#include "panfrost-quirks.h"
-#include "pan_allocate.h"
+#include "pan_pool.h"
#include "pan_bo.h"
#include "pan_cmdstream.h"
#include "pan_context.h"
.scratchpad = panfrost_batch_get_scratchpad(batch, shift, dev->thread_tls_alloc, dev->core_count)->gpu,
.shared_workgroup_count = ~0,
};
- postfix->shared_memory = panfrost_upload_transient(batch, &shared, sizeof(shared));
+ postfix->shared_memory = panfrost_pool_upload(&batch->pool, &shared, sizeof(shared));
}
static void
panfrost_vt_attach_framebuffer(struct panfrost_context *ctx,
struct mali_vertex_tiler_postfix *postfix)
{
- struct panfrost_device *dev = pan_device(ctx->base.screen);
struct panfrost_batch *batch = panfrost_get_batch_for_fbo(ctx);
-
- /* If we haven't, reserve space for the framebuffer */
-
- if (!batch->framebuffer.gpu) {
- unsigned size = (dev->quirks & MIDGARD_SFBD) ?
- sizeof(struct mali_single_framebuffer) :
- sizeof(struct mali_framebuffer);
-
- batch->framebuffer = panfrost_allocate_transient(batch, size);
-
- /* Tag the pointer */
- if (!(dev->quirks & MIDGARD_SFBD))
- batch->framebuffer.gpu |= MALI_MFBD;
- }
-
- postfix->shared_memory = batch->framebuffer.gpu;
+ postfix->shared_memory = panfrost_batch_reserve_framebuffer(batch);
}
static void
struct mali_vertex_tiler_postfix *postfix)
{
SET_BIT(postfix->gl_enables, MALI_OCCLUSION_QUERY, ctx->occlusion_query);
- if (ctx->occlusion_query)
+ if (ctx->occlusion_query) {
postfix->occlusion_counter = ctx->occlusion_query->bo->gpu;
- else
+ panfrost_batch_add_bo(ctx->batch, ctx->occlusion_query->bo,
+ PAN_BO_ACCESS_SHARED |
+ PAN_BO_ACCESS_RW |
+ PAN_BO_ACCESS_FRAGMENT);
+ } else {
postfix->occlusion_counter = 0;
+ }
}
void
} else {
/* Otherwise, we need to upload to transient memory */
const uint8_t *ibuf8 = (const uint8_t *) info->index.user;
- out = panfrost_upload_transient(batch, ibuf8 + offset,
+ out = panfrost_pool_upload(&batch->pool, ibuf8 + offset,
info->count *
info->index_size);
}
else {
/* First clause ATEST |= 0x4000000.
* Less than 32 regs |= 0x200 */
- meta->bifrost1.unk1 = 0x958020;
+ meta->bifrost1.unk1 = 0x950020;
}
meta->bifrost1.uniform_buffer_count = panfrost_ubo_count(ctx, st);
if (st == PIPE_SHADER_VERTEX)
meta->bifrost2.preload_regs = 0xC0;
- else
+ else {
meta->bifrost2.preload_regs = 0x1;
+ SET_BIT(meta->bifrost2.preload_regs, 0x10, ss->reads_frag_coord);
+ }
+
meta->bifrost2.uniform_count = MIN2(ss->uniform_count,
ss->uniform_cutoff);
} else {
meta->midgard1.uniform_count = MIN2(ss->uniform_count,
ss->uniform_cutoff);
meta->midgard1.work_count = ss->work_reg_count;
- meta->midgard1.flags_hi = 0x8; /* XXX */
- meta->midgard1.flags_lo = 0x220;
+
+ /* TODO: This is not conformant on ES3 */
+ meta->midgard1.flags_hi = MALI_SUPPRESS_INF_NAN;
+
+ meta->midgard1.flags_lo = 0x20;
meta->midgard1.uniform_buffer_count = panfrost_ubo_count(ctx, st);
+
+ SET_BIT(meta->midgard1.flags_hi, MALI_WRITES_GLOBAL, ss->writes_global);
}
}
.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_flip_compare_func(func),
+ .compare_func = cso->compare_mode ?
+ panfrost_flip_compare_func(func) :
+ MALI_FUNC_NEVER,
.border_color = {
cso->border_color.f[0],
cso->border_color.f[1],
fragmeta->depth_factor = 0.0f;
SET_BIT(fragmeta->unknown2_4, MALI_DEPTH_RANGE_A, false);
SET_BIT(fragmeta->unknown2_4, MALI_DEPTH_RANGE_B, false);
+ SET_BIT(fragmeta->unknown2_3, MALI_DEPTH_CLIP_NEAR, true);
+ SET_BIT(fragmeta->unknown2_3, MALI_DEPTH_CLIP_FAR, true);
return;
}
- bool msaa = ctx->rasterizer->base.multisample;
+ struct pipe_rasterizer_state *rast = &ctx->rasterizer->base;
+
+ bool msaa = rast->multisample;
/* TODO: Sample size */
SET_BIT(fragmeta->unknown2_3, MALI_HAS_MSAA, msaa);
SET_BIT(fragmeta->unknown2_4, MALI_NO_MSAA, !msaa);
- fragmeta->depth_units = ctx->rasterizer->base.offset_units * 2.0f;
- fragmeta->depth_factor = ctx->rasterizer->base.offset_scale;
+
+ struct panfrost_shader_state *fs;
+ fs = panfrost_get_shader_state(ctx, PIPE_SHADER_FRAGMENT);
+
+ /* EXT_shader_framebuffer_fetch requires the shader to be run
+ * per-sample when outputs are read. */
+ bool per_sample = ctx->min_samples > 1 || fs->outputs_read;
+ SET_BIT(fragmeta->unknown2_3, MALI_PER_SAMPLE, msaa && per_sample);
+
+ fragmeta->depth_units = rast->offset_units * 2.0f;
+ fragmeta->depth_factor = rast->offset_scale;
/* XXX: Which bit is which? Does this maybe allow offseting not-tri? */
- SET_BIT(fragmeta->unknown2_4, MALI_DEPTH_RANGE_A,
- ctx->rasterizer->base.offset_tri);
- SET_BIT(fragmeta->unknown2_4, MALI_DEPTH_RANGE_B,
- ctx->rasterizer->base.offset_tri);
+ SET_BIT(fragmeta->unknown2_4, MALI_DEPTH_RANGE_A, rast->offset_tri);
+ SET_BIT(fragmeta->unknown2_4, MALI_DEPTH_RANGE_B, rast->offset_tri);
+
+ SET_BIT(fragmeta->unknown2_3, MALI_DEPTH_CLIP_NEAR, rast->depth_clip_near);
+ SET_BIT(fragmeta->unknown2_3, MALI_DEPTH_CLIP_FAR, rast->depth_clip_far);
}
static void
fragmeta->unknown2_3 |= MALI_DEPTH_FUNC(panfrost_translate_compare_func(zfunc));
}
+static bool
+panfrost_fs_required(
+ struct panfrost_shader_state *fs,
+ struct panfrost_blend_final *blend,
+ unsigned rt_count)
+{
+ /* If we generally have side effects */
+ if (fs->fs_sidefx)
+ return true;
+
+ /* If colour is written we need to execute */
+ for (unsigned i = 0; i < rt_count; ++i) {
+ if (!blend[i].no_colour)
+ return true;
+ }
+
+ /* If depth is written and not implied we need to execute.
+ * TODO: Predicate on Z/S writes being enabled */
+ return (fs->writes_depth || fs->writes_stencil);
+}
+
static void
panfrost_frag_meta_blend_update(struct panfrost_context *ctx,
struct mali_shader_meta *fragmeta,
void *rts)
{
+ struct panfrost_batch *batch = panfrost_get_batch_for_fbo(ctx);
const struct panfrost_device *dev = pan_device(ctx->base.screen);
+ struct panfrost_shader_state *fs;
+ fs = panfrost_get_shader_state(ctx, PIPE_SHADER_FRAGMENT);
SET_BIT(fragmeta->unknown2_4, MALI_NO_DITHER,
(dev->quirks & MIDGARD_SFBD) && ctx->blend &&
!ctx->blend->base.dither);
+ SET_BIT(fragmeta->unknown2_4, MALI_ALPHA_TO_COVERAGE,
+ ctx->blend->base.alpha_to_coverage);
+
/* Get blending setup */
unsigned rt_count = MAX2(ctx->pipe_framebuffer.nr_cbufs, 1);
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 */
+ /* Disable shader execution if we can */
+ if (dev->quirks & MIDGARD_SHADERLESS
+ && !panfrost_fs_required(fs, blend, rt_count)) {
+ fragmeta->shader = 0;
+ fragmeta->attribute_count = 0;
+ fragmeta->varying_count = 0;
+ fragmeta->texture_count = 0;
+ fragmeta->sampler_count = 0;
+
+ /* This feature is not known to work on Bifrost */
+ fragmeta->midgard1.work_count = 1;
+ fragmeta->midgard1.uniform_count = 0;
+ fragmeta->midgard1.uniform_buffer_count = 0;
+ }
+
+ /* If there is a blend shader, work registers are shared. We impose 8
+ * work registers as a limit for blend shaders. Should be lower XXX */
if (!(dev->quirks & IS_BIFROST)) {
for (unsigned c = 0; c < rt_count; ++c) {
- if (blend[c].is_shader)
- fragmeta->midgard1.work_count = 16;
+ if (blend[c].is_shader) {
+ fragmeta->midgard1.work_count =
+ MAX2(fragmeta->midgard1.work_count, 8);
+ }
}
}
}
SET_BIT(fragmeta->unknown2_3, MALI_CAN_DISCARD,
- !blend[0].no_blending);
+ !blend[0].no_blending || fs->can_discard);
+
+ batch->draws |= PIPE_CLEAR_COLOR0;
return;
}
+ if (dev->quirks & IS_BIFROST) {
+ bool no_blend = true;
+
+ for (unsigned i = 0; i < rt_count; ++i)
+ no_blend &= (blend[i].no_blending | blend[i].no_colour);
+
+ SET_BIT(fragmeta->bifrost1.unk1, MALI_BIFROST_EARLY_Z,
+ !fs->can_discard && !fs->writes_depth && no_blend);
+ }
+
/* Additional blend descriptor tacked on for jobs using MFBD */
for (unsigned i = 0; i < rt_count; ++i) {
+ unsigned flags = 0;
+
+ if (ctx->pipe_framebuffer.nr_cbufs > i && !blend[i].no_colour) {
+ flags = 0x200;
+ batch->draws |= (PIPE_CLEAR_COLOR0 << i);
+
+ bool is_srgb = (ctx->pipe_framebuffer.nr_cbufs > i) &&
+ (ctx->pipe_framebuffer.cbufs[i]) &&
+ util_format_is_srgb(ctx->pipe_framebuffer.cbufs[i]->format);
+
+ SET_BIT(flags, MALI_BLEND_MRT_SHADER, blend[i].is_shader);
+ SET_BIT(flags, MALI_BLEND_LOAD_TIB, !blend[i].no_blending);
+ SET_BIT(flags, MALI_BLEND_SRGB, is_srgb);
+ SET_BIT(flags, MALI_BLEND_NO_DITHER, !ctx->blend->base.dither);
+ }
+
if (dev->quirks & IS_BIFROST) {
struct bifrost_blend_rt *brts = rts;
- struct panfrost_shader_state *fs;
- fs = panfrost_get_shader_state(ctx, PIPE_SHADER_FRAGMENT);
- brts[i].flags = 0x200;
+ brts[i].flags = flags;
+
if (blend[i].is_shader) {
/* The blend shader's address needs to be at
* the same top 32 bit as the fragment shader.
(fs->bo->gpu & (0xffffffffull << 32)));
brts[i].shader = blend[i].shader.gpu;
brts[i].unk2 = 0x0;
- } else {
+ } else if (ctx->pipe_framebuffer.nr_cbufs > i) {
enum pipe_format format = ctx->pipe_framebuffer.cbufs[i]->format;
const struct util_format_description *format_desc;
format_desc = util_format_description(format);
brts[i].constant = blend[i].equation.constant;
brts[i].format = panfrost_format_to_bifrost_blend(format_desc);
- brts[i].unk2 = 0x19;
+ /* 0x19 disables blending and forces REPLACE
+ * mode (equivalent to rgb_mode = alpha_mode =
+ * x122, colour mask = 0xF). 0x1a allows
+ * blending. */
+ brts[i].unk2 = blend[i].no_blending ? 0x19 : 0x1a;
+
+ brts[i].shader_type = fs->blend_types[i];
+ } else {
+ /* Dummy attachment for depth-only */
+ brts[i].unk2 = 0x3;
brts[i].shader_type = fs->blend_types[i];
}
} else {
struct midgard_blend_rt *mrts = rts;
-
- mrts[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);
-
- SET_BIT(mrts[i].flags, MALI_BLEND_MRT_SHADER, blend[i].is_shader);
- SET_BIT(mrts[i].flags, MALI_BLEND_LOAD_TIB, !blend[i].no_blending);
- SET_BIT(mrts[i].flags, MALI_BLEND_SRGB, is_srgb);
- SET_BIT(mrts[i].flags, MALI_BLEND_NO_DITHER, !ctx->blend->base.dither);
+ mrts[i].flags = flags;
if (blend[i].is_shader) {
mrts[i].blend.shader = blend[i].shader.gpu | blend[i].shader.first_tag;
fs = panfrost_get_shader_state(ctx, PIPE_SHADER_FRAGMENT);
- fragmeta->alpha_coverage = ~MALI_ALPHA_COVERAGE(0.000000);
- fragmeta->unknown2_3 = MALI_DEPTH_FUNC(MALI_FUNC_ALWAYS) | 0x3010;
+ bool msaa = ctx->rasterizer && ctx->rasterizer->base.multisample;
+ fragmeta->coverage_mask = (msaa ? ctx->sample_mask : ~0) & 0xF;
+
+ fragmeta->unknown2_3 = MALI_DEPTH_FUNC(MALI_FUNC_ALWAYS) | 0x10;
fragmeta->unknown2_4 = 0x4e0;
/* unknown2_4 has 0x10 bit set on T6XX and T720. We don't know why this
/* TODO */
} else {
/* Depending on whether it's legal to in the given shader, we try to
- * enable early-z testing (or forward-pixel kill?) */
+ * enable early-z testing. TODO: respect e-z force */
SET_BIT(fragmeta->midgard1.flags_lo, MALI_EARLY_Z,
- !fs->can_discard && !fs->writes_depth);
+ !fs->can_discard && !fs->writes_global &&
+ !fs->writes_depth && !fs->writes_stencil &&
+ !ctx->blend->base.alpha_to_coverage);
/* Add the writes Z/S flags if needed. */
SET_BIT(fragmeta->midgard1.flags_lo, MALI_WRITES_Z, fs->writes_depth);
SET_BIT(fragmeta->midgard1.flags_lo, MALI_HELPER_INVOCATIONS,
fs->helper_invocations);
- /* CAN_DISCARD should be set if the fragment shader possibly contains a
- * 'discard' instruction. It is likely this is related to optimizations
- * related to forward-pixel kill, as per "Mali Performance 3: Is
- * EGL_BUFFER_PRESERVED a good thing?" by Peter Harris */
+ const struct pipe_depth_stencil_alpha_state *zsa = ctx->depth_stencil;
- SET_BIT(fragmeta->unknown2_3, MALI_CAN_DISCARD, fs->can_discard);
- SET_BIT(fragmeta->midgard1.flags_lo, 0x400, fs->can_discard);
+ bool depth_enabled = fs->writes_depth ||
+ (zsa && zsa->depth.enabled && zsa->depth.func != PIPE_FUNC_ALWAYS);
+
+ SET_BIT(fragmeta->midgard1.flags_lo, MALI_READS_TILEBUFFER,
+ fs->outputs_read || (!depth_enabled && fs->can_discard));
+ SET_BIT(fragmeta->midgard1.flags_lo, MALI_READS_ZS, depth_enabled && fs->can_discard);
}
panfrost_frag_meta_rasterizer_update(ctx, fragmeta);
panfrost_frag_shader_meta_init(ctx, &meta, rts);
- xfer = panfrost_allocate_transient(batch, desc_size);
+ xfer = panfrost_pool_alloc(&batch->pool, desc_size);
memcpy(xfer.cpu, &meta, sizeof(meta));
memcpy(xfer.cpu + sizeof(meta), rts, rt_size * rt_count);
shader_ptr = xfer.gpu;
} else {
- shader_ptr = panfrost_upload_transient(batch, &meta,
+ shader_ptr = panfrost_pool_upload(&batch->pool, &meta,
sizeof(meta));
}
mvp->viewport0[1] = miny;
mvp->viewport1[1] = MALI_POSITIVE(maxy);
- mvp->clip_minz = minz;
- mvp->clip_maxz = maxz;
+ bool clip_near = true;
+ bool clip_far = true;
+
+ if (ctx->rasterizer) {
+ clip_near = ctx->rasterizer->base.depth_clip_near;
+ clip_far = ctx->rasterizer->base.depth_clip_far;
+ }
+
+ mvp->clip_minz = clip_near ? minz : -INFINITY;
+ mvp->clip_maxz = clip_far ? maxz : INFINITY;
}
void
mvp.viewport1[0] + 1,
mvp.viewport1[1] + 1);
- tiler_postfix->viewport = panfrost_upload_transient(batch, &mvp,
+ tiler_postfix->viewport = panfrost_pool_upload(&batch->pool, &mvp,
sizeof(mvp));
}
* PIPE_CAP_CONSTANT_BUFFER_OFFSET_ALIGNMENT */
return rsrc->bo->gpu + cb->buffer_offset;
} else if (cb->user_buffer) {
- return panfrost_upload_transient(batch,
+ return panfrost_pool_upload(&batch->pool,
cb->user_buffer +
cb->buffer_offset,
cb->buffer_size);
size_t sys_size = sizeof(float) * 4 * ss->sysval_count;
size_t uniform_size = has_uniforms ? (buf->cb[0].buffer_size) : 0;
size_t size = sys_size + uniform_size;
- struct panfrost_transfer transfer = panfrost_allocate_transient(batch,
+ struct panfrost_transfer transfer = panfrost_pool_alloc(&batch->pool,
size);
/* Upload sysvals requested by the shader */
ubos[ubo] = MALI_MAKE_UBO(aligned / bytes_per_field, gpu);
}
- mali_ptr ubufs = panfrost_upload_transient(batch, ubos, sz);
+ mali_ptr ubufs = panfrost_pool_upload(&batch->pool, ubos, sz);
postfix->uniforms = transfer.gpu;
postfix->uniform_buffers = ubufs;
.shared_shift = util_logbase2(single_size) - 1
};
- vtp->postfix.shared_memory = panfrost_upload_transient(batch, &shared,
+ vtp->postfix.shared_memory = panfrost_pool_upload(&batch->pool, &shared,
sizeof(shared));
}
PAN_BO_ACCESS_SHARED | PAN_BO_ACCESS_READ |
panfrost_bo_access_for_stage(st));
- panfrost_batch_add_bo(batch, view->midgard_bo,
+ panfrost_batch_add_bo(batch, view->bo,
PAN_BO_ACCESS_SHARED | PAN_BO_ACCESS_READ |
panfrost_bo_access_for_stage(st));
- return view->midgard_bo->gpu;
+ return view->bo->gpu;
+}
+
+static void
+panfrost_update_sampler_view(struct panfrost_sampler_view *view,
+ struct pipe_context *pctx)
+{
+ struct panfrost_resource *rsrc = pan_resource(view->base.texture);
+ if (view->texture_bo != rsrc->bo->gpu ||
+ view->layout != rsrc->layout) {
+ panfrost_bo_unreference(view->bo);
+ panfrost_create_sampler_view_bo(view, pctx, &rsrc->base);
+ }
}
void
struct panfrost_sampler_view *view = ctx->sampler_views[stage][i];
struct pipe_sampler_view *pview = &view->base;
struct panfrost_resource *rsrc = pan_resource(pview->texture);
+ panfrost_update_sampler_view(view, &ctx->base);
/* Add the BOs to the job so they are retained until the job is done. */
PAN_BO_ACCESS_SHARED | PAN_BO_ACCESS_READ |
panfrost_bo_access_for_stage(stage));
- panfrost_batch_add_bo(batch, view->bifrost_bo,
+ panfrost_batch_add_bo(batch, view->bo,
PAN_BO_ACCESS_SHARED | PAN_BO_ACCESS_READ |
panfrost_bo_access_for_stage(stage));
memcpy(&descriptors[i], view->bifrost_descriptor, sizeof(*view->bifrost_descriptor));
}
- postfix->textures = panfrost_upload_transient(batch,
+ postfix->textures = panfrost_pool_upload(&batch->pool,
descriptors,
sizeof(struct bifrost_texture_descriptor) *
ctx->sampler_view_count[stage]);
} else {
uint64_t trampolines[PIPE_MAX_SHADER_SAMPLER_VIEWS];
- for (int i = 0; i < ctx->sampler_view_count[stage]; ++i)
- trampolines[i] = panfrost_get_tex_desc(batch, stage,
- ctx->sampler_views[stage][i]);
+ for (int i = 0; i < ctx->sampler_view_count[stage]; ++i) {
+ struct panfrost_sampler_view *view = ctx->sampler_views[stage][i];
+
+ panfrost_update_sampler_view(view, &ctx->base);
+
+ trampolines[i] = panfrost_get_tex_desc(batch, stage, view);
+ }
- postfix->textures = panfrost_upload_transient(batch,
+ postfix->textures = panfrost_pool_upload(&batch->pool,
trampolines,
sizeof(uint64_t) *
ctx->sampler_view_count[stage]);
if (device->quirks & IS_BIFROST) {
size_t desc_size = sizeof(struct bifrost_sampler_descriptor);
size_t transfer_size = desc_size * ctx->sampler_count[stage];
- struct panfrost_transfer transfer = panfrost_allocate_transient(batch,
+ struct panfrost_transfer transfer = panfrost_pool_alloc(&batch->pool,
transfer_size);
struct bifrost_sampler_descriptor *desc = (struct bifrost_sampler_descriptor *)transfer.cpu;
} else {
size_t desc_size = sizeof(struct mali_sampler_descriptor);
size_t transfer_size = desc_size * ctx->sampler_count[stage];
- struct panfrost_transfer transfer = panfrost_allocate_transient(batch,
+ struct panfrost_transfer transfer = panfrost_pool_alloc(&batch->pool,
transfer_size);
struct mali_sampler_descriptor *desc = (struct mali_sampler_descriptor *)transfer.cpu;
struct panfrost_vertex_state *so = ctx->vertex;
panfrost_vertex_state_upd_attr_offs(ctx, vertex_postfix);
- vertex_postfix->attribute_meta = panfrost_upload_transient(batch, so->hw,
+ vertex_postfix->attribute_meta = panfrost_pool_upload(&batch->pool, so->hw,
sizeof(*so->hw) *
PAN_MAX_ATTRIBUTE);
}
/* Upload whatever we emitted and go */
- vertex_postfix->attributes = panfrost_upload_transient(batch, attrs,
+ vertex_postfix->attributes = panfrost_pool_upload(&batch->pool, attrs,
k * sizeof(*attrs));
}
slot->size = stride * count;
slot->shift = slot->extra_flags = 0;
- struct panfrost_transfer transfer = panfrost_allocate_transient(batch,
+ struct panfrost_transfer transfer = panfrost_pool_alloc(&batch->pool,
slot->size);
slot->elements = transfer.gpu | MALI_ATTR_LINEAR;
return transfer.gpu;
}
+static unsigned
+panfrost_streamout_offset(unsigned stride, unsigned offset,
+ struct pipe_stream_output_target *target)
+{
+ return (target->buffer_offset + (offset * stride * 4)) & 63;
+}
+
static void
panfrost_emit_streamout(struct panfrost_batch *batch, union mali_attr *slot,
unsigned stride, unsigned offset, unsigned count,
unsigned max_size = target->buffer_size;
unsigned expected_size = slot->stride * count;
- slot->size = MIN2(max_size, expected_size);
-
/* Grab the BO and bind it to the batch */
struct panfrost_bo *bo = pan_resource(target->buffer)->bo;
PAN_BO_ACCESS_VERTEX_TILER |
PAN_BO_ACCESS_FRAGMENT);
+ /* We will have an offset applied to get alignment */
mali_ptr addr = bo->gpu + target->buffer_offset + (offset * slot->stride);
- slot->elements = addr;
-}
-
-/* Given a shader and buffer indices, link varying metadata together */
-
-static bool
-is_special_varying(gl_varying_slot loc)
-{
- switch (loc) {
- case VARYING_SLOT_POS:
- case VARYING_SLOT_PSIZ:
- case VARYING_SLOT_PNTC:
- case VARYING_SLOT_FACE:
- return true;
- default:
- return false;
- }
-}
-
-static void
-panfrost_emit_varying_meta(void *outptr, struct panfrost_shader_state *ss,
- signed general, signed gl_Position,
- signed gl_PointSize, signed gl_PointCoord,
- signed gl_FrontFacing)
-{
- struct mali_attr_meta *out = (struct mali_attr_meta *) outptr;
-
- for (unsigned i = 0; i < ss->varying_count; ++i) {
- gl_varying_slot location = ss->varyings_loc[i];
- int index = -1;
-
- switch (location) {
- case VARYING_SLOT_POS:
- index = gl_Position;
- break;
- case VARYING_SLOT_PSIZ:
- index = gl_PointSize;
- break;
- case VARYING_SLOT_PNTC:
- index = gl_PointCoord;
- break;
- case VARYING_SLOT_FACE:
- index = gl_FrontFacing;
- break;
- default:
- index = general;
- break;
- }
-
- assert(index >= 0);
- out[i].index = index;
- }
+ slot->elements = (addr & ~63) | MALI_ATTR_LINEAR;
+ slot->size = MIN2(max_size, expected_size) + (addr & 63);
}
static bool
unreachable("Varying not captured");
}
-/* TODO: Integers */
-static enum mali_format
-pan_xfb_format(unsigned nr_components)
+static unsigned
+pan_varying_size(enum mali_format fmt)
{
- switch (nr_components) {
- case 1: return MALI_R32F;
- case 2: return MALI_RG32F;
- case 3: return MALI_RGB32F;
- case 4: return MALI_RGBA32F;
- default: unreachable("Invalid format");
+ unsigned type = MALI_EXTRACT_TYPE(fmt);
+ unsigned chan = MALI_EXTRACT_CHANNELS(fmt);
+ unsigned bits = MALI_EXTRACT_BITS(fmt);
+ unsigned bpc = 0;
+
+ if (bits == MALI_CHANNEL_FLOAT) {
+ /* No doubles */
+ bool fp16 = (type == MALI_FORMAT_SINT);
+ assert(fp16 || (type == MALI_FORMAT_UNORM));
+
+ bpc = fp16 ? 2 : 4;
+ } else {
+ assert(type >= MALI_FORMAT_SNORM && type <= MALI_FORMAT_SINT);
+
+ /* See the enums */
+ bits = 1 << bits;
+ assert(bits >= 8);
+ bpc = bits / 8;
}
+
+ return bpc * chan;
}
-void
-panfrost_emit_varying_descriptor(struct panfrost_batch *batch,
- unsigned vertex_count,
- struct mali_vertex_tiler_postfix *vertex_postfix,
- struct mali_vertex_tiler_postfix *tiler_postfix,
- union midgard_primitive_size *primitive_size)
-{
- /* Load the shaders */
- struct panfrost_context *ctx = batch->ctx;
- struct panfrost_shader_state *vs, *fs;
- unsigned int num_gen_varyings = 0;
- size_t vs_size, fs_size;
+/* Indices for named (non-XFB) varyings that are present. These are packed
+ * tightly so they correspond to a bitfield present (P) indexed by (1 <<
+ * PAN_VARY_*). This has the nice property that you can lookup the buffer index
+ * of a given special field given a shift S by:
+ *
+ * idx = popcount(P & ((1 << S) - 1))
+ *
+ * That is... look at all of the varyings that come earlier and count them, the
+ * count is the new index since plus one. Likewise, the total number of special
+ * buffers required is simply popcount(P)
+ */
- /* Allocate the varying descriptor */
+enum pan_special_varying {
+ PAN_VARY_GENERAL = 0,
+ PAN_VARY_POSITION = 1,
+ PAN_VARY_PSIZ = 2,
+ PAN_VARY_PNTCOORD = 3,
+ PAN_VARY_FACE = 4,
+ PAN_VARY_FRAGCOORD = 5,
- vs = panfrost_get_shader_state(ctx, PIPE_SHADER_VERTEX);
- fs = panfrost_get_shader_state(ctx, PIPE_SHADER_FRAGMENT);
- vs_size = sizeof(struct mali_attr_meta) * vs->varying_count;
- fs_size = sizeof(struct mali_attr_meta) * fs->varying_count;
+ /* Keep last */
+ PAN_VARY_MAX,
+};
- struct panfrost_transfer trans = panfrost_allocate_transient(batch,
- vs_size +
- fs_size);
+/* Given a varying, figure out which index it correpsonds to */
- struct pipe_stream_output_info *so = &vs->stream_output;
+static inline unsigned
+pan_varying_index(unsigned present, enum pan_special_varying v)
+{
+ unsigned mask = (1 << v) - 1;
+ return util_bitcount(present & mask);
+}
- /* Check if this varying is linked by us. This is the case for
- * general-purpose, non-captured varyings. If it is, link it. If it's
- * not, use the provided stream out information to determine the
- * offset, since it was already linked for us. */
+/* Get the base offset for XFB buffers, which by convention come after
+ * everything else. Wrapper function for semantic reasons; by construction this
+ * is just popcount. */
- for (unsigned i = 0; i < vs->varying_count; i++) {
- gl_varying_slot loc = vs->varyings_loc[i];
+static inline unsigned
+pan_xfb_base(unsigned present)
+{
+ return util_bitcount(present);
+}
- bool special = is_special_varying(loc);
- bool captured = ((vs->so_mask & (1ll << loc)) ? true : false);
+/* Computes the present mask for varyings so we can start emitting varying records */
- if (captured) {
- struct pipe_stream_output *o = pan_get_so(so, loc);
+static inline unsigned
+pan_varying_present(
+ struct panfrost_shader_state *vs,
+ struct panfrost_shader_state *fs,
+ unsigned quirks)
+{
+ /* At the moment we always emit general and position buffers. Not
+ * strictly necessary but usually harmless */
- unsigned dst_offset = o->dst_offset * 4; /* dwords */
- vs->varyings[i].src_offset = dst_offset;
- } else if (!special) {
- vs->varyings[i].src_offset = 16 * (num_gen_varyings++);
- }
- }
+ unsigned present = (1 << PAN_VARY_GENERAL) | (1 << PAN_VARY_POSITION);
- /* Conversely, we need to set src_offset for the captured varyings.
- * Here, the layout is defined by the stream out info, not us */
+ /* Enable special buffers by the shader info */
- /* Link up with fragment varyings */
- bool reads_point_coord = fs->reads_point_coord;
+ if (vs->writes_point_size)
+ present |= (1 << PAN_VARY_PSIZ);
- for (unsigned i = 0; i < fs->varying_count; i++) {
- gl_varying_slot loc = fs->varyings_loc[i];
- unsigned src_offset;
- signed vs_idx = -1;
-
- /* Link up */
- for (unsigned j = 0; j < vs->varying_count; ++j) {
- if (vs->varyings_loc[j] == loc) {
- vs_idx = j;
- break;
- }
- }
+ if (fs->reads_point_coord)
+ present |= (1 << PAN_VARY_PNTCOORD);
- /* Either assign or reuse */
- if (vs_idx >= 0)
- src_offset = vs->varyings[vs_idx].src_offset;
- else
- src_offset = 16 * (num_gen_varyings++);
+ if (fs->reads_face)
+ present |= (1 << PAN_VARY_FACE);
- fs->varyings[i].src_offset = src_offset;
+ if (fs->reads_frag_coord && !(quirks & IS_BIFROST))
+ present |= (1 << PAN_VARY_FRAGCOORD);
+
+ /* Also, if we have a point sprite, we need a point coord buffer */
+
+ for (unsigned i = 0; i < fs->varying_count; i++) {
+ gl_varying_slot loc = fs->varyings_loc[i];
if (has_point_coord(fs->point_sprite_mask, loc))
- reads_point_coord = true;
+ present |= (1 << PAN_VARY_PNTCOORD);
}
- memcpy(trans.cpu, vs->varyings, vs_size);
- memcpy(trans.cpu + vs_size, fs->varyings, fs_size);
+ return present;
+}
- union mali_attr varyings[PIPE_MAX_ATTRIBS] = {0};
+/* Emitters for varying records */
- /* Figure out how many streamout buffers could be bound */
- unsigned so_count = ctx->streamout.num_targets;
- for (unsigned i = 0; i < vs->varying_count; i++) {
- gl_varying_slot loc = vs->varyings_loc[i];
+static struct mali_attr_meta
+pan_emit_vary(unsigned present, enum pan_special_varying buf,
+ unsigned quirks, enum mali_format format,
+ unsigned offset)
+{
+ unsigned nr_channels = MALI_EXTRACT_CHANNELS(format);
+
+ struct mali_attr_meta meta = {
+ .index = pan_varying_index(present, buf),
+ .unknown1 = quirks & IS_BIFROST ? 0x0 : 0x2,
+ .swizzle = quirks & HAS_SWIZZLES ?
+ panfrost_get_default_swizzle(nr_channels) :
+ panfrost_bifrost_swizzle(nr_channels),
+ .format = format,
+ .src_offset = offset
+ };
- bool captured = ((vs->so_mask & (1ll << loc)) ? true : false);
- if (!captured) continue;
+ return meta;
+}
- struct pipe_stream_output *o = pan_get_so(so, loc);
- so_count = MAX2(so_count, o->output_buffer + 1);
- }
+/* General varying that is unused */
- signed idx = so_count;
- signed general = idx++;
- signed gl_Position = idx++;
- signed gl_PointSize = vs->writes_point_size ? (idx++) : -1;
- signed gl_PointCoord = reads_point_coord ? (idx++) : -1;
- signed gl_FrontFacing = fs->reads_face ? (idx++) : -1;
- signed gl_FragCoord = fs->reads_frag_coord ? (idx++) : -1;
+static struct mali_attr_meta
+pan_emit_vary_only(unsigned present, unsigned quirks)
+{
+ return pan_emit_vary(present, 0, quirks, MALI_VARYING_DISCARD, 0);
+}
- /* Emit the stream out buffers */
+/* Special records */
- unsigned out_count = u_stream_outputs_for_vertices(ctx->active_prim,
- ctx->vertex_count);
+static const enum mali_format pan_varying_formats[PAN_VARY_MAX] = {
+ [PAN_VARY_POSITION] = MALI_VARYING_POS,
+ [PAN_VARY_PSIZ] = MALI_R16F,
+ [PAN_VARY_PNTCOORD] = MALI_R16F,
+ [PAN_VARY_FACE] = MALI_R32I,
+ [PAN_VARY_FRAGCOORD] = MALI_RGBA32F
+};
- for (unsigned i = 0; i < so_count; ++i) {
- if (i < ctx->streamout.num_targets) {
- panfrost_emit_streamout(batch, &varyings[i],
- so->stride[i],
- ctx->streamout.offsets[i],
- out_count,
- ctx->streamout.targets[i]);
- } else {
- /* Emit a dummy buffer */
- panfrost_emit_varyings(batch, &varyings[i],
- so->stride[i] * 4,
- out_count);
+static struct mali_attr_meta
+pan_emit_vary_special(unsigned present, enum pan_special_varying buf,
+ unsigned quirks)
+{
+ assert(buf < PAN_VARY_MAX);
+ return pan_emit_vary(present, buf, quirks, pan_varying_formats[buf], 0);
+}
- /* Clear the attribute type */
- varyings[i].elements &= ~0xF;
- }
- }
+static enum mali_format
+pan_xfb_format(enum mali_format format, unsigned nr)
+{
+ if (MALI_EXTRACT_BITS(format) == MALI_CHANNEL_FLOAT)
+ return MALI_R32F | MALI_NR_CHANNELS(nr);
+ else
+ return MALI_EXTRACT_TYPE(format) | MALI_NR_CHANNELS(nr) | MALI_CHANNEL_32;
+}
- panfrost_emit_varyings(batch, &varyings[general],
- num_gen_varyings * 16,
- vertex_count);
+/* Transform feedback records. Note struct pipe_stream_output is (if packed as
+ * a bitfield) 32-bit, smaller than a 64-bit pointer, so may as well pass by
+ * value. */
+
+static struct mali_attr_meta
+pan_emit_vary_xfb(unsigned present,
+ unsigned max_xfb,
+ unsigned *streamout_offsets,
+ unsigned quirks,
+ enum mali_format format,
+ struct pipe_stream_output o)
+{
+ /* Otherwise construct a record for it */
+ struct mali_attr_meta meta = {
+ /* XFB buffers come after everything else */
+ .index = pan_xfb_base(present) + o.output_buffer,
- mali_ptr varyings_p;
+ /* As usual unknown bit */
+ .unknown1 = quirks & IS_BIFROST ? 0x0 : 0x2,
- /* fp32 vec4 gl_Position */
- varyings_p = panfrost_emit_varyings(batch, &varyings[gl_Position],
- sizeof(float) * 4, vertex_count);
- tiler_postfix->position_varying = varyings_p;
+ /* Override swizzle with number of channels */
+ .swizzle = quirks & HAS_SWIZZLES ?
+ panfrost_get_default_swizzle(o.num_components) :
+ panfrost_bifrost_swizzle(o.num_components),
+ /* Override number of channels and precision to highp */
+ .format = pan_xfb_format(format, o.num_components),
- if (panfrost_writes_point_size(ctx)) {
- varyings_p = panfrost_emit_varyings(batch,
- &varyings[gl_PointSize],
- 2, vertex_count);
- primitive_size->pointer = varyings_p;
- }
+ /* Apply given offsets together */
+ .src_offset = (o.dst_offset * 4) /* dwords */
+ + streamout_offsets[o.output_buffer]
+ };
- if (reads_point_coord)
- varyings[gl_PointCoord].elements = MALI_VARYING_POINT_COORD;
+ return meta;
+}
- if (fs->reads_face)
- varyings[gl_FrontFacing].elements = MALI_VARYING_FRONT_FACING;
+/* Determine if we should capture a varying for XFB. This requires actually
+ * having a buffer for it. If we don't capture it, we'll fallback to a general
+ * varying path (linked or unlinked, possibly discarding the write) */
- if (fs->reads_frag_coord)
- varyings[gl_FragCoord].elements = MALI_VARYING_FRAG_COORD;
+static bool
+panfrost_xfb_captured(struct panfrost_shader_state *xfb,
+ unsigned loc, unsigned max_xfb)
+{
+ if (!(xfb->so_mask & (1ll << loc)))
+ return false;
- struct panfrost_device *device = pan_device(ctx->base.screen);
- assert(!(device->quirks & IS_BIFROST) || !(reads_point_coord));
+ struct pipe_stream_output *o = pan_get_so(&xfb->stream_output, loc);
+ return o->output_buffer < max_xfb;
+}
- /* Let's go ahead and link varying meta to the buffer in question, now
- * that that information is available. VARYING_SLOT_POS is mapped to
- * gl_FragCoord for fragment shaders but gl_Positionf or vertex shaders
- * */
+/* Higher-level wrapper around all of the above, classifying a varying into one
+ * of the above types */
+
+static struct mali_attr_meta
+panfrost_emit_varying(
+ struct panfrost_shader_state *stage,
+ struct panfrost_shader_state *other,
+ struct panfrost_shader_state *xfb,
+ unsigned present,
+ unsigned max_xfb,
+ unsigned *streamout_offsets,
+ unsigned quirks,
+ unsigned *gen_offsets,
+ enum mali_format *gen_formats,
+ unsigned *gen_stride,
+ unsigned idx,
+ bool should_alloc,
+ bool is_fragment)
+{
+ gl_varying_slot loc = stage->varyings_loc[idx];
+ enum mali_format format = stage->varyings[idx];
+
+ /* Override format to match linkage */
+ if (!should_alloc && gen_formats[idx])
+ format = gen_formats[idx];
+
+ if (has_point_coord(stage->point_sprite_mask, loc)) {
+ return pan_emit_vary_special(present, PAN_VARY_PNTCOORD, quirks);
+ } else if (panfrost_xfb_captured(xfb, loc, max_xfb)) {
+ struct pipe_stream_output *o = pan_get_so(&xfb->stream_output, loc);
+ return pan_emit_vary_xfb(present, max_xfb, streamout_offsets, quirks, format, *o);
+ } else if (loc == VARYING_SLOT_POS) {
+ if (is_fragment)
+ return pan_emit_vary_special(present, PAN_VARY_FRAGCOORD, quirks);
+ else
+ return pan_emit_vary_special(present, PAN_VARY_POSITION, quirks);
+ } else if (loc == VARYING_SLOT_PSIZ) {
+ return pan_emit_vary_special(present, PAN_VARY_PSIZ, quirks);
+ } else if (loc == VARYING_SLOT_PNTC) {
+ return pan_emit_vary_special(present, PAN_VARY_PNTCOORD, quirks);
+ } else if (loc == VARYING_SLOT_FACE) {
+ return pan_emit_vary_special(present, PAN_VARY_FACE, quirks);
+ }
- panfrost_emit_varying_meta(trans.cpu, vs, general, gl_Position,
- gl_PointSize, gl_PointCoord,
- gl_FrontFacing);
+ /* We've exhausted special cases, so it's otherwise a general varying. Check if we're linked */
+ signed other_idx = -1;
- panfrost_emit_varying_meta(trans.cpu + vs_size, fs, general,
- gl_FragCoord, gl_PointSize,
- gl_PointCoord, gl_FrontFacing);
+ for (unsigned j = 0; j < other->varying_count; ++j) {
+ if (other->varyings_loc[j] == loc) {
+ other_idx = j;
+ break;
+ }
+ }
- /* Replace streamout */
+ if (other_idx < 0)
+ return pan_emit_vary_only(present, quirks);
- struct mali_attr_meta *ovs = (struct mali_attr_meta *)trans.cpu;
- struct mali_attr_meta *ofs = ovs + vs->varying_count;
+ unsigned offset = gen_offsets[other_idx];
- for (unsigned i = 0; i < vs->varying_count; i++) {
- gl_varying_slot loc = vs->varyings_loc[i];
+ if (should_alloc) {
+ /* We're linked, so allocate a space via a watermark allocation */
+ enum mali_format alt = other->varyings[other_idx];
- bool captured = ((vs->so_mask & (1ll << loc)) ? true : false);
- if (!captured)
- continue;
+ /* Do interpolation at minimum precision */
+ unsigned size_main = pan_varying_size(format);
+ unsigned size_alt = pan_varying_size(alt);
+ unsigned size = MIN2(size_main, size_alt);
- struct pipe_stream_output *o = pan_get_so(so, loc);
- ovs[i].index = o->output_buffer;
+ /* If a varying is marked for XFB but not actually captured, we
+ * should match the format to the format that would otherwise
+ * be used for XFB, since dEQP checks for invariance here. It's
+ * unclear if this is required by the spec. */
- /* Set the type appropriately. TODO: Integer varyings XXX */
- assert(o->stream == 0);
- ovs[i].format = pan_xfb_format(o->num_components);
+ if (xfb->so_mask & (1ull << loc)) {
+ struct pipe_stream_output *o = pan_get_so(&xfb->stream_output, loc);
+ format = pan_xfb_format(format, o->num_components);
+ size = pan_varying_size(format);
+ } else if (size == size_alt) {
+ format = alt;
+ }
- if (device->quirks & HAS_SWIZZLES)
- ovs[i].swizzle = panfrost_get_default_swizzle(o->num_components);
- else
- ovs[i].swizzle = panfrost_bifrost_swizzle(o->num_components);
+ gen_offsets[idx] = *gen_stride;
+ gen_formats[other_idx] = format;
+ offset = *gen_stride;
+ *gen_stride += size;
+ }
- /* Link to the fragment */
- signed fs_idx = -1;
+ return pan_emit_vary(present, PAN_VARY_GENERAL,
+ quirks, format, offset);
+}
- /* Link up */
- for (unsigned j = 0; j < fs->varying_count; ++j) {
- if (fs->varyings_loc[j] == loc) {
- fs_idx = j;
- break;
- }
- }
+static void
+pan_emit_special_input(union mali_attr *varyings,
+ unsigned present,
+ enum pan_special_varying v,
+ mali_ptr addr)
+{
+ if (present & (1 << v)) {
+ /* Ensure we write exactly once for performance and with fields
+ * zeroed appropriately to avoid flakes */
- if (fs_idx >= 0) {
- ofs[fs_idx].index = ovs[i].index;
- ofs[fs_idx].format = ovs[i].format;
- ofs[fs_idx].swizzle = ovs[i].swizzle;
- }
+ union mali_attr s = {
+ .elements = addr
+ };
+
+ varyings[pan_varying_index(present, v)] = s;
}
+}
- /* Replace point sprite */
- for (unsigned i = 0; i < fs->varying_count; i++) {
- /* If we have a point sprite replacement, handle that here. We
- * have to translate location first. TODO: Flip y in shader.
- * We're already keying ... just time crunch .. */
+void
+panfrost_emit_varying_descriptor(struct panfrost_batch *batch,
+ unsigned vertex_count,
+ struct mali_vertex_tiler_postfix *vertex_postfix,
+ struct mali_vertex_tiler_postfix *tiler_postfix,
+ union midgard_primitive_size *primitive_size)
+{
+ /* Load the shaders */
+ struct panfrost_context *ctx = batch->ctx;
+ struct panfrost_device *dev = pan_device(ctx->base.screen);
+ struct panfrost_shader_state *vs, *fs;
+ size_t vs_size, fs_size;
- if (has_point_coord(fs->point_sprite_mask,
- fs->varyings_loc[i])) {
- ofs[i].index = gl_PointCoord;
+ /* Allocate the varying descriptor */
- /* Swizzle out the z/w to 0/1 */
- ofs[i].format = MALI_RG16F;
- ofs[i].swizzle = panfrost_get_default_swizzle(2);
- }
+ vs = panfrost_get_shader_state(ctx, PIPE_SHADER_VERTEX);
+ fs = panfrost_get_shader_state(ctx, PIPE_SHADER_FRAGMENT);
+ vs_size = sizeof(struct mali_attr_meta) * vs->varying_count;
+ fs_size = sizeof(struct mali_attr_meta) * fs->varying_count;
+
+ struct panfrost_transfer trans = panfrost_pool_alloc(&batch->pool,
+ vs_size +
+ fs_size);
+
+ struct pipe_stream_output_info *so = &vs->stream_output;
+ unsigned present = pan_varying_present(vs, fs, dev->quirks);
+
+ /* Check if this varying is linked by us. This is the case for
+ * general-purpose, non-captured varyings. If it is, link it. If it's
+ * not, use the provided stream out information to determine the
+ * offset, since it was already linked for us. */
+
+ unsigned gen_offsets[32];
+ enum mali_format gen_formats[32];
+ memset(gen_offsets, 0, sizeof(gen_offsets));
+ memset(gen_formats, 0, sizeof(gen_formats));
+
+ unsigned gen_stride = 0;
+ assert(vs->varying_count < ARRAY_SIZE(gen_offsets));
+ assert(fs->varying_count < ARRAY_SIZE(gen_offsets));
+
+ unsigned streamout_offsets[32];
+
+ for (unsigned i = 0; i < ctx->streamout.num_targets; ++i) {
+ streamout_offsets[i] = panfrost_streamout_offset(
+ so->stride[i],
+ ctx->streamout.offsets[i],
+ ctx->streamout.targets[i]);
}
- /* Fix up unaligned addresses */
- for (unsigned i = 0; i < so_count; ++i) {
- if (varyings[i].elements < MALI_RECORD_SPECIAL)
- continue;
+ struct mali_attr_meta *ovs = (struct mali_attr_meta *)trans.cpu;
+ struct mali_attr_meta *ofs = ovs + vs->varying_count;
- unsigned align = (varyings[i].elements & 63);
+ for (unsigned i = 0; i < vs->varying_count; i++) {
+ ovs[i] = panfrost_emit_varying(vs, fs, vs, present,
+ ctx->streamout.num_targets, streamout_offsets,
+ dev->quirks,
+ gen_offsets, gen_formats, &gen_stride, i, true, false);
+ }
- /* While we're at it, the SO buffers are linear */
+ for (unsigned i = 0; i < fs->varying_count; i++) {
+ ofs[i] = panfrost_emit_varying(fs, vs, vs, present,
+ ctx->streamout.num_targets, streamout_offsets,
+ dev->quirks,
+ gen_offsets, gen_formats, &gen_stride, i, false, true);
+ }
- if (!align) {
- varyings[i].elements |= MALI_ATTR_LINEAR;
- continue;
- }
+ unsigned xfb_base = pan_xfb_base(present);
+ struct panfrost_transfer T = panfrost_pool_alloc(&batch->pool,
+ sizeof(union mali_attr) * (xfb_base + ctx->streamout.num_targets));
+ union mali_attr *varyings = (union mali_attr *) T.cpu;
- /* We need to adjust alignment */
- varyings[i].elements &= ~63;
- varyings[i].elements |= MALI_ATTR_LINEAR;
- varyings[i].size += align;
+ /* Emit the stream out buffers */
- for (unsigned v = 0; v < vs->varying_count; ++v) {
- if (ovs[v].index != i)
- continue;
+ unsigned out_count = u_stream_outputs_for_vertices(ctx->active_prim,
+ ctx->vertex_count);
- ovs[v].src_offset = vs->varyings[v].src_offset + align;
- }
+ for (unsigned i = 0; i < ctx->streamout.num_targets; ++i) {
+ panfrost_emit_streamout(batch, &varyings[xfb_base + i],
+ so->stride[i],
+ ctx->streamout.offsets[i],
+ out_count,
+ ctx->streamout.targets[i]);
+ }
- for (unsigned f = 0; f < fs->varying_count; ++f) {
- if (ofs[f].index != i)
- continue;
+ panfrost_emit_varyings(batch,
+ &varyings[pan_varying_index(present, PAN_VARY_GENERAL)],
+ gen_stride, vertex_count);
- ofs[f].src_offset = fs->varyings[f].src_offset + align;
- }
+ /* fp32 vec4 gl_Position */
+ tiler_postfix->position_varying = panfrost_emit_varyings(batch,
+ &varyings[pan_varying_index(present, PAN_VARY_POSITION)],
+ sizeof(float) * 4, vertex_count);
+
+ if (present & (1 << PAN_VARY_PSIZ)) {
+ primitive_size->pointer = panfrost_emit_varyings(batch,
+ &varyings[pan_varying_index(present, PAN_VARY_PSIZ)],
+ 2, vertex_count);
}
- varyings_p = panfrost_upload_transient(batch, varyings,
- idx * sizeof(*varyings));
- vertex_postfix->varyings = varyings_p;
- tiler_postfix->varyings = varyings_p;
+ pan_emit_special_input(varyings, present, PAN_VARY_PNTCOORD, MALI_VARYING_POINT_COORD);
+ pan_emit_special_input(varyings, present, PAN_VARY_FACE, MALI_VARYING_FRONT_FACING);
+ pan_emit_special_input(varyings, present, PAN_VARY_FRAGCOORD, MALI_VARYING_FRAG_COORD);
+
+ vertex_postfix->varyings = T.gpu;
+ tiler_postfix->varyings = T.gpu;
vertex_postfix->varying_meta = trans.gpu;
tiler_postfix->varying_meta = trans.gpu + vs_size;
{
struct panfrost_context *ctx = batch->ctx;
struct panfrost_device *device = pan_device(ctx->base.screen);
- bool wallpapering = ctx->wallpaper_batch && batch->tiler_dep;
+ bool wallpapering = ctx->wallpaper_batch && batch->scoreboard.tiler_dep;
struct bifrost_payload_vertex bifrost_vertex = {0,};
struct bifrost_payload_tiler bifrost_tiler = {0,};
struct midgard_payload_vertex_tiler midgard_vertex = {0,};
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, tp, tp_size, true);
- panfrost_new_job(batch, JOB_TYPE_VERTEX, false, 0,
+ panfrost_new_job(&batch->pool, &batch->scoreboard, JOB_TYPE_TILER, false,
+ batch->scoreboard.job_index + 2, tp, tp_size, true);
+ panfrost_new_job(&batch->pool, &batch->scoreboard, JOB_TYPE_VERTEX, false, 0,
vp, vp_size, true);
return;
}
bool rasterizer_discard = ctx->rasterizer &&
ctx->rasterizer->base.rasterizer_discard;
- unsigned vertex = panfrost_new_job(batch, JOB_TYPE_VERTEX, false, 0,
+ unsigned vertex = panfrost_new_job(&batch->pool, &batch->scoreboard, JOB_TYPE_VERTEX, false, 0,
vp, vp_size, false);
if (rasterizer_discard)
return;
- panfrost_new_job(batch, JOB_TYPE_TILER, false, vertex, tp, tp_size,
+ panfrost_new_job(&batch->pool, &batch->scoreboard, JOB_TYPE_TILER, false, vertex, tp, tp_size,
false);
}
0, 0,
};
- return panfrost_upload_transient(batch, locations, 96 * sizeof(uint16_t));
+ return panfrost_pool_upload(&batch->pool, locations, 96 * sizeof(uint16_t));
}