src/intel/dev/gen_device_info.h

   1  /*
   2   * Copyright © 2013 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
  25 #ifndef GEN_DEVICE_INFO_H
  26 #define GEN_DEVICE_INFO_H
  27
  28 #include <stdbool.h>
  29 #include <stdint.h>
  30
  31 #include "util/macros.h"
  32
  33 #ifdef __cplusplus
  34 extern "C" {
  35 #endif
  36
  37 struct drm_i915_query_topology_info;
  38
  39 #define GEN_DEVICE_MAX_SLICES           (6)  /* Maximum on gen10 */
  40 #define GEN_DEVICE_MAX_SUBSLICES        (8)  /* Maximum on gen11 */
  41 #define GEN_DEVICE_MAX_EUS_PER_SUBSLICE (10) /* Maximum on Haswell */
  42 #define GEN_DEVICE_MAX_PIXEL_PIPES      (2)  /* Maximum on gen11 */
  43
  44 /**
  45  * Intel hardware information and quirks
  46  */
  47 struct gen_device_info
  48 {
  49    int gen; /**< Generation number: 4, 5, 6, 7, ... */
  50    int revision;
  51    int gt;
  52
  53    bool is_g4x;
  54    bool is_ivybridge;
  55    bool is_baytrail;
  56    bool is_haswell;
  57    bool is_broadwell;
  58    bool is_cherryview;
  59    bool is_skylake;
  60    bool is_broxton;
  61    bool is_kabylake;
  62    bool is_geminilake;
  63    bool is_coffeelake;
  64    bool is_cannonlake;
  65    bool is_elkhartlake;
  66
  67    bool has_hiz_and_separate_stencil;
  68    bool must_use_separate_stencil;
  69    bool has_sample_with_hiz;
  70    bool has_llc;
  71
  72    bool has_pln;
  73    bool has_64bit_float;
  74    bool has_64bit_int;
  75    bool has_integer_dword_mul;
  76    bool has_compr4;
  77    bool has_surface_tile_offset;
  78    bool supports_simd16_3src;
  79    bool has_resource_streamer;
  80    bool disable_ccs_repack;
  81    bool has_aux_map;
  82
  83    /**
  84     * \name Intel hardware quirks
  85     *  @{
  86     */
  87    bool has_negative_rhw_bug;
  88
  89    /**
  90     * Some versions of Gen hardware don't do centroid interpolation correctly
  91     * on unlit pixels, causing incorrect values for derivatives near triangle
  92     * edges.  Enabling this flag causes the fragment shader to use
  93     * non-centroid interpolation for unlit pixels, at the expense of two extra
  94     * fragment shader instructions.
  95     */
  96    bool needs_unlit_centroid_workaround;
  97    /** @} */
  98
  99    /**
 100     * \name GPU hardware limits
 101     *
 102     * In general, you can find shader thread maximums by looking at the "Maximum
 103     * Number of Threads" field in the Intel PRM description of the 3DSTATE_VS,
 104     * 3DSTATE_GS, 3DSTATE_HS, 3DSTATE_DS, and 3DSTATE_PS commands. URB entry
 105     * limits come from the "Number of URB Entries" field in the
 106     * 3DSTATE_URB_VS command and friends.
 107     *
 108     * These fields are used to calculate the scratch space to allocate.  The
 109     * amount of scratch space can be larger without being harmful on modern
 110     * GPUs, however, prior to Haswell, programming the maximum number of threads
 111     * to greater than the hardware maximum would cause GPU performance to tank.
 112     *
 113     *  @{
 114     */
 115    /**
 116     * Total number of slices present on the device whether or not they've been
 117     * fused off.
 118     *
 119     * XXX: CS thread counts are limited by the inability to do cross subslice
 120     * communication. It is the effectively the number of logical threads which
 121     * can be executed in a subslice. Fuse configurations may cause this number
 122     * to change, so we program @max_cs_threads as the lower maximum.
 123     */
 124    unsigned num_slices;
 125
 126    /**
 127     * Number of subslices for each slice (used to be uniform until CNL).
 128     */
 129    unsigned num_subslices[GEN_DEVICE_MAX_SUBSLICES];
 130
 131    /**
 132     * Number of subslices on each pixel pipe (ICL).
 133     */
 134    unsigned ppipe_subslices[GEN_DEVICE_MAX_PIXEL_PIPES];
 135
 136    /**
 137     * Upper bound of number of EU per subslice (some SKUs might have just 1 EU
 138     * fused across all subslices, like 47 EUs, in which case this number won't
 139     * be acurate for one subslice).
 140     */
 141    unsigned num_eu_per_subslice;
 142
 143    /**
 144     * Number of threads per eu, varies between 4 and 8 between generations.
 145     */
 146    unsigned num_thread_per_eu;
 147
 148    /**
 149     * A bit mask of the slices available.
 150     */
 151    uint8_t slice_masks;
 152
 153    /**
 154     * An array of bit mask of the subslices available, use subslice_slice_stride
 155     * to access this array.
 156     */
 157    uint8_t subslice_masks[GEN_DEVICE_MAX_SLICES *
 158                           DIV_ROUND_UP(GEN_DEVICE_MAX_SUBSLICES, 8)];
 159
 160    /**
 161     * An array of bit mask of EUs available, use eu_slice_stride &
 162     * eu_subslice_stride to access this array.
 163     */
 164    uint8_t eu_masks[GEN_DEVICE_MAX_SLICES *
 165                     GEN_DEVICE_MAX_SUBSLICES *
 166                     DIV_ROUND_UP(GEN_DEVICE_MAX_EUS_PER_SUBSLICE, 8)];
 167
 168    /**
 169     * Stride to access subslice_masks[].
 170     */
 171    uint16_t subslice_slice_stride;
 172
 173    /**
 174     * Strides to access eu_masks[].
 175     */
 176    uint16_t eu_slice_stride;
 177    uint16_t eu_subslice_stride;
 178
 179    unsigned l3_banks;
 180    unsigned max_vs_threads;   /**< Maximum Vertex Shader threads */
 181    unsigned max_tcs_threads;  /**< Maximum Hull Shader threads */
 182    unsigned max_tes_threads;  /**< Maximum Domain Shader threads */
 183    unsigned max_gs_threads;   /**< Maximum Geometry Shader threads. */
 184    /**
 185     * Theoretical maximum number of Pixel Shader threads.
 186     *
 187     * PSD means Pixel Shader Dispatcher. On modern Intel GPUs, hardware will
 188     * automatically scale pixel shader thread count, based on a single value
 189     * programmed into 3DSTATE_PS.
 190     *
 191     * To calculate the maximum number of threads for Gen8 beyond (which have
 192     * multiple Pixel Shader Dispatchers):
 193     *
 194     * - Look up 3DSTATE_PS and find "Maximum Number of Threads Per PSD"
 195     * - Usually there's only one PSD per subslice, so use the number of
 196     *   subslices for number of PSDs.
 197     * - For max_wm_threads, the total should be PSD threads * #PSDs.
 198     */
 199    unsigned max_wm_threads;
 200
 201    /**
 202     * Maximum Compute Shader threads.
 203     *
 204     * Thread count * number of EUs per subslice
 205     */
 206    unsigned max_cs_threads;
 207
 208    struct {
 209       /**
 210        * Fixed size of the URB.
 211        *
 212        * On Gen6 and DG1, this is measured in KB.  Gen4-5 instead measure
 213        * this in 512b blocks, as that's more convenient there.
 214        *
 215        * On most Gen7+ platforms, the URB is a section of the L3 cache,
 216        * and can be resized based on the L3 programming.  For those platforms,
 217        * simply leave this field blank (zero) - it isn't used.
 218        */
 219       unsigned size;
 220
 221       /**
 222        * The minimum number of URB entries.  See the 3DSTATE_URB_<XS> docs.
 223        */
 224       unsigned min_entries[4];
 225
 226       /**
 227        * The maximum number of URB entries.  See the 3DSTATE_URB_<XS> docs.
 228        */
 229       unsigned max_entries[4];
 230    } urb;
 231
 232    /**
 233     * For the longest time the timestamp frequency for Gen's timestamp counter
 234     * could be assumed to be 12.5MHz, where the least significant bit neatly
 235     * corresponded to 80 nanoseconds.
 236     *
 237     * Since Gen9 the numbers aren't so round, with a a frequency of 12MHz for
 238     * SKL (or scale factor of 83.33333333) and a frequency of 19200000Hz for
 239     * BXT.
 240     *
 241     * For simplicty to fit with the current code scaling by a single constant
 242     * to map from raw timestamps to nanoseconds we now do the conversion in
 243     * floating point instead of integer arithmetic.
 244     *
 245     * In general it's probably worth noting that the documented constants we
 246     * have for the per-platform timestamp frequencies aren't perfect and
 247     * shouldn't be trusted for scaling and comparing timestamps with a large
 248     * delta.
 249     *
 250     * E.g. with crude testing on my system using the 'correct' scale factor I'm
 251     * seeing a drift of ~2 milliseconds per second.
 252     */
 253    uint64_t timestamp_frequency;
 254
 255    uint64_t aperture_bytes;
 256
 257    /**
 258     * ID to put into the .aub files.
 259     */
 260    int simulator_id;
 261
 262    /**
 263     * holds the pci device id
 264     */
 265    uint32_t chipset_id;
 266
 267    /**
 268     * no_hw is true when the chipset_id pci device id has been overridden
 269     */
 270    bool no_hw;
 271    /** @} */
 272 };
 273
 274 #define gen_device_info_is_9lp(devinfo) \
 275    ((devinfo)->is_broxton || (devinfo)->is_geminilake)
 276
 277 static inline bool
 278 gen_device_info_subslice_available(const struct gen_device_info *devinfo,
 279                                    int slice, int subslice)
 280 {
 281    return (devinfo->subslice_masks[slice * devinfo->subslice_slice_stride +
 282                                    subslice / 8] & (1U << (subslice % 8))) != 0;
 283 }
 284
 285 int gen_device_name_to_pci_device_id(const char *name);
 286 const char *gen_get_device_name(int devid);
 287
 288 static inline uint64_t
 289 gen_device_info_timebase_scale(const struct gen_device_info *devinfo,
 290                                uint64_t gpu_timestamp)
 291 {
 292    return (1000000000ull * gpu_timestamp) / devinfo->timestamp_frequency;
 293 }
 294
 295 bool gen_get_device_info_from_fd(int fh, struct gen_device_info *devinfo);
 296 bool gen_get_device_info_from_pci_id(int pci_id,
 297                                      struct gen_device_info *devinfo);
 298 int gen_get_aperture_size(int fd, uint64_t *size);
 299
 300 #ifdef __cplusplus
 301 }
 302 #endif
 303
 304 #endif /* GEN_DEVICE_INFO_H */