panfrost: Use nir_foreach_variable_with_modes in pan_compile

[mesa.git] / src / gallium / drivers / panfrost / pan_assemble.c

diff --git a/src/gallium/drivers/panfrost/pan_assemble.c b/src/gallium/drivers/panfrost/pan_assemble.c
index 7d0f07a0dd100644a098ed0ca5e620f42d470d5a..4a3bf793913f3e5791e33a3292eab055d1eadc54 100644 (file)
--- a/src/gallium/drivers/panfrost/pan_assemble.c
+++ b/src/gallium/drivers/panfrost/pan_assemble.c
@@ -77,7 +77,7 @@ pan_format_from_nir_size(nir_alu_type base, unsigned size)
 }
 
 static enum mali_format
-pan_format_from_glsl(const struct glsl_type *type, unsigned frac)
+pan_format_from_glsl(const struct glsl_type *type, unsigned precision, unsigned frac)
 {
         const struct glsl_type *column = glsl_without_array_or_matrix(type);
         enum glsl_base_type glsl_base = glsl_get_base_type(column);
@@ -95,6 +95,19 @@ pan_format_from_glsl(const struct glsl_type *type, unsigned frac)
         unsigned base = nir_alu_type_get_base_type(t);
         unsigned size = nir_alu_type_get_type_size(t);
 
+        /* Demote to fp16 where possible. int16 varyings are TODO as the hw
+         * will saturate instead of wrap which is not conformant, so we need to
+         * insert i2i16/u2u16 instructions before the st_vary_32i/32u to get
+         * the intended behaviour */
+
+        bool is_16 = (precision == GLSL_PRECISION_MEDIUM)
+                || (precision == GLSL_PRECISION_LOW);
+
+        if (is_16 && base == nir_type_float)
+                size = 16;
+        else
+                size = 32;
+
         return pan_format_from_nir_base(base) |
                 pan_format_from_nir_size(base, size) |
                 MALI_NR_CHANNELS(chan);
@@ -119,8 +132,7 @@ bifrost_blend_type_from_nir(nir_alu_type nir_type)
         case nir_type_uint16:
                 return BIFROST_BLEND_U16;
         default:
-                DBG("Unsupported blend shader type for NIR alu type %d", nir_type);
-                assert(0);
+                unreachable("Unsupported blend shader type for NIR alu type");
                 return 0;
         }
 }
@@ -153,11 +165,13 @@ panfrost_shader_compile(struct panfrost_context *ctx,
                 .alpha_ref = state->alpha_state.ref_value
         };
 
+        memcpy(program.rt_formats, state->rt_formats, sizeof(program.rt_formats));
+
         if (dev->quirks & IS_BIFROST) {
                 bifrost_compile_shader_nir(s, &program, dev->gpu_id);
         } else {
                 midgard_compile_shader_nir(s, &program, false, 0, dev->gpu_id,
-                                pan_debug & PAN_DBG_PRECOMPILE);
+                                dev->debug & PAN_DBG_PRECOMPILE, false);
         }
 
         /* Prepare the compiled binary for upload */
@@ -169,7 +183,7 @@ panfrost_shader_compile(struct panfrost_context *ctx,
          * that's how I'd do it. */
 
         if (size) {
-                state->bo = pan_bo_create(dev, size, PAN_BO_EXECUTE);
+                state->bo = panfrost_bo_create(dev, size, PAN_BO_EXECUTE);
                 memcpy(state->bo->cpu, dst, size);
         }
 
@@ -212,6 +226,12 @@ panfrost_shader_compile(struct panfrost_context *ctx,
                 if (s->info.outputs_written & BITFIELD64_BIT(FRAG_RESULT_STENCIL))
                         state->writes_stencil = true;
 
+                uint64_t outputs_read = s->info.outputs_read;
+                if (outputs_read & BITFIELD64_BIT(FRAG_RESULT_COLOR))
+                        outputs_read |= BITFIELD64_BIT(FRAG_RESULT_DATA0);
+
+                state->outputs_read = outputs_read >> FRAG_RESULT_DATA0;
+
                 /* List of reasons we need to execute frag shaders when things
                  * are masked off */
 
@@ -254,16 +274,17 @@ panfrost_shader_compile(struct panfrost_context *ctx,
 
         /* Record the varying mapping for the command stream's bookkeeping */
 
-        struct exec_list *l_varyings =
-                        stage == MESA_SHADER_VERTEX ? &s->outputs : &s->inputs;
+        nir_variable_mode varying_mode =
+                        stage == MESA_SHADER_VERTEX ? nir_var_shader_out : nir_var_shader_in;
 
-        nir_foreach_variable(var, l_varyings) {
+        nir_foreach_variable_with_modes(var, s, varying_mode) {
                 unsigned loc = var->data.driver_location;
                 unsigned sz = glsl_count_attribute_slots(var->type, FALSE);
 
                 for (int c = 0; c < sz; ++c) {
                         state->varyings_loc[loc + c] = var->data.location + c;
-                        state->varyings[loc + c] = pan_format_from_glsl(var->type, var->data.location_frac);
+                        state->varyings[loc + c] = pan_format_from_glsl(var->type,
+                                        var->data.precision, var->data.location_frac);
                 }
         }
 }