unreference(view->bo);
ralloc_free(view);
}
for (unsigned i = 0; i < fb->nr_cbufs; ++i) {
struct pipe_surface *surf = fb->cbufs[i];
struct panfrost_resource *rsrc = pan_resource(surf->texture);
- panfrost_resource_hint_layout(screen, rsrc, PAN_AFBC, 1);
+ panfrost_resource_hint_layout(screen, rsrc, MALI_TEXTURE_AFBC, 1);
}
/* Also hint it to the depth buffer */
if (fb->zsbuf) {
struct panfrost_resource *rsrc = pan_resource(fb->zsbuf->texture);
- panfrost_resource_hint_layout(screen, rsrc, PAN_AFBC, 1);
+ panfrost_resource_hint_layout(screen, rsrc, MALI_TEXTURE_AFBC, 1);
}
}
{
struct panfrost_context *ctx = pan_context(pctx);
- /* Flush when switching framebuffers, but not if the framebuffer
- * state is being restored by u_blitter
- */
-
- struct panfrost_batch *batch = panfrost_get_batch_for_fbo(ctx);
- bool is_scanout = panfrost_batch_is_scanout(batch);
- bool has_draws = batch->last_job.gpu;
-
- /* Bail out early when the current and new states are the same. */
- if (util_framebuffer_state_equal(&ctx->pipe_framebuffer, fb))
- return;
-
- /* The wallpaper logic sets a new FB state before doing the blit and
- * restore the old one when it's done. Those FB states are reported to
- * be different because the surface they are pointing to are different,
- * but those surfaces actually point to the same cbufs/zbufs. In that
- * case we definitely don't want new FB descs to be emitted/attached
- * since the job is expected to be flushed just after the blit is done,
- * so let's just copy the new state and return here.
- */
- if (ctx->wallpaper_batch) {
- util_copy_framebuffer_state(&ctx->pipe_framebuffer, fb);
- return;
- }
-
- if (!is_scanout || has_draws)
- panfrost_flush(pctx, NULL, PIPE_FLUSH_END_OF_FRAME);
- else
- assert(!ctx->payloads[PIPE_SHADER_VERTEX].postfix.framebuffer &&
- !ctx->payloads[PIPE_SHADER_FRAGMENT].postfix.framebuffer);
-
- /* Invalidate the FBO job cache since we've just been assigned a new
- * FB state.
- */
- ctx->batch = NULL;
-
+ panfrost_hint_afbc(pan_screen(pctx->screen), fb);
util_copy_framebuffer_state(&ctx->pipe_framebuffer, fb);
-
- /* Given that we're rendering, we'd love to have compression */
- struct panfrost_screen *screen = pan_screen(ctx->base.screen);
-
- panfrost_hint_afbc(screen, &ctx->pipe_framebuffer);
- for (unsigned i = 0; i < PIPE_SHADER_TYPES; ++i)
- ctx->payloads[i].postfix.framebuffer = 0;
+ ctx->batch = NULL;
+ panfrost_invalidate_frame(ctx);
}
static void *
ctx->fragment_shader_core.stencil_mask_back = depth_stencil->stencil[back_index].writemask;
/* Depth state (TODO: Refactor) */
- SET_BIT(ctx->fragment_shader_core.unknown2_3, MALI_DEPTH_TEST, depth_stencil->depth.enabled);
+ SET_BIT(ctx->fragment_shader_core.unknown2_3, MALI_DEPTH_WRITEMASK,
+ depth_stencil->depth.writemask);
int func = depth_stencil->depth.enabled ? depth_stencil->depth.func : PIPE_FUNC_ALWAYS;
/* Bounds test not implemented */
assert(!depth_stencil->depth.bounds_test);
-
- ctx->dirty |= PAN_DIRTY_FS;
}
static void
panfrost_destroy(struct pipe_context *pipe)
{
struct panfrost_context *panfrost = pan_context(pipe);
- struct panfrost_screen *screen = pan_screen(pipe->screen);
if (panfrost->blitter)
util_blitter_destroy(panfrost->blitter);
if (panfrost->blitter_wallpaper)
util_blitter_destroy(panfrost->blitter_wallpaper);
- panfrost_drm_release_bo(screen, panfrost->scratchpad, false);
- panfrost_drm_release_bo(screen, panfrost->tiler_heap, false);
- panfrost_drm_release_bo(screen, panfrost->tiler_dummy, false);
+ util_unreference_framebuffer_state(&panfrost->pipe_framebuffer);
+ u_upload_destroy(pipe->stream_uploader);
ralloc_free(pipe);
}
static void
panfrost_destroy_query(struct pipe_context *pipe, struct pipe_query *q)
{
+ struct panfrost_query *query = (struct panfrost_query *) q;
+
+ if (query->bo) {
+ panfrost_bo_unreference(query->bo);
+ query->bo = NULL;
+ }
+
ralloc_free(q);
}
{
struct panfrost_context *ctx = pan_context(pipe);
struct panfrost_query *query = (struct panfrost_query *) q;
- struct panfrost_batch *batch = panfrost_get_batch_for_fbo(ctx);
switch (query->type) {
case PIPE_QUERY_OCCLUSION_COUNTER:
case PIPE_QUERY_OCCLUSION_PREDICATE:
case PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE:
- /* Allocate a word for the query results to be stored */
- query->transfer = panfrost_allocate_transient(batch, sizeof(unsigned));
+ /* Allocate a bo for the query results to be stored */
+ if (!query->bo) {
+ query->bo = panfrost_bo_create(
+ pan_screen(ctx->base.screen),
+ sizeof(unsigned), 0);
+ }
+
+ unsigned *result = (unsigned *)query->bo->cpu;
+ *result = 0; /* Default to 0 if nothing at all drawn. */
ctx->occlusion_query = query;
break;
break;
default:
- fprintf(stderr, "Skipping query %u\n", query->type);
+ DBG("Skipping query %u\n", query->type);
break;
}
union pipe_query_result *vresult)
{
struct panfrost_query *query = (struct panfrost_query *) q;
+ struct panfrost_context *ctx = pan_context(pipe);
switch (query->type) {
case PIPE_QUERY_OCCLUSION_PREDICATE:
case PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE:
/* Flush first */
- panfrost_flush(pipe, NULL, PIPE_FLUSH_END_OF_FRAME);
+ panfrost_flush_all_batches(ctx, true);
/* Read back the query results */
- unsigned *result = (unsigned *) query->transfer.cpu;
+ unsigned *result = (unsigned *) query->bo->cpu;
unsigned passed = *result;
if (query->type == PIPE_QUERY_OCCLUSION_COUNTER) {
case PIPE_QUERY_PRIMITIVES_GENERATED:
case PIPE_QUERY_PRIMITIVES_EMITTED:
- panfrost_flush(pipe, NULL, PIPE_FLUSH_END_OF_FRAME);
+ panfrost_flush_all_batches(ctx, true);
vresult->u64 = query->end - query->start;
break;
so->num_targets = num_targets;
}
-static void
-panfrost_setup_hardware(struct panfrost_context *ctx)
-{
- struct pipe_context *gallium = (struct pipe_context *) ctx;
- struct panfrost_screen *screen = pan_screen(gallium->screen);
-
- ctx->scratchpad = panfrost_drm_create_bo(screen, 64 * 4 * 4096, 0);
- ctx->tiler_heap = panfrost_drm_create_bo(screen, 4096 * 4096,
- PAN_ALLOCATE_INVISIBLE |
- PAN_ALLOCATE_GROWABLE);
- ctx->tiler_dummy = panfrost_drm_create_bo(screen, 4096,
- PAN_ALLOCATE_INVISIBLE);
- assert(ctx->scratchpad && ctx->tiler_heap && ctx->tiler_dummy);
-}
-
-/* New context creation, which also does hardware initialisation since I don't
- * know the better way to structure this :smirk: */
-
struct pipe_context *
panfrost_create_context(struct pipe_screen *screen, void *priv, unsigned flags)
{
struct panfrost_context *ctx = rzalloc(screen, struct panfrost_context);
- struct panfrost_screen *pscreen = pan_screen(screen);
- memset(ctx, 0, sizeof(*ctx));
struct pipe_context *gallium = (struct pipe_context *) ctx;
- ctx->is_t6xx = pscreen->gpu_id < 0x0700; /* Literally, "earlier than T700" */
-
gallium->screen = screen;
gallium->destroy = panfrost_destroy;
gallium->bind_vertex_elements_state = panfrost_bind_vertex_elements_state;
gallium->delete_vertex_elements_state = panfrost_generic_cso_delete;
- gallium->create_fs_state = panfrost_create_shader_state;
+ gallium->create_fs_state = panfrost_create_fs_state;
gallium->delete_fs_state = panfrost_delete_shader_state;
gallium->bind_fs_state = panfrost_bind_fs_state;
- gallium->create_vs_state = panfrost_create_shader_state;
+ gallium->create_vs_state = panfrost_create_vs_state;
gallium->delete_vs_state = panfrost_delete_shader_state;
gallium->bind_vs_state = panfrost_bind_vs_state;
panfrost_blend_context_init(gallium);
panfrost_compute_context_init(gallium);
- panfrost_drm_init_context(ctx);
-
- panfrost_setup_hardware(ctx);
-
/* XXX: leaks */
gallium->stream_uploader = u_upload_create_default(gallium);
gallium->const_uploader = gallium->stream_uploader;
panfrost_batch_init(ctx);
panfrost_emit_vertex_payload(ctx);
- panfrost_emit_tiler_payload(ctx);
panfrost_invalidate_frame(ctx);
panfrost_default_shader_backend(ctx);
projects / mesa.git / blobdiff