Mesa (master): panfrost: emit shader image attribute descriptors
GitLab Mirror
gitlab-mirror at kemper.freedesktop.org
Wed Jan 27 12:52:33 UTC 2021
Module: Mesa
Branch: master
Commit: dc85f65e059d3c603210bc3a0d93c34bdcdd6d17
URL: http://cgit.freedesktop.org/mesa/mesa/commit/?id=dc85f65e059d3c603210bc3a0d93c34bdcdd6d17
Author: Italo Nicola <italonicola at collabora.com>
Date: Fri Dec 11 20:55:14 2020 +0000
panfrost: emit shader image attribute descriptors
Each shader image attribute buffer descriptor needs a continuation,
meaning that each image attribute buffer is double the size of a common
attribute buffer.
For compute and fragment shaders, we allocate a single buffer with
enough space for all the attribute descriptors and attribute buffer
descriptors.
For vertex shaders, we do it separately to acomodate for the way
descriptors are currently allocated there.
Signed-off-by: Italo Nicola <italonicola at collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon at collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig at collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8066>
---
src/gallium/drivers/panfrost/pan_assemble.c | 6 +-
src/gallium/drivers/panfrost/pan_cmdstream.c | 141 ++++++++++++++++++++++++++-
src/gallium/drivers/panfrost/pan_cmdstream.h | 5 +
src/gallium/drivers/panfrost/pan_compute.c | 1 +
src/gallium/drivers/panfrost/pan_context.c | 1 +
5 files changed, 148 insertions(+), 6 deletions(-)
diff --git a/src/gallium/drivers/panfrost/pan_assemble.c b/src/gallium/drivers/panfrost/pan_assemble.c
index 06ad1a8103d..af6129e6425 100644
--- a/src/gallium/drivers/panfrost/pan_assemble.c
+++ b/src/gallium/drivers/panfrost/pan_assemble.c
@@ -309,7 +309,8 @@ panfrost_shader_compile(struct panfrost_context *ctx,
switch (stage) {
case MESA_SHADER_VERTEX:
- attribute_count = util_bitcount64(s->info.inputs_read);
+ attribute_count = util_bitcount64(s->info.inputs_read) +
+ util_bitcount(s->info.images_used);
varying_count = util_bitcount64(s->info.outputs_written);
if (vertex_id)
@@ -328,6 +329,7 @@ panfrost_shader_compile(struct panfrost_context *ctx,
program->blend_ret_offsets[i];
assert(!(state->blend_ret_addrs[i] & 0x7));
}
+ attribute_count = util_bitcount(s->info.images_used);
varying_count = util_bitcount64(s->info.inputs_read);
if (s->info.outputs_written & BITFIELD64_BIT(FRAG_RESULT_DEPTH))
state->writes_depth = true;
@@ -349,7 +351,7 @@ panfrost_shader_compile(struct panfrost_context *ctx,
s->info.fs.uses_demote;
break;
case MESA_SHADER_COMPUTE:
- /* TODO: images */
+ attribute_count = util_bitcount(s->info.images_used);
state->shared_size = s->info.cs.shared_size;
break;
default:
diff --git a/src/gallium/drivers/panfrost/pan_cmdstream.c b/src/gallium/drivers/panfrost/pan_cmdstream.c
index 7707696dbbb..e370554ffe3 100644
--- a/src/gallium/drivers/panfrost/pan_cmdstream.c
+++ b/src/gallium/drivers/panfrost/pan_cmdstream.c
@@ -34,6 +34,7 @@
#include "pan_cmdstream.h"
#include "pan_context.h"
#include "pan_job.h"
+#include "pan_texture.h"
/* If a BO is accessed for a particular shader stage, will it be in the primary
* batch (vertex/tiler) or the secondary batch (fragment)? Anything but
@@ -1181,6 +1182,127 @@ panfrost_emit_sampler_descriptors(struct panfrost_batch *batch,
return T.gpu;
}
+/* Packs all image attribute descs and attribute buffer descs.
+ * `first_image_buf_index` must be the index of the first image attribute buffer descriptor.
+ */
+static void
+emit_image_attribs(struct panfrost_batch *batch, enum pipe_shader_type shader,
+ struct mali_attribute_packed *attribs,
+ struct mali_attribute_buffer_packed *bufs,
+ unsigned first_image_buf_index)
+{
+ struct panfrost_context *ctx = batch->ctx;
+ struct panfrost_device *dev = pan_device(ctx->base.screen);
+
+ unsigned k = 0;
+ unsigned last_bit = util_last_bit(ctx->image_mask[shader]);
+ for (unsigned i = 0; i < last_bit; ++i) {
+ struct pipe_image_view *image = &ctx->images[shader][i];
+
+ /* TODO: understand how v3d/freedreno does it */
+ if (!(ctx->image_mask[shader] & (1 << i)) ||
+ !(image->shader_access & PIPE_IMAGE_ACCESS_READ_WRITE)) {
+ /* Unused image bindings */
+ pan_pack(bufs + (k * 2), ATTRIBUTE_BUFFER, cfg);
+ pan_pack(bufs + (k * 2) + 1, ATTRIBUTE_BUFFER, cfg);
+ pan_pack(attribs + k, ATTRIBUTE, cfg);
+ k++;
+ continue;
+ }
+
+ struct panfrost_resource *rsrc = pan_resource(image->resource);
+
+ /* TODO: MSAA */
+ assert(image->resource->nr_samples <= 1 && "MSAA'd images not supported");
+
+ bool is_3d = rsrc->base.target == PIPE_TEXTURE_3D;
+ bool is_linear = rsrc->layout.modifier == DRM_FORMAT_MOD_LINEAR;
+ bool is_buffer = rsrc->base.target == PIPE_BUFFER;
+
+ unsigned offset = is_buffer ? image->u.buf.offset :
+ panfrost_texture_offset(&rsrc->layout,
+ image->u.tex.level,
+ is_3d ? 0 : image->u.tex.first_layer,
+ is_3d ? image->u.tex.first_layer : 0);
+
+ /* AFBC should've been converted to tiled on panfrost_set_shader_image */
+ assert(!drm_is_afbc(rsrc->layout.modifier));
+
+ /* Add a dependency of the batch on the shader image buffer */
+ uint32_t flags = PAN_BO_ACCESS_SHARED | PAN_BO_ACCESS_VERTEX_TILER;
+ if (image->shader_access & PIPE_IMAGE_ACCESS_READ)
+ flags |= PAN_BO_ACCESS_READ;
+ if (image->shader_access & PIPE_IMAGE_ACCESS_WRITE) {
+ flags |= PAN_BO_ACCESS_WRITE;
+ unsigned level = is_buffer ? 0 : image->u.tex.level;
+ rsrc->layout.slices[level].initialized = true;
+ }
+ panfrost_batch_add_bo(batch, rsrc->bo, flags);
+
+ pan_pack(bufs + (k * 2), ATTRIBUTE_BUFFER, cfg) {
+ cfg.type = is_linear ?
+ MALI_ATTRIBUTE_TYPE_3D_LINEAR :
+ MALI_ATTRIBUTE_TYPE_3D_INTERLEAVED;
+
+ cfg.pointer = rsrc->bo->ptr.gpu + offset;
+ cfg.stride = util_format_get_blocksize(image->format);
+ cfg.size = rsrc->bo->size;
+ }
+
+ pan_pack(bufs + (k * 2) + 1, ATTRIBUTE_BUFFER_CONTINUATION_3D, cfg) {
+ cfg.s_dimension = rsrc->base.width0;
+ cfg.t_dimension = rsrc->base.height0;
+ cfg.r_dimension = is_3d ? rsrc->base.depth0 :
+ image->u.tex.last_layer - image->u.tex.first_layer + 1;
+
+ cfg.row_stride =
+ is_buffer ? 0 : rsrc->layout.slices[image->u.tex.level].row_stride;
+
+ if (rsrc->base.target != PIPE_TEXTURE_2D && !is_buffer) {
+ cfg.slice_stride =
+ panfrost_get_layer_stride(&rsrc->layout,
+ image->u.tex.level);
+ }
+ }
+
+ /* We map compute shader attributes 1:2 with attribute buffers, because
+ * every image attribute buffer needs an ATTRIBUTE_BUFFER_CONTINUATION_3D */
+ pan_pack(attribs + k, ATTRIBUTE, cfg) {
+ cfg.buffer_index = first_image_buf_index + (k * 2);
+ cfg.offset_enable = !(dev->quirks & IS_BIFROST);
+ cfg.format =
+ dev->formats[image->format].hw;
+ }
+
+ k++;
+ }
+}
+
+mali_ptr
+panfrost_emit_image_attribs(struct panfrost_batch *batch,
+ mali_ptr *buffers,
+ enum pipe_shader_type type)
+{
+ struct panfrost_context *ctx = batch->ctx;
+ struct panfrost_shader_state *shader = panfrost_get_shader_state(ctx, type);
+
+ /* Images always need a MALI_ATTRIBUTE_BUFFER_CONTINUATION_3D */
+ unsigned attrib_buf_size = MALI_ATTRIBUTE_BUFFER_LENGTH +
+ MALI_ATTRIBUTE_BUFFER_CONTINUATION_3D_LENGTH;
+ unsigned bytes_per_image_desc = MALI_ATTRIBUTE_LENGTH + attrib_buf_size;
+ unsigned attribs_offset = attrib_buf_size * shader->attribute_count;
+
+ struct panfrost_ptr ptr =
+ panfrost_pool_alloc_aligned(&batch->pool,
+ bytes_per_image_desc * shader->attribute_count,
+ util_next_power_of_two(bytes_per_image_desc));
+
+ emit_image_attribs(batch, type, ptr.cpu + attribs_offset, ptr.cpu, 0);
+
+ *buffers = ptr.gpu;
+ return ptr.gpu + attribs_offset;
+}
+
mali_ptr
panfrost_emit_vertex_data(struct panfrost_batch *batch,
mali_ptr *buffers)
@@ -1190,12 +1312,18 @@ panfrost_emit_vertex_data(struct panfrost_batch *batch,
bool is_bifrost = !!(dev->quirks & IS_BIFROST);
struct panfrost_vertex_state *so = ctx->vertex;
struct panfrost_shader_state *vs = panfrost_get_shader_state(ctx, PIPE_SHADER_VERTEX);
+ uint32_t image_mask = ctx->image_mask[PIPE_SHADER_VERTEX];
+ unsigned nr_images = util_bitcount(image_mask);
/* Worst case: everything is NPOT, which is only possible if instancing
- * is enabled. Otherwise single record is gauranteed */
+ * is enabled. Otherwise single record is gauranteed.
+ * Also, we allocate more memory than what's needed here if either instancing
+ * is enabled or images are present, this can be improved. */
+ unsigned bufs_per_attrib = (ctx->instance_count > 1 || nr_images > 0) ? 2 : 1;
+ unsigned nr_bufs = (vs->attribute_count + 1) * bufs_per_attrib;
+
struct panfrost_ptr S = panfrost_pool_alloc_aligned(&batch->pool,
- MALI_ATTRIBUTE_BUFFER_LENGTH * (vs->attribute_count + 1) *
- (ctx->instance_count > 1 ? 2 : 1),
+ MALI_ATTRIBUTE_BUFFER_LENGTH * nr_bufs,
MALI_ATTRIBUTE_BUFFER_LENGTH * 2);
struct panfrost_ptr T = panfrost_pool_alloc_aligned(&batch->pool,
@@ -1206,7 +1334,8 @@ panfrost_emit_vertex_data(struct panfrost_batch *batch,
(struct mali_attribute_buffer_packed *) S.cpu;
/* Determine (n + 1)'th index to suppress prefetch on Bifrost */
- unsigned last = vs->attribute_count * ((ctx->instance_count > 1) ? 2 : 1);
+ unsigned last = (nr_images + vs->attribute_count) *
+ ((ctx->instance_count > 1) ? 2 : 1);
memset(bufs + last, 0, sizeof(*bufs));
struct mali_attribute_packed *out =
@@ -1327,6 +1456,10 @@ panfrost_emit_vertex_data(struct panfrost_batch *batch,
}
}
+ k = ALIGN_POT(k, 2);
+ emit_image_attribs(batch, PIPE_SHADER_VERTEX, out + so->num_elements, bufs + k, k);
+ k += util_bitcount(ctx->image_mask[PIPE_SHADER_VERTEX]);
+
/* We need an empty attrib buf to stop the prefetching on Bifrost */
if (is_bifrost)
pan_pack(&bufs[k], ATTRIBUTE_BUFFER, cfg);
diff --git a/src/gallium/drivers/panfrost/pan_cmdstream.h b/src/gallium/drivers/panfrost/pan_cmdstream.h
index a1c5aea2537..42f4eb8c4ed 100644
--- a/src/gallium/drivers/panfrost/pan_cmdstream.h
+++ b/src/gallium/drivers/panfrost/pan_cmdstream.h
@@ -61,6 +61,11 @@ mali_ptr
panfrost_emit_sampler_descriptors(struct panfrost_batch *batch,
enum pipe_shader_type stage);
+mali_ptr
+panfrost_emit_image_attribs(struct panfrost_batch *batch,
+ mali_ptr *buffers,
+ enum pipe_shader_type type);
+
mali_ptr
panfrost_emit_vertex_data(struct panfrost_batch *batch,
mali_ptr *buffers);
diff --git a/src/gallium/drivers/panfrost/pan_compute.c b/src/gallium/drivers/panfrost/pan_compute.c
index 324b2fe0fbf..dcbf0310f28 100644
--- a/src/gallium/drivers/panfrost/pan_compute.c
+++ b/src/gallium/drivers/panfrost/pan_compute.c
@@ -153,6 +153,7 @@ panfrost_launch_grid(struct pipe_context *pipe,
if (!(dev->quirks & IS_BIFROST))
cfg.texture_descriptor_is_64b = true;
cfg.state = panfrost_emit_compute_shader_meta(batch, PIPE_SHADER_COMPUTE);
+ cfg.attributes = panfrost_emit_image_attribs(batch, &cfg.attribute_buffers, PIPE_SHADER_COMPUTE);
cfg.thread_storage = panfrost_emit_shared_memory(batch, info);
cfg.uniform_buffers = panfrost_emit_const_buf(batch,
PIPE_SHADER_COMPUTE, &cfg.push_uniforms);
diff --git a/src/gallium/drivers/panfrost/pan_context.c b/src/gallium/drivers/panfrost/pan_context.c
index 3382c35c087..36d0f087950 100644
--- a/src/gallium/drivers/panfrost/pan_context.c
+++ b/src/gallium/drivers/panfrost/pan_context.c
@@ -432,6 +432,7 @@ panfrost_draw_emit_tiler(struct panfrost_batch *batch,
cfg.cull_back_face = rast->cull_face & PIPE_FACE_BACK;
cfg.position = pos;
cfg.state = panfrost_emit_frag_shader_meta(batch);
+ cfg.attributes = panfrost_emit_image_attribs(batch, &cfg.attribute_buffers, PIPE_SHADER_FRAGMENT);
cfg.viewport = panfrost_emit_viewport(batch);
cfg.varyings = fs_vary;
cfg.varying_buffers = fs_vary ? varyings : 0;
More information about the mesa-commit
mailing list