nir: Make nir_lower_ubo_vec4() handle non-vec4-aligned loads.
It turns out I had missed a case in my enumeration of why everything
currently was vec4-aligned.
Fixes a simple testcase of loading from a vec3[2] array in freedreno with
IR3_SHADER_DEBUG=nouboopt.
Initial shader-db results look devastating:
total instructions in shared programs: 8019997 -> 12829370 (59.97%)
total cat6 in shared programs: 87683 -> 145840 (66.33%)
Hopefully this will recover once we introduce the i/o vectorizer, but that
was blocked on getting the vec3 case fixed.
Reviewed-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6612>
diff --git a/src/compiler/nir/nir_lower_ubo_vec4.c b/src/compiler/nir/nir_lower_ubo_vec4.c
index 2f739ff..85e780d 100644
--- a/src/compiler/nir/nir_lower_ubo_vec4.c
+++ b/src/compiler/nir/nir_lower_ubo_vec4.c
@@ -26,8 +26,10 @@
* hardware to have, and it gives NIR a chance to optimize the addressing math
* and CSE the loads.
*
- * We assume that the UBO loads do not cross a vec4 boundary. This is true
- * for:
+ * This pass handles lowering for loads that straddle a vec4 alignment
+ * boundary. We try to minimize the extra loads we generate for that case,
+ * and are ensured non-straddling loads with:
+ *
* - std140 (GLSL 1.40, GLSL ES)
* - Vulkan "Extended Layout" (the baseline for UBOs)
*
@@ -36,6 +38,10 @@
* - GLSL 4.30's new packed mode (enabled by PIPE_CAP_LOAD_CONSTBUF) where
* vec3 arrays are packed tightly.
*
+ * - PackedDriverUniformStorage in GL (enabled by PIPE_CAP_PACKED_UNIFORMS)
+ * combined with nir_lower_uniforms_to_ubo, where values in the default
+ * uniform block are packed tightly.
+ *
* - Vulkan's scalarBlockLayout optional feature:
*
* "A member is defined to improperly straddle if either of the following are
@@ -66,6 +72,22 @@
return nir_instr_as_intrinsic(instr)->intrinsic == nir_intrinsic_load_ubo;
}
+static nir_intrinsic_instr *
+nir_load_ubo_vec4(nir_builder *b, nir_ssa_def *block, nir_ssa_def *offset,
+ unsigned bit_size, unsigned num_components)
+{
+ nir_intrinsic_instr *load =
+ nir_intrinsic_instr_create(b->shader, nir_intrinsic_load_ubo_vec4);
+ load->src[0] = nir_src_for_ssa(block);
+ load->src[1] = nir_src_for_ssa(offset);
+
+ nir_ssa_dest_init(&load->instr, &load->dest, num_components, bit_size, NULL);
+ load->num_components = num_components;
+ nir_builder_instr_insert(b, &load->instr);
+
+ return load;
+}
+
static nir_ssa_def *
nir_lower_ubo_vec4_lower(nir_builder *b, nir_instr *instr, void *data)
{
@@ -74,11 +96,7 @@
nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
nir_ssa_def *byte_offset = nir_ssa_for_src(b, intr->src[1], 1);
-
- nir_intrinsic_instr *load =
- nir_intrinsic_instr_create(b->shader, nir_intrinsic_load_ubo_vec4);
- nir_src_copy(&load->src[0], &intr->src[0], &load->instr);
- load->src[1] = nir_src_for_ssa(nir_ushr_imm(b, byte_offset, 4));
+ nir_ssa_def *vec4_offset = nir_ushr_imm(b, byte_offset, 4);
unsigned align_mul = nir_intrinsic_align_mul(intr);
unsigned align_offset = nir_intrinsic_align_offset(intr);
@@ -93,24 +111,16 @@
align_offset &= 15;
assert(align_offset % chan_size_bytes == 0);
- /* We assume that loads don't cross vec4 boundaries, just that we need
- * to extract from within the vec4 when we don't have a good alignment.
- */
- if (intr->num_components == chans_per_vec4) {
- align_mul = 16;
- align_offset = 0;
- }
-
unsigned num_components = intr->num_components;
- bool aligned_mul = align_mul % 16 == 0;
+ bool aligned_mul = (align_mul == 16 &&
+ align_offset + chan_size_bytes * num_components <= 16);
if (!aligned_mul)
num_components = chans_per_vec4;
- nir_ssa_dest_init(&load->instr, &load->dest,
- num_components, intr->dest.ssa.bit_size,
- intr->dest.ssa.name);
- load->num_components = num_components;
- nir_builder_instr_insert(b, &load->instr);
+ nir_intrinsic_instr *load = nir_load_ubo_vec4(b, intr->src[0].ssa,
+ vec4_offset,
+ intr->dest.ssa.bit_size,
+ num_components);
nir_ssa_def *result = &load->dest.ssa;
@@ -120,39 +130,62 @@
* offset's component.
*/
nir_intrinsic_set_component(load, align_chan_offset);
+ } else if (intr->num_components == 1) {
+ /* If we're loading a single component, that component alone won't
+ * straddle a vec4 boundary so we can do this with a single UBO load.
+ */
+ nir_ssa_def *component =
+ nir_iand_imm(b,
+ nir_udiv_imm(b, byte_offset, chan_size_bytes),
+ chans_per_vec4 - 1);
+
+ result = nir_vector_extract(b, result, component);
+ } else if (align_mul == 8 &&
+ align_offset + chan_size_bytes * intr->num_components <= 16) {
+ /* Special case: Loading small vectors from offset % 8 == 0 can be done
+ * with just one load and one bcsel.
+ */
+ nir_component_mask_t low_channels =
+ BITSET_MASK(intr->num_components) << (align_chan_offset);
+ nir_component_mask_t high_channels =
+ low_channels << (8 / chan_size_bytes);
+ result = nir_bcsel(b,
+ nir_i2b(b, nir_iand_imm(b, byte_offset, 8)),
+ nir_channels(b, result, high_channels),
+ nir_channels(b, result, low_channels));
} else {
- if (align_mul == 8) {
- /* Special case: Loading small vectors from offset % 8 == 0 can be
- * done with just one bcsel.
- */
- nir_component_mask_t low_channels =
- BITSET_MASK(intr->num_components) << (align_chan_offset);
- nir_component_mask_t high_channels =
- low_channels << (8 / chan_size_bytes);
- result = nir_bcsel(b,
- nir_i2b(b, nir_iand_imm(b, byte_offset, 8)),
- nir_channels(b, result, high_channels),
- nir_channels(b, result, low_channels));
- } else {
- /* General fallback case: Per-result-channel bcsel-based extraction
- * from the load.
- */
- assert(align_mul == 4);
- assert(align_chan_offset == 0);
+ /* General fallback case: Per-result-channel bcsel-based extraction
+ * from two separate vec4 loads.
+ */
+ assert(num_components == 4);
+ nir_ssa_def *next_vec4_offset = nir_iadd_imm(b, vec4_offset, 1);
+ nir_intrinsic_instr *next_load = nir_load_ubo_vec4(b, intr->src[0].ssa,
+ next_vec4_offset,
+ intr->dest.ssa.bit_size,
+ num_components);
+
+ nir_ssa_def *channels[NIR_MAX_VEC_COMPONENTS];
+ for (unsigned i = 0; i < intr->num_components; i++) {
+ nir_ssa_def *chan_byte_offset = nir_iadd_imm(b, byte_offset, i * chan_size_bytes);
+
+ nir_ssa_def *chan_vec4_offset = nir_ushr_imm(b, chan_byte_offset, 4);
nir_ssa_def *component =
nir_iand_imm(b,
- nir_udiv_imm(b, byte_offset, chan_size_bytes),
+ nir_udiv_imm(b, chan_byte_offset, chan_size_bytes),
chans_per_vec4 - 1);
- nir_ssa_def *channels[NIR_MAX_VEC_COMPONENTS];
- for (unsigned i = 0; i < intr->num_components; i++) {
- channels[i] = nir_vector_extract(b, result,
- nir_iadd_imm(b, component, i));
- }
-
- result = nir_vec(b, channels, intr->num_components);
+ channels[i] = nir_vector_extract(b,
+ nir_bcsel(b,
+ nir_ieq(b,
+ chan_vec4_offset,
+ vec4_offset),
+ &load->dest.ssa,
+ &next_load->dest.ssa),
+ component);
}
+
+ result = nir_vec(b, channels, intr->num_components);
}
return result;