src/intel/blorp/blorp.c

   1 /*
   2  * Copyright © 2012 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  */
  23
  24 #include <errno.h>
  25
  26 #include "program/prog_instruction.h"
  27
  28 #include "blorp_priv.h"
  29 #include "brw_compiler.h"
  30 #include "brw_nir.h"
  31
  32 void
  33 blorp_init(struct blorp_context *blorp, void *driver_ctx,
  34            struct isl_device *isl_dev)
  35 {
  36    blorp->driver_ctx = driver_ctx;
  37    blorp->isl_dev = isl_dev;
  38 }
  39
  40 void
  41 blorp_finish(struct blorp_context *blorp)
  42 {
  43    blorp->driver_ctx = NULL;
  44 }
  45
  46 void
  47 blorp_batch_init(struct blorp_context *blorp,
  48                  struct blorp_batch *batch, void *driver_batch,
  49                  enum blorp_batch_flags flags)
  50 {
  51    batch->blorp = blorp;
  52    batch->driver_batch = driver_batch;
  53    batch->flags = flags;
  54 }
  55
  56 void
  57 blorp_batch_finish(struct blorp_batch *batch)
  58 {
  59    batch->blorp = NULL;
  60 }
  61
  62 void
  63 brw_blorp_surface_info_init(struct blorp_context *blorp,
  64                             struct brw_blorp_surface_info *info,
  65                             const struct blorp_surf *surf,
  66                             unsigned int level, unsigned int layer,
  67                             enum isl_format format, bool is_render_target)
  68 {
  69    info->enabled = true;
  70
  71    if (format == ISL_FORMAT_UNSUPPORTED)
  72       format = surf->surf->format;
  73
  74    if (format == ISL_FORMAT_R24_UNORM_X8_TYPELESS) {
  75       /* Unfortunately, ISL_FORMAT_R24_UNORM_X8_TYPELESS it isn't supported as
  76        * a render target, which would prevent us from blitting to 24-bit
  77        * depth.  The miptree consists of 32 bits per pixel, arranged as 24-bit
  78        * depth values interleaved with 8 "don't care" bits.  Since depth
  79        * values don't require any blending, it doesn't matter how we interpret
  80        * the bit pattern as long as we copy the right amount of data, so just
  81        * map it as 8-bit BGRA.
  82        */
  83       format = ISL_FORMAT_B8G8R8A8_UNORM;
  84    } else if (surf->surf->usage & ISL_SURF_USAGE_STENCIL_BIT) {
  85       assert(surf->surf->format == ISL_FORMAT_R8_UINT);
  86       /* Prior to Broadwell, we can't render to R8_UINT */
  87       if (blorp->isl_dev->info->gen < 8)
  88          format = ISL_FORMAT_R8_UNORM;
  89    }
  90
  91    info->surf = *surf->surf;
  92    info->addr = surf->addr;
  93
  94    info->aux_usage = surf->aux_usage;
  95    if (info->aux_usage != ISL_AUX_USAGE_NONE) {
  96       info->aux_surf = *surf->aux_surf;
  97       info->aux_addr = surf->aux_addr;
  98    }
  99
 100    info->clear_color = surf->clear_color;
 101
 102    info->view = (struct isl_view) {
 103       .usage = is_render_target ? ISL_SURF_USAGE_RENDER_TARGET_BIT :
 104                                   ISL_SURF_USAGE_TEXTURE_BIT,
 105       .format = format,
 106       .base_level = level,
 107       .levels = 1,
 108       .swizzle = ISL_SWIZZLE_IDENTITY,
 109    };
 110
 111    info->view.array_len = MAX2(info->surf.logical_level0_px.depth,
 112                                info->surf.logical_level0_px.array_len);
 113
 114    if (!is_render_target &&
 115        (info->surf.dim == ISL_SURF_DIM_3D ||
 116         info->surf.msaa_layout == ISL_MSAA_LAYOUT_ARRAY)) {
 117       /* 3-D textures don't support base_array layer and neither do 2-D
 118        * multisampled textures on IVB so we need to pass it through the
 119        * sampler in those cases.  These are also two cases where we are
 120        * guaranteed that we won't be doing any funny surface hacks.
 121        */
 122       info->view.base_array_layer = 0;
 123       info->z_offset = layer;
 124    } else {
 125       info->view.base_array_layer = layer;
 126
 127       assert(info->view.array_len >= info->view.base_array_layer);
 128       info->view.array_len -= info->view.base_array_layer;
 129       info->z_offset = 0;
 130    }
 131
 132    /* Sandy Bridge has a limit of a maximum of 512 layers for layered
 133     * rendering.
 134     */
 135    if (is_render_target && blorp->isl_dev->info->gen == 6)
 136       info->view.array_len = MIN2(info->view.array_len, 512);
 137 }
 138
 139
 140 void
 141 blorp_params_init(struct blorp_params *params)
 142 {
 143    memset(params, 0, sizeof(*params));
 144    params->num_draw_buffers = 1;
 145    params->num_layers = 1;
 146 }
 147
 148 void
 149 brw_blorp_init_wm_prog_key(struct brw_wm_prog_key *wm_key)
 150 {
 151    memset(wm_key, 0, sizeof(*wm_key));
 152    wm_key->nr_color_regions = 1;
 153    for (int i = 0; i < MAX_SAMPLERS; i++)
 154       wm_key->tex.swizzles[i] = SWIZZLE_XYZW;
 155 }
 156
 157 static int
 158 nir_uniform_type_size(const struct glsl_type *type)
 159 {
 160    /* Only very basic types are allowed */
 161    assert(glsl_type_is_vector_or_scalar(type));
 162    assert(glsl_get_bit_size(type) == 32);
 163
 164    return glsl_get_vector_elements(type) * 4;
 165 }
 166
 167 const unsigned *
 168 brw_blorp_compile_nir_shader(struct blorp_context *blorp, struct nir_shader *nir,
 169                              const struct brw_wm_prog_key *wm_key,
 170                              bool use_repclear,
 171                              struct brw_blorp_prog_data *prog_data,
 172                              unsigned *program_size)
 173 {
 174    const struct brw_compiler *compiler = blorp->compiler;
 175
 176    void *mem_ctx = ralloc_context(NULL);
 177
 178    /* Calling brw_preprocess_nir and friends is destructive and, if cloning is
 179     * enabled, may end up completely replacing the nir_shader.  Therefore, we
 180     * own it and might as well put it in our context for easy cleanup.
 181     */
 182    ralloc_steal(mem_ctx, nir);
 183    nir->options =
 184       compiler->glsl_compiler_options[MESA_SHADER_FRAGMENT].NirOptions;
 185
 186    struct brw_wm_prog_data wm_prog_data;
 187    memset(&wm_prog_data, 0, sizeof(wm_prog_data));
 188
 189    wm_prog_data.base.nr_params = 0;
 190    wm_prog_data.base.param = NULL;
 191
 192    /* BLORP always just uses the first two binding table entries */
 193    wm_prog_data.binding_table.render_target_start = BLORP_RENDERBUFFER_BT_INDEX;
 194    wm_prog_data.base.binding_table.texture_start = BLORP_TEXTURE_BT_INDEX;
 195
 196    nir = brw_preprocess_nir(compiler, nir);
 197    nir_remove_dead_variables(nir, nir_var_shader_in);
 198    nir_shader_gather_info(nir, nir_shader_get_entrypoint(nir));
 199
 200    /* Uniforms are required to be lowered before going into compile_fs.  For
 201     * BLORP, we'll assume that whoever builds the shader sets the location
 202     * they want so we just need to lower them and figure out how many we have
 203     * in total.
 204     */
 205    nir->num_uniforms = 0;
 206    nir_foreach_variable(var, &nir->uniforms) {
 207       var->data.driver_location = var->data.location;
 208       unsigned end = var->data.location + nir_uniform_type_size(var->type);
 209       nir->num_uniforms = MAX2(nir->num_uniforms, end);
 210    }
 211    nir_lower_io(nir, nir_var_uniform, nir_uniform_type_size, 0);
 212
 213    const unsigned *program =
 214       brw_compile_fs(compiler, blorp->driver_ctx, mem_ctx,
 215                      wm_key, &wm_prog_data, nir,
 216                      NULL, -1, -1, false, use_repclear, program_size, NULL);
 217
 218    /* Copy the relavent bits of wm_prog_data over into the blorp prog data */
 219    prog_data->dispatch_8 = wm_prog_data.dispatch_8;
 220    prog_data->dispatch_16 = wm_prog_data.dispatch_16;
 221    prog_data->first_curbe_grf_0 = wm_prog_data.base.dispatch_grf_start_reg;
 222    prog_data->first_curbe_grf_2 = wm_prog_data.dispatch_grf_start_reg_2;
 223    prog_data->ksp_offset_2 = wm_prog_data.prog_offset_2;
 224    prog_data->persample_msaa_dispatch = wm_prog_data.persample_dispatch;
 225    prog_data->flat_inputs = wm_prog_data.flat_inputs;
 226    prog_data->num_varying_inputs = wm_prog_data.num_varying_inputs;
 227    prog_data->inputs_read = nir->info->inputs_read;
 228
 229    assert(wm_prog_data.base.nr_params == 0);
 230
 231    return program;
 232 }
 233
 234 void
 235 blorp_gen6_hiz_op(struct blorp_batch *batch,
 236                   struct blorp_surf *surf, unsigned level, unsigned layer,
 237                   enum blorp_hiz_op op)
 238 {
 239    struct blorp_params params;
 240    blorp_params_init(&params);
 241
 242    params.hiz_op = op;
 243
 244    brw_blorp_surface_info_init(batch->blorp, &params.depth, surf, level, layer,
 245                                surf->surf->format, true);
 246
 247    /* Align the rectangle primitive to 8x4 pixels.
 248     *
 249     * During fast depth clears, the emitted rectangle primitive  must be
 250     * aligned to 8x4 pixels.  From the Ivybridge PRM, Vol 2 Part 1 Section
 251     * 11.5.3.1 Depth Buffer Clear (and the matching section in the Sandybridge
 252     * PRM):
 253     *     If Number of Multisamples is NUMSAMPLES_1, the rectangle must be
 254     *     aligned to an 8x4 pixel block relative to the upper left corner
 255     *     of the depth buffer [...]
 256     *
 257     * For hiz resolves, the rectangle must also be 8x4 aligned. Item
 258     * WaHizAmbiguate8x4Aligned from the Haswell workarounds page and the
 259     * Ivybridge simulator require the alignment.
 260     *
 261     * To be safe, let's just align the rect for all hiz operations and all
 262     * hardware generations.
 263     *
 264     * However, for some miptree slices of a Z24 texture, emitting an 8x4
 265     * aligned rectangle that covers the slice may clobber adjacent slices if
 266     * we strictly adhered to the texture alignments specified in the PRM.  The
 267     * Ivybridge PRM, Section "Alignment Unit Size", states that
 268     * SURFACE_STATE.Surface_Horizontal_Alignment should be 4 for Z24 surfaces,
 269     * not 8. But commit 1f112cc increased the alignment from 4 to 8, which
 270     * prevents the clobbering.
 271     */
 272    params.x1 = minify(params.depth.surf.logical_level0_px.width,
 273                       params.depth.view.base_level);
 274    params.y1 = minify(params.depth.surf.logical_level0_px.height,
 275                       params.depth.view.base_level);
 276    params.x1 = ALIGN(params.x1, 8);
 277    params.y1 = ALIGN(params.y1, 4);
 278
 279    if (params.depth.view.base_level == 0) {
 280       /* TODO: What about MSAA? */
 281       params.depth.surf.logical_level0_px.width = params.x1;
 282       params.depth.surf.logical_level0_px.height = params.y1;
 283    }
 284
 285    params.dst.surf.samples = params.depth.surf.samples;
 286    params.dst.surf.logical_level0_px = params.depth.surf.logical_level0_px;
 287    params.depth_format = isl_format_get_depth_format(surf->surf->format, false);
 288
 289    batch->blorp->exec(batch, &params);
 290 }