/* TODO: Stub */
struct midgard_payload_vertex_tiler payload = { 0 };
+ struct mali_invocation_packed invocation;
struct mali_draw_packed postfix;
/* We implement OpenCL inputs as uniforms (or a UBO -- same thing), so
/* Invoke according to the grid info */
- panfrost_pack_work_groups_compute(&payload.prefix,
+ panfrost_pack_work_groups_compute(&invocation,
info->grid[0], info->grid[1],
info->grid[2],
info->block[0], info->block[1],
info->block[2],
false);
+ payload.prefix.invocation = invocation;
panfrost_new_job(&batch->pool, &batch->scoreboard,
MALI_JOB_TYPE_COMPUTE, true, 0, &payload,
struct mali_vertex_tiler_prefix vertex_prefix = { 0 }, tiler_prefix = { 0 };
struct mali_draw_packed vertex_postfix, tiler_postfix;
struct mali_primitive_packed primitive;
+ struct mali_invocation_packed invocation;
union midgard_primitive_size primitive_size;
unsigned vertex_count = ctx->vertex_count;
panfrost_statistics_record(ctx, info);
- panfrost_pack_work_groups_fused(&vertex_prefix, &tiler_prefix,
+ panfrost_pack_work_groups_compute(&invocation,
1, vertex_count, info->instance_count,
- 1, 1, 1);
+ 1, 1, 1, true);
+
+ vertex_prefix.invocation = invocation;
+ tiler_prefix.invocation = invocation;
/* Emit all sort of descriptors. */
mali_ptr varyings = 0, vs_vary = 0, fs_vary = 0, pos = 0, psiz = 0;
},
};
- panfrost_pack_work_groups_compute(&payload.prefix,
+ struct mali_invocation_packed invocation;
+
+ panfrost_pack_work_groups_compute(&invocation,
1, 1, 1,
1, 1, 1,
true);
+ payload.prefix.invocation = invocation;
+
struct panfrost_bo *bos[] = {
scratchpad, shmem, shader, shader_desc, ubo, var, attr
};
*/
struct mali_vertex_tiler_prefix {
- /* This is a dynamic bitfield containing the following things in this order:
- *
- * - gl_WorkGroupSize.x
- * - gl_WorkGroupSize.y
- * - gl_WorkGroupSize.z
- * - gl_NumWorkGroups.x
- * - gl_NumWorkGroups.y
- * - gl_NumWorkGroups.z
- *
- * The number of bits allocated for each number is based on the *_shift
- * fields below. For example, workgroups_y_shift gives the bit that
- * gl_NumWorkGroups.y starts at, and workgroups_z_shift gives the bit
- * that gl_NumWorkGroups.z starts at (and therefore one after the bit
- * that gl_NumWorkGroups.y ends at). The actual value for each gl_*
- * value is one more than the stored value, since if any of the values
- * are zero, then there would be no invocations (and hence no job). If
- * there were 0 bits allocated to a given field, then it must be zero,
- * and hence the real value is one.
- *
- * Vertex jobs reuse the same job dispatch mechanism as compute jobs,
- * effectively doing glDispatchCompute(1, vertex_count, instance_count)
- * where vertex count is the number of vertices.
- */
- u32 invocation_count;
-
- /* Bitfield for shifts:
- *
- * size_y_shift : 5
- * size_z_shift : 5
- * workgroups_x_shift : 6
- * workgroups_y_shift : 6
- * workgroups_z_shift : 6
- * workgroups_x_shift_2 : 4
- */
- u32 invocation_shifts;
-
+ struct mali_invocation_packed invocation;
struct mali_primitive_packed primitive;
} __attribute__((packed));
/* Decode invocation_count. See the comment before the definition of
* invocation_count for an explanation.
*/
+ struct MALI_INVOCATION invocation;
+ struct mali_invocation_packed invocation_packed = p->invocation;
+ MALI_INVOCATION_unpack((const uint8_t *) &invocation_packed, &invocation);
- unsigned size_y_shift = bits(p->invocation_shifts, 0, 5);
- unsigned size_z_shift = bits(p->invocation_shifts, 5, 10);
- unsigned workgroups_x_shift = bits(p->invocation_shifts, 10, 16);
- unsigned workgroups_y_shift = bits(p->invocation_shifts, 16, 22);
- unsigned workgroups_z_shift = bits(p->invocation_shifts, 22, 28);
- unsigned workgroups_x_shift_2 = bits(p->invocation_shifts, 28, 32);
+ unsigned size_x = bits(invocation.invocations, 0, invocation.size_y_shift) + 1;
+ unsigned size_y = bits(invocation.invocations, invocation.size_y_shift, invocation.size_z_shift) + 1;
+ unsigned size_z = bits(invocation.invocations, invocation.size_z_shift, invocation.workgroups_x_shift) + 1;
- unsigned size_x = bits(p->invocation_count, 0, size_y_shift) + 1;
- unsigned size_y = bits(p->invocation_count, size_y_shift, size_z_shift) + 1;
- unsigned size_z = bits(p->invocation_count, size_z_shift, workgroups_x_shift) + 1;
-
- unsigned groups_x = bits(p->invocation_count, workgroups_x_shift, workgroups_y_shift) + 1;
- unsigned groups_y = bits(p->invocation_count, workgroups_y_shift, workgroups_z_shift) + 1;
- unsigned groups_z = bits(p->invocation_count, workgroups_z_shift, 32) + 1;
+ unsigned groups_x = bits(invocation.invocations, invocation.workgroups_x_shift, invocation.workgroups_y_shift) + 1;
+ unsigned groups_y = bits(invocation.invocations, invocation.workgroups_y_shift, invocation.workgroups_z_shift) + 1;
+ unsigned groups_z = bits(invocation.invocations, invocation.workgroups_z_shift, 32) + 1;
/* Even though we have this decoded, we want to ensure that the
* representation is "unique" so we don't lose anything by printing only
* decode and pack it ourselves! If it is bit exact with what we
* decoded, we're good to go. */
- struct mali_vertex_tiler_prefix ref;
+ struct mali_invocation_packed ref;
panfrost_pack_work_groups_compute(&ref, groups_x, groups_y, groups_z, size_x, size_y, size_z, graphics);
- bool canonical =
- (p->invocation_count == ref.invocation_count) &&
- (p->invocation_shifts == ref.invocation_shifts);
-
- if (!canonical) {
+ if (memcmp(&ref, &invocation_packed, sizeof(ref))) {
pandecode_msg("XXX: non-canonical workgroups packing\n");
- pandecode_msg("expected: %X, %X",
- ref.invocation_count,
- ref.invocation_shifts);
-
- pandecode_prop("invocation_count = 0x%" PRIx32, p->invocation_count);
- pandecode_prop("size_y_shift = %d", size_y_shift);
- pandecode_prop("size_z_shift = %d", size_z_shift);
- pandecode_prop("workgroups_x_shift = %d", workgroups_x_shift);
- pandecode_prop("workgroups_y_shift = %d", workgroups_y_shift);
- pandecode_prop("workgroups_z_shift = %d", workgroups_z_shift);
- pandecode_prop("workgroups_x_shift_2 = %d", workgroups_x_shift_2);
+ MALI_INVOCATION_print(pandecode_dump_stream, &invocation, 1 * 2);
}
/* Regardless, print the decode */
- pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
+ fprintf(pandecode_dump_stream,
+ "Invocation (%d, %d, %d) x (%d, %d, %d)\n",
size_x, size_y, size_z,
groups_x, groups_y, groups_z);
struct midgard_payload_vertex_tiler payload = {};
struct mali_primitive_packed primitive;
struct mali_draw_packed draw;
+ struct mali_invocation_packed invocation;
pan_pack(&draw, DRAW, cfg) {
cfg.unknown_1 = 0x7;
cfg.unknown_3 = 6;
}
- memcpy(&payload.prefix.primitive, &primitive, MALI_DRAW_LENGTH);
- memcpy(&payload.postfix, &draw, MALI_DRAW_LENGTH);
+ panfrost_pack_work_groups_compute(&invocation, 1, vertex_count, 1, 1, 1, 1, true);
- panfrost_pack_work_groups_compute(&payload.prefix, 1, vertex_count, 1, 1, 1, 1, true);
+ payload.prefix.primitive = primitive;
+ memcpy(&payload.postfix, &draw, MALI_DRAW_LENGTH);
+ payload.prefix.invocation = invocation;
panfrost_new_job(pool, scoreboard, MALI_JOB_TYPE_TILER, false, 0, &payload, sizeof(payload), true);
}
void
panfrost_pack_work_groups_compute(
- struct mali_vertex_tiler_prefix *out,
+ struct mali_invocation_packed *out,
unsigned num_x,
unsigned num_y,
unsigned num_z,
unsigned size_z,
bool quirk_graphics);
-void
-panfrost_pack_work_groups_fused(
- struct mali_vertex_tiler_prefix *vertex,
- struct mali_vertex_tiler_prefix *tiler,
- unsigned num_x,
- unsigned num_y,
- unsigned num_z,
- unsigned size_x,
- unsigned size_y,
- unsigned size_z);
-
/* Tiler structure size computation */
unsigned
void
panfrost_pack_work_groups_compute(
- struct mali_vertex_tiler_prefix *out,
+ struct mali_invocation_packed *out,
unsigned num_x,
unsigned num_y,
unsigned num_z,
shifts[i + 1] = shifts[i] + bit_count;
}
- /* Quirk: for non-instanced graphics, the blob sets workgroups_z_shift
- * = 32. This doesn't appear to matter to the hardware, but it's good
- * to be bit-identical. */
+ pan_pack(out, INVOCATION, cfg) {
+ cfg.invocations = packed;
+ cfg.size_y_shift = shifts[1];
+ cfg.size_z_shift = shifts[2];
+ cfg.workgroups_x_shift = shifts[3];
+ cfg.workgroups_y_shift = shifts[4];
+ cfg.workgroups_z_shift = shifts[5];
- if (quirk_graphics && (num_z <= 1))
- shifts[5] = 32;
+ /* Quirk: for non-instanced graphics, the blob sets
+ * workgroups_z_shift = 32. This doesn't appear to matter to
+ * the hardware, but it's good to be bit-identical. */
- /* Quirk: for graphics, workgroups_x_shift_2 must be at least 2,
- * whereas for OpenCL it is simply equal to workgroups_x_shift. For GL
- * compute, it is always 2 if no barriers are in use, but is equal to
- * workgroups_x_shift is barriers are in use. */
+ if (quirk_graphics && (num_z <= 1))
+ cfg.workgroups_z_shift = 32;
- unsigned shift_2 = shifts[3];
+ /* Quirk: for graphics, >= 2. For compute, 2 without barriers
+ * but equal to workgroups_x_shift with barriers */
- if (quirk_graphics)
- shift_2 = MAX2(shift_2, 2);
-
- /* Pack them in */
- uint32_t packed_shifts =
- (shifts[1] << 0) |
- (shifts[2] << 5) |
- (shifts[3] << 10) |
- (shifts[4] << 16) |
- (shifts[5] << 22) |
- (shift_2 << 28);
-
- /* Upload the packed bitfields */
- out->invocation_count = packed;
- out->invocation_shifts = packed_shifts;
-}
-
-/* Packs vertex/tiler descriptors simultaneously */
-void
-panfrost_pack_work_groups_fused(
- struct mali_vertex_tiler_prefix *vertex,
- struct mali_vertex_tiler_prefix *tiler,
- unsigned num_x,
- unsigned num_y,
- unsigned num_z,
- unsigned size_x,
- unsigned size_y,
- unsigned size_z)
-{
- panfrost_pack_work_groups_compute(vertex, num_x, num_y, num_z, size_x, size_y, size_z, true);
-
- /* Copy results over */
- tiler->invocation_count = vertex->invocation_count;
- tiler->invocation_shifts = vertex->invocation_shifts;
+ cfg.unknown_shift = quirk_graphics ? 2 : cfg.workgroups_x_shift;
+ }
}
-
projects / mesa.git / commitdiff