freedreno/ir3: Stop shifting UBO 1 down to be UBO 0.

It turns out the GL uniforms file is larger than the hardware constant
file, so we need to limit how many UBOs we lower to constbuf loads.  To do
actual UBO loads, we'll need to be able to upload UBO 0's pointer or
descriptor.

No difference on nohw 1 UBO update drawoverhead case (n=35).

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5273>

diff --git a/src/freedreno/ir3/ir3_nir_analyze_ubo_ranges.c b/src/freedreno/ir3/ir3_nir_analyze_ubo_ranges.c
index ba72e7351eff3d32c246c9526fc6f0bb4653d64d..a4733cdb2ef45ac25650c6fb42c97919617484d5 100644 (file)
--- a/src/freedreno/ir3/ir3_nir_analyze_ubo_ranges.c
+++ b/src/freedreno/ir3/ir3_nir_analyze_ubo_ranges.c
@@ -177,32 +177,23 @@ handle_partial_const(nir_builder *b, nir_ssa_def **srcp, int *offp)
        }
 }
 
+/* Tracks the maximum bindful UBO accessed so that we reduce the UBO
+ * descriptors emitted in the fast path for GL.
+ */
 static void
-lower_ubo_block_decrement(nir_intrinsic_instr *instr, nir_builder *b, int *num_ubos)
+track_ubo_use(nir_intrinsic_instr *instr, nir_builder *b, int *num_ubos)
 {
-       /* Skip shifting things for turnip's bindless resources. */
        if (ir3_bindless_resource(instr->src[0])) {
                assert(!b->shader->info.first_ubo_is_default_ubo); /* only set for GL */
                return;
        }
 
-       /* Shift all GL nir_intrinsic_load_ubo UBO indices down by 1, because we
-        * have lowered block 0 off of load_ubo to constbuf and ir3_const only
-        * uploads pointers for block 1-N.  This is also where we update the NIR
-        * num_ubos to reflect the UBOs that remain in use after others got
-        * lowered to constbuf access.
-        */
        if (nir_src_is_const(instr->src[0])) {
-               int block = nir_src_as_uint(instr->src[0]) - 1;
+               int block = nir_src_as_uint(instr->src[0]);
                *num_ubos = MAX2(*num_ubos, block + 1);
        } else {
-               *num_ubos = b->shader->info.num_ubos - 1;
+               *num_ubos = b->shader->info.num_ubos;
        }
-
-       nir_ssa_def *old_idx = nir_ssa_for_src(b, instr->src[0], 1);
-       nir_ssa_def *new_idx = nir_iadd_imm(b, old_idx, -1);
-       nir_instr_rewrite_src(&instr->instr, &instr->src[0],
-                       nir_src_for_ssa(new_idx));
 }
 
 static void
@@ -217,7 +208,7 @@ lower_ubo_load_to_uniform(nir_intrinsic_instr *instr, nir_builder *b,
         */
        struct ir3_ubo_range *range = get_existing_range(instr, state, false);
        if (!range) {
-               lower_ubo_block_decrement(instr, b, num_ubos);
+               track_ubo_use(instr, b, num_ubos);
                return;
        }
 
@@ -227,7 +218,7 @@ lower_ubo_load_to_uniform(nir_intrinsic_instr *instr, nir_builder *b,
                 * access, so for now just fall back to pulling.
                 */
                if (!nir_src_is_const(instr->src[1])) {
-                       lower_ubo_block_decrement(instr, b, num_ubos);
+                       track_ubo_use(instr, b, num_ubos);
                        return;
                }
 
@@ -236,7 +227,7 @@ lower_ubo_load_to_uniform(nir_intrinsic_instr *instr, nir_builder *b,
                 */
                const struct ir3_ubo_range r = get_ubo_load_range(instr, alignment);
                if (!(range->start <= r.start && r.end <= range->end)) {
-                       lower_ubo_block_decrement(instr, b, num_ubos);
+                       track_ubo_use(instr, b, num_ubos);
                        return;
                }
        }

diff --git a/src/gallium/drivers/freedreno/a6xx/fd6_const.c b/src/gallium/drivers/freedreno/a6xx/fd6_const.c
index 5b7f7518ab709efd1e3d1b1d2bce10ebd970cf9e..d8ba060da469145f9c71e68e0d0c84dce7096580 100644 (file)
--- a/src/gallium/drivers/freedreno/a6xx/fd6_const.c
+++ b/src/gallium/drivers/freedreno/a6xx/fd6_const.c
@@ -246,10 +246,7 @@ fd6_emit_ubos(struct fd_context *ctx, const struct ir3_shader_variant *v,
        OUT_RING(ring, CP_LOAD_STATE6_2_EXT_SRC_ADDR_HI(0));
 
        for (int i = 0; i < num_ubos; i++) {
-               /* Note: gallium constbuf 0 was always lowered to hardware constbuf,
-                * and UBO load indices decremented by one.
-                */
-               struct pipe_constant_buffer *cb = &constbuf->cb[i + 1];
+               struct pipe_constant_buffer *cb = &constbuf->cb[i];
 
                /* If we have user pointers (constbuf 0, aka GL uniforms), upload them
                 * to a buffer now, and save it in the constbuf so that we don't have

diff --git a/src/gallium/drivers/freedreno/ir3/ir3_const.h b/src/gallium/drivers/freedreno/ir3/ir3_const.h
index 6ed49baa8faf6e4ac29015224b9052831e41e14f..235c603590d5f1e289a9a4b997ef7023647d4f63 100644 (file)
--- a/src/gallium/drivers/freedreno/ir3/ir3_const.h
+++ b/src/gallium/drivers/freedreno/ir3/ir3_const.h
@@ -142,8 +142,7 @@ ir3_emit_ubos(struct fd_context *ctx, const struct ir3_shader_variant *v,
                struct pipe_resource *prscs[params];
 
                for (uint32_t i = 0; i < params; i++) {
-                       const uint32_t index = i + 1;   /* UBOs start at index 1 */
-                       struct pipe_constant_buffer *cb = &constbuf->cb[index];
+                       struct pipe_constant_buffer *cb = &constbuf->cb[i];
 
                        /* If we have user pointers (constbuf 0, aka GL uniforms), upload
                         * them to a buffer now, and save it in the constbuf so that we
@@ -159,7 +158,7 @@ ir3_emit_ubos(struct fd_context *ctx, const struct ir3_shader_variant *v,
                                cb->user_buffer = NULL;
                        }
 
-                       if ((constbuf->enabled_mask & (1 << index)) && cb->buffer) {
+                       if ((constbuf->enabled_mask & (1 << i)) && cb->buffer) {
                                offsets[i] = cb->buffer_offset;
                                prscs[i] = cb->buffer;
                        } else {