}
}
-static void
-panfrost_shader_meta_init(struct panfrost_context *ctx,
- enum pipe_shader_type st,
- struct mali_shader_meta *meta)
-{
- const struct panfrost_device *dev = pan_device(ctx->base.screen);
- struct panfrost_shader_state *ss = panfrost_get_shader_state(ctx, st);
-
- memset(meta, 0, sizeof(*meta));
- meta->shader = (ss->bo ? ss->bo->gpu : 0) | ss->first_tag;
- meta->attribute_count = ss->attribute_count;
- meta->varying_count = ss->varying_count;
- meta->texture_count = ctx->sampler_view_count[st];
- meta->sampler_count = ctx->sampler_count[st];
-
- if (dev->quirks & IS_BIFROST) {
- if (st == PIPE_SHADER_VERTEX)
- meta->bifrost1.unk1 = 0x800000;
- else {
- /* First clause ATEST |= 0x4000000.
- * Less than 32 regs |= 0x200 */
- 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 {
- 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;
-
- /* 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_lo, MALI_WRITES_GLOBAL, ss->writes_global);
- }
-}
-
static unsigned
translate_tex_wrap(enum pipe_tex_wrap w)
{
}
static void
-panfrost_frag_shader_meta_init(struct panfrost_context *ctx,
+panfrost_emit_frag_shader(struct panfrost_context *ctx,
struct mali_shader_meta *fragmeta,
struct panfrost_blend_final *blend)
{
struct pipe_rasterizer_state *rast = &ctx->rasterizer->base;
const struct panfrost_zsa_state *zsa = ctx->depth_stencil;
+ unsigned rt_count = ctx->pipe_framebuffer.nr_cbufs;
- bool msaa = rast->multisample;
- fragmeta->coverage_mask = msaa ? ctx->sample_mask : ~0;
+ memset(fragmeta, 0, sizeof(*fragmeta));
+ memcpy(&fragmeta->shader, &fs->shader, sizeof(fs->shader));
- fragmeta->unknown2_3 = MALI_DEPTH_FUNC(MALI_FUNC_ALWAYS) | 0x10;
- fragmeta->unknown2_4 = 0x4e0;
+ if (!panfrost_fs_required(fs, blend, rt_count)) {
+ struct mali_shader_packed shader = { 0 };
+ struct mali_midgard_properties_packed prop;
- /* 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 (dev->quirks & IS_BIFROST) {
+ struct mali_preload_packed preload = { 0 };
+ memcpy(&fragmeta->bifrost_preload, &preload, sizeof(preload));
- SET_BIT(fragmeta->unknown2_4, 0x10, dev->quirks & MIDGARD_SFBD);
+ pan_pack(&prop, BIFROST_PROPERTIES, cfg) {
+ cfg.unknown = 0x950020; /* XXX */
+ cfg.early_z_enable = true;
+ }
+ } else {
+ pan_pack(&shader, SHADER, cfg) {
+ cfg.shader = 0x1;
+ }
- if (dev->quirks & IS_BIFROST) {
- /* TODO */
- } else {
- /* Depending on whether it's legal to in the given shader, we try to
- * enable early-z testing. TODO: respect e-z force */
+ pan_pack(&prop, MIDGARD_PROPERTIES, cfg) {
+ cfg.work_register_count = 1;
+ cfg.depth_source = MALI_DEPTH_SOURCE_FIXED_FUNCTION;
+ cfg.early_z_enable = true;
+ }
+ }
- SET_BIT(fragmeta->midgard1.flags_lo, MALI_EARLY_Z,
- !fs->can_discard && !fs->writes_global &&
- !fs->writes_depth && !fs->writes_stencil &&
- !ctx->blend->base.alpha_to_coverage);
+ memcpy(&fragmeta->shader, &shader, sizeof(shader));
+ memcpy(&fragmeta->midgard_props, &prop, sizeof(prop));
+ } else if (dev->quirks & IS_BIFROST) {
+ struct mali_bifrost_properties_packed prop;
- /* 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_hi, MALI_WRITES_S, fs->writes_stencil);
+ bool no_blend = true;
- /* Any time texturing is used, derivatives are implicitly calculated,
- * so we need to enable helper invocations */
+ for (unsigned i = 0; i < rt_count; ++i)
+ no_blend &= (!blend[i].load_dest | blend[i].no_colour);
- SET_BIT(fragmeta->midgard1.flags_lo, MALI_HELPER_INVOCATIONS,
- fs->helper_invocations);
+ pan_pack(&prop, BIFROST_PROPERTIES, cfg) {
+ cfg.early_z_enable = !fs->can_discard && !fs->writes_depth && no_blend;
+ }
+
+ /* Combine with prepacked properties */
+ prop.opaque[0] |= fs->properties.opaque[0];
- /* If discard is enabled, which bit we set to convey this
- * depends on if depth/stencil is used for the draw or not.
- * Just one of depth OR stencil is enough to trigger this. */
+ memcpy(&fragmeta->bifrost_props, &prop, sizeof(prop));
+ memcpy(&fragmeta->bifrost_preload, &fs->preload, sizeof(fs->preload));
+ } else {
+ struct mali_midgard_properties_packed prop;
+ /* Reasons to disable early-Z from a shader perspective */
+ bool late_z = fs->can_discard || fs->writes_global ||
+ fs->writes_depth || fs->writes_stencil;
+
+ /* Reasons to disable early-Z from a CSO perspective */
+ bool alpha_to_coverage = ctx->blend->base.alpha_to_coverage;
+
+ /* If either depth or stencil is enabled, discard matters */
bool zs_enabled =
- fs->writes_depth || fs->writes_stencil ||
(zsa->base.depth.enabled && zsa->base.depth.func != PIPE_FUNC_ALWAYS) ||
zsa->base.stencil[0].enabled;
- SET_BIT(fragmeta->midgard1.flags_lo, MALI_READS_TILEBUFFER,
- fs->outputs_read || (!zs_enabled && fs->can_discard));
- SET_BIT(fragmeta->midgard1.flags_lo, MALI_READS_ZS, zs_enabled && fs->can_discard);
+ bool has_blend_shader = false;
+
+ for (unsigned c = 0; c < rt_count; ++c)
+ has_blend_shader |= blend[c].is_shader;
+
+ pan_pack(&prop, MIDGARD_PROPERTIES, cfg) {
+ /* TODO: Reduce this limit? */
+ if (has_blend_shader)
+ cfg.work_register_count = MAX2(fs->work_reg_count, 8);
+ else
+ cfg.work_register_count = fs->work_reg_count;
+
+ cfg.early_z_enable = !(late_z || alpha_to_coverage);
+ cfg.reads_tilebuffer = fs->outputs_read || (!zs_enabled && fs->can_discard);
+ cfg.reads_depth_stencil = zs_enabled && fs->can_discard;
+ }
+
+ /* Combine with prepacked properties */
+ prop.opaque[0] |= fs->properties.opaque[0];
+ memcpy(&fragmeta->midgard_props, &prop, sizeof(prop));
}
+ bool msaa = rast->multisample;
+ fragmeta->coverage_mask = msaa ? ctx->sample_mask : ~0;
+
+ fragmeta->unknown2_3 = MALI_DEPTH_FUNC(MALI_FUNC_ALWAYS) | 0x10;
+ fragmeta->unknown2_4 = 0x4e0;
+
/* TODO: Sample size */
SET_BIT(fragmeta->unknown2_3, MALI_HAS_MSAA, msaa);
SET_BIT(fragmeta->unknown2_4, MALI_NO_MSAA, !msaa);
fragmeta->unknown2_3 |= MALI_DEPTH_FUNC(panfrost_translate_compare_func(
zsa->base.depth.enabled ? zsa->base.depth.func : PIPE_FUNC_ALWAYS));
- 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 = ctx->pipe_framebuffer.nr_cbufs;
-
- /* 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 =
- MAX2(fragmeta->midgard1.work_count, 8);
- }
- }
- }
-
- /* 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...?).
- * Specify the last RTi with a blend shader. */
-
- fragmeta->blend.shader = 0;
-
- for (signed rt = ((signed) rt_count - 1); rt >= 0; --rt) {
- if (!blend[rt].is_shader)
- continue;
-
- fragmeta->blend.shader = blend[rt].shader.gpu |
- blend[rt].shader.first_tag;
- break;
- }
-
if (dev->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) */
+ if (blend[0].no_colour)
+ return;
+
+ fragmeta->unknown2_4 |= MALI_SFBD_ENABLE;
+
+ SET_BIT(fragmeta->unknown2_4, MALI_SFBD_SRGB,
+ util_format_is_srgb(ctx->pipe_framebuffer.cbufs[0]->format));
+
SET_BIT(fragmeta->unknown2_3, MALI_HAS_BLEND_SHADER,
blend[0].is_shader);
- if (!blend[0].is_shader) {
+ if (blend[0].is_shader) {
+ fragmeta->blend.shader = blend[0].shader.gpu |
+ blend[0].shader.first_tag;
+ } else {
fragmeta->blend.equation = blend[0].equation.equation;
fragmeta->blend.constant = blend[0].equation.constant;
}
SET_BIT(fragmeta->unknown2_3, MALI_CAN_DISCARD,
blend[0].load_dest);
- }
- if (dev->quirks & IS_BIFROST) {
- bool no_blend = true;
+ SET_BIT(fragmeta->unknown2_4, MALI_NO_DITHER, !ctx->blend->base.dither);
+ } else if (!(dev->quirks & IS_BIFROST)) {
+ /* Bug where MRT-capable hw apparently reads the last blend
+ * shader from here instead of the usual location? */
- for (unsigned i = 0; i < rt_count; ++i)
- no_blend &= (!blend[i].load_dest | blend[i].no_colour);
+ for (signed rt = ((signed) rt_count - 1); rt >= 0; --rt) {
+ if (!blend[rt].is_shader)
+ continue;
- SET_BIT(fragmeta->bifrost1.unk1, MALI_BIFROST_EARLY_Z,
- !fs->can_discard && !fs->writes_depth && no_blend);
+ fragmeta->blend.shader = blend[rt].shader.gpu |
+ blend[rt].shader.first_tag;
+ break;
+ }
}
}
-void
-panfrost_emit_shader_meta(struct panfrost_batch *batch,
- enum pipe_shader_type st,
- struct mali_vertex_tiler_postfix *postfix)
+mali_ptr
+panfrost_emit_compute_shader_meta(struct panfrost_batch *batch, enum pipe_shader_type stage)
{
- struct panfrost_context *ctx = batch->ctx;
- struct panfrost_shader_state *ss = panfrost_get_shader_state(ctx, st);
+ struct panfrost_shader_state *ss = panfrost_get_shader_state(batch->ctx, stage);
- if (!ss) {
- postfix->shader = 0;
- return;
- }
+ panfrost_batch_add_bo(batch, ss->bo,
+ PAN_BO_ACCESS_PRIVATE |
+ PAN_BO_ACCESS_READ |
+ PAN_BO_ACCESS_VERTEX_TILER);
- struct mali_shader_meta meta;
+ panfrost_batch_add_bo(batch, pan_resource(ss->upload.rsrc)->bo,
+ PAN_BO_ACCESS_PRIVATE |
+ PAN_BO_ACCESS_READ |
+ PAN_BO_ACCESS_VERTEX_TILER);
- panfrost_shader_meta_init(ctx, st, &meta);
+ return pan_resource(ss->upload.rsrc)->bo->gpu + ss->upload.offset;
+}
+
+mali_ptr
+panfrost_emit_frag_shader_meta(struct panfrost_batch *batch)
+{
+ struct panfrost_context *ctx = batch->ctx;
+ struct panfrost_shader_state *ss = panfrost_get_shader_state(ctx, PIPE_SHADER_FRAGMENT);
+ struct mali_shader_meta meta;
/* 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(st));
-
- mali_ptr shader_ptr;
-
- if (st == PIPE_SHADER_FRAGMENT) {
- struct panfrost_device *dev = pan_device(ctx->base.screen);
- unsigned rt_count = MAX2(ctx->pipe_framebuffer.nr_cbufs, 1);
- size_t desc_size = sizeof(meta);
- void *rts = NULL;
- struct panfrost_transfer xfer;
- unsigned rt_size;
-
- if (dev->quirks & MIDGARD_SFBD)
- rt_size = 0;
- else if (dev->quirks & IS_BIFROST)
- rt_size = sizeof(struct bifrost_blend_rt);
- else
- rt_size = sizeof(struct midgard_blend_rt);
+ PAN_BO_ACCESS_FRAGMENT);
- desc_size += rt_size * rt_count;
+ struct panfrost_device *dev = pan_device(ctx->base.screen);
+ unsigned rt_count = MAX2(ctx->pipe_framebuffer.nr_cbufs, 1);
+ size_t desc_size = sizeof(meta);
+ void *rts = NULL;
+ struct panfrost_transfer xfer;
+ unsigned rt_size;
+
+ if (dev->quirks & MIDGARD_SFBD)
+ rt_size = 0;
+ else if (dev->quirks & IS_BIFROST)
+ rt_size = sizeof(struct bifrost_blend_rt);
+ else
+ rt_size = sizeof(struct midgard_blend_rt);
- if (rt_size)
- rts = rzalloc_size(ctx, rt_size * rt_count);
+ desc_size += rt_size * rt_count;
- struct panfrost_blend_final blend[PIPE_MAX_COLOR_BUFS];
+ if (rt_size)
+ rts = rzalloc_size(ctx, rt_size * rt_count);
- for (unsigned c = 0; c < ctx->pipe_framebuffer.nr_cbufs; ++c)
- blend[c] = panfrost_get_blend_for_context(ctx, c);
+ struct panfrost_blend_final blend[PIPE_MAX_COLOR_BUFS];
- panfrost_frag_shader_meta_init(ctx, &meta, blend);
+ for (unsigned c = 0; c < ctx->pipe_framebuffer.nr_cbufs; ++c)
+ blend[c] = panfrost_get_blend_for_context(ctx, c);
- if (!(dev->quirks & MIDGARD_SFBD))
- panfrost_emit_blend(batch, rts, blend);
- else
- batch->draws |= PIPE_CLEAR_COLOR0;
+ panfrost_emit_frag_shader(ctx, &meta, blend);
- xfer = panfrost_pool_alloc_aligned(&batch->pool, desc_size, sizeof(meta));
+ if (!(dev->quirks & MIDGARD_SFBD))
+ panfrost_emit_blend(batch, rts, blend);
+ else
+ batch->draws |= PIPE_CLEAR_COLOR0;
- memcpy(xfer.cpu, &meta, sizeof(meta));
- memcpy(xfer.cpu + sizeof(meta), rts, rt_size * rt_count);
+ xfer = panfrost_pool_alloc_aligned(&batch->pool, desc_size, sizeof(meta));
- if (rt_size)
- ralloc_free(rts);
+ memcpy(xfer.cpu, &meta, sizeof(meta));
+ memcpy(xfer.cpu + sizeof(meta), rts, rt_size * rt_count);
- shader_ptr = xfer.gpu;
- } else {
- shader_ptr = panfrost_pool_upload(&batch->pool, &meta,
- sizeof(meta));
- }
+ if (rt_size)
+ ralloc_free(rts);
- postfix->shader = shader_ptr;
+ return xfer.gpu;
}
void
}
/* Next up, attach UBOs. UBO #0 is the uniforms we just
- * uploaded */
+ * uploaded, so it's always included. The count is the highest UBO
+ * addressable -- gaps are included. */
- unsigned ubo_count = panfrost_ubo_count(ctx, stage);
- assert(ubo_count >= 1);
+ unsigned ubo_count = 32 - __builtin_clz(buf->enabled_mask | 1);
size_t sz = MALI_UNIFORM_BUFFER_LENGTH * ubo_count;
struct panfrost_transfer ubos =