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 "pan_cmdstream.h"
  31 #include "pan_bo.h"
  32 #include "util/u_memory.h"
  33 #include "nir_serialize.h"
  34
  35 /* Compute CSOs are tracked like graphics shader CSOs, but are
  36  * considerably simpler. We do not implement multiple
  37  * variants/keying. So the CSO create function just goes ahead and
  38  * compiles the thing. */
  39
  40 static void *
  41 panfrost_create_compute_state(
  42         struct pipe_context *pctx,
  43         const struct pipe_compute_state *cso)
  44 {
  45         struct panfrost_context *ctx = pan_context(pctx);
  46
  47         struct panfrost_shader_variants *so = CALLOC_STRUCT(panfrost_shader_variants);
  48         so->cbase = *cso;
  49         so->is_compute = true;
  50
  51         struct panfrost_shader_state *v = calloc(1, sizeof(*v));
  52         so->variants = v;
  53
  54         so->variant_count = 1;
  55         so->active_variant = 0;
  56
  57         if (cso->ir_type == PIPE_SHADER_IR_NIR_SERIALIZED) {
  58                 struct blob_reader reader;
  59                 const struct pipe_binary_program_header *hdr = cso->prog;
  60
  61                 blob_reader_init(&reader, hdr->blob, hdr->num_bytes);
  62                 so->cbase.prog = nir_deserialize(NULL, &midgard_nir_options, &reader);
  63                 so->cbase.ir_type = PIPE_SHADER_IR_NIR;
  64         }
  65
  66         panfrost_shader_compile(ctx, 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         /* TODO: Stub */
 105         struct midgard_payload_vertex_tiler payload;
 106
 107         /* We implement OpenCL inputs as uniforms (or a UBO -- same thing), so
 108          * reuse the graphics path for this by lowering to Gallium */
 109
 110         struct pipe_constant_buffer ubuf = {
 111                 .buffer = NULL,
 112                 .buffer_offset = 0,
 113                 .buffer_size = ctx->shader[PIPE_SHADER_COMPUTE]->cbase.req_input_mem,
 114                 .user_buffer = info->input
 115         };
 116
 117         if (info->input)
 118                 pipe->set_constant_buffer(pipe, PIPE_SHADER_COMPUTE, 0, &ubuf);
 119
 120         panfrost_vt_init(ctx, PIPE_SHADER_COMPUTE, &payload.prefix, &payload.postfix);
 121
 122         panfrost_emit_shader_meta(batch, PIPE_SHADER_COMPUTE, &payload.postfix);
 123         panfrost_emit_const_buf(batch, PIPE_SHADER_COMPUTE, &payload.postfix);
 124         panfrost_emit_shared_memory(batch, info, &payload);
 125
 126         /* Invoke according to the grid info */
 127
 128         panfrost_pack_work_groups_compute(&payload.prefix,
 129                                           info->grid[0], info->grid[1],
 130                                           info->grid[2],
 131                                           info->block[0], info->block[1],
 132                                           info->block[2],
 133                                           false);
 134
 135         panfrost_new_job(&batch->pool, &batch->scoreboard,
 136                         MALI_JOB_TYPE_COMPUTE, true, 0, &payload,
 137                          sizeof(payload), false);
 138         panfrost_flush_all_batches(ctx, 0);
 139 }
 140
 141 static void
 142 panfrost_set_compute_resources(struct pipe_context *pctx,
 143                          unsigned start, unsigned count,
 144                          struct pipe_surface **resources)
 145 {
 146         /* TODO */
 147 }
 148
 149 static void
 150 panfrost_set_global_binding(struct pipe_context *pctx,
 151                       unsigned first, unsigned count,
 152                       struct pipe_resource **resources,
 153                       uint32_t **handles)
 154 {
 155         /* TODO */
 156 }
 157
 158 static void
 159 panfrost_memory_barrier(struct pipe_context *pctx, unsigned flags)
 160 {
 161         /* TODO */
 162 }
 163
 164 void
 165 panfrost_compute_context_init(struct pipe_context *pctx)
 166 {
 167         pctx->create_compute_state = panfrost_create_compute_state;
 168         pctx->bind_compute_state = panfrost_bind_compute_state;
 169         pctx->delete_compute_state = panfrost_delete_compute_state;
 170
 171         pctx->launch_grid = panfrost_launch_grid;
 172
 173         pctx->set_compute_resources = panfrost_set_compute_resources;
 174         pctx->set_global_binding = panfrost_set_global_binding;
 175
 176         pctx->memory_barrier = panfrost_memory_barrier;
 177 }