src/gallium/drivers/panfrost/pan_compute.c

   1 /*
   2  * Copyright (C) 2019 Collabora, Ltd.
   3  * Copyright (C) 2019 Red Hat Inc.
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the "Software"),
   7  * to deal in the Software without restriction, including without limitation
   8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   9  * and/or sell copies of the Software, and to permit persons to whom the
  10  * Software is furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice (including the next
  13  * paragraph) shall be included in all copies or substantial portions of the
  14  * Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  22  * SOFTWARE.
  23  *
  24  * Authors (Collabora):
  25  *   Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
  26  *
  27  */
  28
  29 #include "pan_context.h"
  30 #include "util/u_memory.h"
  31 #include "nir_serialize.h"
  32
  33 /* Compute CSOs are tracked like graphics shader CSOs, but are
  34  * considerably simpler. We do not implement multiple
  35  * variants/keying. So the CSO create function just goes ahead and
  36  * compiles the thing. */
  37
  38 static void *
  39 panfrost_create_compute_state(
  40         struct pipe_context *pctx,
  41         const struct pipe_compute_state *cso)
  42 {
  43         struct panfrost_context *ctx = pan_context(pctx);
  44
  45         struct panfrost_shader_variants *so = CALLOC_STRUCT(panfrost_shader_variants);
  46         so->cbase = *cso;
  47         so->is_compute = true;
  48
  49         struct panfrost_shader_state *v = &so->variants[0];
  50
  51         so->variant_count = 1;
  52         so->active_variant = 0;
  53
  54         v->tripipe = malloc(sizeof(struct mali_shader_meta));
  55
  56         if (cso->ir_type == PIPE_SHADER_IR_NIR_SERIALIZED) {
  57                 struct blob_reader reader;
  58                 const struct pipe_binary_program_header *hdr = cso->prog;
  59
  60                 blob_reader_init(&reader, hdr->blob, hdr->num_bytes);
  61                 so->cbase.prog = nir_deserialize(NULL, &midgard_nir_options, &reader);
  62                 so->cbase.ir_type = PIPE_SHADER_IR_NIR;
  63         }
  64
  65         panfrost_shader_compile(ctx, v->tripipe,
  66                         so->cbase.ir_type, so->cbase.prog,
  67                         MESA_SHADER_COMPUTE, v, NULL);
  68
  69         return so;
  70 }
  71
  72 static void
  73 panfrost_bind_compute_state(struct pipe_context *pipe, void *cso)
  74 {
  75         struct panfrost_context *ctx = pan_context(pipe);
  76
  77         struct panfrost_shader_variants *variants =
  78                 (struct panfrost_shader_variants *) cso;
  79
  80         ctx->shader[PIPE_SHADER_COMPUTE] = variants;
  81 }
  82
  83 static void
  84 panfrost_delete_compute_state(struct pipe_context *pipe, void *cso)
  85 {
  86         free(cso);
  87 }
  88
  89 /* Launch grid is the compute equivalent of draw_vbo, so in this routine, we
  90  * construct the COMPUTE job and some of its payload.
  91  */
  92
  93 static void
  94 panfrost_launch_grid(struct pipe_context *pipe,
  95                 const struct pipe_grid_info *info)
  96 {
  97         struct panfrost_context *ctx = pan_context(pipe);
  98
  99         /* TODO: Do we want a special compute-only batch? */
 100         struct panfrost_batch *batch = panfrost_get_batch_for_fbo(ctx);
 101
 102         ctx->compute_grid = info;
 103
 104         struct mali_job_descriptor_header job = {
 105                 .job_type = JOB_TYPE_COMPUTE,
 106                 .job_descriptor_size = 1,
 107                 .job_barrier = 1
 108         };
 109
 110         /* TODO: Stub */
 111         struct midgard_payload_vertex_tiler *payload = &ctx->payloads[PIPE_SHADER_COMPUTE];
 112
 113         panfrost_emit_for_draw(ctx, false);
 114
 115         /* Compute jobs have a "compute FBD". It's not a real framebuffer
 116          * descriptor - there is no framebuffer - but it takes the place of
 117          * one. As far as I can tell, it's actually the beginning of a
 118          * single-render-target framebuffer descriptor with almost everything
 119          * zeroed out.
 120          */
 121         struct mali_compute_fbd compute_fbd = {
 122                 .unknown1 = {
 123                         0, 0x1F, 0, 0, 0, 0, 0, 0
 124                 }
 125         };
 126
 127         payload->postfix.framebuffer =
 128                 panfrost_upload_transient(batch, &compute_fbd, sizeof(compute_fbd));
 129
 130         /* Invoke according to the grid info */
 131
 132         panfrost_pack_work_groups_compute(&payload->prefix,
 133                         info->grid[0], info->grid[1], info->grid[2],
 134                         info->block[0], info->block[1], info->block[2], false);
 135
 136         /* Upload the payload */
 137
 138         struct panfrost_transfer transfer = panfrost_allocate_transient(batch, sizeof(job) + sizeof(*payload));
 139         memcpy(transfer.cpu, &job, sizeof(job));
 140         memcpy(transfer.cpu + sizeof(job), payload, sizeof(*payload));
 141
 142         /* Queue the job */
 143         panfrost_scoreboard_queue_compute_job(batch, transfer);
 144
 145         panfrost_flush_all_batches(ctx, true);
 146 }
 147
 148 void
 149 panfrost_compute_context_init(struct pipe_context *pctx)
 150 {
 151         pctx->create_compute_state = panfrost_create_compute_state;
 152         pctx->bind_compute_state = panfrost_bind_compute_state;
 153         pctx->delete_compute_state = panfrost_delete_compute_state;
 154
 155         pctx->launch_grid = panfrost_launch_grid;
 156 }
 157
 158