1/*
2 * Copyright (C) 2017-2018 Rob Clark <robclark@freedesktop.org>
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 FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21 * SOFTWARE.
22 *
23 * Authors:
24 *    Rob Clark <robclark@freedesktop.org>
25 */
26
27#define GPU 400
28
29#include "ir3_context.h"
30#include "ir3_image.h"
31
32/*
33 * Handlers for instructions changed/added in a4xx:
34 */
35
36
37/* src[] = { buffer_index, offset }. No const_index */
38static void
39emit_intrinsic_load_ssbo(struct ir3_context *ctx, nir_intrinsic_instr *intr,
40		struct ir3_instruction **dst)
41{
42	struct ir3_block *b = ctx->block;
43	struct ir3_instruction *ldgb, *src0, *src1, *byte_offset, *offset;
44
45	/* can this be non-const buffer_index?  how do we handle that? */
46	int ibo_idx = ir3_ssbo_to_ibo(&ctx->so->image_mapping, nir_src_as_uint(intr->src[0]));
47
48	byte_offset = ir3_get_src(ctx, &intr->src[1])[0];
49	offset = ir3_get_src(ctx, &intr->src[2])[0];
50
51	/* src0 is uvec2(offset*4, 0), src1 is offset.. nir already *= 4: */
52	src0 = ir3_create_collect(ctx, (struct ir3_instruction*[]){
53		byte_offset,
54		create_immed(b, 0),
55	}, 2);
56	src1 = offset;
57
58	ldgb = ir3_LDGB(b, create_immed(b, ibo_idx), 0,
59			src0, 0, src1, 0);
60	ldgb->regs[0]->wrmask = MASK(intr->num_components);
61	ldgb->cat6.iim_val = intr->num_components;
62	ldgb->cat6.d = 4;
63	ldgb->cat6.type = TYPE_U32;
64	ldgb->barrier_class = IR3_BARRIER_BUFFER_R;
65	ldgb->barrier_conflict = IR3_BARRIER_BUFFER_W;
66
67	ir3_split_dest(b, dst, ldgb, 0, intr->num_components);
68}
69
70/* src[] = { value, block_index, offset }. const_index[] = { write_mask } */
71static void
72emit_intrinsic_store_ssbo(struct ir3_context *ctx, nir_intrinsic_instr *intr)
73{
74	struct ir3_block *b = ctx->block;
75	struct ir3_instruction *stgb, *src0, *src1, *src2, *byte_offset, *offset;
76	/* TODO handle wrmask properly, see _store_shared().. but I think
77	 * it is more a PITA than that, since blob ends up loading the
78	 * masked components and writing them back out.
79	 */
80	unsigned wrmask = intr->const_index[0];
81	unsigned ncomp = ffs(~wrmask) - 1;
82
83	/* can this be non-const buffer_index?  how do we handle that? */
84	int ibo_idx = ir3_ssbo_to_ibo(&ctx->so->image_mapping, nir_src_as_uint(intr->src[1]));
85
86	byte_offset = ir3_get_src(ctx, &intr->src[2])[0];
87	offset = ir3_get_src(ctx, &intr->src[3])[0];
88
89	/* src0 is value, src1 is offset, src2 is uvec2(offset*4, 0)..
90	 * nir already *= 4:
91	 */
92	src0 = ir3_create_collect(ctx, ir3_get_src(ctx, &intr->src[0]), ncomp);
93	src1 = offset;
94	src2 = ir3_create_collect(ctx, (struct ir3_instruction*[]){
95		byte_offset,
96		create_immed(b, 0),
97	}, 2);
98
99	stgb = ir3_STGB(b, create_immed(b, ibo_idx), 0, src0, 0, src1, 0, src2, 0);
100	stgb->cat6.iim_val = ncomp;
101	stgb->cat6.d = 4;
102	stgb->cat6.type = TYPE_U32;
103	stgb->barrier_class = IR3_BARRIER_BUFFER_W;
104	stgb->barrier_conflict = IR3_BARRIER_BUFFER_R | IR3_BARRIER_BUFFER_W;
105
106	array_insert(b, b->keeps, stgb);
107}
108
109/*
110 * SSBO atomic intrinsics
111 *
112 * All of the SSBO atomic memory operations read a value from memory,
113 * compute a new value using one of the operations below, write the new
114 * value to memory, and return the original value read.
115 *
116 * All operations take 3 sources except CompSwap that takes 4. These
117 * sources represent:
118 *
119 * 0: The SSBO buffer index.
120 * 1: The offset into the SSBO buffer of the variable that the atomic
121 *    operation will operate on.
122 * 2: The data parameter to the atomic function (i.e. the value to add
123 *    in ssbo_atomic_add, etc).
124 * 3: For CompSwap only: the second data parameter.
125 */
126static struct ir3_instruction *
127emit_intrinsic_atomic_ssbo(struct ir3_context *ctx, nir_intrinsic_instr *intr)
128{
129	struct ir3_block *b = ctx->block;
130	struct ir3_instruction *atomic, *ssbo, *src0, *src1, *src2, *byte_offset,
131		*offset;
132	type_t type = TYPE_U32;
133
134	/* can this be non-const buffer_index?  how do we handle that? */
135	int ibo_idx = ir3_ssbo_to_ibo(&ctx->so->image_mapping, nir_src_as_uint(intr->src[0]));
136	ssbo = create_immed(b, ibo_idx);
137
138	byte_offset = ir3_get_src(ctx, &intr->src[1])[0];
139	offset = ir3_get_src(ctx, &intr->src[3])[0];
140
141	/* src0 is data (or uvec2(data, compare))
142	 * src1 is offset
143	 * src2 is uvec2(offset*4, 0) (appears to be 64b byte offset)
144	 *
145	 * Note that nir already multiplies the offset by four
146	 */
147	src0 = ir3_get_src(ctx, &intr->src[2])[0];
148	src1 = offset;
149	src2 = ir3_create_collect(ctx, (struct ir3_instruction*[]){
150		byte_offset,
151		create_immed(b, 0),
152	}, 2);
153
154	switch (intr->intrinsic) {
155	case nir_intrinsic_ssbo_atomic_add_ir3:
156		atomic = ir3_ATOMIC_ADD_G(b, ssbo, 0, src0, 0, src1, 0, src2, 0);
157		break;
158	case nir_intrinsic_ssbo_atomic_imin_ir3:
159		atomic = ir3_ATOMIC_MIN_G(b, ssbo, 0, src0, 0, src1, 0, src2, 0);
160		type = TYPE_S32;
161		break;
162	case nir_intrinsic_ssbo_atomic_umin_ir3:
163		atomic = ir3_ATOMIC_MIN_G(b, ssbo, 0, src0, 0, src1, 0, src2, 0);
164		break;
165	case nir_intrinsic_ssbo_atomic_imax_ir3:
166		atomic = ir3_ATOMIC_MAX_G(b, ssbo, 0, src0, 0, src1, 0, src2, 0);
167		type = TYPE_S32;
168		break;
169	case nir_intrinsic_ssbo_atomic_umax_ir3:
170		atomic = ir3_ATOMIC_MAX_G(b, ssbo, 0, src0, 0, src1, 0, src2, 0);
171		break;
172	case nir_intrinsic_ssbo_atomic_and_ir3:
173		atomic = ir3_ATOMIC_AND_G(b, ssbo, 0, src0, 0, src1, 0, src2, 0);
174		break;
175	case nir_intrinsic_ssbo_atomic_or_ir3:
176		atomic = ir3_ATOMIC_OR_G(b, ssbo, 0, src0, 0, src1, 0, src2, 0);
177		break;
178	case nir_intrinsic_ssbo_atomic_xor_ir3:
179		atomic = ir3_ATOMIC_XOR_G(b, ssbo, 0, src0, 0, src1, 0, src2, 0);
180		break;
181	case nir_intrinsic_ssbo_atomic_exchange_ir3:
182		atomic = ir3_ATOMIC_XCHG_G(b, ssbo, 0, src0, 0, src1, 0, src2, 0);
183		break;
184	case nir_intrinsic_ssbo_atomic_comp_swap_ir3:
185		/* for cmpxchg, src0 is [ui]vec2(data, compare): */
186		src0 = ir3_create_collect(ctx, (struct ir3_instruction*[]){
187			ir3_get_src(ctx, &intr->src[3])[0],
188			src0,
189		}, 2);
190		src1 = ir3_get_src(ctx, &intr->src[4])[0];
191		atomic = ir3_ATOMIC_CMPXCHG_G(b, ssbo, 0, src0, 0, src1, 0, src2, 0);
192		break;
193	default:
194		unreachable("boo");
195	}
196
197	atomic->cat6.iim_val = 1;
198	atomic->cat6.d = 4;
199	atomic->cat6.type = type;
200	atomic->barrier_class = IR3_BARRIER_BUFFER_W;
201	atomic->barrier_conflict = IR3_BARRIER_BUFFER_R | IR3_BARRIER_BUFFER_W;
202
203	/* even if nothing consume the result, we can't DCE the instruction: */
204	array_insert(b, b->keeps, atomic);
205
206	return atomic;
207}
208
209static struct ir3_instruction *
210get_image_offset(struct ir3_context *ctx, const nir_variable *var,
211		struct ir3_instruction * const *coords, bool byteoff)
212{
213	struct ir3_block *b = ctx->block;
214	struct ir3_instruction *offset;
215	unsigned ncoords = ir3_get_image_coords(var, NULL);
216
217	/* to calculate the byte offset (yes, uggg) we need (up to) three
218	 * const values to know the bytes per pixel, and y and z stride:
219	 */
220	unsigned cb = regid(ctx->so->constbase.image_dims, 0) +
221		ctx->so->const_layout.image_dims.off[var->data.driver_location];
222
223	debug_assert(ctx->so->const_layout.image_dims.mask &
224			(1 << var->data.driver_location));
225
226	/* offset = coords.x * bytes_per_pixel: */
227	offset = ir3_MUL_S(b, coords[0], 0, create_uniform(b, cb + 0), 0);
228	if (ncoords > 1) {
229		/* offset += coords.y * y_pitch: */
230		offset = ir3_MAD_S24(b, create_uniform(b, cb + 1), 0,
231				coords[1], 0, offset, 0);
232	}
233	if (ncoords > 2) {
234		/* offset += coords.z * z_pitch: */
235		offset = ir3_MAD_S24(b, create_uniform(b, cb + 2), 0,
236				coords[2], 0, offset, 0);
237	}
238
239	if (!byteoff) {
240		/* Some cases, like atomics, seem to use dword offset instead
241		 * of byte offsets.. blob just puts an extra shr.b in there
242		 * in those cases:
243		 */
244		offset = ir3_SHR_B(b, offset, 0, create_immed(b, 2), 0);
245	}
246
247	return ir3_create_collect(ctx, (struct ir3_instruction*[]){
248		offset,
249		create_immed(b, 0),
250	}, 2);
251}
252
253/* src[] = { deref, coord, sample_index, value }. const_index[] = {} */
254static void
255emit_intrinsic_store_image(struct ir3_context *ctx, nir_intrinsic_instr *intr)
256{
257	struct ir3_block *b = ctx->block;
258	const nir_variable *var = nir_intrinsic_get_var(intr, 0);
259	struct ir3_instruction *stib, *offset;
260	struct ir3_instruction * const *value = ir3_get_src(ctx, &intr->src[3]);
261	struct ir3_instruction * const *coords = ir3_get_src(ctx, &intr->src[1]);
262	unsigned ncoords = ir3_get_image_coords(var, NULL);
263	unsigned slot = ir3_get_image_slot(nir_src_as_deref(intr->src[0]));
264	unsigned ibo_idx = ir3_image_to_ibo(&ctx->so->image_mapping, slot);
265	unsigned ncomp = ir3_get_num_components_for_glformat(var->data.image.format);
266
267	/* src0 is value
268	 * src1 is coords
269	 * src2 is 64b byte offset
270	 */
271
272	offset = get_image_offset(ctx, var, coords, true);
273
274	/* NOTE: stib seems to take byte offset, but stgb.typed can be used
275	 * too and takes a dword offset.. not quite sure yet why blob uses
276	 * one over the other in various cases.
277	 */
278
279	stib = ir3_STIB(b, create_immed(b, ibo_idx), 0,
280			ir3_create_collect(ctx, value, ncomp), 0,
281			ir3_create_collect(ctx, coords, ncoords), 0,
282			offset, 0);
283	stib->cat6.iim_val = ncomp;
284	stib->cat6.d = ncoords;
285	stib->cat6.type = ir3_get_image_type(var);
286	stib->cat6.typed = true;
287	stib->barrier_class = IR3_BARRIER_IMAGE_W;
288	stib->barrier_conflict = IR3_BARRIER_IMAGE_R | IR3_BARRIER_IMAGE_W;
289
290	array_insert(b, b->keeps, stib);
291}
292
293/* src[] = { deref, coord, sample_index, value, compare }. const_index[] = {} */
294static struct ir3_instruction *
295emit_intrinsic_atomic_image(struct ir3_context *ctx, nir_intrinsic_instr *intr)
296{
297	struct ir3_block *b = ctx->block;
298	const nir_variable *var = nir_intrinsic_get_var(intr, 0);
299	struct ir3_instruction *atomic, *image, *src0, *src1, *src2;
300	struct ir3_instruction * const *coords = ir3_get_src(ctx, &intr->src[1]);
301	unsigned ncoords = ir3_get_image_coords(var, NULL);
302	unsigned slot = ir3_get_image_slot(nir_src_as_deref(intr->src[0]));
303	unsigned ibo_idx = ir3_image_to_ibo(&ctx->so->image_mapping, slot);
304
305	image = create_immed(b, ibo_idx);
306
307	/* src0 is value (or uvec2(value, compare))
308	 * src1 is coords
309	 * src2 is 64b byte offset
310	 */
311	src0 = ir3_get_src(ctx, &intr->src[3])[0];
312	src1 = ir3_create_collect(ctx, coords, ncoords);
313	src2 = get_image_offset(ctx, var, coords, false);
314
315	switch (intr->intrinsic) {
316	case nir_intrinsic_image_deref_atomic_add:
317		atomic = ir3_ATOMIC_ADD_G(b, image, 0, src0, 0, src1, 0, src2, 0);
318		break;
319	case nir_intrinsic_image_deref_atomic_min:
320		atomic = ir3_ATOMIC_MIN_G(b, image, 0, src0, 0, src1, 0, src2, 0);
321		break;
322	case nir_intrinsic_image_deref_atomic_max:
323		atomic = ir3_ATOMIC_MAX_G(b, image, 0, src0, 0, src1, 0, src2, 0);
324		break;
325	case nir_intrinsic_image_deref_atomic_and:
326		atomic = ir3_ATOMIC_AND_G(b, image, 0, src0, 0, src1, 0, src2, 0);
327		break;
328	case nir_intrinsic_image_deref_atomic_or:
329		atomic = ir3_ATOMIC_OR_G(b, image, 0, src0, 0, src1, 0, src2, 0);
330		break;
331	case nir_intrinsic_image_deref_atomic_xor:
332		atomic = ir3_ATOMIC_XOR_G(b, image, 0, src0, 0, src1, 0, src2, 0);
333		break;
334	case nir_intrinsic_image_deref_atomic_exchange:
335		atomic = ir3_ATOMIC_XCHG_G(b, image, 0, src0, 0, src1, 0, src2, 0);
336		break;
337	case nir_intrinsic_image_deref_atomic_comp_swap:
338		/* for cmpxchg, src0 is [ui]vec2(data, compare): */
339		src0 = ir3_create_collect(ctx, (struct ir3_instruction*[]){
340			ir3_get_src(ctx, &intr->src[4])[0],
341			src0,
342		}, 2);
343		atomic = ir3_ATOMIC_CMPXCHG_G(b, image, 0, src0, 0, src1, 0, src2, 0);
344		break;
345	default:
346		unreachable("boo");
347	}
348
349	atomic->cat6.iim_val = 1;
350	atomic->cat6.d = ncoords;
351	atomic->cat6.type = ir3_get_image_type(var);
352	atomic->cat6.typed = true;
353	atomic->barrier_class = IR3_BARRIER_IMAGE_W;
354	atomic->barrier_conflict = IR3_BARRIER_IMAGE_R | IR3_BARRIER_IMAGE_W;
355
356	/* even if nothing consume the result, we can't DCE the instruction: */
357	array_insert(b, b->keeps, atomic);
358
359	return atomic;
360}
361
362const struct ir3_context_funcs ir3_a4xx_funcs = {
363		.emit_intrinsic_load_ssbo = emit_intrinsic_load_ssbo,
364		.emit_intrinsic_store_ssbo = emit_intrinsic_store_ssbo,
365		.emit_intrinsic_atomic_ssbo = emit_intrinsic_atomic_ssbo,
366		.emit_intrinsic_store_image = emit_intrinsic_store_image,
367		.emit_intrinsic_atomic_image = emit_intrinsic_atomic_image,
368};
369