• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright © 2016 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 #ifndef BLORP_GENX_EXEC_H
25 #define BLORP_GENX_EXEC_H
26 
27 #include "blorp_priv.h"
28 #include "common/gen_device_info.h"
29 #include "common/gen_sample_positions.h"
30 #include "genxml/gen_macros.h"
31 
32 /**
33  * This file provides the blorp pipeline setup and execution functionality.
34  * It defines the following function:
35  *
36  * static void
37  * blorp_exec(struct blorp_context *blorp, void *batch_data,
38  *            const struct blorp_params *params);
39  *
40  * It is the job of whoever includes this header to wrap this in something
41  * to get an externally visible symbol.
42  *
43  * In order for the blorp_exec function to work, the driver must provide
44  * implementations of the following static helper functions.
45  */
46 
47 static void *
48 blorp_emit_dwords(struct blorp_batch *batch, unsigned n);
49 
50 static uint64_t
51 blorp_emit_reloc(struct blorp_batch *batch,
52                  void *location, struct blorp_address address, uint32_t delta);
53 
54 static void *
55 blorp_alloc_dynamic_state(struct blorp_batch *batch,
56                           uint32_t size,
57                           uint32_t alignment,
58                           uint32_t *offset);
59 static void *
60 blorp_alloc_vertex_buffer(struct blorp_batch *batch, uint32_t size,
61                           struct blorp_address *addr);
62 
63 #if GEN_GEN >= 8
64 static struct blorp_address
65 blorp_get_workaround_page(struct blorp_batch *batch);
66 #endif
67 
68 static void
69 blorp_alloc_binding_table(struct blorp_batch *batch, unsigned num_entries,
70                           unsigned state_size, unsigned state_alignment,
71                           uint32_t *bt_offset, uint32_t *surface_offsets,
72                           void **surface_maps);
73 
74 static void
75 blorp_flush_range(struct blorp_batch *batch, void *start, size_t size);
76 
77 static void
78 blorp_surface_reloc(struct blorp_batch *batch, uint32_t ss_offset,
79                     struct blorp_address address, uint32_t delta);
80 
81 #if GEN_GEN >= 7
82 static struct blorp_address
83 blorp_get_surface_base_address(struct blorp_batch *batch);
84 #endif
85 
86 static void
87 blorp_emit_urb_config(struct blorp_batch *batch,
88                       unsigned vs_entry_size, unsigned sf_entry_size);
89 
90 static void
91 blorp_emit_pipeline(struct blorp_batch *batch,
92                     const struct blorp_params *params);
93 
94 /***** BEGIN blorp_exec implementation ******/
95 
96 static uint64_t
_blorp_combine_address(struct blorp_batch * batch,void * location,struct blorp_address address,uint32_t delta)97 _blorp_combine_address(struct blorp_batch *batch, void *location,
98                        struct blorp_address address, uint32_t delta)
99 {
100    if (address.buffer == NULL) {
101       return address.offset + delta;
102    } else {
103       return blorp_emit_reloc(batch, location, address, delta);
104    }
105 }
106 
107 #define __gen_address_type struct blorp_address
108 #define __gen_user_data struct blorp_batch
109 #define __gen_combine_address _blorp_combine_address
110 
111 #include "genxml/genX_pack.h"
112 
113 #define _blorp_cmd_length(cmd) cmd ## _length
114 #define _blorp_cmd_length_bias(cmd) cmd ## _length_bias
115 #define _blorp_cmd_header(cmd) cmd ## _header
116 #define _blorp_cmd_pack(cmd) cmd ## _pack
117 
118 #define blorp_emit(batch, cmd, name)                              \
119    for (struct cmd name = { _blorp_cmd_header(cmd) },             \
120         *_dst = blorp_emit_dwords(batch, _blorp_cmd_length(cmd)); \
121         __builtin_expect(_dst != NULL, 1);                        \
122         _blorp_cmd_pack(cmd)(batch, (void *)_dst, &name),         \
123         _dst = NULL)
124 
125 #define blorp_emitn(batch, cmd, n) ({                       \
126       uint32_t *_dw = blorp_emit_dwords(batch, n);          \
127       if (_dw) {                                            \
128          struct cmd template = {                            \
129             _blorp_cmd_header(cmd),                         \
130             .DWordLength = n - _blorp_cmd_length_bias(cmd), \
131          };                                                 \
132          _blorp_cmd_pack(cmd)(batch, _dw, &template);       \
133       }                                                     \
134       _dw ? _dw + 1 : NULL; /* Array starts at dw[1] */     \
135    })
136 
137 #define STRUCT_ZERO(S) ({ struct S t; memset(&t, 0, sizeof(t)); t; })
138 
139 #define blorp_emit_dynamic(batch, state, name, align, offset)      \
140    for (struct state name = STRUCT_ZERO(state),                         \
141         *_dst = blorp_alloc_dynamic_state(batch,                   \
142                                           _blorp_cmd_length(state) * 4, \
143                                           align, offset);               \
144         __builtin_expect(_dst != NULL, 1);                              \
145         _blorp_cmd_pack(state)(batch, (void *)_dst, &name),             \
146         blorp_flush_range(batch, _dst, _blorp_cmd_length(state) * 4),   \
147         _dst = NULL)
148 
149 /* 3DSTATE_URB
150  * 3DSTATE_URB_VS
151  * 3DSTATE_URB_HS
152  * 3DSTATE_URB_DS
153  * 3DSTATE_URB_GS
154  *
155  * Assign the entire URB to the VS. Even though the VS disabled, URB space
156  * is still needed because the clipper loads the VUE's from the URB. From
157  * the Sandybridge PRM, Volume 2, Part 1, Section 3DSTATE,
158  * Dword 1.15:0 "VS Number of URB Entries":
159  *     This field is always used (even if VS Function Enable is DISABLED).
160  *
161  * The warning below appears in the PRM (Section 3DSTATE_URB), but we can
162  * safely ignore it because this batch contains only one draw call.
163  *     Because of URB corruption caused by allocating a previous GS unit
164  *     URB entry to the VS unit, software is required to send a “GS NULL
165  *     Fence” (Send URB fence with VS URB size == 1 and GS URB size == 0)
166  *     plus a dummy DRAW call before any case where VS will be taking over
167  *     GS URB space.
168  *
169  * If the 3DSTATE_URB_VS is emitted, than the others must be also.
170  * From the Ivybridge PRM, Volume 2 Part 1, section 1.7.1 3DSTATE_URB_VS:
171  *
172  *     3DSTATE_URB_HS, 3DSTATE_URB_DS, and 3DSTATE_URB_GS must also be
173  *     programmed in order for the programming of this state to be
174  *     valid.
175  */
176 static void
emit_urb_config(struct blorp_batch * batch,const struct blorp_params * params)177 emit_urb_config(struct blorp_batch *batch,
178                 const struct blorp_params *params)
179 {
180    /* Once vertex fetcher has written full VUE entries with complete
181     * header the space requirement is as follows per vertex (in bytes):
182     *
183     *     Header    Position    Program constants
184     *   +--------+------------+-------------------+
185     *   |   16   |     16     |      n x 16       |
186     *   +--------+------------+-------------------+
187     *
188     * where 'n' stands for number of varying inputs expressed as vec4s.
189     */
190     const unsigned num_varyings =
191        params->wm_prog_data ? params->wm_prog_data->num_varying_inputs : 0;
192     const unsigned total_needed = 16 + 16 + num_varyings * 16;
193 
194    /* The URB size is expressed in units of 64 bytes (512 bits) */
195    const unsigned vs_entry_size = DIV_ROUND_UP(total_needed, 64);
196 
197    const unsigned sf_entry_size =
198       params->sf_prog_data ? params->sf_prog_data->urb_entry_size : 0;
199 
200    blorp_emit_urb_config(batch, vs_entry_size, sf_entry_size);
201 }
202 
203 static void
blorp_emit_vertex_data(struct blorp_batch * batch,const struct blorp_params * params,struct blorp_address * addr,uint32_t * size)204 blorp_emit_vertex_data(struct blorp_batch *batch,
205                        const struct blorp_params *params,
206                        struct blorp_address *addr,
207                        uint32_t *size)
208 {
209    const float vertices[] = {
210       /* v0 */ (float)params->x1, (float)params->y1, params->z,
211       /* v1 */ (float)params->x0, (float)params->y1, params->z,
212       /* v2 */ (float)params->x0, (float)params->y0, params->z,
213    };
214 
215    void *data = blorp_alloc_vertex_buffer(batch, sizeof(vertices), addr);
216    memcpy(data, vertices, sizeof(vertices));
217    *size = sizeof(vertices);
218    blorp_flush_range(batch, data, *size);
219 }
220 
221 static void
blorp_emit_input_varying_data(struct blorp_batch * batch,const struct blorp_params * params,struct blorp_address * addr,uint32_t * size)222 blorp_emit_input_varying_data(struct blorp_batch *batch,
223                               const struct blorp_params *params,
224                               struct blorp_address *addr,
225                               uint32_t *size)
226 {
227    const unsigned vec4_size_in_bytes = 4 * sizeof(float);
228    const unsigned max_num_varyings =
229       DIV_ROUND_UP(sizeof(params->wm_inputs), vec4_size_in_bytes);
230    const unsigned num_varyings =
231       params->wm_prog_data ? params->wm_prog_data->num_varying_inputs : 0;
232 
233    *size = 16 + num_varyings * vec4_size_in_bytes;
234 
235    const uint32_t *const inputs_src = (const uint32_t *)&params->wm_inputs;
236    void *data = blorp_alloc_vertex_buffer(batch, *size, addr);
237    uint32_t *inputs = data;
238 
239    /* Copy in the VS inputs */
240    assert(sizeof(params->vs_inputs) == 16);
241    memcpy(inputs, &params->vs_inputs, sizeof(params->vs_inputs));
242    inputs += 4;
243 
244    if (params->wm_prog_data) {
245       /* Walk over the attribute slots, determine if the attribute is used by
246        * the program and when necessary copy the values from the input storage
247        * to the vertex data buffer.
248        */
249       for (unsigned i = 0; i < max_num_varyings; i++) {
250          const gl_varying_slot attr = VARYING_SLOT_VAR0 + i;
251 
252          const int input_index = params->wm_prog_data->urb_setup[attr];
253          if (input_index < 0)
254             continue;
255 
256          memcpy(inputs, inputs_src + i * 4, vec4_size_in_bytes);
257 
258          inputs += 4;
259       }
260    }
261 
262    blorp_flush_range(batch, data, *size);
263 }
264 
265 static void
blorp_emit_vertex_buffers(struct blorp_batch * batch,const struct blorp_params * params)266 blorp_emit_vertex_buffers(struct blorp_batch *batch,
267                           const struct blorp_params *params)
268 {
269    struct GENX(VERTEX_BUFFER_STATE) vb[2];
270    memset(vb, 0, sizeof(vb));
271 
272    uint32_t size;
273    blorp_emit_vertex_data(batch, params, &vb[0].BufferStartingAddress, &size);
274    vb[0].VertexBufferIndex = 0;
275    vb[0].BufferPitch = 3 * sizeof(float);
276 #if GEN_GEN >= 6
277    vb[0].VertexBufferMOCS = vb[0].BufferStartingAddress.mocs;
278 #endif
279 #if GEN_GEN >= 7
280    vb[0].AddressModifyEnable = true;
281 #endif
282 #if GEN_GEN >= 8
283    vb[0].BufferSize = size;
284 #elif GEN_GEN >= 5
285    vb[0].BufferAccessType = VERTEXDATA;
286    vb[0].EndAddress = vb[0].BufferStartingAddress;
287    vb[0].EndAddress.offset += size - 1;
288 #elif GEN_GEN == 4
289    vb[0].BufferAccessType = VERTEXDATA;
290    vb[0].MaxIndex = 2;
291 #endif
292 
293    blorp_emit_input_varying_data(batch, params,
294                                  &vb[1].BufferStartingAddress, &size);
295    vb[1].VertexBufferIndex = 1;
296    vb[1].BufferPitch = 0;
297 #if GEN_GEN >= 6
298    vb[1].VertexBufferMOCS = vb[1].BufferStartingAddress.mocs;
299 #endif
300 #if GEN_GEN >= 7
301    vb[1].AddressModifyEnable = true;
302 #endif
303 #if GEN_GEN >= 8
304    vb[1].BufferSize = size;
305 #elif GEN_GEN >= 5
306    vb[1].BufferAccessType = INSTANCEDATA;
307    vb[1].EndAddress = vb[1].BufferStartingAddress;
308    vb[1].EndAddress.offset += size - 1;
309 #elif GEN_GEN == 4
310    vb[1].BufferAccessType = INSTANCEDATA;
311    vb[1].MaxIndex = 0;
312 #endif
313 
314    const unsigned num_dwords = 1 + GENX(VERTEX_BUFFER_STATE_length) * 2;
315    uint32_t *dw = blorp_emitn(batch, GENX(3DSTATE_VERTEX_BUFFERS), num_dwords);
316    if (!dw)
317       return;
318 
319    for (unsigned i = 0; i < 2; i++) {
320       GENX(VERTEX_BUFFER_STATE_pack)(batch, dw, &vb[i]);
321       dw += GENX(VERTEX_BUFFER_STATE_length);
322    }
323 }
324 
325 static void
blorp_emit_vertex_elements(struct blorp_batch * batch,const struct blorp_params * params)326 blorp_emit_vertex_elements(struct blorp_batch *batch,
327                            const struct blorp_params *params)
328 {
329    const unsigned num_varyings =
330       params->wm_prog_data ? params->wm_prog_data->num_varying_inputs : 0;
331    bool need_ndc = batch->blorp->compiler->devinfo->gen <= 5;
332    const unsigned num_elements = 2 + need_ndc + num_varyings;
333 
334    struct GENX(VERTEX_ELEMENT_STATE) ve[num_elements];
335    memset(ve, 0, num_elements * sizeof(*ve));
336 
337    /* Setup VBO for the rectangle primitive..
338     *
339     * A rectangle primitive (3DPRIM_RECTLIST) consists of only three
340     * vertices. The vertices reside in screen space with DirectX
341     * coordinates (that is, (0, 0) is the upper left corner).
342     *
343     *   v2 ------ implied
344     *    |        |
345     *    |        |
346     *   v1 ----- v0
347     *
348     * Since the VS is disabled, the clipper loads each VUE directly from
349     * the URB. This is controlled by the 3DSTATE_VERTEX_BUFFERS and
350     * 3DSTATE_VERTEX_ELEMENTS packets below. The VUE contents are as follows:
351     *   dw0: Reserved, MBZ.
352     *   dw1: Render Target Array Index. Below vertex fetcher gets programmed
353     *        to assign this with primitive instance identifier which will be
354     *        used for layered clears. All other renders have only one instance
355     *        and therefore the value will be effectively zero.
356     *   dw2: Viewport Index. The HiZ op disables viewport mapping and
357     *        scissoring, so set the dword to 0.
358     *   dw3: Point Width: The HiZ op does not emit the POINTLIST primitive,
359     *        so set the dword to 0.
360     *   dw4: Vertex Position X.
361     *   dw5: Vertex Position Y.
362     *   dw6: Vertex Position Z.
363     *   dw7: Vertex Position W.
364     *
365     *   dw8: Flat vertex input 0
366     *   dw9: Flat vertex input 1
367     *   ...
368     *   dwn: Flat vertex input n - 8
369     *
370     * For details, see the Sandybridge PRM, Volume 2, Part 1, Section 1.5.1
371     * "Vertex URB Entry (VUE) Formats".
372     *
373     * Only vertex position X and Y are going to be variable, Z is fixed to
374     * zero and W to one. Header words dw0,2,3 are zero. There is no need to
375     * include the fixed values in the vertex buffer. Vertex fetcher can be
376     * instructed to fill vertex elements with constant values of one and zero
377     * instead of reading them from the buffer.
378     * Flat inputs are program constants that are not interpolated. Moreover
379     * their values will be the same between vertices.
380     *
381     * See the vertex element setup below.
382     */
383    unsigned slot = 0;
384 
385    ve[slot] = (struct GENX(VERTEX_ELEMENT_STATE)) {
386       .VertexBufferIndex = 1,
387       .Valid = true,
388       .SourceElementFormat = (enum GENX(SURFACE_FORMAT)) ISL_FORMAT_R32G32B32A32_FLOAT,
389       .SourceElementOffset = 0,
390       .Component0Control = VFCOMP_STORE_SRC,
391 
392       /* From Gen8 onwards hardware is no more instructed to overwrite
393        * components using an element specifier. Instead one has separate
394        * 3DSTATE_VF_SGVS (System Generated Value Setup) state packet for it.
395        */
396 #if GEN_GEN >= 8
397       .Component1Control = VFCOMP_STORE_0,
398 #elif GEN_GEN >= 5
399       .Component1Control = VFCOMP_STORE_IID,
400 #else
401       .Component1Control = VFCOMP_STORE_0,
402 #endif
403       .Component2Control = VFCOMP_STORE_0,
404       .Component3Control = VFCOMP_STORE_0,
405 #if GEN_GEN <= 5
406       .DestinationElementOffset = slot * 4,
407 #endif
408    };
409    slot++;
410 
411 #if GEN_GEN <= 5
412    /* On Iron Lake and earlier, a native device coordinates version of the
413     * position goes right after the normal VUE header and before position.
414     * Since w == 1 for all of our coordinates, this is just a copy of the
415     * position.
416     */
417    ve[slot] = (struct GENX(VERTEX_ELEMENT_STATE)) {
418       .VertexBufferIndex = 0,
419       .Valid = true,
420       .SourceElementFormat = (enum GENX(SURFACE_FORMAT)) ISL_FORMAT_R32G32B32_FLOAT,
421       .SourceElementOffset = 0,
422       .Component0Control = VFCOMP_STORE_SRC,
423       .Component1Control = VFCOMP_STORE_SRC,
424       .Component2Control = VFCOMP_STORE_SRC,
425       .Component3Control = VFCOMP_STORE_1_FP,
426       .DestinationElementOffset = slot * 4,
427    };
428    slot++;
429 #endif
430 
431    ve[slot] = (struct GENX(VERTEX_ELEMENT_STATE)) {
432       .VertexBufferIndex = 0,
433       .Valid = true,
434       .SourceElementFormat = (enum GENX(SURFACE_FORMAT)) ISL_FORMAT_R32G32B32_FLOAT,
435       .SourceElementOffset = 0,
436       .Component0Control = VFCOMP_STORE_SRC,
437       .Component1Control = VFCOMP_STORE_SRC,
438       .Component2Control = VFCOMP_STORE_SRC,
439       .Component3Control = VFCOMP_STORE_1_FP,
440 #if GEN_GEN <= 5
441       .DestinationElementOffset = slot * 4,
442 #endif
443    };
444    slot++;
445 
446    for (unsigned i = 0; i < num_varyings; ++i) {
447       ve[slot] = (struct GENX(VERTEX_ELEMENT_STATE)) {
448          .VertexBufferIndex = 1,
449          .Valid = true,
450          .SourceElementFormat = (enum GENX(SURFACE_FORMAT)) ISL_FORMAT_R32G32B32A32_FLOAT,
451          .SourceElementOffset = 16 + i * 4 * sizeof(float),
452          .Component0Control = VFCOMP_STORE_SRC,
453          .Component1Control = VFCOMP_STORE_SRC,
454          .Component2Control = VFCOMP_STORE_SRC,
455          .Component3Control = VFCOMP_STORE_SRC,
456 #if GEN_GEN <= 5
457          .DestinationElementOffset = slot * 4,
458 #endif
459       };
460       slot++;
461    }
462 
463    const unsigned num_dwords =
464       1 + GENX(VERTEX_ELEMENT_STATE_length) * num_elements;
465    uint32_t *dw = blorp_emitn(batch, GENX(3DSTATE_VERTEX_ELEMENTS), num_dwords);
466    if (!dw)
467       return;
468 
469    for (unsigned i = 0; i < num_elements; i++) {
470       GENX(VERTEX_ELEMENT_STATE_pack)(batch, dw, &ve[i]);
471       dw += GENX(VERTEX_ELEMENT_STATE_length);
472    }
473 
474 #if GEN_GEN >= 8
475    /* Overwrite Render Target Array Index (2nd dword) in the VUE header with
476     * primitive instance identifier. This is used for layered clears.
477     */
478    blorp_emit(batch, GENX(3DSTATE_VF_SGVS), sgvs) {
479       sgvs.InstanceIDEnable = true;
480       sgvs.InstanceIDComponentNumber = COMP_1;
481       sgvs.InstanceIDElementOffset = 0;
482    }
483 
484    for (unsigned i = 0; i < num_elements; i++) {
485       blorp_emit(batch, GENX(3DSTATE_VF_INSTANCING), vf) {
486          vf.VertexElementIndex = i;
487          vf.InstancingEnable = false;
488       }
489    }
490 
491    blorp_emit(batch, GENX(3DSTATE_VF_TOPOLOGY), topo) {
492       topo.PrimitiveTopologyType = _3DPRIM_RECTLIST;
493    }
494 #endif
495 }
496 
497 /* 3DSTATE_VIEWPORT_STATE_POINTERS */
498 static uint32_t
blorp_emit_cc_viewport(struct blorp_batch * batch,const struct blorp_params * params)499 blorp_emit_cc_viewport(struct blorp_batch *batch,
500                        const struct blorp_params *params)
501 {
502    uint32_t cc_vp_offset;
503    blorp_emit_dynamic(batch, GENX(CC_VIEWPORT), vp, 32, &cc_vp_offset) {
504       vp.MinimumDepth = 0.0;
505       vp.MaximumDepth = 1.0;
506    }
507 
508 #if GEN_GEN >= 7
509    blorp_emit(batch, GENX(3DSTATE_VIEWPORT_STATE_POINTERS_CC), vsp) {
510       vsp.CCViewportPointer = cc_vp_offset;
511    }
512 #elif GEN_GEN == 6
513    blorp_emit(batch, GENX(3DSTATE_VIEWPORT_STATE_POINTERS), vsp) {
514       vsp.CCViewportStateChange = true;
515       vsp.PointertoCC_VIEWPORT = cc_vp_offset;
516    }
517 #endif
518 
519    return cc_vp_offset;
520 }
521 
522 static uint32_t
blorp_emit_sampler_state(struct blorp_batch * batch,const struct blorp_params * params)523 blorp_emit_sampler_state(struct blorp_batch *batch,
524                          const struct blorp_params *params)
525 {
526    uint32_t offset;
527    blorp_emit_dynamic(batch, GENX(SAMPLER_STATE), sampler, 32, &offset) {
528       sampler.MipModeFilter = MIPFILTER_NONE;
529       sampler.MagModeFilter = MAPFILTER_LINEAR;
530       sampler.MinModeFilter = MAPFILTER_LINEAR;
531       sampler.MinLOD = 0;
532       sampler.MaxLOD = 0;
533       sampler.TCXAddressControlMode = TCM_CLAMP;
534       sampler.TCYAddressControlMode = TCM_CLAMP;
535       sampler.TCZAddressControlMode = TCM_CLAMP;
536       sampler.MaximumAnisotropy = RATIO21;
537       sampler.RAddressMinFilterRoundingEnable = true;
538       sampler.RAddressMagFilterRoundingEnable = true;
539       sampler.VAddressMinFilterRoundingEnable = true;
540       sampler.VAddressMagFilterRoundingEnable = true;
541       sampler.UAddressMinFilterRoundingEnable = true;
542       sampler.UAddressMagFilterRoundingEnable = true;
543 #if GEN_GEN > 6
544       sampler.NonnormalizedCoordinateEnable = true;
545 #endif
546    }
547 
548 #if GEN_GEN >= 7
549    blorp_emit(batch, GENX(3DSTATE_SAMPLER_STATE_POINTERS_PS), ssp) {
550       ssp.PointertoPSSamplerState = offset;
551    }
552 #elif GEN_GEN == 6
553    blorp_emit(batch, GENX(3DSTATE_SAMPLER_STATE_POINTERS), ssp) {
554       ssp.VSSamplerStateChange = true;
555       ssp.GSSamplerStateChange = true;
556       ssp.PSSamplerStateChange = true;
557       ssp.PointertoPSSamplerState = offset;
558    }
559 #endif
560 
561    return offset;
562 }
563 
564 /* What follows is the code for setting up a "pipeline" on Sandy Bridge and
565  * later hardware.  This file will be included by i965 for gen4-5 as well, so
566  * this code is guarded by GEN_GEN >= 6.
567  */
568 #if GEN_GEN >= 6
569 
570 static void
blorp_emit_vs_config(struct blorp_batch * batch,const struct blorp_params * params)571 blorp_emit_vs_config(struct blorp_batch *batch,
572                      const struct blorp_params *params)
573 {
574    struct brw_vs_prog_data *vs_prog_data = params->vs_prog_data;
575 
576    blorp_emit(batch, GENX(3DSTATE_VS), vs) {
577       if (vs_prog_data) {
578          vs.Enable = true;
579 
580          vs.KernelStartPointer = params->vs_prog_kernel;
581 
582          vs.DispatchGRFStartRegisterForURBData =
583             vs_prog_data->base.base.dispatch_grf_start_reg;
584          vs.VertexURBEntryReadLength =
585             vs_prog_data->base.urb_read_length;
586          vs.VertexURBEntryReadOffset = 0;
587 
588          vs.MaximumNumberofThreads =
589             batch->blorp->isl_dev->info->max_vs_threads - 1;
590 
591 #if GEN_GEN >= 8
592          vs.SIMD8DispatchEnable =
593             vs_prog_data->base.dispatch_mode == DISPATCH_MODE_SIMD8;
594 #endif
595       }
596    }
597 }
598 
599 static void
blorp_emit_sf_config(struct blorp_batch * batch,const struct blorp_params * params)600 blorp_emit_sf_config(struct blorp_batch *batch,
601                      const struct blorp_params *params)
602 {
603    const struct brw_wm_prog_data *prog_data = params->wm_prog_data;
604 
605    /* 3DSTATE_SF
606     *
607     * Disable ViewportTransformEnable (dw2.1)
608     *
609     * From the SandyBridge PRM, Volume 2, Part 1, Section 1.3, "3D
610     * Primitives Overview":
611     *     RECTLIST: Viewport Mapping must be DISABLED (as is typical with the
612     *     use of screen- space coordinates).
613     *
614     * A solid rectangle must be rendered, so set FrontFaceFillMode (dw2.4:3)
615     * and BackFaceFillMode (dw2.5:6) to SOLID(0).
616     *
617     * From the Sandy Bridge PRM, Volume 2, Part 1, Section
618     * 6.4.1.1 3DSTATE_SF, Field FrontFaceFillMode:
619     *     SOLID: Any triangle or rectangle object found to be front-facing
620     *     is rendered as a solid object. This setting is required when
621     *     (rendering rectangle (RECTLIST) objects.
622     */
623 
624 #if GEN_GEN >= 8
625 
626    blorp_emit(batch, GENX(3DSTATE_SF), sf);
627 
628    blorp_emit(batch, GENX(3DSTATE_RASTER), raster) {
629       raster.CullMode = CULLMODE_NONE;
630    }
631 
632    blorp_emit(batch, GENX(3DSTATE_SBE), sbe) {
633       sbe.VertexURBEntryReadOffset = 1;
634       if (prog_data) {
635          sbe.NumberofSFOutputAttributes = prog_data->num_varying_inputs;
636          sbe.VertexURBEntryReadLength = brw_blorp_get_urb_length(prog_data);
637          sbe.ConstantInterpolationEnable = prog_data->flat_inputs;
638       } else {
639          sbe.NumberofSFOutputAttributes = 0;
640          sbe.VertexURBEntryReadLength = 1;
641       }
642       sbe.ForceVertexURBEntryReadLength = true;
643       sbe.ForceVertexURBEntryReadOffset = true;
644 
645 #if GEN_GEN >= 9
646       for (unsigned i = 0; i < 32; i++)
647          sbe.AttributeActiveComponentFormat[i] = ACF_XYZW;
648 #endif
649    }
650 
651 #elif GEN_GEN >= 7
652 
653    blorp_emit(batch, GENX(3DSTATE_SF), sf) {
654       sf.FrontFaceFillMode = FILL_MODE_SOLID;
655       sf.BackFaceFillMode = FILL_MODE_SOLID;
656 
657       sf.MultisampleRasterizationMode = params->num_samples > 1 ?
658          MSRASTMODE_ON_PATTERN : MSRASTMODE_OFF_PIXEL;
659 
660 #if GEN_GEN == 7
661       sf.DepthBufferSurfaceFormat = params->depth_format;
662 #endif
663    }
664 
665    blorp_emit(batch, GENX(3DSTATE_SBE), sbe) {
666       sbe.VertexURBEntryReadOffset = 1;
667       if (prog_data) {
668          sbe.NumberofSFOutputAttributes = prog_data->num_varying_inputs;
669          sbe.VertexURBEntryReadLength = brw_blorp_get_urb_length(prog_data);
670          sbe.ConstantInterpolationEnable = prog_data->flat_inputs;
671       } else {
672          sbe.NumberofSFOutputAttributes = 0;
673          sbe.VertexURBEntryReadLength = 1;
674       }
675    }
676 
677 #else /* GEN_GEN <= 6 */
678 
679    blorp_emit(batch, GENX(3DSTATE_SF), sf) {
680       sf.FrontFaceFillMode = FILL_MODE_SOLID;
681       sf.BackFaceFillMode = FILL_MODE_SOLID;
682 
683       sf.MultisampleRasterizationMode = params->num_samples > 1 ?
684          MSRASTMODE_ON_PATTERN : MSRASTMODE_OFF_PIXEL;
685 
686       sf.VertexURBEntryReadOffset = 1;
687       if (prog_data) {
688          sf.NumberofSFOutputAttributes = prog_data->num_varying_inputs;
689          sf.VertexURBEntryReadLength = brw_blorp_get_urb_length(prog_data);
690          sf.ConstantInterpolationEnable = prog_data->flat_inputs;
691       } else {
692          sf.NumberofSFOutputAttributes = 0;
693          sf.VertexURBEntryReadLength = 1;
694       }
695    }
696 
697 #endif /* GEN_GEN */
698 }
699 
700 static void
blorp_emit_ps_config(struct blorp_batch * batch,const struct blorp_params * params)701 blorp_emit_ps_config(struct blorp_batch *batch,
702                      const struct blorp_params *params)
703 {
704    const struct brw_wm_prog_data *prog_data = params->wm_prog_data;
705 
706    /* Even when thread dispatch is disabled, max threads (dw5.25:31) must be
707     * nonzero to prevent the GPU from hanging.  While the documentation doesn't
708     * mention this explicitly, it notes that the valid range for the field is
709     * [1,39] = [2,40] threads, which excludes zero.
710     *
711     * To be safe (and to minimize extraneous code) we go ahead and fully
712     * configure the WM state whether or not there is a WM program.
713     */
714 
715 #if GEN_GEN >= 8
716 
717    blorp_emit(batch, GENX(3DSTATE_WM), wm);
718 
719    blorp_emit(batch, GENX(3DSTATE_PS), ps) {
720       if (params->src.enabled) {
721          ps.SamplerCount = 1; /* Up to 4 samplers */
722          ps.BindingTableEntryCount = 2;
723       } else {
724          ps.BindingTableEntryCount = 1;
725       }
726 
727       if (prog_data) {
728          ps.DispatchGRFStartRegisterForConstantSetupData0 =
729             prog_data->base.dispatch_grf_start_reg;
730          ps.DispatchGRFStartRegisterForConstantSetupData2 =
731             prog_data->dispatch_grf_start_reg_2;
732 
733          ps._8PixelDispatchEnable = prog_data->dispatch_8;
734          ps._16PixelDispatchEnable = prog_data->dispatch_16;
735 
736          ps.KernelStartPointer0 = params->wm_prog_kernel;
737          ps.KernelStartPointer2 =
738             params->wm_prog_kernel + prog_data->prog_offset_2;
739       }
740 
741       /* 3DSTATE_PS expects the number of threads per PSD, which is always 64;
742        * it implicitly scales for different GT levels (which have some # of
743        * PSDs).
744        *
745        * In Gen8 the format is U8-2 whereas in Gen9 it is U8-1.
746        */
747       if (GEN_GEN >= 9)
748          ps.MaximumNumberofThreadsPerPSD = 64 - 1;
749       else
750          ps.MaximumNumberofThreadsPerPSD = 64 - 2;
751 
752       switch (params->fast_clear_op) {
753       case BLORP_FAST_CLEAR_OP_NONE:
754          break;
755 #if GEN_GEN >= 9
756       case BLORP_FAST_CLEAR_OP_RESOLVE_PARTIAL:
757          ps.RenderTargetResolveType = RESOLVE_PARTIAL;
758          break;
759       case BLORP_FAST_CLEAR_OP_RESOLVE_FULL:
760          ps.RenderTargetResolveType = RESOLVE_FULL;
761          break;
762 #else
763       case BLORP_FAST_CLEAR_OP_RESOLVE_FULL:
764          ps.RenderTargetResolveEnable = true;
765          break;
766 #endif
767       case BLORP_FAST_CLEAR_OP_CLEAR:
768          ps.RenderTargetFastClearEnable = true;
769          break;
770       default:
771          unreachable("Invalid fast clear op");
772       }
773    }
774 
775    blorp_emit(batch, GENX(3DSTATE_PS_EXTRA), psx) {
776       if (prog_data) {
777          psx.PixelShaderValid = true;
778          psx.AttributeEnable = prog_data->num_varying_inputs > 0;
779          psx.PixelShaderIsPerSample = prog_data->persample_dispatch;
780       }
781 
782       if (params->src.enabled)
783          psx.PixelShaderKillsPixel = true;
784    }
785 
786 #elif GEN_GEN >= 7
787 
788    blorp_emit(batch, GENX(3DSTATE_WM), wm) {
789       switch (params->hiz_op) {
790       case BLORP_HIZ_OP_DEPTH_CLEAR:
791          wm.DepthBufferClear = true;
792          break;
793       case BLORP_HIZ_OP_DEPTH_RESOLVE:
794          wm.DepthBufferResolveEnable = true;
795          break;
796       case BLORP_HIZ_OP_HIZ_RESOLVE:
797          wm.HierarchicalDepthBufferResolveEnable = true;
798          break;
799       case BLORP_HIZ_OP_NONE:
800          break;
801       default:
802          unreachable("not reached");
803       }
804 
805       if (prog_data)
806          wm.ThreadDispatchEnable = true;
807 
808       if (params->src.enabled)
809          wm.PixelShaderKillsPixel = true;
810 
811       if (params->num_samples > 1) {
812          wm.MultisampleRasterizationMode = MSRASTMODE_ON_PATTERN;
813          wm.MultisampleDispatchMode =
814             (prog_data && prog_data->persample_dispatch) ?
815             MSDISPMODE_PERSAMPLE : MSDISPMODE_PERPIXEL;
816       } else {
817          wm.MultisampleRasterizationMode = MSRASTMODE_OFF_PIXEL;
818          wm.MultisampleDispatchMode = MSDISPMODE_PERSAMPLE;
819       }
820    }
821 
822    blorp_emit(batch, GENX(3DSTATE_PS), ps) {
823       ps.MaximumNumberofThreads =
824          batch->blorp->isl_dev->info->max_wm_threads - 1;
825 
826 #if GEN_IS_HASWELL
827       ps.SampleMask = 1;
828 #endif
829 
830       if (prog_data) {
831          ps.DispatchGRFStartRegisterForConstantSetupData0 =
832             prog_data->base.dispatch_grf_start_reg;
833          ps.DispatchGRFStartRegisterForConstantSetupData2 =
834             prog_data->dispatch_grf_start_reg_2;
835 
836          ps.KernelStartPointer0 = params->wm_prog_kernel;
837          ps.KernelStartPointer2 =
838             params->wm_prog_kernel + prog_data->prog_offset_2;
839 
840          ps._8PixelDispatchEnable = prog_data->dispatch_8;
841          ps._16PixelDispatchEnable = prog_data->dispatch_16;
842 
843          ps.AttributeEnable = prog_data->num_varying_inputs > 0;
844       } else {
845          /* Gen7 hardware gets angry if we don't enable at least one dispatch
846           * mode, so just enable 16-pixel dispatch if we don't have a program.
847           */
848          ps._16PixelDispatchEnable = true;
849       }
850 
851       if (params->src.enabled)
852          ps.SamplerCount = 1; /* Up to 4 samplers */
853 
854       switch (params->fast_clear_op) {
855       case BLORP_FAST_CLEAR_OP_NONE:
856          break;
857       case BLORP_FAST_CLEAR_OP_RESOLVE_FULL:
858          ps.RenderTargetResolveEnable = true;
859          break;
860       case BLORP_FAST_CLEAR_OP_CLEAR:
861          ps.RenderTargetFastClearEnable = true;
862          break;
863       default:
864          unreachable("Invalid fast clear op");
865       }
866    }
867 
868 #else /* GEN_GEN <= 6 */
869 
870    blorp_emit(batch, GENX(3DSTATE_WM), wm) {
871       wm.MaximumNumberofThreads =
872          batch->blorp->isl_dev->info->max_wm_threads - 1;
873 
874       switch (params->hiz_op) {
875       case BLORP_HIZ_OP_DEPTH_CLEAR:
876          wm.DepthBufferClear = true;
877          break;
878       case BLORP_HIZ_OP_DEPTH_RESOLVE:
879          wm.DepthBufferResolveEnable = true;
880          break;
881       case BLORP_HIZ_OP_HIZ_RESOLVE:
882          wm.HierarchicalDepthBufferResolveEnable = true;
883          break;
884       case BLORP_HIZ_OP_NONE:
885          break;
886       default:
887          unreachable("not reached");
888       }
889 
890       if (prog_data) {
891          wm.ThreadDispatchEnable = true;
892 
893          wm.DispatchGRFStartRegisterForConstantSetupData0 =
894             prog_data->base.dispatch_grf_start_reg;
895          wm.DispatchGRFStartRegisterForConstantSetupData2 =
896             prog_data->dispatch_grf_start_reg_2;
897 
898          wm.KernelStartPointer0 = params->wm_prog_kernel;
899          wm.KernelStartPointer2 =
900             params->wm_prog_kernel + prog_data->prog_offset_2;
901 
902          wm._8PixelDispatchEnable = prog_data->dispatch_8;
903          wm._16PixelDispatchEnable = prog_data->dispatch_16;
904 
905          wm.NumberofSFOutputAttributes = prog_data->num_varying_inputs;
906       }
907 
908       if (params->src.enabled) {
909          wm.SamplerCount = 1; /* Up to 4 samplers */
910          wm.PixelShaderKillsPixel = true; /* TODO: temporarily smash on */
911       }
912 
913       if (params->num_samples > 1) {
914          wm.MultisampleRasterizationMode = MSRASTMODE_ON_PATTERN;
915          wm.MultisampleDispatchMode =
916             (prog_data && prog_data->persample_dispatch) ?
917             MSDISPMODE_PERSAMPLE : MSDISPMODE_PERPIXEL;
918       } else {
919          wm.MultisampleRasterizationMode = MSRASTMODE_OFF_PIXEL;
920          wm.MultisampleDispatchMode = MSDISPMODE_PERSAMPLE;
921       }
922    }
923 
924 #endif /* GEN_GEN */
925 }
926 
927 static uint32_t
blorp_emit_blend_state(struct blorp_batch * batch,const struct blorp_params * params)928 blorp_emit_blend_state(struct blorp_batch *batch,
929                        const struct blorp_params *params)
930 {
931    struct GENX(BLEND_STATE) blend;
932    memset(&blend, 0, sizeof(blend));
933 
934    uint32_t offset;
935    int size = GENX(BLEND_STATE_length) * 4;
936    size += GENX(BLEND_STATE_ENTRY_length) * 4 * params->num_draw_buffers;
937    uint32_t *state = blorp_alloc_dynamic_state(batch, size, 64, &offset);
938    uint32_t *pos = state;
939 
940    GENX(BLEND_STATE_pack)(NULL, pos, &blend);
941    pos += GENX(BLEND_STATE_length);
942 
943    for (unsigned i = 0; i < params->num_draw_buffers; ++i) {
944       struct GENX(BLEND_STATE_ENTRY) entry = {
945          .PreBlendColorClampEnable = true,
946          .PostBlendColorClampEnable = true,
947          .ColorClampRange = COLORCLAMP_RTFORMAT,
948 
949          .WriteDisableRed = params->color_write_disable[0],
950          .WriteDisableGreen = params->color_write_disable[1],
951          .WriteDisableBlue = params->color_write_disable[2],
952          .WriteDisableAlpha = params->color_write_disable[3],
953       };
954       GENX(BLEND_STATE_ENTRY_pack)(NULL, pos, &entry);
955       pos += GENX(BLEND_STATE_ENTRY_length);
956    }
957 
958    blorp_flush_range(batch, state, size);
959 
960 #if GEN_GEN >= 7
961    blorp_emit(batch, GENX(3DSTATE_BLEND_STATE_POINTERS), sp) {
962       sp.BlendStatePointer = offset;
963 #if GEN_GEN >= 8
964       sp.BlendStatePointerValid = true;
965 #endif
966    }
967 #endif
968 
969 #if GEN_GEN >= 8
970    blorp_emit(batch, GENX(3DSTATE_PS_BLEND), ps_blend) {
971       ps_blend.HasWriteableRT = true;
972    }
973 #endif
974 
975    return offset;
976 }
977 
978 static uint32_t
blorp_emit_color_calc_state(struct blorp_batch * batch,const struct blorp_params * params)979 blorp_emit_color_calc_state(struct blorp_batch *batch,
980                             const struct blorp_params *params)
981 {
982    uint32_t offset;
983    blorp_emit_dynamic(batch, GENX(COLOR_CALC_STATE), cc, 64, &offset) {
984 #if GEN_GEN <= 8
985       cc.StencilReferenceValue = params->stencil_ref;
986 #endif
987    }
988 
989 #if GEN_GEN >= 7
990    blorp_emit(batch, GENX(3DSTATE_CC_STATE_POINTERS), sp) {
991       sp.ColorCalcStatePointer = offset;
992 #if GEN_GEN >= 8
993       sp.ColorCalcStatePointerValid = true;
994 #endif
995    }
996 #endif
997 
998    return offset;
999 }
1000 
1001 static uint32_t
blorp_emit_depth_stencil_state(struct blorp_batch * batch,const struct blorp_params * params)1002 blorp_emit_depth_stencil_state(struct blorp_batch *batch,
1003                                const struct blorp_params *params)
1004 {
1005 #if GEN_GEN >= 8
1006    struct GENX(3DSTATE_WM_DEPTH_STENCIL) ds = {
1007       GENX(3DSTATE_WM_DEPTH_STENCIL_header),
1008    };
1009 #else
1010    struct GENX(DEPTH_STENCIL_STATE) ds = { 0 };
1011 #endif
1012 
1013    if (params->depth.enabled) {
1014       ds.DepthBufferWriteEnable = true;
1015 
1016       switch (params->hiz_op) {
1017       case BLORP_HIZ_OP_NONE:
1018          ds.DepthTestEnable = true;
1019          ds.DepthTestFunction = COMPAREFUNCTION_ALWAYS;
1020          break;
1021 
1022       /* See the following sections of the Sandy Bridge PRM, Volume 2, Part1:
1023        *   - 7.5.3.1 Depth Buffer Clear
1024        *   - 7.5.3.2 Depth Buffer Resolve
1025        *   - 7.5.3.3 Hierarchical Depth Buffer Resolve
1026        */
1027       case BLORP_HIZ_OP_DEPTH_RESOLVE:
1028          ds.DepthTestEnable = true;
1029          ds.DepthTestFunction = COMPAREFUNCTION_NEVER;
1030          break;
1031 
1032       case BLORP_HIZ_OP_DEPTH_CLEAR:
1033       case BLORP_HIZ_OP_HIZ_RESOLVE:
1034          ds.DepthTestEnable = false;
1035          break;
1036       }
1037    }
1038 
1039    if (params->stencil.enabled) {
1040       ds.StencilBufferWriteEnable = true;
1041       ds.StencilTestEnable = true;
1042       ds.DoubleSidedStencilEnable = false;
1043 
1044       ds.StencilTestFunction = COMPAREFUNCTION_ALWAYS;
1045       ds.StencilPassDepthPassOp = STENCILOP_REPLACE;
1046 
1047       ds.StencilWriteMask = params->stencil_mask;
1048 #if GEN_GEN >= 9
1049       ds.StencilReferenceValue = params->stencil_ref;
1050 #endif
1051    }
1052 
1053 #if GEN_GEN >= 8
1054    uint32_t offset = 0;
1055    uint32_t *dw = blorp_emit_dwords(batch,
1056                                     GENX(3DSTATE_WM_DEPTH_STENCIL_length));
1057    if (!dw)
1058       return 0;
1059 
1060    GENX(3DSTATE_WM_DEPTH_STENCIL_pack)(NULL, dw, &ds);
1061 #else
1062    uint32_t offset;
1063    void *state = blorp_alloc_dynamic_state(batch,
1064                                            GENX(DEPTH_STENCIL_STATE_length) * 4,
1065                                            64, &offset);
1066    GENX(DEPTH_STENCIL_STATE_pack)(NULL, state, &ds);
1067    blorp_flush_range(batch, state, GENX(DEPTH_STENCIL_STATE_length) * 4);
1068 #endif
1069 
1070 #if GEN_GEN == 7
1071    blorp_emit(batch, GENX(3DSTATE_DEPTH_STENCIL_STATE_POINTERS), sp) {
1072       sp.PointertoDEPTH_STENCIL_STATE = offset;
1073    }
1074 #endif
1075 
1076    return offset;
1077 }
1078 
1079 static void
blorp_emit_3dstate_multisample(struct blorp_batch * batch,const struct blorp_params * params)1080 blorp_emit_3dstate_multisample(struct blorp_batch *batch,
1081                                const struct blorp_params *params)
1082 {
1083    blorp_emit(batch, GENX(3DSTATE_MULTISAMPLE), ms) {
1084       ms.NumberofMultisamples       = __builtin_ffs(params->num_samples) - 1;
1085 
1086 #if GEN_GEN >= 8
1087       /* The PRM says that this bit is valid only for DX9:
1088        *
1089        *    SW can choose to set this bit only for DX9 API. DX10/OGL API's
1090        *    should not have any effect by setting or not setting this bit.
1091        */
1092       ms.PixelPositionOffsetEnable  = false;
1093 #elif GEN_GEN >= 7
1094 
1095       switch (params->num_samples) {
1096       case 1:
1097          GEN_SAMPLE_POS_1X(ms.Sample);
1098          break;
1099       case 2:
1100          GEN_SAMPLE_POS_2X(ms.Sample);
1101          break;
1102       case 4:
1103          GEN_SAMPLE_POS_4X(ms.Sample);
1104          break;
1105       case 8:
1106          GEN_SAMPLE_POS_8X(ms.Sample);
1107          break;
1108       default:
1109          break;
1110       }
1111 #else
1112       GEN_SAMPLE_POS_4X(ms.Sample);
1113 #endif
1114       ms.PixelLocation              = CENTER;
1115    }
1116 }
1117 
1118 static void
blorp_emit_pipeline(struct blorp_batch * batch,const struct blorp_params * params)1119 blorp_emit_pipeline(struct blorp_batch *batch,
1120                     const struct blorp_params *params)
1121 {
1122    uint32_t blend_state_offset = 0;
1123    uint32_t color_calc_state_offset;
1124    uint32_t depth_stencil_state_offset;
1125 
1126    emit_urb_config(batch, params);
1127 
1128    if (params->wm_prog_data) {
1129       blend_state_offset = blorp_emit_blend_state(batch, params);
1130    }
1131    color_calc_state_offset = blorp_emit_color_calc_state(batch, params);
1132    depth_stencil_state_offset = blorp_emit_depth_stencil_state(batch, params);
1133 
1134 #if GEN_GEN == 6
1135    /* 3DSTATE_CC_STATE_POINTERS
1136     *
1137     * The pointer offsets are relative to
1138     * CMD_STATE_BASE_ADDRESS.DynamicStateBaseAddress.
1139     *
1140     * The HiZ op doesn't use BLEND_STATE or COLOR_CALC_STATE.
1141     *
1142     * The dynamic state emit helpers emit their own STATE_POINTERS packets on
1143     * gen7+.  However, on gen6 and earlier, they're all lumpped together in
1144     * one CC_STATE_POINTERS packet so we have to emit that here.
1145     */
1146    blorp_emit(batch, GENX(3DSTATE_CC_STATE_POINTERS), cc) {
1147       cc.BLEND_STATEChange = true;
1148       cc.ColorCalcStatePointerValid = true;
1149       cc.DEPTH_STENCIL_STATEChange = true;
1150       cc.PointertoBLEND_STATE = blend_state_offset;
1151       cc.ColorCalcStatePointer = color_calc_state_offset;
1152       cc.PointertoDEPTH_STENCIL_STATE = depth_stencil_state_offset;
1153    }
1154 #else
1155    (void)blend_state_offset;
1156    (void)color_calc_state_offset;
1157    (void)depth_stencil_state_offset;
1158 #endif
1159 
1160    blorp_emit(batch, GENX(3DSTATE_CONSTANT_VS), vs);
1161 #if GEN_GEN >= 7
1162    blorp_emit(batch, GENX(3DSTATE_CONSTANT_HS), hs);
1163    blorp_emit(batch, GENX(3DSTATE_CONSTANT_DS), DS);
1164 #endif
1165    blorp_emit(batch, GENX(3DSTATE_CONSTANT_GS), gs);
1166    blorp_emit(batch, GENX(3DSTATE_CONSTANT_PS), ps);
1167 
1168    if (params->src.enabled)
1169       blorp_emit_sampler_state(batch, params);
1170 
1171    blorp_emit_3dstate_multisample(batch, params);
1172 
1173    blorp_emit(batch, GENX(3DSTATE_SAMPLE_MASK), mask) {
1174       mask.SampleMask = (1 << params->num_samples) - 1;
1175    }
1176 
1177    /* From the BSpec, 3D Pipeline > Geometry > Vertex Shader > State,
1178     * 3DSTATE_VS, Dword 5.0 "VS Function Enable":
1179     *
1180     *   [DevSNB] A pipeline flush must be programmed prior to a
1181     *   3DSTATE_VS command that causes the VS Function Enable to
1182     *   toggle. Pipeline flush can be executed by sending a PIPE_CONTROL
1183     *   command with CS stall bit set and a post sync operation.
1184     *
1185     * We've already done one at the start of the BLORP operation.
1186     */
1187    blorp_emit_vs_config(batch, params);
1188 #if GEN_GEN >= 7
1189    blorp_emit(batch, GENX(3DSTATE_HS), hs);
1190    blorp_emit(batch, GENX(3DSTATE_TE), te);
1191    blorp_emit(batch, GENX(3DSTATE_DS), DS);
1192    blorp_emit(batch, GENX(3DSTATE_STREAMOUT), so);
1193 #endif
1194    blorp_emit(batch, GENX(3DSTATE_GS), gs);
1195 
1196    blorp_emit(batch, GENX(3DSTATE_CLIP), clip) {
1197       clip.PerspectiveDivideDisable = true;
1198    }
1199 
1200    blorp_emit_sf_config(batch, params);
1201    blorp_emit_ps_config(batch, params);
1202 
1203    blorp_emit_cc_viewport(batch, params);
1204 }
1205 
1206 /******** This is the end of the pipeline setup code ********/
1207 
1208 #endif /* GEN_GEN >= 6 */
1209 
1210 #if GEN_GEN >= 7 && GEN_GEN <= 10
1211 static void
blorp_emit_memcpy(struct blorp_batch * batch,struct blorp_address dst,struct blorp_address src,uint32_t size)1212 blorp_emit_memcpy(struct blorp_batch *batch,
1213                   struct blorp_address dst,
1214                   struct blorp_address src,
1215                   uint32_t size)
1216 {
1217    assert(size % 4 == 0);
1218 
1219    for (unsigned dw = 0; dw < size; dw += 4) {
1220 #if GEN_GEN >= 8
1221       blorp_emit(batch, GENX(MI_COPY_MEM_MEM), cp) {
1222          cp.DestinationMemoryAddress = dst;
1223          cp.SourceMemoryAddress = src;
1224       }
1225 #else
1226       /* IVB does not have a general purpose register for command streamer
1227        * commands. Therefore, we use an alternate temporary register.
1228        */
1229 #define BLORP_TEMP_REG 0x2440 /* GEN7_3DPRIM_BASE_VERTEX */
1230       blorp_emit(batch, GENX(MI_LOAD_REGISTER_MEM), load) {
1231          load.RegisterAddress = BLORP_TEMP_REG;
1232          load.MemoryAddress = src;
1233       }
1234       blorp_emit(batch, GENX(MI_STORE_REGISTER_MEM), store) {
1235          store.RegisterAddress = BLORP_TEMP_REG;
1236          store.MemoryAddress = dst;
1237       }
1238 #undef BLORP_TEMP_REG
1239 #endif
1240       dst.offset += 4;
1241       src.offset += 4;
1242    }
1243 }
1244 #endif
1245 
1246 static void
blorp_emit_surface_state(struct blorp_batch * batch,const struct brw_blorp_surface_info * surface,void * state,uint32_t state_offset,const bool color_write_disables[4],bool is_render_target)1247 blorp_emit_surface_state(struct blorp_batch *batch,
1248                          const struct brw_blorp_surface_info *surface,
1249                          void *state, uint32_t state_offset,
1250                          const bool color_write_disables[4],
1251                          bool is_render_target)
1252 {
1253    const struct isl_device *isl_dev = batch->blorp->isl_dev;
1254    struct isl_surf surf = surface->surf;
1255 
1256    if (surf.dim == ISL_SURF_DIM_1D &&
1257        surf.dim_layout == ISL_DIM_LAYOUT_GEN4_2D) {
1258       assert(surf.logical_level0_px.height == 1);
1259       surf.dim = ISL_SURF_DIM_2D;
1260    }
1261 
1262    /* Blorp doesn't support HiZ in any of the blit or slow-clear paths */
1263    enum isl_aux_usage aux_usage = surface->aux_usage;
1264    if (aux_usage == ISL_AUX_USAGE_HIZ)
1265       aux_usage = ISL_AUX_USAGE_NONE;
1266 
1267    isl_channel_mask_t write_disable_mask = 0;
1268    if (is_render_target && GEN_GEN <= 5) {
1269       if (color_write_disables[0])
1270          write_disable_mask |= ISL_CHANNEL_RED_BIT;
1271       if (color_write_disables[1])
1272          write_disable_mask |= ISL_CHANNEL_GREEN_BIT;
1273       if (color_write_disables[2])
1274          write_disable_mask |= ISL_CHANNEL_BLUE_BIT;
1275       if (color_write_disables[3])
1276          write_disable_mask |= ISL_CHANNEL_ALPHA_BIT;
1277    }
1278 
1279    isl_surf_fill_state(batch->blorp->isl_dev, state,
1280                        .surf = &surf, .view = &surface->view,
1281                        .aux_surf = &surface->aux_surf, .aux_usage = aux_usage,
1282                        .mocs = surface->addr.mocs,
1283                        .clear_color = surface->clear_color,
1284                        .write_disables = write_disable_mask);
1285 
1286    blorp_surface_reloc(batch, state_offset + isl_dev->ss.addr_offset,
1287                        surface->addr, 0);
1288 
1289    if (aux_usage != ISL_AUX_USAGE_NONE) {
1290       /* On gen7 and prior, the bottom 12 bits of the MCS base address are
1291        * used to store other information.  This should be ok, however, because
1292        * surface buffer addresses are always 4K page alinged.
1293        */
1294       assert((surface->aux_addr.offset & 0xfff) == 0);
1295       uint32_t *aux_addr = state + isl_dev->ss.aux_addr_offset;
1296       blorp_surface_reloc(batch, state_offset + isl_dev->ss.aux_addr_offset,
1297                           surface->aux_addr, *aux_addr);
1298    }
1299 
1300    blorp_flush_range(batch, state, GENX(RENDER_SURFACE_STATE_length) * 4);
1301 
1302    if (surface->clear_color_addr.buffer) {
1303 #if GEN_GEN > 10
1304       unreachable("Implement indirect clear support on gen11+");
1305 #elif GEN_GEN >= 7 && GEN_GEN <= 10
1306       struct blorp_address dst_addr = blorp_get_surface_base_address(batch);
1307       dst_addr.offset += state_offset + isl_dev->ss.clear_value_offset;
1308       blorp_emit_memcpy(batch, dst_addr, surface->clear_color_addr,
1309                         isl_dev->ss.clear_value_size);
1310 #else
1311       unreachable("Fast clears are only supported on gen7+");
1312 #endif
1313    }
1314 }
1315 
1316 static void
blorp_emit_null_surface_state(struct blorp_batch * batch,const struct brw_blorp_surface_info * surface,uint32_t * state)1317 blorp_emit_null_surface_state(struct blorp_batch *batch,
1318                               const struct brw_blorp_surface_info *surface,
1319                               uint32_t *state)
1320 {
1321    struct GENX(RENDER_SURFACE_STATE) ss = {
1322       .SurfaceType = SURFTYPE_NULL,
1323       .SurfaceFormat = (enum GENX(SURFACE_FORMAT)) ISL_FORMAT_R8G8B8A8_UNORM,
1324       .Width = surface->surf.logical_level0_px.width - 1,
1325       .Height = surface->surf.logical_level0_px.height - 1,
1326       .MIPCountLOD = surface->view.base_level,
1327       .MinimumArrayElement = surface->view.base_array_layer,
1328       .Depth = surface->view.array_len - 1,
1329       .RenderTargetViewExtent = surface->view.array_len - 1,
1330 #if GEN_GEN >= 6
1331       .NumberofMultisamples = ffs(surface->surf.samples) - 1,
1332 #endif
1333 
1334 #if GEN_GEN >= 7
1335       .SurfaceArray = surface->surf.dim != ISL_SURF_DIM_3D,
1336 #endif
1337 
1338 #if GEN_GEN >= 8
1339       .TileMode = YMAJOR,
1340 #else
1341       .TiledSurface = true,
1342 #endif
1343    };
1344 
1345    GENX(RENDER_SURFACE_STATE_pack)(NULL, state, &ss);
1346 
1347    blorp_flush_range(batch, state, GENX(RENDER_SURFACE_STATE_length) * 4);
1348 }
1349 
1350 static void
blorp_emit_surface_states(struct blorp_batch * batch,const struct blorp_params * params)1351 blorp_emit_surface_states(struct blorp_batch *batch,
1352                           const struct blorp_params *params)
1353 {
1354    const struct isl_device *isl_dev = batch->blorp->isl_dev;
1355    uint32_t bind_offset, surface_offsets[2];
1356    void *surface_maps[2];
1357 
1358    MAYBE_UNUSED bool has_indirect_clear_color = false;
1359    if (params->use_pre_baked_binding_table) {
1360       bind_offset = params->pre_baked_binding_table_offset;
1361    } else {
1362       unsigned num_surfaces = 1 + params->src.enabled;
1363       blorp_alloc_binding_table(batch, num_surfaces,
1364                                 isl_dev->ss.size, isl_dev->ss.align,
1365                                 &bind_offset, surface_offsets, surface_maps);
1366 
1367       if (params->dst.enabled) {
1368          blorp_emit_surface_state(batch, &params->dst,
1369                                   surface_maps[BLORP_RENDERBUFFER_BT_INDEX],
1370                                   surface_offsets[BLORP_RENDERBUFFER_BT_INDEX],
1371                                   params->color_write_disable, true);
1372          if (params->dst.clear_color_addr.buffer != NULL)
1373             has_indirect_clear_color = true;
1374       } else {
1375          assert(params->depth.enabled || params->stencil.enabled);
1376          const struct brw_blorp_surface_info *surface =
1377             params->depth.enabled ? &params->depth : &params->stencil;
1378          blorp_emit_null_surface_state(batch, surface,
1379                                        surface_maps[BLORP_RENDERBUFFER_BT_INDEX]);
1380       }
1381 
1382       if (params->src.enabled) {
1383          blorp_emit_surface_state(batch, &params->src,
1384                                   surface_maps[BLORP_TEXTURE_BT_INDEX],
1385                                   surface_offsets[BLORP_TEXTURE_BT_INDEX],
1386                                   NULL, false);
1387          if (params->src.clear_color_addr.buffer != NULL)
1388             has_indirect_clear_color = true;
1389       }
1390    }
1391 
1392 #if GEN_GEN >= 7 && GEN_GEN <= 10
1393    if (has_indirect_clear_color) {
1394       /* Updating a surface state object may require that the state cache be
1395        * invalidated. From the SKL PRM, Shared Functions -> State -> State
1396        * Caching:
1397        *
1398        *    Whenever the RENDER_SURFACE_STATE object in memory pointed to by
1399        *    the Binding Table Pointer (BTP) and Binding Table Index (BTI) is
1400        *    modified [...], the L1 state cache must be invalidated to ensure
1401        *    the new surface or sampler state is fetched from system memory.
1402        */
1403       blorp_emit(batch, GENX(PIPE_CONTROL), pipe) {
1404          pipe.StateCacheInvalidationEnable = true;
1405       }
1406    }
1407 #endif
1408 
1409 #if GEN_GEN >= 7
1410    blorp_emit(batch, GENX(3DSTATE_BINDING_TABLE_POINTERS_VS), bt);
1411    blorp_emit(batch, GENX(3DSTATE_BINDING_TABLE_POINTERS_HS), bt);
1412    blorp_emit(batch, GENX(3DSTATE_BINDING_TABLE_POINTERS_DS), bt);
1413    blorp_emit(batch, GENX(3DSTATE_BINDING_TABLE_POINTERS_GS), bt);
1414 
1415    blorp_emit(batch, GENX(3DSTATE_BINDING_TABLE_POINTERS_PS), bt) {
1416       bt.PointertoPSBindingTable = bind_offset;
1417    }
1418 #elif GEN_GEN >= 6
1419    blorp_emit(batch, GENX(3DSTATE_BINDING_TABLE_POINTERS), bt) {
1420       bt.PSBindingTableChange = true;
1421       bt.PointertoPSBindingTable = bind_offset;
1422    }
1423 #else
1424    blorp_emit(batch, GENX(3DSTATE_BINDING_TABLE_POINTERS), bt) {
1425       bt.PointertoPSBindingTable = bind_offset;
1426    }
1427 #endif
1428 }
1429 
1430 static void
blorp_emit_depth_stencil_config(struct blorp_batch * batch,const struct blorp_params * params)1431 blorp_emit_depth_stencil_config(struct blorp_batch *batch,
1432                                 const struct blorp_params *params)
1433 {
1434    const struct isl_device *isl_dev = batch->blorp->isl_dev;
1435 
1436    uint32_t *dw = blorp_emit_dwords(batch, isl_dev->ds.size / 4);
1437    if (dw == NULL)
1438       return;
1439 
1440    struct isl_depth_stencil_hiz_emit_info info = { };
1441 
1442    if (params->depth.enabled) {
1443       info.view = &params->depth.view;
1444       info.mocs = params->depth.addr.mocs;
1445    } else if (params->stencil.enabled) {
1446       info.view = &params->stencil.view;
1447       info.mocs = params->stencil.addr.mocs;
1448    }
1449 
1450    if (params->depth.enabled) {
1451       info.depth_surf = &params->depth.surf;
1452 
1453       info.depth_address =
1454          blorp_emit_reloc(batch, dw + isl_dev->ds.depth_offset / 4,
1455                           params->depth.addr, 0);
1456 
1457       info.hiz_usage = params->depth.aux_usage;
1458       if (info.hiz_usage == ISL_AUX_USAGE_HIZ) {
1459          info.hiz_surf = &params->depth.aux_surf;
1460 
1461          struct blorp_address hiz_address = params->depth.aux_addr;
1462 #if GEN_GEN == 6
1463          /* Sandy bridge hardware does not technically support mipmapped HiZ.
1464           * However, we have a special layout that allows us to make it work
1465           * anyway by manually offsetting to the specified miplevel.
1466           */
1467          assert(info.hiz_surf->dim_layout == ISL_DIM_LAYOUT_GEN6_STENCIL_HIZ);
1468          uint32_t offset_B;
1469          isl_surf_get_image_offset_B_tile_sa(info.hiz_surf,
1470                                              info.view->base_level, 0, 0,
1471                                              &offset_B, NULL, NULL);
1472          hiz_address.offset += offset_B;
1473 #endif
1474 
1475          info.hiz_address =
1476             blorp_emit_reloc(batch, dw + isl_dev->ds.hiz_offset / 4,
1477                              hiz_address, 0);
1478 
1479          info.depth_clear_value = params->depth.clear_color.f32[0];
1480       }
1481    }
1482 
1483    if (params->stencil.enabled) {
1484       info.stencil_surf = &params->stencil.surf;
1485 
1486       struct blorp_address stencil_address = params->stencil.addr;
1487 #if GEN_GEN == 6
1488       /* Sandy bridge hardware does not technically support mipmapped stencil.
1489        * However, we have a special layout that allows us to make it work
1490        * anyway by manually offsetting to the specified miplevel.
1491        */
1492       assert(info.stencil_surf->dim_layout == ISL_DIM_LAYOUT_GEN6_STENCIL_HIZ);
1493       uint32_t offset_B;
1494       isl_surf_get_image_offset_B_tile_sa(info.stencil_surf,
1495                                           info.view->base_level, 0, 0,
1496                                           &offset_B, NULL, NULL);
1497       stencil_address.offset += offset_B;
1498 #endif
1499 
1500       info.stencil_address =
1501          blorp_emit_reloc(batch, dw + isl_dev->ds.stencil_offset / 4,
1502                           stencil_address, 0);
1503    }
1504 
1505    isl_emit_depth_stencil_hiz_s(isl_dev, dw, &info);
1506 }
1507 
1508 #if GEN_GEN >= 8
1509 /* Emits the Optimized HiZ sequence specified in the BDW+ PRMs. The
1510  * depth/stencil buffer extents are ignored to handle APIs which perform
1511  * clearing operations without such information.
1512  * */
1513 static void
blorp_emit_gen8_hiz_op(struct blorp_batch * batch,const struct blorp_params * params)1514 blorp_emit_gen8_hiz_op(struct blorp_batch *batch,
1515                        const struct blorp_params *params)
1516 {
1517    /* We should be performing an operation on a depth or stencil buffer.
1518     */
1519    assert(params->depth.enabled || params->stencil.enabled);
1520 
1521    /* The stencil buffer should only be enabled if a fast clear operation is
1522     * requested.
1523     */
1524    if (params->stencil.enabled)
1525       assert(params->hiz_op == BLORP_HIZ_OP_DEPTH_CLEAR);
1526 
1527    /* From the BDW PRM Volume 2, 3DSTATE_WM_HZ_OP:
1528     *
1529     * 3DSTATE_MULTISAMPLE packet must be used prior to this packet to change
1530     * the Number of Multisamples. This packet must not be used to change
1531     * Number of Multisamples in a rendering sequence.
1532     *
1533     * Since HIZ may be the first thing in a batch buffer, play safe and always
1534     * emit 3DSTATE_MULTISAMPLE.
1535     */
1536    blorp_emit_3dstate_multisample(batch, params);
1537 
1538    /* If we can't alter the depth stencil config and multiple layers are
1539     * involved, the HiZ op will fail. This is because the op requires that a
1540     * new config is emitted for each additional layer.
1541     */
1542    if (batch->flags & BLORP_BATCH_NO_EMIT_DEPTH_STENCIL) {
1543       assert(params->num_layers <= 1);
1544    } else {
1545       blorp_emit_depth_stencil_config(batch, params);
1546    }
1547 
1548    blorp_emit(batch, GENX(3DSTATE_WM_HZ_OP), hzp) {
1549       switch (params->hiz_op) {
1550       case BLORP_HIZ_OP_DEPTH_CLEAR:
1551          hzp.StencilBufferClearEnable = params->stencil.enabled;
1552          hzp.DepthBufferClearEnable = params->depth.enabled;
1553          hzp.StencilClearValue = params->stencil_ref;
1554          hzp.FullSurfaceDepthandStencilClear = params->full_surface_hiz_op;
1555          break;
1556       case BLORP_HIZ_OP_DEPTH_RESOLVE:
1557          assert(params->full_surface_hiz_op);
1558          hzp.DepthBufferResolveEnable = true;
1559          break;
1560       case BLORP_HIZ_OP_HIZ_RESOLVE:
1561          assert(params->full_surface_hiz_op);
1562          hzp.HierarchicalDepthBufferResolveEnable = true;
1563          break;
1564       case BLORP_HIZ_OP_NONE:
1565          unreachable("Invalid HIZ op");
1566       }
1567 
1568       hzp.NumberofMultisamples = ffs(params->num_samples) - 1;
1569       hzp.SampleMask = 0xFFFF;
1570 
1571       /* Due to a hardware issue, this bit MBZ */
1572       assert(hzp.ScissorRectangleEnable == false);
1573 
1574       /* Contrary to the HW docs both fields are inclusive */
1575       hzp.ClearRectangleXMin = params->x0;
1576       hzp.ClearRectangleYMin = params->y0;
1577 
1578       /* Contrary to the HW docs both fields are exclusive */
1579       hzp.ClearRectangleXMax = params->x1;
1580       hzp.ClearRectangleYMax = params->y1;
1581    }
1582 
1583    /* PIPE_CONTROL w/ all bits clear except for “Post-Sync Operation” must set
1584     * to “Write Immediate Data” enabled.
1585     */
1586    blorp_emit(batch, GENX(PIPE_CONTROL), pc) {
1587       pc.PostSyncOperation = WriteImmediateData;
1588       pc.Address = blorp_get_workaround_page(batch);
1589    }
1590 
1591    blorp_emit(batch, GENX(3DSTATE_WM_HZ_OP), hzp);
1592 }
1593 #endif
1594 
1595 /**
1596  * \brief Execute a blit or render pass operation.
1597  *
1598  * To execute the operation, this function manually constructs and emits a
1599  * batch to draw a rectangle primitive. The batchbuffer is flushed before
1600  * constructing and after emitting the batch.
1601  *
1602  * This function alters no GL state.
1603  */
1604 static void
blorp_exec(struct blorp_batch * batch,const struct blorp_params * params)1605 blorp_exec(struct blorp_batch *batch, const struct blorp_params *params)
1606 {
1607 #if GEN_GEN >= 8
1608    if (params->hiz_op != BLORP_HIZ_OP_NONE) {
1609       blorp_emit_gen8_hiz_op(batch, params);
1610       return;
1611    }
1612 #endif
1613 
1614    blorp_emit_vertex_buffers(batch, params);
1615    blorp_emit_vertex_elements(batch, params);
1616 
1617    blorp_emit_pipeline(batch, params);
1618 
1619    blorp_emit_surface_states(batch, params);
1620 
1621    if (!(batch->flags & BLORP_BATCH_NO_EMIT_DEPTH_STENCIL))
1622       blorp_emit_depth_stencil_config(batch, params);
1623 
1624    blorp_emit(batch, GENX(3DPRIMITIVE), prim) {
1625       prim.VertexAccessType = SEQUENTIAL;
1626       prim.PrimitiveTopologyType = _3DPRIM_RECTLIST;
1627 #if GEN_GEN >= 7
1628       prim.PredicateEnable = batch->flags & BLORP_BATCH_PREDICATE_ENABLE;
1629 #endif
1630       prim.VertexCountPerInstance = 3;
1631       prim.InstanceCount = params->num_layers;
1632    }
1633 }
1634 
1635 #endif /* BLORP_GENX_EXEC_H */
1636