src/mesa/drivers/dri/i965/brw_lower_texture_gradients.cpp

   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
  21  * DEALINGS IN THE SOFTWARE.
  22  */
  23
  24 /**
  25  * \file brw_lower_texture_gradients.cpp
  26  */
  27
  28 #include "compiler/glsl/ir.h"
  29 #include "compiler/glsl/ir_builder.h"
  30 #include "program/prog_instruction.h"
  31 #include "brw_context.h"
  32
  33 using namespace ir_builder;
  34
  35 class lower_texture_grad_visitor : public ir_hierarchical_visitor {
  36 public:
  37    lower_texture_grad_visitor(bool has_sample_d_c)
  38       : has_sample_d_c(has_sample_d_c)
  39    {
  40       progress = false;
  41    }
  42
  43    ir_visitor_status visit_leave(ir_texture *ir);
  44
  45
  46    bool progress;
  47    bool has_sample_d_c;
  48
  49 private:
  50    void emit(ir_variable *, ir_rvalue *);
  51    ir_variable *temp(void *ctx, const glsl_type *type, const char *name);
  52 };
  53
  54 /**
  55  * Emit a variable declaration and an assignment to initialize it.
  56  */
  57 void
  58 lower_texture_grad_visitor::emit(ir_variable *var, ir_rvalue *value)
  59 {
  60    base_ir->insert_before(var);
  61    base_ir->insert_before(assign(var, value));
  62 }
  63
  64 /**
  65  * Emit a temporary variable declaration
  66  */
  67 ir_variable *
  68 lower_texture_grad_visitor::temp(void *ctx, const glsl_type *type, const char *name)
  69 {
  70    ir_variable *var = new(ctx) ir_variable(type, name, ir_var_temporary);
  71    base_ir->insert_before(var);
  72    return var;
  73 }
  74
  75 static const glsl_type *
  76 txs_type(const glsl_type *type)
  77 {
  78    unsigned dims;
  79    switch (type->sampler_dimensionality) {
  80    case GLSL_SAMPLER_DIM_1D:
  81       dims = 1;
  82       break;
  83    case GLSL_SAMPLER_DIM_2D:
  84    case GLSL_SAMPLER_DIM_RECT:
  85    case GLSL_SAMPLER_DIM_CUBE:
  86       dims = 2;
  87       break;
  88    case GLSL_SAMPLER_DIM_3D:
  89       dims = 3;
  90       break;
  91    default:
  92       unreachable("Should not get here: invalid sampler dimensionality");
  93    }
  94
  95    if (type->sampler_array)
  96       dims++;
  97
  98    return glsl_type::get_instance(GLSL_TYPE_INT, dims, 1);
  99 }
 100
 101 ir_visitor_status
 102 lower_texture_grad_visitor::visit_leave(ir_texture *ir)
 103 {
 104    /* Only lower textureGrad with cube maps or shadow samplers */
 105    if (ir->op != ir_txd ||
 106       (ir->sampler->type->sampler_dimensionality != GLSL_SAMPLER_DIM_CUBE &&
 107        !ir->shadow_comparitor))
 108       return visit_continue;
 109
 110    /* Lower textureGrad() with samplerCube* even if we have the sample_d_c
 111     * message.  GLSL provides gradients for the 'r' coordinate.  Unfortunately:
 112     *
 113     * From the Ivybridge PRM, Volume 4, Part 1, sample_d message description:
 114     * "The r coordinate contains the faceid, and the r gradients are ignored
 115     *  by hardware."
 116     */
 117    bool need_lowering = !has_sample_d_c ||
 118       ir->sampler->type->sampler_dimensionality == GLSL_SAMPLER_DIM_CUBE;
 119
 120    if (!need_lowering)
 121       return visit_continue;
 122
 123    void *mem_ctx = ralloc_parent(ir);
 124
 125    const glsl_type *grad_type = ir->lod_info.grad.dPdx->type;
 126
 127    /* Use textureSize() to get the width and height of LOD 0; swizzle away
 128     * the depth/number of array slices.
 129     */
 130    ir_texture *txs = new(mem_ctx) ir_texture(ir_txs);
 131    txs->set_sampler(ir->sampler->clone(mem_ctx, NULL),
 132                     txs_type(ir->sampler->type));
 133    txs->lod_info.lod = new(mem_ctx) ir_constant(0);
 134    ir_variable *size =
 135       new(mem_ctx) ir_variable(grad_type, "size", ir_var_temporary);
 136    if (ir->sampler->type->sampler_dimensionality == GLSL_SAMPLER_DIM_CUBE) {
 137       base_ir->insert_before(size);
 138       base_ir->insert_before(assign(size,
 139                                     swizzle_for_size(expr(ir_unop_i2f, txs), 2),
 140                                     WRITEMASK_XY));
 141       base_ir->insert_before(assign(size, new(mem_ctx) ir_constant(1.0f), WRITEMASK_Z));
 142    } else {
 143       emit(size, expr(ir_unop_i2f,
 144                       swizzle_for_size(txs, grad_type->vector_elements)));
 145    }
 146
 147    /* Scale the gradients by width and height.  Effectively, the incoming
 148     * gradients are s'(x,y), t'(x,y), and r'(x,y) from equation 3.19 in the
 149     * GL 3.0 spec; we want u'(x,y), which is w_t * s'(x,y).
 150     */
 151    ir_variable *dPdx =
 152       new(mem_ctx) ir_variable(grad_type, "dPdx", ir_var_temporary);
 153    emit(dPdx, mul(size, ir->lod_info.grad.dPdx));
 154
 155    ir_variable *dPdy =
 156       new(mem_ctx) ir_variable(grad_type, "dPdy", ir_var_temporary);
 157    emit(dPdy, mul(size, ir->lod_info.grad.dPdy));
 158
 159    ir->op = ir_txl;
 160    if (ir->sampler->type->sampler_dimensionality == GLSL_SAMPLER_DIM_CUBE) {
 161       /* Cubemap texture lookups first generate a texture coordinate normalized
 162        * to [-1, 1] on the appropiate face. The appropiate face is determined
 163        * by which component has largest magnitude and its sign. The texture
 164        * coordinate is the quotient of the remaining texture coordinates against
 165        * that absolute value of the component of largest magnitude. This
 166        * division requires that the computing of the derivative of the texel
 167        * coordinate must use the quotient rule. The high level GLSL code is as
 168        * follows:
 169        *
 170        * Step 1: selection
 171        *
 172        * vec3 abs_p, Q, dQdx, dQdy;
 173        * abs_p = abs(ir->coordinate);
 174        * if (abs_p.x >= max(abs_p.y, abs_p.z)) {
 175        *    Q = ir->coordinate.yzx;
 176        *    dQdx = ir->lod_info.grad.dPdx.yzx;
 177        *    dQdy = ir->lod_info.grad.dPdy.yzx;
 178        * }
 179        * if (abs_p.y >= max(abs_p.x, abs_p.z)) {
 180        *    Q = ir->coordinate.xzy;
 181        *    dQdx = ir->lod_info.grad.dPdx.xzy;
 182        *    dQdy = ir->lod_info.grad.dPdy.xzy;
 183        * }
 184        * if (abs_p.z >= max(abs_p.x, abs_p.y)) {
 185        *    Q = ir->coordinate;
 186        *    dQdx = ir->lod_info.grad.dPdx;
 187        *    dQdy = ir->lod_info.grad.dPdy;
 188        * }
 189        *
 190        * Step 2: use quotient rule to compute derivative. The normalized to
 191        * [-1, 1] texel coordinate is given by Q.xy / (sign(Q.z) * Q.z). We are
 192        * only concerned with the magnitudes of the derivatives whose values are
 193        * not affected by the sign. We drop the sign from the computation.
 194        *
 195        * vec2 dx, dy;
 196        * float recip;
 197        *
 198        * recip = 1.0 / Q.z;
 199        * dx = recip * ( dQdx.xy - Q.xy * (dQdx.z * recip) );
 200        * dy = recip * ( dQdy.xy - Q.xy * (dQdy.z * recip) );
 201        *
 202        * Step 3: compute LOD. At this point we have the derivatives of the
 203        * texture coordinates normalized to [-1,1]. We take the LOD to be
 204        *  result = log2(max(sqrt(dot(dx, dx)), sqrt(dy, dy)) * 0.5 * L)
 205        *         = -1.0 + log2(max(sqrt(dot(dx, dx)), sqrt(dy, dy)) * L)
 206        *         = -1.0 + log2(sqrt(max(dot(dx, dx), dot(dy,dy))) * L)
 207        *         = -1.0 + log2(sqrt(L * L * max(dot(dx, dx), dot(dy,dy))))
 208        *         = -1.0 + 0.5 * log2(L * L * max(dot(dx, dx), dot(dy,dy)))
 209        * where L is the dimension of the cubemap. The code is:
 210        *
 211        * float M, result;
 212        * M = max(dot(dx, dx), dot(dy, dy));
 213        * L = textureSize(sampler, 0).x;
 214        * result = -1.0 + 0.5 * log2(L * L * M);
 215        */
 216
 217 /* Helpers to make code more human readable. */
 218 #define EMIT(instr) base_ir->insert_before(instr)
 219 #define THEN(irif, instr) irif->then_instructions.push_tail(instr)
 220 #define CLONE(x) x->clone(mem_ctx, NULL)
 221
 222       ir_variable *abs_p = temp(mem_ctx, glsl_type::vec3_type, "abs_p");
 223
 224       EMIT(assign(abs_p, swizzle_for_size(abs(CLONE(ir->coordinate)), 3)));
 225
 226       ir_variable *Q = temp(mem_ctx, glsl_type::vec3_type, "Q");
 227       ir_variable *dQdx = temp(mem_ctx, glsl_type::vec3_type, "dQdx");
 228       ir_variable *dQdy = temp(mem_ctx, glsl_type::vec3_type, "dQdy");
 229
 230       /* unmodified dPdx, dPdy values */
 231       ir_rvalue *dPdx = ir->lod_info.grad.dPdx;
 232       ir_rvalue *dPdy = ir->lod_info.grad.dPdy;
 233
 234       /* 1. compute selector */
 235
 236       /* if (abs_p.x >= max(abs_p.y, abs_p.z))  ... */
 237       ir_if *branch_x =
 238          new(mem_ctx) ir_if(gequal(swizzle_x(abs_p),
 239                                    max2(swizzle_y(abs_p), swizzle_z(abs_p))));
 240
 241       /* Q = p.yzx;
 242        * dQdx = dPdx.yzx;
 243        * dQdy = dPdy.yzx;
 244        */
 245       int yzx = MAKE_SWIZZLE4(SWIZZLE_Y, SWIZZLE_Z, SWIZZLE_X, 0);
 246       THEN(branch_x, assign(Q, swizzle(CLONE(ir->coordinate), yzx, 3)));
 247       THEN(branch_x, assign(dQdx, swizzle(CLONE(dPdx), yzx, 3)));
 248       THEN(branch_x, assign(dQdy, swizzle(CLONE(dPdy), yzx, 3)));
 249       EMIT(branch_x);
 250
 251       /* if (abs_p.y >= max(abs_p.x, abs_p.z)) */
 252       ir_if *branch_y =
 253          new(mem_ctx) ir_if(gequal(swizzle_y(abs_p),
 254                                    max2(swizzle_x(abs_p), swizzle_z(abs_p))));
 255
 256       /* Q = p.xzy;
 257        * dQdx = dPdx.xzy;
 258        * dQdy = dPdy.xzy;
 259        */
 260       int xzy = MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Z, SWIZZLE_Y, 0);
 261       THEN(branch_y, assign(Q, swizzle(CLONE(ir->coordinate), xzy, 3)));
 262       THEN(branch_y, assign(dQdx, swizzle(CLONE(dPdx), xzy, 3)));
 263       THEN(branch_y, assign(dQdy, swizzle(CLONE(dPdy), xzy, 3)));
 264       EMIT(branch_y);
 265
 266       /* if (abs_p.z >= max(abs_p.x, abs_p.y)) */
 267       ir_if *branch_z =
 268          new(mem_ctx) ir_if(gequal(swizzle_z(abs_p),
 269                             max2(swizzle_x(abs_p), swizzle_y(abs_p))));
 270
 271       /* Q = p;
 272        * dQdx = dPdx;
 273        * dQdy = dPdy;
 274        */
 275       THEN(branch_z, assign(Q, swizzle_for_size(CLONE(ir->coordinate), 3)));
 276       THEN(branch_z, assign(dQdx, CLONE(dPdx)));
 277       THEN(branch_z, assign(dQdy, CLONE(dPdy)));
 278       EMIT(branch_z);
 279
 280       /* 2. quotient rule */
 281       ir_variable *recip = temp(mem_ctx, glsl_type::float_type, "recip");
 282       EMIT(assign(recip, expr(ir_unop_rcp, swizzle_z(Q))));
 283
 284       ir_variable *dx = temp(mem_ctx, glsl_type::vec2_type, "dx");
 285       ir_variable *dy = temp(mem_ctx, glsl_type::vec2_type, "dy");
 286
 287       /* tmp = Q.xy * recip;
 288        * dx = recip * ( dQdx.xy - (tmp * dQdx.z) );
 289        * dy = recip * ( dQdy.xy - (tmp * dQdy.z) );
 290        */
 291       ir_variable *tmp = temp(mem_ctx, glsl_type::vec2_type, "tmp");
 292       EMIT(assign(tmp, mul(swizzle_xy(Q), recip)));
 293       EMIT(assign(dx, mul(recip, sub(swizzle_xy(dQdx),
 294                                      mul(tmp, swizzle_z(dQdx))))));
 295       EMIT(assign(dy, mul(recip, sub(swizzle_xy(dQdy),
 296                                      mul(tmp, swizzle_z(dQdy))))));
 297
 298       /* M = max(dot(dx, dx), dot(dy, dy)); */
 299       ir_variable *M = temp(mem_ctx, glsl_type::float_type, "M");
 300       EMIT(assign(M, max2(dot(dx, dx), dot(dy, dy))));
 301
 302       /* size has textureSize() of LOD 0 */
 303       ir_variable *L = temp(mem_ctx, glsl_type::float_type, "L");
 304       EMIT(assign(L, swizzle_x(size)));
 305
 306       ir_variable *result = temp(mem_ctx, glsl_type::float_type, "result");
 307
 308       /* result = -1.0 + 0.5 * log2(L * L * M); */
 309       EMIT(assign(result,
 310                   add(new(mem_ctx)ir_constant(-1.0f),
 311                       mul(new(mem_ctx)ir_constant(0.5f),
 312                           expr(ir_unop_log2, mul(mul(L, L), M))))));
 313
 314       /* 3. final assignment of parameters to textureLod call */
 315       ir->lod_info.lod = new (mem_ctx) ir_dereference_variable(result);
 316
 317 #undef THEN
 318 #undef EMIT
 319
 320    } else {
 321       /* Calculate rho from equation 3.20 of the GL 3.0 specification. */
 322       ir_rvalue *rho;
 323       if (dPdx->type->is_scalar()) {
 324          rho = expr(ir_binop_max, expr(ir_unop_abs, dPdx),
 325                     expr(ir_unop_abs, dPdy));
 326       } else {
 327          rho = expr(ir_binop_max, expr(ir_unop_sqrt, dot(dPdx, dPdx)),
 328                     expr(ir_unop_sqrt, dot(dPdy, dPdy)));
 329       }
 330
 331       /* lambda_base = log2(rho).  We're ignoring GL state biases for now. */
 332       ir->lod_info.lod = expr(ir_unop_log2, rho);
 333    }
 334
 335    progress = true;
 336    return visit_continue;
 337 }
 338
 339 extern "C" {
 340
 341 bool
 342 brw_lower_texture_gradients(struct brw_context *brw,
 343                             struct exec_list *instructions)
 344 {
 345    bool has_sample_d_c = brw->gen >= 8 || brw->is_haswell;
 346    lower_texture_grad_visitor v(has_sample_d_c);
 347
 348    visit_list_elements(&v, instructions);
 349
 350    return v.progress;
 351 }
 352
 353 }