<div dir="auto"><div><div class="gmail_extra"><div class="gmail_quote">On Jan 4, 2017 12:02, "Lionel Landwerlin" <<a href="mailto:lionel.g.landwerlin@intel.com">lionel.g.landwerlin@intel.com</a>> wrote:<br type="attribution"><blockquote class="quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">v2: Move relative push constant relative offset computation down to<br>
_vtn_load_store_tail() (Jason)<br></blockquote></div></div></div><div dir="auto"><br></div><div dir="auto">Hm... I may not have been particularly clear. I meant, more or less, to have get_io_offset (is that what its called?) return the offer in terms of base, offset, and range like the push constant intrinsic expects and then, for everything else, do an add to combine the base and indirect parts. Does that make sense? There's even a comment in there about how I was lazy and just trusted constant folding before.</div><div dir="auto"><br></div><div dir="auto"><div class="gmail_extra"><div class="gmail_quote"><blockquote class="quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<div class="quoted-text"><br>
Signed-off-by: Lionel Landwerlin <<a href="mailto:lionel.g.landwerlin@intel.com">lionel.g.landwerlin@intel.com</a><wbr>><br>
---<br>
</div> src/compiler/spirv/vtn_<wbr>variables.c | 78 ++++++++++++++++++++-----<br>
src/intel/vulkan/anv_nir_<wbr>lower_push_constants.c | 1 -<br>
2 files changed, 65 insertions(+), 14 deletions(-)<br>
<br>
diff --git a/src/compiler/spirv/vtn_<wbr>variables.c b/src/compiler/spirv/vtn_<wbr>variables.c<br>
index 040c65061c..e3845365bd 100644<br>
--- a/src/compiler/spirv/vtn_<wbr>variables.c<br>
+++ b/src/compiler/spirv/vtn_<wbr>variables.c<br>
@@ -434,8 +434,37 @@ vtn_type_block_size(struct vtn_type *type)<br>
<div class="elided-text"> }<br>
<br>
static void<br>
+vtn_access_chain_get_offset_<wbr>size(struct vtn_access_chain *chain,<br>
+ unsigned *access_offset,<br>
+ unsigned *access_size)<br>
+{<br>
+ /* Only valid for push constants accesses now. */<br>
+ assert(chain->var->mode == vtn_variable_mode_push_<wbr>constant);<br>
+<br>
+ struct vtn_type *type = chain->var->type;<br>
+<br>
+ *access_offset = 0;<br>
+<br>
+ for (unsigned i = 0; i < chain->length; i++) {<br>
+ if (chain->link[i].mode != vtn_access_mode_literal)<br>
+ break;<br>
+<br>
+ if (glsl_type_is_struct(type-><wbr>type)) {<br>
+ *access_offset += type->offsets[chain->link[i].<wbr>id];<br>
+ type = type->members[chain->link[i].<wbr>id];<br>
+ } else {<br>
+ *access_offset += type->stride * chain->link[i].id;<br>
+ type = type->array_element;<br>
+ }<br>
+ }<br>
+<br>
+ *access_size = vtn_type_block_size(type);<br>
+}<br>
+<br>
+static void<br>
_vtn_load_store_tail(struct vtn_builder *b, nir_intrinsic_op op, bool load,<br>
nir_ssa_def *index, nir_ssa_def *offset,<br>
</div><div class="quoted-text">+ unsigned access_offset, unsigned access_size,<br>
</div><div class="quoted-text"> struct vtn_ssa_value **inout, const struct glsl_type *type)<br>
</div><div class="quoted-text"> {<br>
nir_intrinsic_instr *instr = nir_intrinsic_instr_create(b-><wbr>nb.shader, op);<br>
</div>@@ -447,18 +476,25 @@ _vtn_load_store_tail(struct vtn_builder *b, nir_intrinsic_op op, bool load,<br>
<div class="quoted-text"> instr->src[src++] = nir_src_for_ssa((*inout)->def)<wbr>;<br>
}<br>
<br>
- /* We set the base and size for push constant load to the entire push<br>
- * constant block for now.<br>
- */<br>
if (op == nir_intrinsic_load_push_<wbr>constant) {<br>
- nir_intrinsic_set_base(instr, 0);<br>
- nir_intrinsic_set_range(instr, 128);<br>
+ assert(access_offset % 4 == 0);<br>
+<br>
+ nir_intrinsic_set_base(instr, access_offset);<br>
+ nir_intrinsic_set_range(instr, access_size);<br>
}<br>
<br>
if (index)<br>
</div> instr->src[src++] = nir_src_for_ssa(index);<br>
<br>
- instr->src[src++] = nir_src_for_ssa(offset);<br>
+ if (op == nir_intrinsic_load_push_<wbr>constant) {<br>
<div class="quoted-text">+ /* We need to subtract the offset from where the intrinsic will load the<br>
+ * data. */<br>
</div>+ instr->src[src++] =<br>
+ nir_src_for_ssa(nir_isub(&b-><wbr>nb, offset,<br>
+ nir_imm_int(&b->nb, access_offset)));<br>
+ } else {<br>
+ instr->src[src++] = nir_src_for_ssa(offset);<br>
+ }<br>
<br>
if (load) {<br>
nir_ssa_dest_init(&instr-><wbr>instr, &instr->dest,<br>
@@ -476,6 +512,7 @@ _vtn_load_store_tail(struct vtn_builder *b, nir_intrinsic_op op, bool load,<br>
<div class="quoted-text"> static void<br>
_vtn_block_load_store(struct vtn_builder *b, nir_intrinsic_op op, bool load,<br>
nir_ssa_def *index, nir_ssa_def *offset,<br>
+ unsigned access_offset, unsigned access_size,<br>
struct vtn_access_chain *chain, unsigned chain_idx,<br>
struct vtn_type *type, struct vtn_ssa_value **inout)<br>
{<br>
</div>@@ -520,6 +557,7 @@ _vtn_block_load_store(struct vtn_builder *b, nir_intrinsic_op op, bool load,<br>
<div class="quoted-text"> nir_iadd(&b->nb, offset,<br>
nir_imm_int(&b->nb, i * type->stride));<br>
</div><div class="quoted-text"> _vtn_load_store_tail(b, op, load, index, elem_offset,<br>
</div>+ access_offset, access_size,<br>
&(*inout)->elems[i],<br>
glsl_vector_type(base_type, vec_width));<br>
}<br>
@@ -541,8 +579,9 @@ _vtn_block_load_store(struct vtn_builder *b, nir_intrinsic_op op, bool load,<br>
offset = nir_iadd(&b->nb, offset, row_offset);<br>
<div class="quoted-text"> if (load)<br>
*inout = vtn_create_ssa_value(b, glsl_scalar_type(base_type));<br>
</div><div class="quoted-text">- _vtn_load_store_tail(b, op, load, index, offset, inout,<br>
</div>- glsl_scalar_type(base_type));<br>
<div class="quoted-text">+ _vtn_load_store_tail(b, op, load, index, offset,<br>
</div>+ access_offset, access_size,<br>
+ inout, glsl_scalar_type(base_type));<br>
<div class="quoted-text"> } else {<br>
/* Grabbing a column; picking one element off each row */<br>
unsigned num_comps = glsl_get_vector_elements(type-<wbr>>type);<br>
</div>@@ -562,6 +601,7 @@ _vtn_block_load_store(struct vtn_builder *b, nir_intrinsic_op op, bool load,<br>
<div class="quoted-text"> }<br>
comp = &temp_val;<br>
_vtn_load_store_tail(b, op, load, index, elem_offset,<br>
</div>+ access_offset, access_size,<br>
&comp, glsl_scalar_type(base_type));<br>
<div class="quoted-text"> comps[i] = comp->def;<br>
}<br>
</div>@@ -580,20 +620,25 @@ _vtn_block_load_store(struct vtn_builder *b, nir_intrinsic_op op, bool load,<br>
<div class="quoted-text"> offset = nir_iadd(&b->nb, offset, col_offset);<br>
<br>
_vtn_block_load_store(b, op, load, index, offset,<br>
+ access_offset, access_size,<br>
chain, chain_idx + 1,<br>
type->array_element, inout);<br>
}<br>
} else if (chain == NULL) {<br>
/* Single whole vector */<br>
assert(glsl_type_is_vector_or_<wbr>scalar(type->type));<br>
- _vtn_load_store_tail(b, op, load, index, offset, inout, type->type);<br>
+ _vtn_load_store_tail(b, op, load, index, offset,<br>
</div>+ access_offset, access_size,<br>
+ inout, type->type);<br>
<div class="quoted-text"> } else {<br>
/* Single component of a vector. Fall through to array case. */<br>
nir_ssa_def *elem_offset =<br>
vtn_access_link_as_ssa(b, chain->link[chain_idx], type->stride);<br>
offset = nir_iadd(&b->nb, offset, elem_offset);<br>
<br>
- _vtn_block_load_store(b, op, load, index, offset, NULL, 0,<br>
+ _vtn_block_load_store(b, op, load, index, offset,<br>
+ access_offset, access_size,<br>
+ NULL, 0,<br>
type->array_element, inout);<br>
}<br>
return;<br>
</div>@@ -603,7 +648,9 @@ _vtn_block_load_store(struct vtn_builder *b, nir_intrinsic_op op, bool load,<br>
<div class="quoted-text"> for (unsigned i = 0; i < elems; i++) {<br>
nir_ssa_def *elem_off =<br>
nir_iadd(&b->nb, offset, nir_imm_int(&b->nb, i * type->stride));<br>
- _vtn_block_load_store(b, op, load, index, elem_off, NULL, 0,<br>
+ _vtn_block_load_store(b, op, load, index, elem_off,<br>
+ access_offset, access_size,<br>
+ NULL, 0,<br>
type->array_element, &(*inout)->elems[i]);<br>
}<br>
return;<br>
</div>@@ -614,7 +661,9 @@ _vtn_block_load_store(struct vtn_builder *b, nir_intrinsic_op op, bool load,<br>
<div class="quoted-text"> for (unsigned i = 0; i < elems; i++) {<br>
nir_ssa_def *elem_off =<br>
nir_iadd(&b->nb, offset, nir_imm_int(&b->nb, type->offsets[i]));<br>
- _vtn_block_load_store(b, op, load, index, elem_off, NULL, 0,<br>
+ _vtn_block_load_store(b, op, load, index, elem_off,<br>
+ access_offset, access_size,<br>
+ NULL, 0,<br>
type->members[i], &(*inout)->elems[i]);<br>
}<br>
return;<br>
</div>@@ -629,6 +678,7 @@ static struct vtn_ssa_value *<br>
<div class="quoted-text"> vtn_block_load(struct vtn_builder *b, struct vtn_access_chain *src)<br>
{<br>
</div><div class="quoted-text"> nir_intrinsic_op op;<br>
+ unsigned access_offset = 0, access_size = 0;<br>
switch (src->var->mode) {<br>
case vtn_variable_mode_ubo:<br>
op = nir_intrinsic_load_ubo;<br>
</div>@@ -638,6 +688,7 @@ vtn_block_load(struct vtn_builder *b, struct vtn_access_chain *src)<br>
<div class="quoted-text"> break;<br>
case vtn_variable_mode_push_<wbr>constant:<br>
op = nir_intrinsic_load_push_<wbr>constant;<br>
+ vtn_access_chain_get_offset_<wbr>size(src, &access_offset, &access_size);<br>
</div><div class="quoted-text"> break;<br>
default:<br>
assert(!"Invalid block variable mode");<br>
</div>@@ -650,6 +701,7 @@ vtn_block_load(struct vtn_builder *b, struct vtn_access_chain *src)<br>
<div class="quoted-text"><br>
struct vtn_ssa_value *value = NULL;<br>
_vtn_block_load_store(b, op, true, index, offset,<br>
+ access_offset, access_size,<br>
src, chain_idx, type, &value);<br>
return value;<br>
}<br>
</div>@@ -664,7 +716,7 @@ vtn_block_store(struct vtn_builder *b, struct vtn_ssa_value *src,<br>
<div class="elided-text"> offset = vtn_access_chain_to_offset(b, dst, &index, &type, &chain_idx, true);<br>
<br>
_vtn_block_load_store(b, nir_intrinsic_store_ssbo, false, index, offset,<br>
- dst, chain_idx, type, &src);<br>
+ 0, 0, dst, chain_idx, type, &src);<br>
}<br>
<br>
static bool<br>
diff --git a/src/intel/vulkan/anv_nir_<wbr>lower_push_constants.c b/src/intel/vulkan/anv_nir_<wbr>lower_push_constants.c<br>
index 64c155372f..b66552825b 100644<br>
--- a/src/intel/vulkan/anv_nir_<wbr>lower_push_constants.c<br>
+++ b/src/intel/vulkan/anv_nir_<wbr>lower_push_constants.c<br>
@@ -42,7 +42,6 @@ anv_nir_lower_push_constants(<wbr>nir_shader *shader)<br>
continue;<br>
<br>
assert(intrin->const_index[0] % 4 == 0);<br>
- assert(intrin->const_index[1] == 128);<br>
<br>
/* We just turn them into uniform loads */<br>
intrin->intrinsic = nir_intrinsic_load_uniform;<br>
--<br>
2.11.0<br>
</div></blockquote></div><br></div></div></div>