src/panfrost/bifrost/bifrost.h

   1 /*
   2  * Copyright (C) 2019 Connor Abbott <cwabbott0@gmail.com>
   3  * Copyright (C) 2019 Lyude Paul <thatslyude@gmail.com>
   4  * Copyright (C) 2019 Ryan Houdek <Sonicadvance1@gmail.com>
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the "Software"),
   8  * to deal in the Software without restriction, including without limitation
   9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  10  * and/or sell copies of the Software, and to permit persons to whom the
  11  * Software is furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice (including the next
  14  * paragraph) shall be included in all copies or substantial portions of the
  15  * Software.
  16  *
  17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  18  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  19  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  20  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  21  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  22  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  23  * SOFTWARE.
  24  */
  25
  26 #ifndef __bifrost_h__
  27 #define __bifrost_h__
  28
  29 #include <stdint.h>
  30 #include <stdbool.h>
  31
  32 struct bifrost_header {
  33         unsigned unk0 : 7;
  34         // If true, convert any infinite result of any floating-point operation to
  35         // the biggest representable number.
  36         unsigned suppress_inf: 1;
  37         // Convert any NaN results to 0.
  38         unsigned suppress_nan : 1;
  39         unsigned unk1 : 2;
  40         // true if the execution mask of the next clause is the same as the mask of
  41         // the current clause.
  42         unsigned back_to_back : 1;
  43         unsigned no_end_of_shader: 1;
  44         unsigned unk2 : 2;
  45         // Set to true for fragment shaders, to implement this bit of spec text
  46         // from section 7.1.5 of the GLSL ES spec:
  47         //
  48         // "Stores to image and buffer variables performed by helper invocations
  49         // have no effect on the underlying image or buffer memory."
  50         //
  51         // Helper invocations are threads (invocations) corresponding to pixels in
  52         // a quad that aren't actually part of the triangle, but are included to
  53         // make derivatives work correctly. They're usually turned on, but they
  54         // need to be masked off for GLSL-level stores. This bit seems to be the
  55         // only bit that's actually different between fragment shaders and other
  56         // shaders, so this is probably what it's doing.
  57         unsigned elide_writes : 1;
  58         // If backToBack is off:
  59         // - true for conditional branches and fallthrough
  60         // - false for unconditional branches
  61         // The blob seems to always set it to true if back-to-back is on.
  62         unsigned branch_cond : 1;
  63         // This bit is set when the next clause writes to the data register of some
  64         // previous clause.
  65         unsigned datareg_writebarrier: 1;
  66         unsigned datareg : 6;
  67         unsigned scoreboard_deps: 8;
  68         unsigned scoreboard_index: 3;
  69         unsigned clause_type: 4;
  70         unsigned unk3 : 1; // part of clauseType?
  71         unsigned next_clause_type: 4;
  72         unsigned unk4 : 1; // part of nextClauseType?
  73 };
  74
  75 struct bifrost_fma_inst {
  76         unsigned src0 : 3;
  77         unsigned op   : 20;
  78 };
  79
  80 struct bifrost_add_inst {
  81         unsigned src0 : 3;
  82         unsigned op   : 17;
  83 };
  84
  85 enum bifrost_outmod {
  86         BIFROST_NONE = 0x0,
  87         BIFROST_POS = 0x1,
  88         BIFROST_SAT_SIGNED = 0x2,
  89         BIFROST_SAT = 0x3,
  90 };
  91
  92 enum bifrost_roundmode {
  93         BIFROST_RTE = 0x0,
  94         BIFROST_RTP = 0x1,
  95         BIFROST_RTN = 0x2,
  96         BIFROST_RTZ = 0x3
  97 };
  98
  99 enum bifrost_minmax_mode {
 100         BIFROST_MINMAX_NONE = 0x0,
 101         BIFROST_NAN_WINS    = 0x1,
 102         BIFROST_SRC1_WINS   = 0x2,
 103         BIFROST_SRC0_WINS   = 0x3,
 104 };
 105
 106 struct bifrost_fma_add {
 107         unsigned src0 : 3;
 108         unsigned src1 : 3;
 109         unsigned src1_abs : 1;
 110         unsigned src0_neg : 1;
 111         unsigned src1_neg : 1;
 112         unsigned unk : 3;
 113         unsigned src0_abs : 1;
 114         enum bifrost_outmod outmod : 2;
 115         enum bifrost_roundmode roundmode : 2;
 116         unsigned op : 6;
 117 };
 118
 119 enum bifrost_csel_cond {
 120         BIFROST_FEQ_F = 0x0,
 121         BIFROST_FGT_F = 0x1,
 122         BIFROST_FGE_F = 0x2,
 123         BIFROST_IEQ_F = 0x3,
 124         BIFROST_IGT_I = 0x4,
 125         BIFROST_IGE_I = 0x5,
 126         BIFROST_UGT_I = 0x6,
 127         BIFROST_UGE_I = 0x7
 128 };
 129
 130 struct bifrost_csel4 {
 131         unsigned src0 : 3;
 132         unsigned src1 : 3;
 133         unsigned src2 : 3;
 134         unsigned src3 : 3;
 135         enum bifrost_csel_cond cond : 3;
 136         unsigned op   : 8;
 137 };
 138
 139 struct bifrost_shift_fma {
 140         unsigned src0 : 3;
 141         unsigned src1 : 3;
 142         unsigned src2 : 3;
 143         unsigned half : 3; /* 000 for i32, 100 for i8, 111 for v2i16 */
 144         unsigned unk  : 1; /* always set? */
 145         unsigned invert_1 : 1; /* Inverts sources to combining op */
 146         /* For XOR, switches RSHIFT to LSHIFT since only one invert needed */
 147         unsigned invert_2 : 1;
 148         unsigned op : 8;
 149 };
 150
 151 struct bifrost_shift_add {
 152         unsigned src0 : 3;
 153         unsigned src1 : 3;
 154         unsigned src2 : 3;
 155         unsigned zero : 2;
 156
 157         unsigned invert_1 : 1;
 158         unsigned invert_2 : 1;
 159
 160         unsigned op : 7;
 161 };
 162
 163 enum bifrost_ldst_type {
 164         BIFROST_LDST_F16 = 0,
 165         BIFROST_LDST_F32 = 1,
 166         BIFROST_LDST_I32 = 2,
 167         BIFROST_LDST_U32 = 3
 168 };
 169
 170 struct bifrost_ld_var_addr {
 171         unsigned src0 : 3;
 172         unsigned src1 : 3;
 173         unsigned location : 5;
 174         enum bifrost_ldst_type type : 2;
 175         unsigned op : 7;
 176 };
 177
 178 struct bifrost_ld_attr {
 179         unsigned src0 : 3;
 180         unsigned src1 : 3;
 181         unsigned location : 5;
 182         unsigned channels : 2; /* MALI_POSITIVE */
 183         enum bifrost_ldst_type type : 2;
 184         unsigned op : 5;
 185 };
 186
 187 enum bifrost_interp_mode {
 188         BIFROST_INTERP_PER_FRAG = 0x0,
 189         BIFROST_INTERP_CENTROID = 0x1,
 190         BIFROST_INTERP_DEFAULT  = 0x2,
 191         BIFROST_INTERP_EXPLICIT = 0x3
 192 };
 193
 194 struct bifrost_ld_var {
 195         unsigned src0 : 3;
 196
 197         /* If top two bits set, indirect with src in bottom three */
 198         unsigned addr : 5;
 199
 200         unsigned channels : 2; /* MALI_POSITIVE */
 201         enum bifrost_interp_mode interp_mode : 2;
 202         unsigned reuse : 1;
 203         unsigned flat : 1;
 204         unsigned op : 6;
 205 };
 206
 207 struct bifrost_tex_ctrl {
 208         unsigned sampler_index : 4; // also used to signal indirects
 209         unsigned tex_index : 7;
 210         bool no_merge_index : 1; // whether to merge (direct) sampler & texture indices
 211         bool filter : 1; // use the usual filtering pipeline (0 for texelFetch & textureGather)
 212         unsigned unk0 : 2;
 213         bool texel_offset : 1; // *Offset()
 214         bool is_shadow : 1;
 215         bool is_array : 1;
 216         unsigned tex_type : 2; // 2D, 3D, Cube, Buffer
 217         bool compute_lod : 1; // 0 for *Lod()
 218         bool not_supply_lod : 1; // 0 for *Lod() or when a bias is applied
 219         bool calc_gradients : 1; // 0 for *Grad()
 220         unsigned unk1 : 1;
 221         unsigned result_type : 4; // integer, unsigned, float TODO: why is this 4 bits?
 222         unsigned unk2 : 4;
 223 };
 224
 225 struct bifrost_dual_tex_ctrl {
 226         unsigned sampler_index0 : 2;
 227         unsigned unk0 : 2;
 228         unsigned tex_index0 : 2;
 229         unsigned sampler_index1 : 2;
 230         unsigned tex_index1 : 2;
 231         unsigned unk1 : 22;
 232 };
 233
 234 enum branch_bit_size {
 235         BR_SIZE_32 = 0,
 236         BR_SIZE_16XX = 1,
 237         BR_SIZE_16YY = 2,
 238         // For the above combinations of bitsize and location, an extra bit is
 239         // encoded via comparing the sources. The only possible source of ambiguity
 240         // would be if the sources were the same, but then the branch condition
 241         // would be always true or always false anyways, so we can ignore it. But
 242         // this no longer works when comparing the y component to the x component,
 243         // since it's valid to compare the y component of a source against its own
 244         // x component. Instead, the extra bit is encoded via an extra bitsize.
 245         BR_SIZE_16YX0 = 3,
 246         BR_SIZE_16YX1 = 4,
 247         BR_SIZE_32_AND_16X = 5,
 248         BR_SIZE_32_AND_16Y = 6,
 249         // Used for comparisons with zero and always-true, see below. I think this
 250         // only works for integer comparisons.
 251         BR_SIZE_ZERO = 7,
 252 };
 253
 254 enum bifrost_reg_write_unit {
 255         REG_WRITE_NONE = 0, // don't write
 256         REG_WRITE_TWO, // write using reg2
 257         REG_WRITE_THREE, // write using reg3
 258 };
 259
 260 struct bifrost_regs {
 261         unsigned uniform_const : 8;
 262         unsigned reg2 : 6;
 263         unsigned reg3 : 6;
 264         unsigned reg0 : 5;
 265         unsigned reg1 : 6;
 266         unsigned ctrl : 4;
 267 };
 268
 269 enum bifrost_branch_cond {
 270         BR_COND_LT = 0,
 271         BR_COND_LE = 1,
 272         BR_COND_GE = 2,
 273         BR_COND_GT = 3,
 274         // Equal vs. not-equal determined by src0/src1 comparison
 275         BR_COND_EQ = 4,
 276         // floating-point comparisons
 277         // Becomes UNE when you flip the arguments
 278         BR_COND_OEQ = 5,
 279         // TODO what happens when you flip the arguments?
 280         BR_COND_OGT = 6,
 281         BR_COND_OLT = 7,
 282 };
 283
 284 enum bifrost_branch_code {
 285         BR_ALWAYS = 63,
 286 };
 287
 288 struct bifrost_branch {
 289         unsigned src0 : 3;
 290
 291         /* For BR_SIZE_ZERO, upper two bits become ctrl */
 292         unsigned src1 : 3;
 293
 294         /* Offset source -- always uniform/const but
 295          * theoretically could support indirect jumps? */
 296         unsigned src2 : 3;
 297
 298         enum bifrost_branch_cond cond : 3;
 299         enum branch_bit_size size : 3;
 300
 301         unsigned op : 5;
 302 };
 303
 304 #endif