freedreno/ir3: optimize shader key comparision

Signed-off-by: Rob Clark <[email protected]>

2 files changed, 55 insertions, 30 deletions

diff --git a/src/gallium/drivers/freedreno/ir3/ir3_shader.c b/src/gallium/drivers/freedreno/ir3/ir3_shader.c
index c77cec10cc7..1f7e869d9f3 100644
--- a/src/gallium/drivers/freedreno/ir3/ir3_shader.c
+++ b/src/gallium/drivers/freedreno/ir3/ir3_shader.c
@@ -182,23 +182,30 @@ ir3_shader_variant(struct ir3_shader *shader, struct ir3_shader_key key)
 	 * so normalize the key to avoid constructing multiple identical
 	 * variants:
 	 */
-	if (shader->type == SHADER_FRAGMENT) {
+	switch (shader->type) {
+	case SHADER_FRAGMENT:
+	case SHADER_COMPUTE:
 		key.binning_pass = false;
-		key.vsaturate_s = 0;
-		key.vsaturate_t = 0;
-		key.vsaturate_r = 0;
-	}
-	if (shader->type == SHADER_VERTEX) {
+		if (key.has_per_samp) {
+			key.vsaturate_s = 0;
+			key.vsaturate_t = 0;
+			key.vsaturate_r = 0;
+		}
+		break;
+	case SHADER_VERTEX:
 		key.color_two_side = false;
 		key.half_precision = false;
 		key.alpha = false;
-		key.fsaturate_s = 0;
-		key.fsaturate_t = 0;
-		key.fsaturate_r = 0;
+		if (key.has_per_samp) {
+			key.fsaturate_s = 0;
+			key.fsaturate_t = 0;
+			key.fsaturate_r = 0;
+		}
+		break;
 	}
 
 	for (v = shader->variants; v; v = v->next)
-		if (!memcmp(&key, &v->key, sizeof(key)))
+		if (ir3_shader_key_equal(&key, &v->key))
 			return v;
 
 	/* compile new variant if it doesn't exist already: */
diff --git a/src/gallium/drivers/freedreno/ir3/ir3_shader.h b/src/gallium/drivers/freedreno/ir3/ir3_shader.h
index c531ad704cc..628c09e1be3 100644
--- a/src/gallium/drivers/freedreno/ir3/ir3_shader.h
+++ b/src/gallium/drivers/freedreno/ir3/ir3_shader.h
@@ -54,36 +54,54 @@ static inline uint16_t sem2idx(ir3_semantic sem)
  * in hw (two sided color), binning-pass vertex shader, etc.
  */
 struct ir3_shader_key {
+	union {
+		struct {
+			/* do we need to check {v,f}saturate_{s,t,r}? */
+			unsigned has_per_samp : 1;
+
+			/*
+			 * Vertex shader variant parameters:
+			 */
+			unsigned binning_pass : 1;
+
+			/*
+			 * Fragment shader variant parameters:
+			 */
+			unsigned color_two_side : 1;
+			unsigned half_precision : 1;
+			/* For rendering to alpha, we need a bit of special handling
+			 * since the hw always takes gl_FragColor starting from x
+			 * component, rather than figuring out to take the w component.
+			 * We could be more clever and generate variants for other
+			 * render target formats (ie. luminance formats are xxx1), but
+			 * let's start with this and see how it goes:
+			 */
+			unsigned alpha : 1;
+		};
+		uint32_t global;
+	};
+
 	/* bitmask of sampler which needs coords clamped for vertex
 	 * shader:
 	 */
-	unsigned vsaturate_s, vsaturate_t, vsaturate_r;
+	uint16_t vsaturate_s, vsaturate_t, vsaturate_r;
 
 	/* bitmask of sampler which needs coords clamped for frag
 	 * shader:
 	 */
-	unsigned fsaturate_s, fsaturate_t, fsaturate_r;
-
-	/*
-	 * Vertex shader variant parameters:
-	 */
-	unsigned binning_pass : 1;
+	uint16_t fsaturate_s, fsaturate_t, fsaturate_r;
 
-	/*
-	 * Fragment shader variant parameters:
-	 */
-	unsigned color_two_side : 1;
-	unsigned half_precision : 1;
-	/* For rendering to alpha, we need a bit of special handling
-	 * since the hw always takes gl_FragColor starting from x
-	 * component, rather than figuring out to take the w component.
-	 * We could be more clever and generate variants for other
-	 * render target formats (ie. luminance formats are xxx1), but
-	 * let's start with this and see how it goes:
-	 */
-	unsigned alpha : 1;
 };
 
+static inline bool
+ir3_shader_key_equal(struct ir3_shader_key *a, struct ir3_shader_key *b)
+{
+	/* slow-path if we need to check {v,f}saturate_{s,t,r} */
+	if (a->has_per_samp || b->has_per_samp)
+		return memcmp(a, b, sizeof(struct ir3_shader_key)) == 0;
+	return a->global == b->global;
+}
+
 struct ir3_shader_variant {
 	struct fd_bo *bo;