freedreno/ir3: fb read support

Lower load_output to txf_ms_fb and add support for the new texture fetch
instruction.

Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>

diff --git a/src/freedreno/ir3/ir3_compiler_nir.c b/src/freedreno/ir3/ir3_compiler_nir.c
index 3c813c73ae010611772b9a75019d536788d7a690..f1d9b53c7c4a36055f51d9cbe8f10826896a396b 100644 (file)
--- a/src/freedreno/ir3/ir3_compiler_nir.c
+++ b/src/freedreno/ir3/ir3_compiler_nir.c
@@ -1756,12 +1756,9 @@ emit_tex(struct ir3_context *ctx, nir_tex_instr *tex)
                case 3:              opc = OPC_GATHER4A; break;
                }
                break;
+       case nir_texop_txf_ms_fb:
        case nir_texop_txf_ms:   opc = OPC_ISAMM;    break;
-       case nir_texop_txs:
-       case nir_texop_query_levels:
-       case nir_texop_texture_samples:
-       case nir_texop_samples_identical:
-       case nir_texop_txf_ms_mcs:
+       default:
                ir3_context_error(ctx, "Unhandled NIR tex type: %d\n", tex->op);
                return;
        }
@@ -1838,7 +1835,7 @@ emit_tex(struct ir3_context *ctx, nir_tex_instr *tex)
        /* NOTE a3xx (and possibly a4xx?) might be different, using isaml
         * with scaled x coord according to requested sample:
         */
-       if (tex->op == nir_texop_txf_ms) {
+       if (opc == OPC_ISAMM) {
                if (ctx->compiler->txf_ms_with_isaml) {
                        /* the samples are laid out in x dimension as
                         *     0 1 2 3
@@ -1897,7 +1894,24 @@ emit_tex(struct ir3_context *ctx, nir_tex_instr *tex)
        if (opc == OPC_GETLOD)
                type = TYPE_U32;
 
-       struct ir3_instruction *samp_tex = get_tex_samp_tex_src(ctx, tex);
+       struct ir3_instruction *samp_tex;
+
+       if (tex->op == nir_texop_txf_ms_fb) {
+               /* only expect a single txf_ms_fb per shader: */
+               compile_assert(ctx, !ctx->so->fb_read);
+               compile_assert(ctx, ctx->so->type == MESA_SHADER_FRAGMENT);
+
+               ctx->so->fb_read = true;
+               samp_tex = ir3_create_collect(ctx, (struct ir3_instruction*[]){
+                       create_immed_typed(ctx->block, ctx->so->num_samp, TYPE_U16),
+                       create_immed_typed(ctx->block, ctx->so->num_samp, TYPE_U16),
+               }, 2);
+
+               ctx->so->num_samp++;
+       } else {
+               samp_tex = get_tex_samp_tex_src(ctx, tex);
+       }
+
        struct ir3_instruction *col0 = ir3_create_collect(ctx, src0, nsrc0);
        struct ir3_instruction *col1 = ir3_create_collect(ctx, src1, nsrc1);
 

diff --git a/src/freedreno/ir3/ir3_nir.c b/src/freedreno/ir3/ir3_nir.c
index 43b9c4d72878b6e69c115e710b8a2eabfd196cec..dffcf5f5ab7d76721614705f0a4f2a95117552b5 100644 (file)
--- a/src/freedreno/ir3/ir3_nir.c
+++ b/src/freedreno/ir3/ir3_nir.c
@@ -216,6 +216,12 @@ ir3_optimize_nir(struct ir3_shader *shader, nir_shader *s,
                 * and not again on any potential 2nd variant lowering pass:
                 */
                OPT_V(s, ir3_nir_apply_trig_workarounds);
+
+               /* This wouldn't hurt to run multiple times, but there is
+                * no need to:
+                */
+               if (shader->type == MESA_SHADER_FRAGMENT)
+                       OPT_V(s, nir_lower_fb_read);
        }
 
        OPT_V(s, nir_lower_tex, &tex_options);

diff --git a/src/freedreno/ir3/ir3_shader.h b/src/freedreno/ir3/ir3_shader.h
index 4e8ab085d7e5fb683c70364d50a3da17e2f1bd6a..7c1dc38de236b29c3565221dc0cad2e5a288beaa 100644 (file)
--- a/src/freedreno/ir3/ir3_shader.h
+++ b/src/freedreno/ir3/ir3_shader.h
@@ -434,6 +434,12 @@ struct ir3_shader_variant {
        /* number of samplers/textures (which are currently 1:1): */
        int num_samp;
 
+       /* is there an implicit sampler to read framebuffer (FS only).. if
+        * so the sampler-idx is 'num_samp - 1' (ie. it is appended after
+        * the last "real" texture)
+        */
+       bool fb_read;
+
        /* do we have one or more SSBO instructions: */
        bool has_ssbo;