| 1 |
/* |
| 2 |
* Copyright © 2025 Imagination Technologies Ltd. |
| 3 |
* |
| 4 |
* SPDX-License-Identifier: MIT |
| 5 |
*/ |
| 6 |
|
| 7 |
#ifndef PVR_USC_H |
| 8 |
#define PVR_USC_H |
| 9 |
|
| 10 |
/** |
| 11 |
* \file pvr_usc.h |
| 12 |
* |
| 13 |
* \brief USC internal shader generation header. |
| 14 |
*/ |
| 15 |
|
| 16 |
#include "common/pvr_iface.h" |
| 17 |
#include "compiler/shader_enums.h" |
| 18 |
#include "pco/pco.h" |
| 19 |
#include "pvr_common.h" |
| 20 |
#include "pvr_formats.h" |
| 21 |
|
| 22 |
struct pvr_load_op; |
| 23 |
|
| 24 |
/* EOT shader generation. */ |
| 25 |
struct pvr_eot_props { |
| 26 |
unsigned emit_count; |
| 27 |
|
| 28 |
bool shared_words; |
| 29 |
union { |
| 30 |
const uint32_t *state_words; |
| 31 |
const unsigned *state_regs; |
| 32 |
}; |
| 33 |
|
| 34 |
unsigned msaa_samples; |
| 35 |
unsigned num_output_regs; |
| 36 |
|
| 37 |
uint64_t tile_buffer_addrs[PVR_MAX_COLOR_ATTACHMENTS]; |
| 38 |
}; |
| 39 |
|
| 40 |
nir_shader *pvr_usc_fs_pfo_passthrough_nir (pco_ctx *ctx); |
| 41 |
|
| 42 |
pco_shader *pvr_usc_eot(pco_ctx *ctx, |
| 43 |
struct pvr_eot_props *props, |
| 44 |
const struct pvr_device_info *dev_info); |
| 45 |
|
| 46 |
/* Transfer queue shader generation. */ |
| 47 |
struct pvr_tq_props {}; |
| 48 |
|
| 49 |
pco_shader *pvr_usc_tq(pco_ctx *ctx, struct pvr_tq_props *props); |
| 50 |
|
| 51 |
/* Legacy transfer queue and loadop shader gen. */ |
| 52 |
enum pvr_int_coord_set_floats { |
| 53 |
PVR_INT_COORD_SET_FLOATS_0 = 0, |
| 54 |
PVR_INT_COORD_SET_FLOATS_4 = 1, |
| 55 |
/* For rate changes to 0 base screen space. */ |
| 56 |
PVR_INT_COORD_SET_FLOATS_6 = 2, |
| 57 |
PVR_INT_COORD_SET_FLOATS_NUM = 3 |
| 58 |
}; |
| 59 |
|
| 60 |
struct pvr_tq_shader_properties { |
| 61 |
/* Controls whether this is an iterated shader. */ |
| 62 |
bool iterated; |
| 63 |
|
| 64 |
/* Controls whether this is meant to be running at full rate. */ |
| 65 |
bool full_rate; |
| 66 |
|
| 67 |
/* Sample specific channel of pixel. */ |
| 68 |
bool pick_component; |
| 69 |
|
| 70 |
struct pvr_tq_layer_properties { |
| 71 |
/* Controls whether we need to send the sample count to the TPU. */ |
| 72 |
bool msaa; |
| 73 |
|
| 74 |
/* In case we run pixel rate, to do an USC resolve - but still in MSAA TPU |
| 75 |
* samples. |
| 76 |
*/ |
| 77 |
uint32_t sample_count; |
| 78 |
|
| 79 |
enum pvr_resolve_op resolve_op; |
| 80 |
|
| 81 |
/* Selects the pixel conversion that we have to perform. */ |
| 82 |
enum pvr_transfer_pbe_pixel_src pbe_format; |
| 83 |
|
| 84 |
/* Sampling from a 3D texture with a constant Z position. */ |
| 85 |
bool sample; |
| 86 |
|
| 87 |
/* Number of float coefficients to get from screen space to texture space. |
| 88 |
*/ |
| 89 |
enum pvr_int_coord_set_floats layer_floats; |
| 90 |
|
| 91 |
/* Unaligned texture address in bytes. */ |
| 92 |
uint32_t byte_unwind; |
| 93 |
|
| 94 |
/* Enable bilinear filter in shader. */ |
| 95 |
bool linear; |
| 96 |
} layer_props; |
| 97 |
}; |
| 98 |
|
| 99 |
/* All offsets are in dwords. */ |
| 100 |
/* Devices may have more than 256 sh regs but we're expecting to use vary few so |
| 101 |
* let's use uint8_t. |
| 102 |
*/ |
| 103 |
struct pvr_tq_frag_sh_reg_layout { |
| 104 |
struct { |
| 105 |
/* How many image sampler descriptors are present. */ |
| 106 |
uint8_t count; |
| 107 |
/* TODO: See if we ever need more than one combined image sampler |
| 108 |
* descriptor. If this is linked to the amount of layers used, we only |
| 109 |
* ever use one layer so this wouldn't need to be an array. |
| 110 |
*/ |
| 111 |
struct { |
| 112 |
uint8_t image; |
| 113 |
uint8_t sampler; |
| 114 |
} offsets[PVR_TRANSFER_MAX_IMAGES]; |
| 115 |
} combined_image_samplers; |
| 116 |
|
| 117 |
/* TODO: Dynamic consts are used for various things so do this properly by |
| 118 |
* having an actual layout instead of chucking them all together using an |
| 119 |
* implicit layout. |
| 120 |
*/ |
| 121 |
struct { |
| 122 |
/* How many dynamic consts regs have been allocated. */ |
| 123 |
uint8_t count; |
| 124 |
uint8_t offset; |
| 125 |
} dynamic_consts; |
| 126 |
|
| 127 |
/* Total sh regs allocated by the driver. It does not include the regs |
| 128 |
* necessary for compiler_out. |
| 129 |
*/ |
| 130 |
uint8_t driver_total; |
| 131 |
|
| 132 |
/* Provided by the compiler to the driver to be appended to the shareds. */ |
| 133 |
/* No offset field since these will be appended at the end so driver_total |
| 134 |
* can be used instead. |
| 135 |
*/ |
| 136 |
struct { |
| 137 |
struct { |
| 138 |
/* TODO: Remove this count and just use `compiler_out_total`? Or remove |
| 139 |
* that one and use this one? |
| 140 |
*/ |
| 141 |
uint8_t count; |
| 142 |
/* TODO: The array size is chosen arbitrarily based on the max |
| 143 |
* constants currently produced by the compiler. Make this dynamic? |
| 144 |
*/ |
| 145 |
/* Values to fill in into each shared reg used for usc constants. */ |
| 146 |
uint32_t values[10]; |
| 147 |
} usc_constants; |
| 148 |
} compiler_out; |
| 149 |
|
| 150 |
/* Total extra sh regs needed by the compiler that need to be appended to the |
| 151 |
* shareds by the driver. |
| 152 |
*/ |
| 153 |
uint8_t compiler_out_total; |
| 154 |
}; |
| 155 |
|
| 156 |
pco_shader *pvr_uscgen_tq(pco_ctx *ctx, |
| 157 |
const struct pvr_tq_shader_properties *shader_props, |
| 158 |
struct pvr_tq_frag_sh_reg_layout *sh_reg_layout); |
| 159 |
|
| 160 |
pco_shader *pvr_uscgen_loadop(pco_ctx *ctx, struct pvr_load_op *load_op); |
| 161 |
|
| 162 |
/* Clear attachment shader generation. */ |
| 163 |
struct pvr_clear_attach_props { |
| 164 |
unsigned dword_count; |
| 165 |
unsigned offset; |
| 166 |
bool uses_tile_buffer; |
| 167 |
}; |
| 168 |
|
| 169 |
pco_shader *pvr_uscgen_clear_attach(pco_ctx *ctx, |
| 170 |
struct pvr_clear_attach_props *props); |
| 171 |
|
| 172 |
#define INDEX(d_c, o, u_t_b, i) \ |
| 173 |
if (props->dword_count == d_c && props->offset == o && \ |
| 174 |
props->uses_tile_buffer == u_t_b) \ |
| 175 |
return i |
| 176 |
|
| 177 |
inline static unsigned |
| 178 |
pvr_uscgen_clear_attach_index(struct pvr_clear_attach_props *props) |
| 179 |
{ |
| 180 |
INDEX(1, 0, false, 0); |
| 181 |
INDEX(1, 1, false, 1); |
| 182 |
INDEX(1, 2, false, 2); |
| 183 |
INDEX(1, 3, false, 3); |
| 184 |
INDEX(2, 0, false, 4); |
| 185 |
INDEX(2, 1, false, 5); |
| 186 |
INDEX(2, 2, false, 6); |
| 187 |
INDEX(3, 0, false, 7); |
| 188 |
INDEX(3, 1, false, 8); |
| 189 |
INDEX(4, 0, false, 9); |
| 190 |
|
| 191 |
INDEX(1, 0, true, 10); |
| 192 |
INDEX(1, 1, true, 11); |
| 193 |
INDEX(1, 2, true, 12); |
| 194 |
INDEX(1, 3, true, 13); |
| 195 |
INDEX(2, 0, true, 14); |
| 196 |
INDEX(2, 1, true, 15); |
| 197 |
INDEX(2, 2, true, 16); |
| 198 |
INDEX(3, 0, true, 17); |
| 199 |
INDEX(3, 1, true, 18); |
| 200 |
INDEX(4, 0, true, 19); |
| 201 |
|
| 202 |
UNREACHABLE("Invalid clear attachment shader properties."); |
| 203 |
} |
| 204 |
#undef INDEX |
| 205 |
|
| 206 |
pco_shader * |
| 207 |
pvr_usc_zero_init_wg_mem(pco_ctx *ctx, unsigned start, unsigned count); |
| 208 |
|
| 209 |
/* SPM load shader generation. */ |
| 210 |
struct pvr_spm_load_props { |
| 211 |
unsigned output_reg_count; |
| 212 |
unsigned tile_buffer_count; |
| 213 |
bool is_multisampled; |
| 214 |
}; |
| 215 |
|
| 216 |
static inline unsigned pvr_uscgen_spm_buffer_data(unsigned buffer_index, |
| 217 |
bool addr) |
| 218 |
{ |
| 219 |
switch (buffer_index) { |
| 220 |
case 0: |
| 221 |
return addr ? PVR_SPM_LOAD_DATA_BUF_ADDR_0 : PVR_SPM_LOAD_DATA_BUF_TEX_0; |
| 222 |
|
| 223 |
case 1: |
| 224 |
return addr ? PVR_SPM_LOAD_DATA_BUF_ADDR_1 : PVR_SPM_LOAD_DATA_BUF_TEX_1; |
| 225 |
|
| 226 |
case 2: |
| 227 |
return addr ? PVR_SPM_LOAD_DATA_BUF_ADDR_2 : PVR_SPM_LOAD_DATA_BUF_TEX_2; |
| 228 |
|
| 229 |
case 3: |
| 230 |
return addr ? PVR_SPM_LOAD_DATA_BUF_ADDR_3 : PVR_SPM_LOAD_DATA_BUF_TEX_3; |
| 231 |
|
| 232 |
case 4: |
| 233 |
return addr ? PVR_SPM_LOAD_DATA_BUF_ADDR_4 : PVR_SPM_LOAD_DATA_BUF_TEX_4; |
| 234 |
|
| 235 |
case 5: |
| 236 |
return addr ? PVR_SPM_LOAD_DATA_BUF_ADDR_5 : PVR_SPM_LOAD_DATA_BUF_TEX_5; |
| 237 |
|
| 238 |
case 6: |
| 239 |
return addr ? PVR_SPM_LOAD_DATA_BUF_ADDR_6 : PVR_SPM_LOAD_DATA_BUF_TEX_6; |
| 240 |
|
| 241 |
default: |
| 242 |
break; |
| 243 |
} |
| 244 |
|
| 245 |
UNREACHABLE(""); |
| 246 |
} |
| 247 |
|
| 248 |
static inline unsigned |
| 249 |
pvr_uscgen_spm_load_data_size(struct pvr_spm_load_props *props) |
| 250 |
{ |
| 251 |
return PVR_SPM_LOAD_DATA_BUF_TEX_0 + |
| 252 |
props->tile_buffer_count * (ROGUE_NUM_TEXSTATE_DWORDS + |
| 253 |
(sizeof(uint64_t) / sizeof(uint32_t))); |
| 254 |
} |
| 255 |
|
| 256 |
pco_shader *pvr_uscgen_spm_load(pco_ctx *ctx, struct pvr_spm_load_props *props); |
| 257 |
|
| 258 |
#define INDEX(o_r_c, t_b_c, i_m, i) \ |
| 259 |
if (props->output_reg_count == o_r_c && \ |
| 260 |
props->tile_buffer_count == t_b_c && props->is_multisampled == i_m) \ |
| 261 |
return i |
| 262 |
|
| 263 |
inline static unsigned |
| 264 |
pvr_uscgen_spm_load_index(struct pvr_spm_load_props *props) |
| 265 |
{ |
| 266 |
INDEX(1, 0, false, 0); |
| 267 |
INDEX(2, 0, false, 1); |
| 268 |
INDEX(4, 0, false, 2); |
| 269 |
|
| 270 |
INDEX(4, 1, false, 3); |
| 271 |
INDEX(4, 2, false, 4); |
| 272 |
INDEX(4, 3, false, 5); |
| 273 |
INDEX(4, 4, false, 6); |
| 274 |
INDEX(4, 5, false, 7); |
| 275 |
INDEX(4, 6, false, 8); |
| 276 |
INDEX(4, 7, false, 9); |
| 277 |
|
| 278 |
INDEX(1, 0, true, 10); |
| 279 |
INDEX(2, 0, true, 11); |
| 280 |
INDEX(4, 0, true, 12); |
| 281 |
|
| 282 |
INDEX(4, 1, true, 13); |
| 283 |
INDEX(4, 2, true, 14); |
| 284 |
INDEX(4, 3, true, 15); |
| 285 |
INDEX(4, 4, true, 16); |
| 286 |
INDEX(4, 5, true, 17); |
| 287 |
INDEX(4, 6, true, 18); |
| 288 |
INDEX(4, 7, true, 19); |
| 289 |
|
| 290 |
UNREACHABLE("Invalid SPM load shader properties."); |
| 291 |
} |
| 292 |
#undef INDEX |
| 293 |
|
| 294 |
#define PVR_NUM_SPM_LOAD_SHADERS 20U |
| 295 |
|
| 296 |
#endif /* PVR_USC_H */ |