nir/i965/freedreno/vc4: add a bindless bool to type size functions

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

diff --git a/src/gallium/drivers/panfrost/pan_swizzle.c b/src/gallium/drivers/panfrost/pan_swizzle.c
index c021843846d59b16c04ce7a7dc2404e56cd798ac..52a907ddd55d14c26f2b89d37b2c2718f2b873dc 100644 (file)
--- a/src/gallium/drivers/panfrost/pan_swizzle.c
+++ b/src/gallium/drivers/panfrost/pan_swizzle.c
@@ -97,6 +97,9 @@ swizzle_bpp1_align16(int width, int height, int source_stride, int block_pitch,
 
                 ++y;
 
+                if (y >= height)
+                        break;
+
                 {
                         int block_y = y & ~(0x0f);
                         int rem_y = y & 0x0f;
@@ -147,33 +150,40 @@ swizzle_bpp4_align16(int width, int height, int source_stride, int block_pitch,
 }
 
 void
-panfrost_texture_swizzle(int width, int height, int bytes_per_pixel, int source_stride,
+panfrost_texture_swizzle(unsigned off_x,
+                         unsigned off_y,
+                         int width, int height, int bytes_per_pixel, int dest_width,
                          const uint8_t *pixels,
                          uint8_t *ldest)
 {
         /* Calculate maximum size, overestimating a bit */
-        int block_pitch = ALIGN(width, 16) >> 4;
+        int block_pitch = ALIGN(dest_width, 16) >> 4;
+
+        /* Strides must be tight, since we're only ever called indirectly */
+        int source_stride = width * bytes_per_pixel;
 
         /* Use fast path if available */
-        if (bytes_per_pixel == 4 /* && (ALIGN(width, 16) == width) */) {
-                swizzle_bpp4_align16(width, height, source_stride >> 2, (block_pitch * 256 >> 4), (const uint32_t *) pixels, (uint32_t *) ldest);
-                return;
-        } else if (bytes_per_pixel == 1 /* && (ALIGN(width, 16) == width) */) {
-                swizzle_bpp1_align16(width, height, source_stride, (block_pitch * 256 >> 4), pixels, (uint8_t *) ldest);
-                return;
+        if (!(off_x || off_y) && (width == dest_width)) {
+                if (bytes_per_pixel == 4 /* && (ALIGN(width, 16) == width) */) {
+                        swizzle_bpp4_align16(width, height, source_stride >> 2, (block_pitch * 256 >> 4), (const uint32_t *) pixels, (uint32_t *) ldest);
+                        return;
+                } else if (bytes_per_pixel == 1 /* && (ALIGN(width, 16) == width) */) {
+                        swizzle_bpp1_align16(width, height, source_stride, (block_pitch * 256 >> 4), pixels, (uint8_t *) ldest);
+                        return;
+                }
         }
 
         /* Otherwise, default back on generic path */
 
         for (int y = 0; y < height; ++y) {
-                int block_y = y >> 4;
-                int rem_y = y & 0x0F;
+                int block_y = (y + off_y) >> 4;
+                int rem_y = (y + off_y) & 0x0F;
                 int block_start_s = block_y * block_pitch * 256;
                 int source_start = y * source_stride;
 
                 for (int x = 0; x < width; ++x) {
-                        int block_x_s = (x >> 4) * 256;
-                        int rem_x = x & 0x0F;
+                        int block_x_s = ((x + off_x) >> 4) * 256;
+                        int rem_x = (x + off_x) & 0x0F;
 
                         int index = space_filler[rem_y][rem_x];
                         const uint8_t *source = &pixels[source_start + bytes_per_pixel * x];
@@ -184,51 +194,3 @@ panfrost_texture_swizzle(int width, int height, int bytes_per_pixel, int source_
                 }
         }
 }
-
-
-unsigned
-panfrost_swizzled_size(int width, int height, int bytes_per_pixel)
-{
-        /* Calculate maximum size, overestimating a bit */
-        int block_pitch = ALIGN(width, 16) >> 4;
-        unsigned sz = bytes_per_pixel * 256 * ((height >> 4) + 1) * block_pitch;
-
-        return sz;
-}
-
-#if 0
-#include <stdio.h>
-#include <stdlib.h>
-#include <memory.h>
-#define TW 1920
-#define TH 1080
-void
-main()
-{
-        panfrost_generate_space_filler_indices();
-
-        uint8_t in[TW * TH * 4];
-
-        for (int i = 0; i < TW * TH * 4; ++i) in[i] = i;
-
-        uint8_t *out = malloc(TW * TH * 4 * 2);
-
-        for (int i = 0; i < 60; ++i) {
-                //swizzle_bpp4_align16(TW, TH, TW*4, TW>>4, (uint32_t *) in, (uint32_t *) out);
-                //panfrost_texture_swizzle_bpp4(TW, TH, TW*4, (uint32_t *) in, (uint32_t *) out);
-                //panfrost_texture_swizzle(TW, TH, 4, TW*4, (uint32_t *) in, (uint32_t *) out);
-
-                int block_pitch = ALIGN(TW, 16) >> 4;
-                swizzle_bpp1_align16(TW, TH, TW, (block_pitch * 256 >> 4), in, (uint8_t *) out);
-        }
-
-#if 0
-        uint8_t *reference = malloc(TW * TH * 4 * 2);
-        panfrost_texture_swizzle(TW, TH, 1, TW, (uint8_t *) in, (uint8_t *) reference);
-
-        if (memcmp(reference, out, TW * TH * 4)) printf("XXX\n");
-
-#endif
-        printf("ref %X\n", out[0]);
-}
-#endif