isl: Don't use surface format R32_FLOAT for typed atomic integer operations
[mesa.git] / src / intel / isl / isl_storage_image.c
1 /*
2 * Copyright 2015 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 #include "isl_priv.h"
25 #include "compiler/brw_compiler.h"
26
27 bool
28 isl_is_storage_image_format(enum isl_format format)
29 {
30 /* XXX: Maybe we should put this in the CSV? */
31
32 switch (format) {
33 case ISL_FORMAT_R32G32B32A32_UINT:
34 case ISL_FORMAT_R32G32B32A32_SINT:
35 case ISL_FORMAT_R32G32B32A32_FLOAT:
36 case ISL_FORMAT_R32_UINT:
37 case ISL_FORMAT_R32_SINT:
38 case ISL_FORMAT_R32_FLOAT:
39 case ISL_FORMAT_R16G16B16A16_UINT:
40 case ISL_FORMAT_R16G16B16A16_SINT:
41 case ISL_FORMAT_R16G16B16A16_FLOAT:
42 case ISL_FORMAT_R32G32_UINT:
43 case ISL_FORMAT_R32G32_SINT:
44 case ISL_FORMAT_R32G32_FLOAT:
45 case ISL_FORMAT_R8G8B8A8_UINT:
46 case ISL_FORMAT_R8G8B8A8_SINT:
47 case ISL_FORMAT_R16G16_UINT:
48 case ISL_FORMAT_R16G16_SINT:
49 case ISL_FORMAT_R16G16_FLOAT:
50 case ISL_FORMAT_R8G8_UINT:
51 case ISL_FORMAT_R8G8_SINT:
52 case ISL_FORMAT_R16_UINT:
53 case ISL_FORMAT_R16_FLOAT:
54 case ISL_FORMAT_R16_SINT:
55 case ISL_FORMAT_R8_UINT:
56 case ISL_FORMAT_R8_SINT:
57 case ISL_FORMAT_R10G10B10A2_UINT:
58 case ISL_FORMAT_R10G10B10A2_UNORM:
59 case ISL_FORMAT_R11G11B10_FLOAT:
60 case ISL_FORMAT_R16G16B16A16_UNORM:
61 case ISL_FORMAT_R16G16B16A16_SNORM:
62 case ISL_FORMAT_R8G8B8A8_UNORM:
63 case ISL_FORMAT_R8G8B8A8_SNORM:
64 case ISL_FORMAT_R16G16_UNORM:
65 case ISL_FORMAT_R16G16_SNORM:
66 case ISL_FORMAT_R8G8_UNORM:
67 case ISL_FORMAT_R8G8_SNORM:
68 case ISL_FORMAT_R16_UNORM:
69 case ISL_FORMAT_R16_SNORM:
70 case ISL_FORMAT_R8_UNORM:
71 case ISL_FORMAT_R8_SNORM:
72 return true;
73 default:
74 return false;
75 }
76 }
77
78 enum isl_format
79 isl_lower_storage_image_format(const struct gen_device_info *devinfo,
80 enum isl_format format)
81 {
82 switch (format) {
83 /* These are never lowered. Up to BDW we'll have to fall back to untyped
84 * surface access for 128bpp formats.
85 */
86 case ISL_FORMAT_R32G32B32A32_UINT:
87 case ISL_FORMAT_R32G32B32A32_SINT:
88 case ISL_FORMAT_R32G32B32A32_FLOAT:
89 case ISL_FORMAT_R32_UINT:
90 case ISL_FORMAT_R32_SINT:
91 return format;
92
93 /* The Skylake PRM's "Surface Formats" section says:
94 *
95 * "The surface format for the typed atomic integer operations must
96 * be R32_UINT or R32_SINT."
97 */
98 case ISL_FORMAT_R32_FLOAT:
99 return ISL_FORMAT_R32_UINT;
100
101 /* From HSW to BDW the only 64bpp format supported for typed access is
102 * RGBA_UINT16. IVB falls back to untyped.
103 */
104 case ISL_FORMAT_R16G16B16A16_UINT:
105 case ISL_FORMAT_R16G16B16A16_SINT:
106 case ISL_FORMAT_R16G16B16A16_FLOAT:
107 case ISL_FORMAT_R32G32_UINT:
108 case ISL_FORMAT_R32G32_SINT:
109 case ISL_FORMAT_R32G32_FLOAT:
110 return (devinfo->gen >= 9 ? format :
111 devinfo->gen >= 8 || devinfo->is_haswell ?
112 ISL_FORMAT_R16G16B16A16_UINT :
113 ISL_FORMAT_R32G32_UINT);
114
115 /* Up to BDW no SINT or FLOAT formats of less than 32 bits per component
116 * are supported. IVB doesn't support formats with more than one component
117 * for typed access. For 8 and 16 bpp formats IVB relies on the
118 * undocumented behavior that typed reads from R_UINT8 and R_UINT16
119 * surfaces actually do a 32-bit misaligned read. The alternative would be
120 * to use two surface state entries with different formats for each image,
121 * one for reading (using R_UINT32) and another one for writing (using
122 * R_UINT8 or R_UINT16), but that would complicate the shaders we generate
123 * even more.
124 */
125 case ISL_FORMAT_R8G8B8A8_UINT:
126 case ISL_FORMAT_R8G8B8A8_SINT:
127 return (devinfo->gen >= 9 ? format :
128 devinfo->gen >= 8 || devinfo->is_haswell ?
129 ISL_FORMAT_R8G8B8A8_UINT : ISL_FORMAT_R32_UINT);
130
131 case ISL_FORMAT_R16G16_UINT:
132 case ISL_FORMAT_R16G16_SINT:
133 case ISL_FORMAT_R16G16_FLOAT:
134 return (devinfo->gen >= 9 ? format :
135 devinfo->gen >= 8 || devinfo->is_haswell ?
136 ISL_FORMAT_R16G16_UINT : ISL_FORMAT_R32_UINT);
137
138 case ISL_FORMAT_R8G8_UINT:
139 case ISL_FORMAT_R8G8_SINT:
140 return (devinfo->gen >= 9 ? format :
141 devinfo->gen >= 8 || devinfo->is_haswell ?
142 ISL_FORMAT_R8G8_UINT : ISL_FORMAT_R16_UINT);
143
144 case ISL_FORMAT_R16_UINT:
145 case ISL_FORMAT_R16_FLOAT:
146 case ISL_FORMAT_R16_SINT:
147 return (devinfo->gen >= 9 ? format : ISL_FORMAT_R16_UINT);
148
149 case ISL_FORMAT_R8_UINT:
150 case ISL_FORMAT_R8_SINT:
151 return (devinfo->gen >= 9 ? format : ISL_FORMAT_R8_UINT);
152
153 /* Neither the 2/10/10/10 nor the 11/11/10 packed formats are supported
154 * by the hardware.
155 */
156 case ISL_FORMAT_R10G10B10A2_UINT:
157 case ISL_FORMAT_R10G10B10A2_UNORM:
158 case ISL_FORMAT_R11G11B10_FLOAT:
159 return ISL_FORMAT_R32_UINT;
160
161 /* No normalized fixed-point formats are supported by the hardware. */
162 case ISL_FORMAT_R16G16B16A16_UNORM:
163 case ISL_FORMAT_R16G16B16A16_SNORM:
164 return (devinfo->gen >= 8 || devinfo->is_haswell ?
165 ISL_FORMAT_R16G16B16A16_UINT :
166 ISL_FORMAT_R32G32_UINT);
167
168 case ISL_FORMAT_R8G8B8A8_UNORM:
169 case ISL_FORMAT_R8G8B8A8_SNORM:
170 return (devinfo->gen >= 8 || devinfo->is_haswell ?
171 ISL_FORMAT_R8G8B8A8_UINT : ISL_FORMAT_R32_UINT);
172
173 case ISL_FORMAT_R16G16_UNORM:
174 case ISL_FORMAT_R16G16_SNORM:
175 return (devinfo->gen >= 8 || devinfo->is_haswell ?
176 ISL_FORMAT_R16G16_UINT : ISL_FORMAT_R32_UINT);
177
178 case ISL_FORMAT_R8G8_UNORM:
179 case ISL_FORMAT_R8G8_SNORM:
180 return (devinfo->gen >= 8 || devinfo->is_haswell ?
181 ISL_FORMAT_R8G8_UINT : ISL_FORMAT_R16_UINT);
182
183 case ISL_FORMAT_R16_UNORM:
184 case ISL_FORMAT_R16_SNORM:
185 return ISL_FORMAT_R16_UINT;
186
187 case ISL_FORMAT_R8_UNORM:
188 case ISL_FORMAT_R8_SNORM:
189 return ISL_FORMAT_R8_UINT;
190
191 default:
192 assert(!"Unknown image format");
193 return ISL_FORMAT_UNSUPPORTED;
194 }
195 }
196
197 bool
198 isl_has_matching_typed_storage_image_format(const struct gen_device_info *devinfo,
199 enum isl_format fmt)
200 {
201 if (devinfo->gen >= 9) {
202 return true;
203 } else if (devinfo->gen >= 8 || devinfo->is_haswell) {
204 return isl_format_get_layout(fmt)->bpb <= 64;
205 } else {
206 return isl_format_get_layout(fmt)->bpb <= 32;
207 }
208 }
209
210 static const struct brw_image_param image_param_defaults = {
211 /* Set the swizzling shifts to all-ones to effectively disable
212 * swizzling -- See emit_address_calculation() in
213 * brw_fs_surface_builder.cpp for a more detailed explanation of
214 * these parameters.
215 */
216 .swizzling = { 0xff, 0xff },
217 };
218
219 void
220 isl_surf_fill_image_param(const struct isl_device *dev,
221 struct brw_image_param *param,
222 const struct isl_surf *surf,
223 const struct isl_view *view)
224 {
225 *param = image_param_defaults;
226
227 param->size[0] = isl_minify(surf->logical_level0_px.w, view->base_level);
228 param->size[1] = isl_minify(surf->logical_level0_px.h, view->base_level);
229 if (surf->dim == ISL_SURF_DIM_3D) {
230 param->size[2] = isl_minify(surf->logical_level0_px.d, view->base_level);
231 } else {
232 param->size[2] = surf->logical_level0_px.array_len -
233 view->base_array_layer;
234 }
235
236 isl_surf_get_image_offset_el(surf, view->base_level,
237 surf->dim == ISL_SURF_DIM_3D ?
238 0 : view->base_array_layer,
239 surf->dim == ISL_SURF_DIM_3D ?
240 view->base_array_layer : 0,
241 &param->offset[0], &param->offset[1]);
242
243 const int cpp = isl_format_get_layout(surf->format)->bpb / 8;
244 param->stride[0] = cpp;
245 param->stride[1] = surf->row_pitch / cpp;
246
247 const struct isl_extent3d image_align_sa =
248 isl_surf_get_image_alignment_sa(surf);
249 if (ISL_DEV_GEN(dev) < 9 && surf->dim == ISL_SURF_DIM_3D) {
250 param->stride[2] = isl_align_npot(param->size[0], image_align_sa.w);
251 param->stride[3] = isl_align_npot(param->size[1], image_align_sa.h);
252 } else {
253 param->stride[2] = 0;
254 param->stride[3] = isl_surf_get_array_pitch_el_rows(surf);
255 }
256
257 switch (surf->tiling) {
258 case ISL_TILING_LINEAR:
259 /* image_param_defaults is good enough */
260 break;
261
262 case ISL_TILING_X:
263 /* An X tile is a rectangular block of 512x8 bytes. */
264 param->tiling[0] = isl_log2u(512 / cpp);
265 param->tiling[1] = isl_log2u(8);
266
267 if (dev->has_bit6_swizzling) {
268 /* Right shifts required to swizzle bits 9 and 10 of the memory
269 * address with bit 6.
270 */
271 param->swizzling[0] = 3;
272 param->swizzling[1] = 4;
273 }
274 break;
275
276 case ISL_TILING_Y0:
277 /* The layout of a Y-tiled surface in memory isn't really fundamentally
278 * different to the layout of an X-tiled surface, we simply pretend that
279 * the surface is broken up in a number of smaller 16Bx32 tiles, each
280 * one arranged in X-major order just like is the case for X-tiling.
281 */
282 param->tiling[0] = isl_log2u(16 / cpp);
283 param->tiling[1] = isl_log2u(32);
284
285 if (dev->has_bit6_swizzling) {
286 /* Right shift required to swizzle bit 9 of the memory address with
287 * bit 6.
288 */
289 param->swizzling[0] = 3;
290 param->swizzling[1] = 0xff;
291 }
292 break;
293
294 default:
295 assert(!"Unhandled storage image tiling");
296 }
297
298 /* 3D textures are arranged in 2D in memory with 2^lod slices per row. The
299 * address calculation algorithm (emit_address_calculation() in
300 * brw_fs_surface_builder.cpp) handles this as a sort of tiling with
301 * modulus equal to the LOD.
302 */
303 param->tiling[2] = (ISL_DEV_GEN(dev) < 9 && surf->dim == ISL_SURF_DIM_3D ?
304 view->base_level : 0);
305 }
306
307 void
308 isl_buffer_fill_image_param(const struct isl_device *dev,
309 struct brw_image_param *param,
310 enum isl_format format,
311 uint64_t size)
312 {
313 *param = image_param_defaults;
314
315 param->stride[0] = isl_format_layouts[format].bpb / 8;
316 param->size[0] = size / param->stride[0];
317 }