freedreno/ir3: add a6xx+ SSBO/image support

Signed-off-by: Rob Clark <[email protected]>

7 files changed, 484 insertions, 2 deletions

diff --git a/src/freedreno/Makefile.sources b/src/freedreno/Makefile.sources
index 2d83f70f13e..cbb5de934f7 100644
--- a/src/freedreno/Makefile.sources
+++ b/src/freedreno/Makefile.sources
@@ -20,6 +20,7 @@ ir3_SOURCES := \
 	ir3/instr-a3xx.h \
 	ir3/ir3.c \
 	ir3/ir3_a4xx.c \
+	ir3/ir3_a6xx.c \
 	ir3/ir3_compiler.c \
 	ir3/ir3_compiler.h \
 	ir3/ir3_compiler_nir.c \
diff --git a/src/freedreno/ir3/ir3.h b/src/freedreno/ir3/ir3.h
index deede024dee..e710022f5d2 100644
--- a/src/freedreno/ir3/ir3.h
+++ b/src/freedreno/ir3/ir3.h
@@ -1001,6 +1001,8 @@ void ir3_group(struct ir3 *ir);
 void ir3_sched_add_deps(struct ir3 *ir);
 int ir3_sched(struct ir3 *ir);
 
+void ir3_a6xx_fixup_atomic_dests(struct ir3 *ir, struct ir3_shader_variant *so);
+
 /* register assignment: */
 struct ir3_ra_reg_set * ir3_ra_alloc_reg_set(struct ir3_compiler *compiler);
 int ir3_ra(struct ir3 *ir3, gl_shader_stage type,
@@ -1169,6 +1171,23 @@ ir3_##name(struct ir3_block *block,                                      \
 	return instr;                                                        \
 }
 
+#define INSTR3F(f, name)                                                 \
+static inline struct ir3_instruction *                                   \
+ir3_##name##_##f(struct ir3_block *block,                                \
+		struct ir3_instruction *a, unsigned aflags,                      \
+		struct ir3_instruction *b, unsigned bflags,                      \
+		struct ir3_instruction *c, unsigned cflags)                      \
+{                                                                        \
+	struct ir3_instruction *instr =                                      \
+		ir3_instr_create2(block, OPC_##name, 5);                         \
+	ir3_reg_create(instr, 0, 0);   /* dst */                             \
+	__ssa_src(instr, a, aflags);                                         \
+	__ssa_src(instr, b, bflags);                                         \
+	__ssa_src(instr, c, cflags);                                         \
+	instr->flags |= IR3_INSTR_##f;                                       \
+	return instr;                                                        \
+}
+
 #define INSTR4(name)                                                     \
 static inline struct ir3_instruction *                                   \
 ir3_##name(struct ir3_block *block,                                      \
@@ -1338,7 +1357,20 @@ INSTR2(ATOMIC_MAX)
 INSTR2(ATOMIC_AND)
 INSTR2(ATOMIC_OR)
 INSTR2(ATOMIC_XOR)
-#if GPU >= 400
+#if GPU >= 600
+INSTR3(STIB);
+INSTR3F(G, ATOMIC_ADD)
+INSTR3F(G, ATOMIC_SUB)
+INSTR3F(G, ATOMIC_XCHG)
+INSTR3F(G, ATOMIC_INC)
+INSTR3F(G, ATOMIC_DEC)
+INSTR3F(G, ATOMIC_CMPXCHG)
+INSTR3F(G, ATOMIC_MIN)
+INSTR3F(G, ATOMIC_MAX)
+INSTR3F(G, ATOMIC_AND)
+INSTR3F(G, ATOMIC_OR)
+INSTR3F(G, ATOMIC_XOR)
+#elif GPU >= 400
 INSTR3(LDGB)
 INSTR4(STGB)
 INSTR4(STIB)
diff --git a/src/freedreno/ir3/ir3_a6xx.c b/src/freedreno/ir3/ir3_a6xx.c
new file mode 100644
index 00000000000..7330acc7063
--- /dev/null
+++ b/src/freedreno/ir3/ir3_a6xx.c
@@ -0,0 +1,441 @@
+/*
+ * Copyright (C) 2017-2018 Rob Clark <[email protected]>
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ * Authors:
+ *    Rob Clark <[email protected]>
+ */
+
+#define GPU 600
+
+#include "ir3_context.h"
+#include "ir3_image.h"
+
+/*
+ * Handlers for instructions changed/added in a6xx:
+ *
+ * Starting with a6xx, isam and stbi is used for SSBOs as well; stbi and the
+ * atomic instructions (used for both SSBO and image) use a new instruction
+ * encoding compared to a4xx/a5xx.
+ */
+
+
+static struct ir3_instruction *
+ssbo_offset(struct ir3_block *b, struct ir3_instruction *byte_offset)
+{
+	/* TODO hardware wants offset in terms of elements, not bytes.  Which
+	 * is kinda nice but opposite of what nir does.  It would be nice if
+	 * we had a way to request the units of the offset to avoid the extra
+	 * shift instructions..
+	 */
+	return ir3_SHR_B(b, byte_offset, 0, create_immed(b, 2), 0);
+}
+
+/* src[] = { buffer_index, offset }. No const_index */
+static void
+emit_intrinsic_load_ssbo(struct ir3_context *ctx, nir_intrinsic_instr *intr,
+		struct ir3_instruction **dst)
+{
+	struct ir3_block *b = ctx->block;
+	struct ir3_instruction *offset;
+	struct ir3_instruction *sam;
+	nir_const_value *buffer_index;
+
+	/* can this be non-const buffer_index?  how do we handle that? */
+	buffer_index = nir_src_as_const_value(intr->src[0]);
+	compile_assert(ctx, buffer_index);
+
+	int tex_idx = ir3_ssbo_to_tex(&ctx->so->image_mapping, buffer_index->u32[0]);
+
+	offset = ssbo_offset(b, ir3_get_src(ctx, &intr->src[1])[0]);
+
+	/* Because texture state for SSBO read is setup as a single component
+	 * format (ie. R32_UINT, etc), we can't read more than the .x component
+	 * in one shot.  Maybe there is some way we could mangle the state to
+	 * read more than one component at a shot, which would result is some-
+	 * what less register usage (given how we have to stick in the dummy
+	 * .y coord) and less alu instructions to calc offsets.  But this is
+	 * also what blob does, so meh?
+	 */
+	for (unsigned i; i < intr->num_components; i++) {
+		struct ir3_instruction *coords[2];
+
+		coords[0] = (i == 0) ? offset :
+				ir3_ADD_U(b, offset, 0, create_immed(b, i), 0);
+		coords[1] = create_immed(b, 0);
+
+		sam = ir3_SAM(b, OPC_ISAM, TYPE_U32, 0b1, 0,
+				tex_idx, tex_idx, ir3_create_collect(ctx, coords, 2), NULL);
+
+		sam->barrier_class = IR3_BARRIER_IMAGE_R;
+		sam->barrier_conflict = IR3_BARRIER_IMAGE_W;
+
+		dst[i] = sam;
+	}
+}
+
+/* src[] = { value, block_index, offset }. const_index[] = { write_mask } */
+static void
+emit_intrinsic_store_ssbo(struct ir3_context *ctx, nir_intrinsic_instr *intr)
+{
+	struct ir3_block *b = ctx->block;
+	struct ir3_instruction *stib, *val, *offset;
+	nir_const_value *buffer_index;
+	/* TODO handle wrmask properly, see _store_shared().. but I think
+	 * it is more a PITA than that, since blob ends up loading the
+	 * masked components and writing them back out.
+	 */
+	unsigned wrmask = intr->const_index[0];
+	unsigned ncomp = ffs(~wrmask) - 1;
+
+	/* can this be non-const buffer_index?  how do we handle that? */
+	buffer_index = nir_src_as_const_value(intr->src[1]);
+	compile_assert(ctx, buffer_index);
+
+	int ibo_idx = ir3_ssbo_to_ibo(&ctx->so->image_mapping,  buffer_index->u32[0]);
+
+	/* src0 is offset, src1 is value:
+	 */
+	val = ir3_create_collect(ctx, ir3_get_src(ctx, &intr->src[0]), ncomp);
+	offset = ssbo_offset(b, ir3_get_src(ctx, &intr->src[2])[0]);
+
+	stib = ir3_STIB(b, create_immed(b, ibo_idx), 0, offset, 0, val, 0);
+	stib->cat6.iim_val = ncomp;
+	stib->cat6.d = 1;
+	stib->cat6.type = TYPE_U32;
+	stib->barrier_class = IR3_BARRIER_BUFFER_W;
+	stib->barrier_conflict = IR3_BARRIER_BUFFER_R | IR3_BARRIER_BUFFER_W;
+
+	array_insert(b, b->keeps, stib);
+}
+
+/*
+ * SSBO atomic intrinsics
+ *
+ * All of the SSBO atomic memory operations read a value from memory,
+ * compute a new value using one of the operations below, write the new
+ * value to memory, and return the original value read.
+ *
+ * All operations take 3 sources except CompSwap that takes 4. These
+ * sources represent:
+ *
+ * 0: The SSBO buffer index.
+ * 1: The offset into the SSBO buffer of the variable that the atomic
+ *    operation will operate on.
+ * 2: The data parameter to the atomic function (i.e. the value to add
+ *    in ssbo_atomic_add, etc).
+ * 3: For CompSwap only: the second data parameter.
+ */
+static struct ir3_instruction *
+emit_intrinsic_atomic_ssbo(struct ir3_context *ctx, nir_intrinsic_instr *intr)
+{
+	struct ir3_block *b = ctx->block;
+	struct ir3_instruction *atomic, *ibo, *src0, *src1, *offset, *data, *dummy;
+	nir_const_value *buffer_index;
+	type_t type = TYPE_U32;
+
+	/* can this be non-const buffer_index?  how do we handle that? */
+	buffer_index = nir_src_as_const_value(intr->src[0]);
+	compile_assert(ctx, buffer_index);
+
+	int ibo_idx = ir3_ssbo_to_ibo(&ctx->so->image_mapping,  buffer_index->u32[0]);
+	ibo = create_immed(b, ibo_idx);
+
+	offset = ir3_get_src(ctx, &intr->src[1])[0];
+	data   = ir3_get_src(ctx, &intr->src[2])[0];
+
+	/* So this gets a bit creative:
+	 *
+	 *    src0    - vecN offset/coords
+	 *    src1.x  - is actually destination register
+	 *    src1.y  - is 'data' except for cmpxchg where src2.y is 'compare'
+	 *    src1.z  - is 'data' for cmpxchg
+	 *
+	 * The combining src and dest kinda doesn't work out so well with how
+	 * scheduling and RA work.  So for now we create a dummy src2.x, and
+	 * then in a later fixup path, insert an extra MOV out of src1.x.
+	 * See ir3_a6xx_fixup_atomic_dests().
+	 *
+	 * Note that nir already multiplies the offset by four
+	 */
+	dummy = create_immed(b, 0);
+	src0 = ssbo_offset(b, offset);
+
+	if (intr->intrinsic == nir_intrinsic_ssbo_atomic_comp_swap) {
+		struct ir3_instruction *compare = ir3_get_src(ctx, &intr->src[3])[0];
+		src1 = ir3_create_collect(ctx, (struct ir3_instruction*[]){
+			dummy, compare, data
+		}, 3);
+	} else {
+		src1 = ir3_create_collect(ctx, (struct ir3_instruction*[]){
+			dummy, data
+		}, 2);
+	}
+
+	switch (intr->intrinsic) {
+	case nir_intrinsic_ssbo_atomic_add:
+		atomic = ir3_ATOMIC_ADD_G(b, ibo, 0, src0, 0, src1, 0);
+		break;
+	case nir_intrinsic_ssbo_atomic_imin:
+		atomic = ir3_ATOMIC_MIN_G(b, ibo, 0, src0, 0, src1, 0);
+		type = TYPE_S32;
+		break;
+	case nir_intrinsic_ssbo_atomic_umin:
+		atomic = ir3_ATOMIC_MIN_G(b, ibo, 0, src0, 0, src1, 0);
+		break;
+	case nir_intrinsic_ssbo_atomic_imax:
+		atomic = ir3_ATOMIC_MAX_G(b, ibo, 0, src0, 0, src1, 0);
+		type = TYPE_S32;
+		break;
+	case nir_intrinsic_ssbo_atomic_umax:
+		atomic = ir3_ATOMIC_MAX_G(b, ibo, 0, src0, 0, src1, 0);
+		break;
+	case nir_intrinsic_ssbo_atomic_and:
+		atomic = ir3_ATOMIC_AND_G(b, ibo, 0, src0, 0, src1, 0);
+		break;
+	case nir_intrinsic_ssbo_atomic_or:
+		atomic = ir3_ATOMIC_OR_G(b, ibo, 0, src0, 0, src1, 0);
+		break;
+	case nir_intrinsic_ssbo_atomic_xor:
+		atomic = ir3_ATOMIC_XOR_G(b, ibo, 0, src0, 0, src1, 0);
+		break;
+	case nir_intrinsic_ssbo_atomic_exchange:
+		atomic = ir3_ATOMIC_XCHG_G(b, ibo, 0, src0, 0, src1, 0);
+		break;
+	case nir_intrinsic_ssbo_atomic_comp_swap:
+		atomic = ir3_ATOMIC_CMPXCHG_G(b, ibo, 0, src0, 0, src1, 0);
+		break;
+	default:
+		unreachable("boo");
+	}
+
+	atomic->cat6.iim_val = 1;
+	atomic->cat6.d = 1;
+	atomic->cat6.type = type;
+	atomic->barrier_class = IR3_BARRIER_BUFFER_W;
+	atomic->barrier_conflict = IR3_BARRIER_BUFFER_R | IR3_BARRIER_BUFFER_W;
+
+	/* even if nothing consume the result, we can't DCE the instruction: */
+	array_insert(b, b->keeps, atomic);
+
+	return atomic;
+}
+
+/* src[] = { deref, coord, sample_index, value }. const_index[] = {} */
+static void
+emit_intrinsic_store_image(struct ir3_context *ctx, nir_intrinsic_instr *intr)
+{
+	struct ir3_block *b = ctx->block;
+	const nir_variable *var = nir_intrinsic_get_var(intr, 0);
+	struct ir3_instruction *stib;
+	struct ir3_instruction * const *value = ir3_get_src(ctx, &intr->src[3]);
+	struct ir3_instruction * const *coords = ir3_get_src(ctx, &intr->src[1]);
+	unsigned ncoords = ir3_get_image_coords(var, NULL);
+	unsigned slot = ir3_get_image_slot(nir_src_as_deref(intr->src[0]));
+	unsigned ibo_idx = ir3_image_to_ibo(&ctx->so->image_mapping, slot);
+	unsigned ncomp = ir3_get_num_components_for_glformat(var->data.image.format);
+
+	/* src0 is offset, src1 is value:
+	 */
+	stib = ir3_STIB(b, create_immed(b, ibo_idx), 0,
+			ir3_create_collect(ctx, coords, ncoords), 0,
+			ir3_create_collect(ctx, value, ncomp), 0);
+	stib->cat6.iim_val = ncomp;
+	stib->cat6.d = ncoords;
+	stib->cat6.type = ir3_get_image_type(var);
+	stib->cat6.typed = true;
+	stib->barrier_class = IR3_BARRIER_IMAGE_W;
+	stib->barrier_conflict = IR3_BARRIER_IMAGE_R | IR3_BARRIER_IMAGE_W;
+
+	array_insert(b, b->keeps, stib);
+}
+
+/* src[] = { deref, coord, sample_index, value, compare }. const_index[] = {} */
+static struct ir3_instruction *
+emit_intrinsic_atomic_image(struct ir3_context *ctx, nir_intrinsic_instr *intr)
+{
+	struct ir3_block *b = ctx->block;
+	const nir_variable *var = nir_intrinsic_get_var(intr, 0);
+	struct ir3_instruction *atomic, *ibo, *src0, *src1, *dummy;
+	struct ir3_instruction * const *coords = ir3_get_src(ctx, &intr->src[1]);
+	struct ir3_instruction *value = ir3_get_src(ctx, &intr->src[3])[0];
+	unsigned ncoords = ir3_get_image_coords(var, NULL);
+	unsigned slot = ir3_get_image_slot(nir_src_as_deref(intr->src[0]));
+	unsigned ibo_idx = ir3_image_to_ibo(&ctx->so->image_mapping, slot);
+
+	ibo = create_immed(b, ibo_idx);
+
+	/* So this gets a bit creative:
+	 *
+	 *    src0    - vecN offset/coords
+	 *    src1.x  - is actually destination register
+	 *    src1.y  - is 'value' except for cmpxchg where src2.y is 'compare'
+	 *    src1.z  - is 'value' for cmpxchg
+	 *
+	 * The combining src and dest kinda doesn't work out so well with how
+	 * scheduling and RA work.  So for now we create a dummy src2.x, and
+	 * then in a later fixup path, insert an extra MOV out of src1.x.
+	 * See ir3_a6xx_fixup_atomic_dests().
+	 */
+	dummy = create_immed(b, 0);
+	src0 = ir3_create_collect(ctx, coords, ncoords);
+
+	if (intr->intrinsic == nir_intrinsic_image_deref_atomic_comp_swap) {
+		struct ir3_instruction *compare = ir3_get_src(ctx, &intr->src[4])[0];
+		src1 = ir3_create_collect(ctx, (struct ir3_instruction*[]){
+			dummy, compare, value
+		}, 3);
+	} else {
+		src1 = ir3_create_collect(ctx, (struct ir3_instruction*[]){
+			dummy, value
+		}, 2);
+	}
+
+	switch (intr->intrinsic) {
+	case nir_intrinsic_image_deref_atomic_add:
+		atomic = ir3_ATOMIC_ADD_G(b, ibo, 0, src0, 0, src1, 0);
+		break;
+	case nir_intrinsic_image_deref_atomic_min:
+		atomic = ir3_ATOMIC_MIN_G(b, ibo, 0, src0, 0, src1, 0);
+		break;
+	case nir_intrinsic_image_deref_atomic_max:
+		atomic = ir3_ATOMIC_MAX_G(b, ibo, 0, src0, 0, src1, 0);
+		break;
+	case nir_intrinsic_image_deref_atomic_and:
+		atomic = ir3_ATOMIC_AND_G(b, ibo, 0, src0, 0, src1, 0);
+		break;
+	case nir_intrinsic_image_deref_atomic_or:
+		atomic = ir3_ATOMIC_OR_G(b, ibo, 0, src0, 0, src1, 0);
+		break;
+	case nir_intrinsic_image_deref_atomic_xor:
+		atomic = ir3_ATOMIC_XOR_G(b, ibo, 0, src0, 0, src1, 0);
+		break;
+	case nir_intrinsic_image_deref_atomic_exchange:
+		atomic = ir3_ATOMIC_XCHG_G(b, ibo, 0, src0, 0, src1, 0);
+		break;
+	case nir_intrinsic_image_deref_atomic_comp_swap:
+		atomic = ir3_ATOMIC_CMPXCHG_G(b, ibo, 0, src0, 0, src1, 0);
+		break;
+	default:
+		unreachable("boo");
+	}
+
+	atomic->cat6.iim_val = 1;
+	atomic->cat6.d = ncoords;
+	atomic->cat6.type = ir3_get_image_type(var);
+	atomic->cat6.typed = true;
+	atomic->barrier_class = IR3_BARRIER_IMAGE_W;
+	atomic->barrier_conflict = IR3_BARRIER_IMAGE_R | IR3_BARRIER_IMAGE_W;
+
+	/* even if nothing consume the result, we can't DCE the instruction: */
+	array_insert(b, b->keeps, atomic);
+
+	return atomic;
+}
+
+const struct ir3_context_funcs ir3_a6xx_funcs = {
+		.emit_intrinsic_load_ssbo = emit_intrinsic_load_ssbo,
+		.emit_intrinsic_store_ssbo = emit_intrinsic_store_ssbo,
+		.emit_intrinsic_atomic_ssbo = emit_intrinsic_atomic_ssbo,
+		.emit_intrinsic_store_image = emit_intrinsic_store_image,
+		.emit_intrinsic_atomic_image = emit_intrinsic_atomic_image,
+};
+
+/*
+ * Special pass to run after instruction scheduling to insert an
+ * extra mov from src1.x to dst.  This way the other compiler passes
+ * can ignore this quirk of the new instruction encoding.
+ *
+ * This might cause extra complication in the future when we support
+ * spilling, as I think we'd want to re-run the scheduling pass.  One
+ * possible alternative might be to do this in the RA pass after
+ * ra_allocate() but before destroying the SSA links.  (Ie. we do
+ * want to know if anything consumes the result of the atomic instr,
+ * if there is no consumer then inserting the extra mov is pointless.
+ */
+
+static struct ir3_instruction *
+get_atomic_dest_mov(struct ir3_instruction *atomic)
+{
+	/* if we've already created the mov-out, then re-use it: */
+	if (atomic->data)
+		return atomic->data;
+
+	/* extract back out the 'dummy' which serves as stand-in for dest: */
+	struct ir3_instruction *src = ssa(atomic->regs[3]);
+	debug_assert(src->opc == OPC_META_FI);
+	struct ir3_instruction *dummy = ssa(src->regs[1]);
+
+	struct ir3_instruction *mov = ir3_MOV(atomic->block, dummy, TYPE_U32);
+
+	mov->flags |= IR3_INSTR_SY;
+
+	/* it will have already been appended to the end of the block, which
+	 * isn't where we want it, so fix-up the location:
+	 */
+	list_delinit(&mov->node);
+	list_add(&mov->node, &atomic->node);
+
+	/* And because this is after instruction scheduling, we don't really
+	 * have a good way to know if extra delay slots are needed.  For
+	 * example, if the result is consumed by an stib (storeImage()) there
+	 * would be no extra delay slots in place already, but 5 are needed.
+	 * Just plan for the worst and hope nobody looks at the resulting
+	 * code that is generated :-(
+	 */
+	struct ir3_instruction *nop = ir3_NOP(atomic->block);
+	nop->repeat = 5;
+
+	list_delinit(&nop->node);
+	list_add(&nop->node, &mov->node);
+
+	return atomic->data = mov;
+}
+
+void
+ir3_a6xx_fixup_atomic_dests(struct ir3 *ir, struct ir3_shader_variant *so)
+{
+	if (so->image_mapping.num_ibo == 0)
+		return;
+
+	list_for_each_entry (struct ir3_block, block, &ir->block_list, node) {
+		list_for_each_entry (struct ir3_instruction, instr, &block->instr_list, node) {
+			instr->data = NULL;
+		}
+	}
+
+	list_for_each_entry (struct ir3_block, block, &ir->block_list, node) {
+		list_for_each_entry_safe (struct ir3_instruction, instr, &block->instr_list, node) {
+			struct ir3_register *reg;
+
+			foreach_src(reg, instr) {
+				struct ir3_instruction *src = ssa(reg);
+
+				if (!src)
+					continue;
+
+				if (is_atomic(src->opc) && (src->flags & IR3_INSTR_G))
+					reg->instr = get_atomic_dest_mov(src);
+			}
+		}
+	}
+
+}
diff --git a/src/freedreno/ir3/ir3_compiler_nir.c b/src/freedreno/ir3/ir3_compiler_nir.c
index fbc1b63c746..22da1508f8b 100644
--- a/src/freedreno/ir3/ir3_compiler_nir.c
+++ b/src/freedreno/ir3/ir3_compiler_nir.c
@@ -2612,6 +2612,10 @@ ir3_compile_shader_nir(struct ir3_compiler *compiler,
 		goto out;
 	}
 
+	if (compiler->gpu_id >= 600) {
+		ir3_a6xx_fixup_atomic_dests(ir, so);
+	}
+
 	if (ir3_shader_debug & IR3_DBG_OPTMSGS) {
 		printf("AFTER SCHED:\n");
 		ir3_print(ir);
diff --git a/src/freedreno/ir3/ir3_context.c b/src/freedreno/ir3/ir3_context.c
index 94264043886..01e313fecec 100644
--- a/src/freedreno/ir3/ir3_context.c
+++ b/src/freedreno/ir3/ir3_context.c
@@ -53,7 +53,9 @@ ir3_context_init(struct ir3_compiler *compiler,
 		}
 	}
 
-	if (compiler->gpu_id >= 400) {
+	if (compiler->gpu_id >= 600) {
+		ctx->funcs = &ir3_a6xx_funcs;
+	} else if (compiler->gpu_id >= 400) {
 		ctx->funcs = &ir3_a4xx_funcs;
 	}
 
diff --git a/src/freedreno/ir3/ir3_context.h b/src/freedreno/ir3/ir3_context.h
index c49b41dbd56..c79ea0ef422 100644
--- a/src/freedreno/ir3/ir3_context.h
+++ b/src/freedreno/ir3/ir3_context.h
@@ -138,6 +138,7 @@ struct ir3_context_funcs {
 };
 
 extern const struct ir3_context_funcs ir3_a4xx_funcs;
+extern const struct ir3_context_funcs ir3_a6xx_funcs;
 
 struct ir3_context * ir3_context_init(struct ir3_compiler *compiler,
 		struct ir3_shader_variant *so);
diff --git a/src/freedreno/ir3/meson.build b/src/freedreno/ir3/meson.build
index 2fd5d362e3b..bf99476311d 100644
--- a/src/freedreno/ir3/meson.build
+++ b/src/freedreno/ir3/meson.build
@@ -35,6 +35,7 @@ libfreedreno_ir3_files = files(
   'instr-a3xx.h',
   'ir3.c',
   'ir3_a4xx.c',
+  'ir3_a6xx.c',
   'ir3_compiler_nir.c',
   'ir3_compiler.c',
   'ir3_compiler.h',